diff --git a/Gopkg.lock b/Gopkg.lock
new file mode 100644
index 0000000..6a9680d
--- /dev/null
+++ b/Gopkg.lock
@@ -0,0 +1,420 @@
+# This file is autogenerated, do not edit; changes may be undone by the next 'dep ensure'.
+
+
+[[projects]]
+  digest = "1:ed77032e4241e3b8329c9304d66452ed196e795876e14be677a546f36b94e67a"
+  name = "github.com/DataDog/zstd"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "c7161f8c63c045cbc7ca051dcc969dd0e4054de2"
+  version = "v1.3.5"
+
+[[projects]]
+  digest = "1:82a18170c9c41e36939cb5d26da1546b2cfa786aa030a978d3bf183519849230"
+  name = "github.com/Shopify/sarama"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "4602b5a8c6e826f9e0737865818dd43b2339a092"
+  version = "v1.21.0"
+
+[[projects]]
+  branch = "master"
+  digest = "1:ef5b0622d834c139454148b8fd0c92bb314828900532b267ae62da9fec109866"
+  name = "github.com/armon/go-metrics"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "f0300d1749da6fa982027e449ec0c7a145510c3c"
+
+[[projects]]
+  digest = "1:526d64d0a3ac6c24875724a9355895be56a21f89a5d3ab5ba88d91244269a7d8"
+  name = "github.com/bsm/sarama-cluster"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "c618e605e15c0d7535f6c96ff8efbb0dba4fd66c"
+  version = "v2.1.15"
+
+[[projects]]
+  digest = "1:bcf36df8d43860bfde913d008301aef27c6e9a303582118a837c4a34c0d18167"
+  name = "github.com/coreos/go-systemd"
+  packages = ["journal"]
+  pruneopts = "UT"
+  revision = "95778dfbb74eb7e4dbaf43bf7d71809650ef8076"
+  version = "v19"
+
+[[projects]]
+  digest = "1:39ff4d4d6baca1ac63c6a03404b761d664be14a71bb5a4aca8fef643d0f66fef"
+  name = "github.com/coreos/pkg"
+  packages = ["capnslog"]
+  pruneopts = "UT"
+  revision = "97fdf19511ea361ae1c100dd393cc47f8dcfa1e1"
+  version = "v4"
+
+[[projects]]
+  digest = "1:ffe9824d294da03b391f44e1ae8281281b4afc1bdaa9588c9097785e3af10cec"
+  name = "github.com/davecgh/go-spew"
+  packages = ["spew"]
+  pruneopts = "UT"
+  revision = "8991bc29aa16c548c550c7ff78260e27b9ab7c73"
+  version = "v1.1.1"
+
+[[projects]]
+  digest = "1:1f0c7ab489b407a7f8f9ad16c25a504d28ab461517a971d341388a56156c1bd7"
+  name = "github.com/eapache/go-resiliency"
+  packages = ["breaker"]
+  pruneopts = "UT"
+  revision = "ea41b0fad31007accc7f806884dcdf3da98b79ce"
+  version = "v1.1.0"
+
+[[projects]]
+  branch = "master"
+  digest = "1:79f16588b5576b1b3cd90e48d2374cc9a1a8776862d28d8fd0f23b0e15534967"
+  name = "github.com/eapache/go-xerial-snappy"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "776d5712da21bc4762676d614db1d8a64f4238b0"
+
+[[projects]]
+  digest = "1:444b82bfe35c83bbcaf84e310fb81a1f9ece03edfed586483c869e2c046aef69"
+  name = "github.com/eapache/queue"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "44cc805cf13205b55f69e14bcb69867d1ae92f98"
+  version = "v1.1.0"
+
+[[projects]]
+  digest = "1:4652941215f23c7ad6546bab86964c72248a1134baa53f0df618661344ba5cac"
+  name = "github.com/gogo/protobuf"
+  packages = [
+    "gogoproto",
+    "proto",
+    "protoc-gen-gogo/descriptor",
+  ]
+  pruneopts = "UT"
+  revision = "ba06b47c162d49f2af050fb4c75bcbc86a159d5c"
+  version = "v1.2.1"
+
+[[projects]]
+  digest = "1:f2bc519aefa5511928d70d067ae2828cfa60fd3947cc8d166f070a0ce705f782"
+  name = "github.com/golang/protobuf"
+  packages = [
+    "proto",
+    "protoc-gen-go/descriptor",
+    "ptypes",
+    "ptypes/any",
+    "ptypes/duration",
+    "ptypes/empty",
+    "ptypes/timestamp",
+  ]
+  pruneopts = "UT"
+  revision = "b5d812f8a3706043e23a9cd5babf2e5423744d30"
+  version = "v1.3.1"
+
+[[projects]]
+  digest = "1:e4f5819333ac698d294fe04dbf640f84719658d5c7ce195b10060cc37292ce79"
+  name = "github.com/golang/snappy"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "2a8bb927dd31d8daada140a5d09578521ce5c36a"
+  version = "v0.0.1"
+
+[[projects]]
+  digest = "1:582b704bebaa06b48c29b0cec224a6058a09c86883aaddabde889cd1a5f73e1b"
+  name = "github.com/google/uuid"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "0cd6bf5da1e1c83f8b45653022c74f71af0538a4"
+  version = "v1.1.1"
+
+[[projects]]
+  digest = "1:9537873f2fc92467f451178506405fae087efc624d8d92644042209a03a2bfa0"
+  name = "github.com/hashicorp/consul"
+  packages = ["api"]
+  pruneopts = "UT"
+  revision = "ea5210a30e154f4da9a4c8e729b45b8ce7b9b92c"
+  version = "v1.4.4"
+
+[[projects]]
+  digest = "1:af105c7c5dc0b4ae41991f122cae860b9600f7d226072c2a83127048c991660c"
+  name = "github.com/hashicorp/go-cleanhttp"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "eda1e5db218aad1db63ca4642c8906b26bcf2744"
+  version = "v0.5.1"
+
+[[projects]]
+  digest = "1:2be5a35f0c5b35162c41bb24971e5dcf6ce825403296ee435429cdcc4e1e847e"
+  name = "github.com/hashicorp/go-immutable-radix"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "27df80928bb34bb1b0d6d0e01b9e679902e7a6b5"
+  version = "v1.0.0"
+
+[[projects]]
+  digest = "1:a54ada9beb59fdc35b69322979e870ff0b780e03f4dc309c4c8674b94927df75"
+  name = "github.com/hashicorp/go-rootcerts"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "63503fb4e1eca22f9ae0f90b49c5d5538a0e87eb"
+  version = "v1.0.0"
+
+[[projects]]
+  digest = "1:67474f760e9ac3799f740db2c489e6423a4cde45520673ec123ac831ad849cb8"
+  name = "github.com/hashicorp/golang-lru"
+  packages = ["simplelru"]
+  pruneopts = "UT"
+  revision = "7087cb70de9f7a8bc0a10c375cb0d2280a8edf9c"
+  version = "v0.5.1"
+
+[[projects]]
+  digest = "1:acc81e4e4289587b257ccdfccbc6eaf16d4c2fb57dda73c6bb349bf50f02501f"
+  name = "github.com/hashicorp/serf"
+  packages = ["coordinate"]
+  pruneopts = "UT"
+  revision = "b89a09ebd4b1b570e0076d5097272e67c10ac4f6"
+  version = "v0.8.2"
+
+[[projects]]
+  digest = "1:5d231480e1c64a726869bc4142d270184c419749d34f167646baa21008eb0a79"
+  name = "github.com/mitchellh/go-homedir"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "af06845cf3004701891bf4fdb884bfe4920b3727"
+  version = "v1.1.0"
+
+[[projects]]
+  digest = "1:53bc4cd4914cd7cd52139990d5170d6dc99067ae31c56530621b18b35fc30318"
+  name = "github.com/mitchellh/mapstructure"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "3536a929edddb9a5b34bd6861dc4a9647cb459fe"
+  version = "v1.1.2"
+
+[[projects]]
+  branch = "master"
+  digest = "1:724ce67c43f6b392e3e57a05f566ebe339354a9b3245cba64515769d8506e22f"
+  name = "github.com/opencord/voltha-go"
+  packages = [
+    "adapters",
+    "adapters/common",
+    "common/log",
+    "db/kvstore",
+    "kafka",
+  ]
+  pruneopts = "UT"
+  revision = "c91266ef204b7747a2cbe9bb53005d92d88ea497"
+
+[[projects]]
+  branch = "master"
+  digest = "1:9075f267bf479a68085b08d33b7df7f938b6f813ad51de98d890ea95fc4c81cf"
+  name = "github.com/opencord/voltha-protos"
+  packages = [
+    "go/common",
+    "go/inter_container",
+    "go/omci",
+    "go/openflow_13",
+    "go/openolt",
+    "go/voltha",
+  ]
+  pruneopts = "UT"
+  revision = "61e94873aa4b4cfd5e9277c666402214b031c05b"
+
+[[projects]]
+  digest = "1:d886a3c32c8c1a770d07e36340f061d3afc948d065ffc3c9a19b01b34d4f0b65"
+  name = "github.com/pierrec/lz4"
+  packages = [
+    ".",
+    "internal/xxh32",
+  ]
+  pruneopts = "UT"
+  revision = "315a67e90e415bcdaff33057da191569bf4d8479"
+  version = "v2.1.1"
+
+[[projects]]
+  branch = "master"
+  digest = "1:d38f81081a389f1466ec98192cf9115a82158854d6f01e1c23e2e7554b97db71"
+  name = "github.com/rcrowley/go-metrics"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "3113b8401b8a98917cde58f8bbd42a1b1c03b1fd"
+
+[[projects]]
+  branch = "master"
+  digest = "1:254ca48be90c0a91c39b2be976a8b7a96e59c47e43e892cb8449f279d8951d70"
+  name = "go.etcd.io/etcd"
+  packages = [
+    "auth/authpb",
+    "clientv3",
+    "clientv3/balancer",
+    "clientv3/balancer/picker",
+    "clientv3/balancer/resolver/endpoint",
+    "clientv3/concurrency",
+    "etcdserver/api/v3rpc/rpctypes",
+    "etcdserver/etcdserverpb",
+    "mvcc/mvccpb",
+    "pkg/logutil",
+    "pkg/systemd",
+    "pkg/types",
+    "raft",
+    "raft/raftpb",
+  ]
+  pruneopts = "UT"
+  revision = "a621d807f061e1dd635033a8d6bc261461429e27"
+
+[[projects]]
+  digest = "1:3c1a69cdae3501bf75e76d0d86dc6f2b0a7421bc205c0cb7b96b19eed464a34d"
+  name = "go.uber.org/atomic"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "1ea20fb1cbb1cc08cbd0d913a96dead89aa18289"
+  version = "v1.3.2"
+
+[[projects]]
+  digest = "1:60bf2a5e347af463c42ed31a493d817f8a72f102543060ed992754e689805d1a"
+  name = "go.uber.org/multierr"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "3c4937480c32f4c13a875a1829af76c98ca3d40a"
+  version = "v1.1.0"
+
+[[projects]]
+  digest = "1:c52caf7bd44f92e54627a31b85baf06a68333a196b3d8d241480a774733dcf8b"
+  name = "go.uber.org/zap"
+  packages = [
+    ".",
+    "buffer",
+    "internal/bufferpool",
+    "internal/color",
+    "internal/exit",
+    "zapcore",
+  ]
+  pruneopts = "UT"
+  revision = "ff33455a0e382e8a81d14dd7c922020b6b5e7982"
+  version = "v1.9.1"
+
+[[projects]]
+  branch = "master"
+  digest = "1:a7550c619a7f9701bc2f0b2c866668a207cfb693a2834796f723d2c78b514372"
+  name = "golang.org/x/net"
+  packages = [
+    "context",
+    "http/httpguts",
+    "http2",
+    "http2/hpack",
+    "idna",
+    "internal/timeseries",
+    "trace",
+  ]
+  pruneopts = "UT"
+  revision = "b630fd6fe46bcfc98f989005d8b8ec1400e60a6e"
+
+[[projects]]
+  branch = "master"
+  digest = "1:e18c1645adfe83d66f80d3676040e478895669f9456ca8bafa6a7290656c1b48"
+  name = "golang.org/x/sys"
+  packages = ["unix"]
+  pruneopts = "UT"
+  revision = "81d4e9dc473e5e8c933f2aaeba2a3d81efb9aed2"
+
+[[projects]]
+  digest = "1:a2ab62866c75542dd18d2b069fec854577a20211d7c0ea6ae746072a1dccdd18"
+  name = "golang.org/x/text"
+  packages = [
+    "collate",
+    "collate/build",
+    "internal/colltab",
+    "internal/gen",
+    "internal/tag",
+    "internal/triegen",
+    "internal/ucd",
+    "language",
+    "secure/bidirule",
+    "transform",
+    "unicode/bidi",
+    "unicode/cldr",
+    "unicode/norm",
+    "unicode/rangetable",
+  ]
+  pruneopts = "UT"
+  revision = "f21a4dfb5e38f5895301dc265a8def02365cc3d0"
+  version = "v0.3.0"
+
+[[projects]]
+  branch = "master"
+  digest = "1:08aea4eff69cb424f2ce0d69f88c2f0d351fc44e82cab057ddf3946f41f02108"
+  name = "google.golang.org/genproto"
+  packages = [
+    "googleapis/api/annotations",
+    "googleapis/rpc/status",
+  ]
+  pruneopts = "UT"
+  revision = "f467c93bbac2133ff463e1f93d18d8f9f3f04451"
+
+[[projects]]
+  digest = "1:a0bdff6a1dd68e7fa6edbffe3bef8e061db79543975d1c53afcae3a8d8a6c445"
+  name = "google.golang.org/grpc"
+  packages = [
+    ".",
+    "balancer",
+    "balancer/base",
+    "balancer/roundrobin",
+    "binarylog/grpc_binarylog_v1",
+    "codes",
+    "connectivity",
+    "credentials",
+    "credentials/internal",
+    "encoding",
+    "encoding/proto",
+    "grpclog",
+    "health/grpc_health_v1",
+    "internal",
+    "internal/backoff",
+    "internal/binarylog",
+    "internal/channelz",
+    "internal/envconfig",
+    "internal/grpcrand",
+    "internal/grpcsync",
+    "internal/syscall",
+    "internal/transport",
+    "keepalive",
+    "metadata",
+    "naming",
+    "peer",
+    "resolver",
+    "resolver/dns",
+    "resolver/passthrough",
+    "stats",
+    "status",
+    "tap",
+  ]
+  pruneopts = "UT"
+  revision = "3507fb8e1a5ad030303c106fef3a47c9fdad16ad"
+  version = "v1.19.1"
+
+[[projects]]
+  digest = "1:82a18170c9c41e36939cb5d26da1546b2cfa786aa030a978d3bf183519849230"
+  name = "gopkg.in/Shopify/sarama.v1"
+  packages = ["."]
+  pruneopts = "UT"
+  revision = "4602b5a8c6e826f9e0737865818dd43b2339a092"
+  version = "v1.21.0"
+
+[solve-meta]
+  analyzer-name = "dep"
+  analyzer-version = 1
+  input-imports = [
+    "github.com/gogo/protobuf/proto",
+    "github.com/golang/protobuf/ptypes",
+    "github.com/opencord/voltha-go/adapters",
+    "github.com/opencord/voltha-go/adapters/common",
+    "github.com/opencord/voltha-go/common/log",
+    "github.com/opencord/voltha-go/db/kvstore",
+    "github.com/opencord/voltha-go/kafka",
+    "github.com/opencord/voltha-protos/go/common",
+    "github.com/opencord/voltha-protos/go/inter_container",
+    "github.com/opencord/voltha-protos/go/openflow_13",
+    "github.com/opencord/voltha-protos/go/openolt",
+    "github.com/opencord/voltha-protos/go/voltha",
+    "google.golang.org/grpc",
+  ]
+  solver-name = "gps-cdcl"
+  solver-version = 1
diff --git a/Gopkg.toml b/Gopkg.toml
new file mode 100644
index 0000000..413989e
--- /dev/null
+++ b/Gopkg.toml
@@ -0,0 +1,54 @@
+# Gopkg.toml example
+#
+# Refer to https://golang.github.io/dep/docs/Gopkg.toml.html
+# for detailed Gopkg.toml documentation.
+#
+# required = ["github.com/user/thing/cmd/thing"]
+# ignored = ["github.com/user/project/pkgX", "bitbucket.org/user/project/pkgA/pkgY"]
+#
+# [[constraint]]
+#   name = "github.com/user/project"
+#   version = "1.0.0"
+#
+# [[constraint]]
+#   name = "github.com/user/project2"
+#   branch = "dev"
+#   source = "github.com/myfork/project2"
+#
+# [[override]]
+#   name = "github.com/x/y"
+#   version = "2.4.0"
+#
+# [prune]
+#   non-go = false
+#   go-tests = true
+#   unused-packages = true
+
+
+[[constraint]]
+  name = "github.com/gogo/protobuf"
+  version = "1.2.1"
+
+[[constraint]]
+  name = "github.com/golang/protobuf"
+  version = "1.3.1"
+
+[[constraint]]
+  branch = "master"
+  name = "github.com/opencord/voltha-go"
+
+[[constraint]]
+  branch = "master"
+  name = "github.com/opencord/voltha-protos"
+
+[[constraint]]
+  name = "google.golang.org/grpc"
+  version = "1.19.1"
+
+[[override]]
+  branch = "master"
+  name = "go.etcd.io/etcd"
+
+[prune]
+  go-tests = true
+  unused-packages = true
diff --git a/Makefile b/Makefile
index 08fddff..f272c3c 100644
--- a/Makefile
+++ b/Makefile
@@ -82,6 +82,15 @@
 containers: adapter_openolt_go
 
 adapter_openolt_go:
+ifdef LOCAL_PROTOS
+	mkdir -p vendor/github.com/opencord/voltha-protos/go
+	cp -rf ${GOPATH}/src/github.com/opencord/voltha-protos/go/ vendor/github.com/opencord/voltha-protos
+endif
+ifdef LOCAL_VOLTHAGO
+	mkdir -p vendor/github.com/opencord/voltha-go/
+	cp -rf ${GOPATH}/src/github.com/opencord/voltha-go/ vendor/github.com/opencord/
+	rm -rf vendor/github.com/opencord/voltha-go/vendor
+endif
 	docker build $(DOCKER_BUILD_ARGS) -t ${REGISTRY}${REPOSITORY}voltha-openolt-adapter-go:${TAG} -f docker/Dockerfile.openolt .
 
 # end file
diff --git a/README.md b/README.md
index 0c3609b..77d9f22 100644
--- a/README.md
+++ b/README.md
@@ -3,7 +3,30 @@
 
 Assuming the VOLTHA2.0 environment is made using the quickstart.md in voltha-go.
 
+Dependencies are committed to the repos as per current standard practice.
+If you need to update them you can do so with dep. This includes the voltha-protos and voltha-go dependencies.
+
+Ensure your GOPATH variable is set according to the quickstart
+Create a symbolic link in the $GOPATH/src tree to the voltha-openolt-adapter repository:
+```sh
+mkdir -p $GOPATH/src/github.com/opencord
+ln -s ~/repos/voltha-openolt-adapter $GOPATH/src/github.com/opencord/voltha-openolt-adapter
 ```
+
+Install dep for fetching dependencies
+```sh
+mkdir -p $GOPATH/bin
+curl https://raw.githubusercontent.com/golang/dep/master/install.sh | sh
+
+Pull and build dependencies for the project.  This may take a while.  This may likely update versions of 3rd party packages in the vendor folder.   This may introduce unexpected changes.   If everything still works feel free to check in the vendor updates as needed.
+
+cd $GOPATH/src/github.com/opencord/voltha-go/
+dep ensure
+
+If you are using a custom local copy of protos or voltha-go
+
+Just export LOCAL_PROTOS=true or LOCAL_VOLTHAGO=true to use that instead, assuming you have set them up on GOPATH. See the quickstart.
+
 cd ~/source/voltha-openolt-adapter
 ```
 
diff --git a/adaptercore/device_handler.go b/adaptercore/device_handler.go
index b626242..1c2fdf9 100644
--- a/adaptercore/device_handler.go
+++ b/adaptercore/device_handler.go
@@ -33,7 +33,7 @@
 	ic "github.com/opencord/voltha-protos/go/inter_container"
 	of "github.com/opencord/voltha-protos/go/openflow_13"
 	oop "github.com/opencord/voltha-protos/go/openolt"
-        rsrcMgr "github.com/opencord/voltha-go/adapters/openolt/adaptercore/resourcemanager"
+	rsrcMgr "github.com/opencord/voltha-openolt-adapter/adaptercore/resourcemanager"
 	"github.com/opencord/voltha-protos/go/voltha"
 	"google.golang.org/grpc"
 )
diff --git a/adaptercore/openolt_flowmgr.go b/adaptercore/openolt_flowmgr.go
index 8d90fd8..18efb33 100644
--- a/adaptercore/openolt_flowmgr.go
+++ b/adaptercore/openolt_flowmgr.go
@@ -23,8 +23,7 @@
 	"errors"
 	"fmt"
 	"math/big"
-
-	rsrcMgr "github.com/opencord/voltha-go/adapters/openolt/adaptercore/resourcemanager"
+	rsrcMgr "github.com/opencord/voltha-openolt-adapter/adaptercore/resourcemanager"
 	"github.com/opencord/voltha-go/common/log"
 	tp "github.com/opencord/voltha-go/common/techprofile"
 	fd "github.com/opencord/voltha-go/rw_core/flow_decomposition"
diff --git a/docker/Dockerfile.openolt b/docker/Dockerfile.openolt
index c975591..ed20905 100644
--- a/docker/Dockerfile.openolt
+++ b/docker/Dockerfile.openolt
@@ -6,44 +6,22 @@
 # Install required packages
 RUN apk add --no-cache wget git make build-base protobuf protobuf-dev
 
-# Install protobuf requirements
-RUN git clone https://github.com/googleapis/googleapis.git /usr/local/include/googleapis
-RUN go get google.golang.org/genproto/googleapis/api/annotations
-
 # Prepare directory structure
-RUN ["mkdir", "-p", "/src", "src/protos"]
+RUN ["mkdir", "-p", "/src"]
 RUN ["mkdir", "-p", "$GOPATH/src", "$GOPATH/pkg", "$GOPATH/bin"]
 RUN ["mkdir", "-p", "$GOPATH/src/github.com/opencord"]
-RUN ["mkdir", "-p", "$GOPATH/src/github.com/opencord/voltha-go"]
+RUN ["mkdir", "-p", "$GOPATH/src/github.com/opencord/voltha-openolt-adapter"]
 
-WORKDIR $GOPATH/src/github.com/opencord
+WORKDIR $GOPATH/src/github.com/opencord/voltha-openolt-adapter
 
-RUN git clone https://gerrit.opencord.org/voltha-go.git
-
-WORKDIR $GOPATH/src/github.com/opencord/voltha-go
-
-
-# Copy files
-ADD adaptercore ./adapters/openolt/adaptercore
-ADD config      ./adapters/openolt/config
-ADD *.go        ./adapters/openolt
-
-RUN ls ./adapters/openolt
-
-# Copy required proto files
-ADD openolt.proto ./protos
-
-# Install the protoc-gen-go
-RUN go install ./vendor/github.com/golang/protobuf/protoc-gen-go
-
-# Compile protobuf files
-RUN sh protos/scripts/build_protos.sh protos
-
-RUN protoc --go_out=Mgoogle/protobuf/descriptor.proto=github.com/golang/protobuf/protoc-gen-go/descriptor,plugins=grpc:$GOPATH/src -I protos -I /usr/local/include/googleapis  protos/openolt.proto
+COPY adaptercore ./adaptercore
+COPY config ./config
+COPY vendor ./vendor
+COPY *.go ./
 
 # Build openolt
+RUN go build -o /src/openolt
 
-RUN cd adapters/openolt && go build -o /src/openolt
 
 # -------------
 # Image creation stage
diff --git a/main.go b/main.go
index b349fa8..e4f22fa 100644
--- a/main.go
+++ b/main.go
@@ -21,8 +21,8 @@
 	"fmt"
 	"github.com/opencord/voltha-go/adapters"
 	com "github.com/opencord/voltha-go/adapters/common"
-	ac "github.com/opencord/voltha-go/adapters/openolt/adaptercore"
-	"github.com/opencord/voltha-go/adapters/openolt/config"
+	ac "github.com/opencord/voltha-openolt-adapter/adaptercore"
+	"github.com/opencord/voltha-openolt-adapter/config"
 	"github.com/opencord/voltha-go/common/log"
 	"github.com/opencord/voltha-go/db/kvstore"
 	"github.com/opencord/voltha-go/kafka"
diff --git a/openolt.proto b/openolt.proto
deleted file mode 100644
index 709105b..0000000
--- a/openolt.proto
+++ /dev/null
@@ -1,560 +0,0 @@
-// Copyright (c) 2018 Open Networking Foundation
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at:
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-syntax = "proto3";
-
-option go_package = "github.com/opencord/voltha-go/protos/openolt";
-
-package openolt;
-
-import "google/api/annotations.proto";
-
-service Openolt {
-
-    rpc DisableOlt(Empty) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/Disable"
-          body: "*"
-        };
-    }
-
-    rpc ReenableOlt(Empty) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/Reenable"
-          body: "*"
-        };
-    }
-
-    rpc ActivateOnu(Onu) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/EnableOnu"
-          body: "*"
-        };
-    }
-
-    rpc DeactivateOnu(Onu) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/DisableOnu"
-          body: "*"
-        };
-    }
-
-    rpc DeleteOnu(Onu) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/DeleteOnu"
-          body: "*"
-        };
-    }
-
-    rpc OmciMsgOut(OmciMsg) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/OmciMsgOut"
-          body: "*"
-        };
-    }
-
-    rpc OnuPacketOut(OnuPacket) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/OnuPacketOut"
-          body: "*"
-        };
-    }
-
-    rpc UplinkPacketOut(UplinkPacket) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/UplinkPacketOut"
-          body: "*"
-        };
-    }
-
-    rpc FlowAdd(Flow) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/FlowAdd"
-          body: "*"
-        };
-    }
-
-    rpc FlowRemove(Flow) returns (Empty) {
-        option (google.api.http) = {
-          post: "/v1/FlowRemove"
-          body: "*"
-        };
-    }
-
-    rpc HeartbeatCheck(Empty) returns (Heartbeat) {
-        option (google.api.http) = {
-          post: "/v1/HeartbeatCheck"
-          body: "*"
-        };
-    }
-
-    rpc EnablePonIf(Interface) returns (Empty) {
-        option (google.api.http) = {
-            post: "/v1/EnablePonIf"
-            body: "*"
-        };
-    }
-
-    rpc DisablePonIf(Interface) returns (Empty) {
-        option (google.api.http) = {
-            post: "/v1/DisablePonIf"
-            body: "*"
-        };
-    }
-
-    rpc GetDeviceInfo(Empty) returns (DeviceInfo) {
-        option (google.api.http) = {
-            post: "/v1/GetDeviceInfo"
-            body: "*"
-        };
-    }
-
-    rpc Reboot(Empty) returns (Empty) {
-         option (google.api.http) = {
-            post: "/v1/Reboot"
-            body: "*"
-        };
-    }
-
-    rpc CollectStatistics(Empty) returns (Empty) {
-        option (google.api.http) = {
-            post: "/v1/CollectStatistics"
-            body: "*"
-        };
-    }
-
-    rpc CreateTconts(Tconts) returns (Empty) {
-        option (google.api.http) = {
-            post: "/v1/CreateTconts"
-            body: "*"
-        };
-    }
-
-    rpc RemoveTconts(Tconts) returns (Empty) {
-        option (google.api.http) = {
-            post: "/v1/RemoveTconts"
-            body: "*"
-        };
-    }
-
-    rpc EnableIndication(Empty) returns (stream Indication) {}
-}
-
-message Indication {
-    oneof data {
-        OltIndication olt_ind = 1;
-        IntfIndication intf_ind = 2;
-        IntfOperIndication intf_oper_ind = 3;
-        OnuDiscIndication onu_disc_ind = 4;
-        OnuIndication onu_ind = 5;
-        OmciIndication omci_ind = 6;
-        PacketIndication pkt_ind = 7;
-        PortStatistics port_stats = 8;
-        FlowStatistics flow_stats = 9;
-        AlarmIndication alarm_ind= 10;
-    }
-}
-
-message AlarmIndication {
-    oneof data {
-        LosIndication los_ind = 1;
-        DyingGaspIndication dying_gasp_ind = 2;
-        OnuAlarmIndication onu_alarm_ind = 3;
-        OnuStartupFailureIndication onu_startup_fail_ind = 4;
-        OnuSignalDegradeIndication onu_signal_degrade_ind = 5;
-        OnuDriftOfWindowIndication onu_drift_of_window_ind = 6;
-        OnuLossOfOmciChannelIndication onu_loss_omci_ind = 7;
-        OnuSignalsFailureIndication onu_signals_fail_ind = 8;
-        OnuTransmissionInterferenceWarning onu_tiwi_ind = 9;
-        OnuActivationFailureIndication onu_activation_fail_ind = 10;
-        OnuProcessingErrorIndication onu_processing_error_ind = 11;
-    }
-}
-
-message OltIndication {
-    string oper_state = 1;	// up, down
-}
-
-message IntfIndication {
-    fixed32 intf_id = 1;
-    string oper_state = 2;      // up, down
-}
-
-message OnuDiscIndication {
-    fixed32 intf_id = 1;
-    SerialNumber serial_number = 2;
-}
-
-message OnuIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string oper_state = 3;      // up, down
-    string admin_state = 5;     // up, down
-    SerialNumber serial_number = 4;
-}
-
-message IntfOperIndication {
-    string type = 1;		// nni, pon
-    fixed32 intf_id = 2;
-    string oper_state = 3;      // up, down
-}
-
-message OmciIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    bytes pkt = 3;
-}
-
-message PacketIndication {
-    string intf_type = 5;		// nni, pon, unknown
-    fixed32 intf_id = 1;
-    fixed32 gemport_id = 2;
-    fixed32 flow_id = 3;
-    fixed32 port_no = 6;
-    fixed64 cookie = 7;
-    bytes pkt = 4;
-}
-
-message Interface {
-    fixed32 intf_id = 1;
-}
-
-message Heartbeat {
-    fixed32 heartbeat_signature = 1;
-}
-
-message Onu {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    SerialNumber serial_number = 3;
-    fixed32 pir = 4;   // peak information rate assigned to onu
-}
-
-message OmciMsg {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    bytes pkt = 3;
-}
-
-message OnuPacket {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    fixed32 port_no = 4;
-    bytes pkt = 3;
-}
-
-message UplinkPacket {
-    fixed32 intf_id = 1;
-    bytes pkt = 2;
-}
-
-message DeviceInfo {
-    string vendor = 1;
-    string model = 2;
-    string hardware_version = 3;
-    string firmware_version = 4;
-    string device_id = 16;
-    string device_serial_number = 17;
-
-    // Total number of pon intf ports on the device
-    fixed32 pon_ports = 12;
-
-    // If using global per-device technology profile. To be deprecated
-    string technology = 5;
-    fixed32 onu_id_start = 6;
-    fixed32 onu_id_end = 7;
-    fixed32 alloc_id_start = 8;
-    fixed32 alloc_id_end = 9;
-    fixed32 gemport_id_start = 10;
-    fixed32 gemport_id_end = 11;
-    fixed32 flow_id_start = 13;
-    fixed32 flow_id_end = 14;
-
-    message DeviceResourceRanges {
-
-        // List of 0 or more intf_ids that use the same technology and pools.
-        // If 0 intf_ids supplied, it implies ALL interfaces
-        repeated fixed32 intf_ids = 1;
-
-        // Technology profile for this pool
-        string technology = 2;
-
-        message Pool {
-            enum PoolType {
-                ONU_ID = 0;
-                ALLOC_ID = 1;
-                GEMPORT_ID = 2;
-                FLOW_ID = 3;
-            }
-
-            enum SharingType {
-                DEDICATED_PER_INTF = 0;
-                SHARED_BY_ALL_INTF_ALL_TECH = 1; // Shared across all interfaces in all technologies in all ranges
-                SHARED_BY_ALL_INTF_SAME_TECH = 2; // Shared across all interfaces of the same technology used in this range
-            }
-
-            PoolType type = 1;
-	    SharingType sharing = 2;
-	    fixed32 start = 3; // lower bound on IDs allocated from this pool
-	    fixed32 end = 4; // upper bound on IDs allocated from this pool
-	}
-        repeated Pool pools = 3;
-    }
-    repeated DeviceResourceRanges ranges = 15;
-}
-
-message Classifier {
-    fixed32 o_tpid = 1;
-    fixed32 o_vid = 2;
-    fixed32 i_tpid = 3;
-    fixed32 i_vid = 4;
-    fixed32 o_pbits = 5;
-    fixed32 i_pbits = 6;
-    fixed32 eth_type = 7;
-    bytes dst_mac = 8;
-    bytes src_mac = 9;
-    fixed32 ip_proto = 10;
-    fixed32 dst_ip = 11;
-    fixed32 src_ip = 12;
-    fixed32 src_port = 13;
-    fixed32 dst_port = 14;
-    string pkt_tag_type = 15;	// untagged, single_tag, double_tag
-}
-
-message ActionCmd {
-    bool add_outer_tag = 1;
-    bool remove_outer_tag = 2;
-    bool trap_to_host = 3;
-}
-
-message Action {
-    ActionCmd cmd = 1;
-    fixed32 o_vid = 2;
-    fixed32 o_pbits = 3;
-    fixed32 o_tpid = 4;
-    fixed32 i_vid = 5;
-    fixed32 i_pbits = 6;
-    fixed32 i_tpid = 7;
-}
-
-message Flow {
-    sfixed32 access_intf_id = 1;
-    sfixed32 onu_id = 2;
-    sfixed32 uni_id = 11;
-    fixed32 flow_id = 3;
-    string flow_type = 4;	// upstream, downstream, broadcast, multicast
-    sfixed32 alloc_id = 10;
-    sfixed32 network_intf_id = 5;
-    sfixed32 gemport_id = 6;
-    Classifier classifier = 7;
-    Action action = 8;
-    sfixed32 priority = 9;
-    fixed64 cookie = 12; // must be provided for any flow with trap_to_host action. Returned in PacketIndication
-    fixed32 port_no = 13; // must be provided for any flow with trap_to_host action. Returned in PacketIndication
-}
-
-message SerialNumber {
-    bytes vendor_id = 1;
-    bytes vendor_specific = 2;
-}
-
-message PortStatistics {
-    fixed32 intf_id = 1;
-    fixed64 rx_bytes = 2;
-    fixed64 rx_packets = 3;
-    fixed64 rx_ucast_packets = 4;
-    fixed64 rx_mcast_packets = 5;
-    fixed64 rx_bcast_packets = 6;
-    fixed64 rx_error_packets = 7;
-    fixed64 tx_bytes = 8;
-    fixed64 tx_packets = 9;
-    fixed64 tx_ucast_packets = 10;
-    fixed64 tx_mcast_packets = 11;
-    fixed64 tx_bcast_packets = 12;
-    fixed64 tx_error_packets = 13;
-    fixed64 rx_crc_errors = 14;
-    fixed64 bip_errors = 15;
-    fixed32 timestamp = 16;
-}
-
-message FlowStatistics {
-    fixed32 flow_id = 1;
-    fixed64 rx_bytes = 2;
-    fixed64 rx_packets = 3;
-    fixed64 tx_bytes = 8;
-    fixed64 tx_packets = 9;
-    fixed32 timestamp = 16;
-}
-
-message LosIndication {
-    fixed32 intf_id = 1;
-    string status = 2;
-}
-
-message DyingGaspIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-}
-
-message OnuAlarmIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string los_status = 3;
-    string lob_status = 4;
-    string lopc_miss_status = 5;
-    string lopc_mic_error_status = 6;
-}
-
-message OnuStartupFailureIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-}
-
-message OnuSignalDegradeIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-    fixed32 inverse_bit_error_rate = 4;
-}
-
-message OnuDriftOfWindowIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-    fixed32 drift = 4;
-    fixed32 new_eqd = 5;
-}
-
-message OnuLossOfOmciChannelIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-}
-
-message OnuSignalsFailureIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-    fixed32 inverse_bit_error_rate = 4;
-}
-
-message OnuTransmissionInterferenceWarning {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    string status = 3;
-    fixed32 drift = 4;
-}
-
-message OnuActivationFailureIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-}
-
-message OnuProcessingErrorIndication {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-}
-
-enum Direction {
-    UPSTREAM = 0;
-    DOWNSTREAM = 1;
-    BIDIRECTIONAL = 2;
-}
-
-enum SchedulingPolicy {
-    WRR = 0;
-    StrictPriority = 1;
-    Hybrid = 2;
-}
-
-enum AdditionalBW {
-    AdditionalBW_None = 0;
-    AdditionalBW_NA = 1;
-    AdditionalBW_BestEffort = 2;
-    AdditionalBW_Auto = 3;
-}
-
-enum DiscardPolicy {
-    TailDrop = 0;
-    WTailDrop = 1;
-    Red = 2;
-    WRed = 3;
-}
-
-enum InferredAdditionBWIndication {
-    InferredAdditionBWIndication_None = 0;
-    InferredAdditionBWIndication_Assured = 1;
-    InferredAdditionBWIndication_BestEffort = 2;
-}
-
-message Scheduler {
-    Direction direction = 1;
-    AdditionalBW additional_bw = 2; // Valid on for “direction == Upstream”.
-    fixed32 priority = 3;
-    fixed32 weight = 4;
-    SchedulingPolicy sched_policy = 5;
-}
-
-message TrafficShapingInfo {
-    fixed32 cir = 1;
-    fixed32 cbs = 2;
-    fixed32 pir = 3;
-    fixed32 pbs = 4;
-    fixed32 gir = 5; // only if “direction == Upstream ”
-    InferredAdditionBWIndication add_bw_ind = 6; // only if “direction == Upstream”
-}
-
-message Tcont {
-    Direction direction = 1;
-    fixed32 alloc_id = 2; // valid only if “direction == Upstream ”
-    Scheduler scheduler = 3;
-    TrafficShapingInfo traffic_shaping_info = 4;
-}
-
-message Tconts {
-    fixed32 intf_id = 1;
-    fixed32 onu_id = 2;
-    fixed32 uni_id = 4;
-    fixed32 port_no = 5;
-    repeated Tcont tconts = 3;
-}
-
-message TailDropDiscardConfig {
-    fixed32 queue_size = 1;
-}
-
-message RedDiscardConfig {
-    fixed32 min_threshold = 1;
-    fixed32 max_threshold = 2;
-    fixed32 max_probability = 3;
-}
-
-message WRedDiscardConfig {
-    RedDiscardConfig green = 1;
-    RedDiscardConfig yellow = 2;
-    RedDiscardConfig red = 3;
-}
-
-message DiscardConfig {
-    DiscardPolicy discard_policy = 1;
-    oneof discard_config {
-        TailDropDiscardConfig tail_drop_discard_config = 2;
-        RedDiscardConfig red_discard_config = 3;
-        WRedDiscardConfig wred_discard_config = 4;
-    }
-}
-
-message Empty {}
diff --git a/vendor/github.com/DataDog/zstd/.travis.yml b/vendor/github.com/DataDog/zstd/.travis.yml
new file mode 100644
index 0000000..c5aa33d
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/.travis.yml
@@ -0,0 +1,30 @@
+language: go
+
+go:
+  - 1.9.x
+  - 1.10.x
+  - 1.11.x
+
+os:
+  - linux
+  - osx
+
+matrix:
+  include:
+    name: "Go 1.11.x CentOS 32bits"
+    language: go
+    go: 1.11.x
+    os: linux
+    services:
+      - docker
+    script:
+      # Please update Go version in travis_test_32 as needed
+      - "docker run -i -v \"${PWD}:/zstd\" toopher/centos-i386:centos6 /bin/bash -c \"linux32 --32bit i386 /zstd/travis_test_32.sh\""
+
+install:
+  - "wget https://github.com/DataDog/zstd/files/2246767/mr.zip"
+  - "unzip mr.zip"
+script:
+  - "go build"
+  - "PAYLOAD=`pwd`/mr go test -v"
+  - "PAYLOAD=`pwd`/mr go test -bench ."
diff --git a/vendor/github.com/DataDog/zstd/LICENSE b/vendor/github.com/DataDog/zstd/LICENSE
new file mode 100644
index 0000000..345c1eb
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/LICENSE
@@ -0,0 +1,27 @@
+Simplified BSD License
+
+Copyright (c) 2016, Datadog <info@datadoghq.com>
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+      this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice,
+      this list of conditions and the following disclaimer in the documentation
+      and/or other materials provided with the distribution.
+    * Neither the name of the copyright holder nor the names of its contributors
+      may be used to endorse or promote products derived from this software
+      without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/vendor/github.com/DataDog/zstd/README.md b/vendor/github.com/DataDog/zstd/README.md
new file mode 100644
index 0000000..6c02e16
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/README.md
@@ -0,0 +1,120 @@
+# Zstd Go Wrapper
+
+[C Zstd Homepage](https://github.com/Cyan4973/zstd)
+
+The current headers and C files are from *v1.3.4* (Commit
+[2555975](https://github.com/facebook/zstd/releases/tag/v1.3.4)).
+
+## Usage
+
+There are two main APIs:
+
+* simple Compress/Decompress
+* streaming API (io.Reader/io.Writer)
+
+The compress/decompress APIs mirror that of lz4, while the streaming API was
+designed to be a drop-in replacement for zlib.
+
+### Simple `Compress/Decompress`
+
+
+```go
+// Compress compresses the byte array given in src and writes it to dst.
+// If you already have a buffer allocated, you can pass it to prevent allocation
+// If not, you can pass nil as dst.
+// If the buffer is too small, it will be reallocated, resized, and returned bu the function
+// If dst is nil, this will allocate the worst case size (CompressBound(src))
+Compress(dst, src []byte) ([]byte, error)
+```
+
+```go
+// CompressLevel is the same as Compress but you can pass another compression level
+CompressLevel(dst, src []byte, level int) ([]byte, error)
+```
+
+```go
+// Decompress will decompress your payload into dst.
+// If you already have a buffer allocated, you can pass it to prevent allocation
+// If not, you can pass nil as dst (allocates a 4*src size as default).
+// If the buffer is too small, it will retry 3 times by doubling the dst size
+// After max retries, it will switch to the slower stream API to be sure to be able
+// to decompress. Currently switches if compression ratio > 4*2**3=32.
+Decompress(dst, src []byte) ([]byte, error)
+```
+
+### Stream API
+
+```go
+// NewWriter creates a new object that can optionally be initialized with
+// a precomputed dictionary. If dict is nil, compress without a dictionary.
+// The dictionary array should not be changed during the use of this object.
+// You MUST CALL Close() to write the last bytes of a zstd stream and free C objects.
+NewWriter(w io.Writer) *Writer
+NewWriterLevel(w io.Writer, level int) *Writer
+NewWriterLevelDict(w io.Writer, level int, dict []byte) *Writer
+
+// Write compresses the input data and write it to the underlying writer
+(w *Writer) Write(p []byte) (int, error)
+
+// Close flushes the buffer and frees C zstd objects
+(w *Writer) Close() error
+```
+
+```go
+// NewReader returns a new io.ReadCloser that will decompress data from the
+// underlying reader.  If a dictionary is provided to NewReaderDict, it must
+// not be modified until Close is called.  It is the caller's responsibility
+// to call Close, which frees up C objects.
+NewReader(r io.Reader) io.ReadCloser
+NewReaderDict(r io.Reader, dict []byte) io.ReadCloser
+```
+
+### Benchmarks (benchmarked with v0.5.0)
+
+The author of Zstd also wrote lz4. Zstd is intended to occupy a speed/ratio
+level similar to what zlib currently provides.  In our tests, the can always
+be made to be better than zlib by chosing an appropriate level while still
+keeping compression and decompression time faster than zlib.
+
+You can run the benchmarks against your own payloads by using the Go benchmarks tool.
+Just export your payload filepath as the `PAYLOAD` environment variable and run the benchmarks:
+
+```go
+go test -bench .
+```
+
+Compression of a 7Mb pdf zstd (this wrapper) vs [czlib](https://github.com/DataDog/czlib):
+```
+BenchmarkCompression               5     221056624 ns/op      67.34 MB/s
+BenchmarkDecompression           100      18370416 ns/op     810.32 MB/s
+
+BenchmarkFzlibCompress             2     610156603 ns/op      24.40 MB/s
+BenchmarkFzlibDecompress          20      81195246 ns/op     183.33 MB/s
+```
+
+Ratio is also better by a margin of ~20%.
+Compression speed is always better than zlib on all the payloads we tested;
+However, [czlib](https://github.com/DataDog/czlib) has optimisations that make it
+faster at decompressiong small payloads:
+
+```
+Testing with size: 11... czlib: 8.97 MB/s, zstd: 3.26 MB/s
+Testing with size: 27... czlib: 23.3 MB/s, zstd: 8.22 MB/s
+Testing with size: 62... czlib: 31.6 MB/s, zstd: 19.49 MB/s
+Testing with size: 141... czlib: 74.54 MB/s, zstd: 42.55 MB/s
+Testing with size: 323... czlib: 155.14 MB/s, zstd: 99.39 MB/s
+Testing with size: 739... czlib: 235.9 MB/s, zstd: 216.45 MB/s
+Testing with size: 1689... czlib: 116.45 MB/s, zstd: 345.64 MB/s
+Testing with size: 3858... czlib: 176.39 MB/s, zstd: 617.56 MB/s
+Testing with size: 8811... czlib: 254.11 MB/s, zstd: 824.34 MB/s
+Testing with size: 20121... czlib: 197.43 MB/s, zstd: 1339.11 MB/s
+Testing with size: 45951... czlib: 201.62 MB/s, zstd: 1951.57 MB/s
+```
+
+zstd starts to shine with payloads > 1KB
+
+### Stability - Current state: STABLE
+
+The C library seems to be pretty stable and according to the author has been tested and fuzzed.
+
+For the Go wrapper, the test cover most usual cases and we have succesfully tested it on all staging and prod data.
diff --git a/vendor/github.com/DataDog/zstd/ZSTD_LICENSE b/vendor/github.com/DataDog/zstd/ZSTD_LICENSE
new file mode 100644
index 0000000..a793a80
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/ZSTD_LICENSE
@@ -0,0 +1,30 @@
+BSD License
+
+For Zstandard software
+
+Copyright (c) 2016-present, Facebook, Inc. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+ * Neither the name Facebook nor the names of its contributors may be used to
+   endorse or promote products derived from this software without specific
+   prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/vendor/github.com/DataDog/zstd/bitstream.h b/vendor/github.com/DataDog/zstd/bitstream.h
new file mode 100644
index 0000000..f7f389f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/bitstream.h
@@ -0,0 +1,471 @@
+/* ******************************************************************
+   bitstream
+   Part of FSE library
+   header file (to include)
+   Copyright (C) 2013-2017, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*
+*  This API consists of small unitary functions, which must be inlined for best performance.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+/*-****************************************
+*  Dependencies
+******************************************/
+#include "mem.h"            /* unaligned access routines */
+#include "error_private.h"  /* error codes and messages */
+
+
+/*-*************************************
+*  Debug
+***************************************/
+#if defined(BIT_DEBUG) && (BIT_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
+
+/*=========================================
+*  Target specific
+=========================================*/
+#if defined(__BMI__) && defined(__GNUC__)
+#  include <immintrin.h>   /* support for bextr (experimental) */
+#endif
+
+#define STREAM_ACCUMULATOR_MIN_32  25
+#define STREAM_ACCUMULATOR_MIN_64  57
+#define STREAM_ACCUMULATOR_MIN    ((U32)(MEM_32bits() ? STREAM_ACCUMULATOR_MIN_32 : STREAM_ACCUMULATOR_MIN_64))
+
+
+/*-******************************************
+*  bitStream encoding API (write forward)
+********************************************/
+/* bitStream can mix input from multiple sources.
+ * A critical property of these streams is that they encode and decode in **reverse** direction.
+ * So the first bit sequence you add will be the last to be read, like a LIFO stack.
+ */
+typedef struct
+{
+    size_t bitContainer;
+    unsigned bitPos;
+    char*  startPtr;
+    char*  ptr;
+    char*  endPtr;
+} BIT_CStream_t;
+
+MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC, void* dstBuffer, size_t dstCapacity);
+MEM_STATIC void   BIT_addBits(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+MEM_STATIC void   BIT_flushBits(BIT_CStream_t* bitC);
+MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC);
+
+/* Start with initCStream, providing the size of buffer to write into.
+*  bitStream will never write outside of this buffer.
+*  `dstCapacity` must be >= sizeof(bitD->bitContainer), otherwise @return will be an error code.
+*
+*  bits are first added to a local register.
+*  Local register is size_t, hence 64-bits on 64-bits systems, or 32-bits on 32-bits systems.
+*  Writing data into memory is an explicit operation, performed by the flushBits function.
+*  Hence keep track how many bits are potentially stored into local register to avoid register overflow.
+*  After a flushBits, a maximum of 7 bits might still be stored into local register.
+*
+*  Avoid storing elements of more than 24 bits if you want compatibility with 32-bits bitstream readers.
+*
+*  Last operation is to close the bitStream.
+*  The function returns the final size of CStream in bytes.
+*  If data couldn't fit into `dstBuffer`, it will return a 0 ( == not storable)
+*/
+
+
+/*-********************************************
+*  bitStream decoding API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+    const char* limitPtr;
+} BIT_DStream_t;
+
+typedef enum { BIT_DStream_unfinished = 0,
+               BIT_DStream_endOfBuffer = 1,
+               BIT_DStream_completed = 2,
+               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
+
+
+/* Start by invoking BIT_initDStream().
+*  A chunk of the bitStream is then stored into a local register.
+*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
+*  You can then retrieve bitFields stored into the local register, **in reverse order**.
+*  Local register is explicitly reloaded from memory by the BIT_reloadDStream() method.
+*  A reload guarantee a minimum of ((8*sizeof(bitD->bitContainer))-7) bits when its result is BIT_DStream_unfinished.
+*  Otherwise, it can be less than that, so proceed accordingly.
+*  Checking if DStream has reached its end can be performed with BIT_endOfDStream().
+*/
+
+
+/*-****************************************
+*  unsafe API
+******************************************/
+MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+/* faster, but works only if value is "clean", meaning all high bits above nbBits are 0 */
+
+MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
+/* unsafe version; does not check buffer overflow */
+
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/*-**************************************************************
+*  Internal functions
+****************************************************************/
+MEM_STATIC unsigned BIT_highbit32 (U32 val)
+{
+    assert(val != 0);
+    {
+#   if defined(_MSC_VER)   /* Visual */
+        unsigned long r=0;
+        _BitScanReverse ( &r, val );
+        return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+        return 31 - __builtin_clz (val);
+#   else   /* Software version */
+        static const unsigned DeBruijnClz[32] = { 0,  9,  1, 10, 13, 21,  2, 29,
+                                                 11, 14, 16, 18, 22, 25,  3, 30,
+                                                  8, 12, 20, 28, 15, 17, 24,  7,
+                                                 19, 27, 23,  6, 26,  5,  4, 31 };
+        U32 v = val;
+        v |= v >> 1;
+        v |= v >> 2;
+        v |= v >> 4;
+        v |= v >> 8;
+        v |= v >> 16;
+        return DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+#   endif
+    }
+}
+
+/*=====    Local Constants   =====*/
+static const unsigned BIT_mask[] = {
+    0,          1,         3,         7,         0xF,       0x1F,
+    0x3F,       0x7F,      0xFF,      0x1FF,     0x3FF,     0x7FF,
+    0xFFF,      0x1FFF,    0x3FFF,    0x7FFF,    0xFFFF,    0x1FFFF,
+    0x3FFFF,    0x7FFFF,   0xFFFFF,   0x1FFFFF,  0x3FFFFF,  0x7FFFFF,
+    0xFFFFFF,   0x1FFFFFF, 0x3FFFFFF, 0x7FFFFFF, 0xFFFFFFF, 0x1FFFFFFF,
+    0x3FFFFFFF, 0x7FFFFFFF}; /* up to 31 bits */
+#define BIT_MASK_SIZE (sizeof(BIT_mask) / sizeof(BIT_mask[0]))
+
+/*-**************************************************************
+*  bitStream encoding
+****************************************************************/
+/*! BIT_initCStream() :
+ *  `dstCapacity` must be > sizeof(size_t)
+ *  @return : 0 if success,
+ *            otherwise an error code (can be tested using ERR_isError()) */
+MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
+                                  void* startPtr, size_t dstCapacity)
+{
+    bitC->bitContainer = 0;
+    bitC->bitPos = 0;
+    bitC->startPtr = (char*)startPtr;
+    bitC->ptr = bitC->startPtr;
+    bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer);
+    if (dstCapacity <= sizeof(bitC->bitContainer)) return ERROR(dstSize_tooSmall);
+    return 0;
+}
+
+/*! BIT_addBits() :
+ *  can add up to 31 bits into `bitC`.
+ *  Note : does not check for register overflow ! */
+MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
+                            size_t value, unsigned nbBits)
+{
+    MEM_STATIC_ASSERT(BIT_MASK_SIZE == 32);
+    assert(nbBits < BIT_MASK_SIZE);
+    assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+    bitC->bitContainer |= (value & BIT_mask[nbBits]) << bitC->bitPos;
+    bitC->bitPos += nbBits;
+}
+
+/*! BIT_addBitsFast() :
+ *  works only if `value` is _clean_, meaning all high bits above nbBits are 0 */
+MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC,
+                                size_t value, unsigned nbBits)
+{
+    assert((value>>nbBits) == 0);
+    assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+    bitC->bitContainer |= value << bitC->bitPos;
+    bitC->bitPos += nbBits;
+}
+
+/*! BIT_flushBitsFast() :
+ *  assumption : bitContainer has not overflowed
+ *  unsafe version; does not check buffer overflow */
+MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC)
+{
+    size_t const nbBytes = bitC->bitPos >> 3;
+    assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+    MEM_writeLEST(bitC->ptr, bitC->bitContainer);
+    bitC->ptr += nbBytes;
+    assert(bitC->ptr <= bitC->endPtr);
+    bitC->bitPos &= 7;
+    bitC->bitContainer >>= nbBytes*8;
+}
+
+/*! BIT_flushBits() :
+ *  assumption : bitContainer has not overflowed
+ *  safe version; check for buffer overflow, and prevents it.
+ *  note : does not signal buffer overflow.
+ *  overflow will be revealed later on using BIT_closeCStream() */
+MEM_STATIC void BIT_flushBits(BIT_CStream_t* bitC)
+{
+    size_t const nbBytes = bitC->bitPos >> 3;
+    assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+    MEM_writeLEST(bitC->ptr, bitC->bitContainer);
+    bitC->ptr += nbBytes;
+    if (bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
+    bitC->bitPos &= 7;
+    bitC->bitContainer >>= nbBytes*8;
+}
+
+/*! BIT_closeCStream() :
+ *  @return : size of CStream, in bytes,
+ *            or 0 if it could not fit into dstBuffer */
+MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC)
+{
+    BIT_addBitsFast(bitC, 1, 1);   /* endMark */
+    BIT_flushBits(bitC);
+    if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
+    return (bitC->ptr - bitC->startPtr) + (bitC->bitPos > 0);
+}
+
+
+/*-********************************************************
+*  bitStream decoding
+**********************************************************/
+/*! BIT_initDStream() :
+ *  Initialize a BIT_DStream_t.
+ * `bitD` : a pointer to an already allocated BIT_DStream_t structure.
+ * `srcSize` must be the *exact* size of the bitStream, in bytes.
+ * @return : size of stream (== srcSize), or an errorCode if a problem is detected
+ */
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    bitD->start = (const char*)srcBuffer;
+    bitD->limitPtr = bitD->start + sizeof(bitD->bitContainer);
+
+    if (srcSize >=  sizeof(bitD->bitContainer)) {  /* normal case */
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+          bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
+          if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
+    } else {
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+        case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
+                /* fall-through */
+
+        case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
+                /* fall-through */
+
+        case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
+                /* fall-through */
+
+        case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
+                /* fall-through */
+
+        case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
+                /* fall-through */
+
+        case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
+                /* fall-through */
+
+        default: break;
+        }
+        {   BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+            bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;
+            if (lastByte == 0) return ERROR(corruption_detected);  /* endMark not present */
+        }
+        bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+MEM_STATIC size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
+{
+    return bitContainer >> start;
+}
+
+MEM_STATIC size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
+{
+#if defined(__BMI__) && defined(__GNUC__) && __GNUC__*1000+__GNUC_MINOR__ >= 4008  /* experimental */
+#  if defined(__x86_64__)
+    if (sizeof(bitContainer)==8)
+        return _bextr_u64(bitContainer, start, nbBits);
+    else
+#  endif
+        return _bextr_u32(bitContainer, start, nbBits);
+#else
+    assert(nbBits < BIT_MASK_SIZE);
+    return (bitContainer >> start) & BIT_mask[nbBits];
+#endif
+}
+
+MEM_STATIC size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
+{
+    assert(nbBits < BIT_MASK_SIZE);
+    return bitContainer & BIT_mask[nbBits];
+}
+
+/*! BIT_lookBits() :
+ *  Provides next n bits from local register.
+ *  local register is not modified.
+ *  On 32-bits, maxNbBits==24.
+ *  On 64-bits, maxNbBits==56.
+ * @return : value extracted */
+MEM_STATIC size_t BIT_lookBits(const BIT_DStream_t* bitD, U32 nbBits)
+{
+#if defined(__BMI__) && defined(__GNUC__)   /* experimental; fails if bitD->bitsConsumed + nbBits > sizeof(bitD->bitContainer)*8 */
+    return BIT_getMiddleBits(bitD->bitContainer, (sizeof(bitD->bitContainer)*8) - bitD->bitsConsumed - nbBits, nbBits);
+#else
+    U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> 1) >> ((regMask-nbBits) & regMask);
+#endif
+}
+
+/*! BIT_lookBitsFast() :
+ *  unsafe version; only works if nbBits >= 1 */
+MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
+{
+    U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
+    assert(nbBits >= 1);
+    return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
+}
+
+MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+/*! BIT_readBits() :
+ *  Read (consume) next n bits from local register and update.
+ *  Pay attention to not read more than nbBits contained into local register.
+ * @return : extracted value. */
+MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BIT_lookBits(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*! BIT_readBitsFast() :
+ *  unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BIT_lookBitsFast(bitD, nbBits);
+    assert(nbBits >= 1);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*! BIT_reloadDStream() :
+ *  Refill `bitD` from buffer previously set in BIT_initDStream() .
+ *  This function is safe, it guarantees it will not read beyond src buffer.
+ * @return : status of `BIT_DStream_t` internal register.
+ *           when status == BIT_DStream_unfinished, internal register is filled with at least 25 or 57 bits */
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* overflow detected, like end of stream */
+        return BIT_DStream_overflow;
+
+    if (bitD->ptr >= bitD->limitPtr) {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BIT_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start) {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
+        return BIT_DStream_completed;
+    }
+    /* start < ptr < limitPtr */
+    {   U32 nbBytes = bitD->bitsConsumed >> 3;
+        BIT_DStream_status result = BIT_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start) {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BIT_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD->bitContainer), otherwise bitD->ptr == bitD->start */
+        return result;
+    }
+}
+
+/*! BIT_endOfDStream() :
+ * @return : 1 if DStream has _exactly_ reached its end (all bits consumed).
+ */
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
diff --git a/vendor/github.com/DataDog/zstd/compiler.h b/vendor/github.com/DataDog/zstd/compiler.h
new file mode 100644
index 0000000..e90a3bc
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/compiler.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_COMPILER_H
+#define ZSTD_COMPILER_H
+
+/*-*******************************************************
+*  Compiler specifics
+*********************************************************/
+/* force inlining */
+#if defined (__GNUC__) || defined(__cplusplus) || defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#  define INLINE_KEYWORD inline
+#else
+#  define INLINE_KEYWORD
+#endif
+
+#if defined(__GNUC__)
+#  define FORCE_INLINE_ATTR __attribute__((always_inline))
+#elif defined(_MSC_VER)
+#  define FORCE_INLINE_ATTR __forceinline
+#else
+#  define FORCE_INLINE_ATTR
+#endif
+
+/**
+ * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
+ * parameters. They must be inlined for the compiler to elimininate the constant
+ * branches.
+ */
+#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR
+/**
+ * HINT_INLINE is used to help the compiler generate better code. It is *not*
+ * used for "templates", so it can be tweaked based on the compilers
+ * performance.
+ *
+ * gcc-4.8 and gcc-4.9 have been shown to benefit from leaving off the
+ * always_inline attribute.
+ *
+ * clang up to 5.0.0 (trunk) benefit tremendously from the always_inline
+ * attribute.
+ */
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 8 && __GNUC__ < 5
+#  define HINT_INLINE static INLINE_KEYWORD
+#else
+#  define HINT_INLINE static INLINE_KEYWORD FORCE_INLINE_ATTR
+#endif
+
+/* force no inlining */
+#ifdef _MSC_VER
+#  define FORCE_NOINLINE static __declspec(noinline)
+#else
+#  ifdef __GNUC__
+#    define FORCE_NOINLINE static __attribute__((__noinline__))
+#  else
+#    define FORCE_NOINLINE static
+#  endif
+#endif
+
+/* target attribute */
+#ifndef __has_attribute
+  #define __has_attribute(x) 0  /* Compatibility with non-clang compilers. */
+#endif
+#if defined(__GNUC__)
+#  define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
+#else
+#  define TARGET_ATTRIBUTE(target)
+#endif
+
+/* Enable runtime BMI2 dispatch based on the CPU.
+ * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ */
+#ifndef DYNAMIC_BMI2
+  #if (defined(__clang__) && __has_attribute(__target__)) \
+      || (defined(__GNUC__) \
+          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))) \
+      && (defined(__x86_64__) || defined(_M_X86)) \
+      && !defined(__BMI2__)
+  #  define DYNAMIC_BMI2 1
+  #else
+  #  define DYNAMIC_BMI2 0
+  #endif
+#endif
+
+/* prefetch */
+#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
+#  include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#  define PREFETCH(ptr)   _mm_prefetch((const char*)ptr, _MM_HINT_T0)
+#elif defined(__GNUC__)
+#  define PREFETCH(ptr)   __builtin_prefetch(ptr, 0, 0)
+#else
+#  define PREFETCH(ptr)   /* disabled */
+#endif
+
+/* disable warnings */
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4100)        /* disable: C4100: unreferenced formal parameter */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4204)        /* disable: C4204: non-constant aggregate initializer */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+#endif /* ZSTD_COMPILER_H */
diff --git a/vendor/github.com/DataDog/zstd/cover.c b/vendor/github.com/DataDog/zstd/cover.c
new file mode 100644
index 0000000..b5a3957
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/cover.c
@@ -0,0 +1,1048 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* *****************************************************************************
+ * Constructs a dictionary using a heuristic based on the following paper:
+ *
+ * Liao, Petri, Moffat, Wirth
+ * Effective Construction of Relative Lempel-Ziv Dictionaries
+ * Published in WWW 2016.
+ *
+ * Adapted from code originally written by @ot (Giuseppe Ottaviano).
+ ******************************************************************************/
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include <stdio.h>  /* fprintf */
+#include <stdlib.h> /* malloc, free, qsort */
+#include <string.h> /* memset */
+#include <time.h>   /* clock */
+
+#include "mem.h" /* read */
+#include "pool.h"
+#include "threading.h"
+#include "zstd_internal.h" /* includes zstd.h */
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#define ZDICT_STATIC_LINKING_ONLY
+#endif
+#include "zdict.h"
+
+/*-*************************************
+*  Constants
+***************************************/
+#define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((U32)-1) : ((U32)1 GB))
+
+/*-*************************************
+*  Console display
+***************************************/
+static int g_displayLevel = 2;
+#define DISPLAY(...)                                                           \
+  {                                                                            \
+    fprintf(stderr, __VA_ARGS__);                                              \
+    fflush(stderr);                                                            \
+  }
+#define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
+  if (displayLevel >= l) {                                                     \
+    DISPLAY(__VA_ARGS__);                                                      \
+  } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
+#define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)
+
+#define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
+  if (displayLevel >= l) {                                                     \
+    if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) {             \
+      g_time = clock();                                                        \
+      DISPLAY(__VA_ARGS__);                                                    \
+    }                                                                          \
+  }
+#define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)
+static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
+static clock_t g_time = 0;
+
+/*-*************************************
+* Hash table
+***************************************
+* A small specialized hash map for storing activeDmers.
+* The map does not resize, so if it becomes full it will loop forever.
+* Thus, the map must be large enough to store every value.
+* The map implements linear probing and keeps its load less than 0.5.
+*/
+
+#define MAP_EMPTY_VALUE ((U32)-1)
+typedef struct COVER_map_pair_t_s {
+  U32 key;
+  U32 value;
+} COVER_map_pair_t;
+
+typedef struct COVER_map_s {
+  COVER_map_pair_t *data;
+  U32 sizeLog;
+  U32 size;
+  U32 sizeMask;
+} COVER_map_t;
+
+/**
+ * Clear the map.
+ */
+static void COVER_map_clear(COVER_map_t *map) {
+  memset(map->data, MAP_EMPTY_VALUE, map->size * sizeof(COVER_map_pair_t));
+}
+
+/**
+ * Initializes a map of the given size.
+ * Returns 1 on success and 0 on failure.
+ * The map must be destroyed with COVER_map_destroy().
+ * The map is only guaranteed to be large enough to hold size elements.
+ */
+static int COVER_map_init(COVER_map_t *map, U32 size) {
+  map->sizeLog = ZSTD_highbit32(size) + 2;
+  map->size = (U32)1 << map->sizeLog;
+  map->sizeMask = map->size - 1;
+  map->data = (COVER_map_pair_t *)malloc(map->size * sizeof(COVER_map_pair_t));
+  if (!map->data) {
+    map->sizeLog = 0;
+    map->size = 0;
+    return 0;
+  }
+  COVER_map_clear(map);
+  return 1;
+}
+
+/**
+ * Internal hash function
+ */
+static const U32 prime4bytes = 2654435761U;
+static U32 COVER_map_hash(COVER_map_t *map, U32 key) {
+  return (key * prime4bytes) >> (32 - map->sizeLog);
+}
+
+/**
+ * Helper function that returns the index that a key should be placed into.
+ */
+static U32 COVER_map_index(COVER_map_t *map, U32 key) {
+  const U32 hash = COVER_map_hash(map, key);
+  U32 i;
+  for (i = hash;; i = (i + 1) & map->sizeMask) {
+    COVER_map_pair_t *pos = &map->data[i];
+    if (pos->value == MAP_EMPTY_VALUE) {
+      return i;
+    }
+    if (pos->key == key) {
+      return i;
+    }
+  }
+}
+
+/**
+ * Returns the pointer to the value for key.
+ * If key is not in the map, it is inserted and the value is set to 0.
+ * The map must not be full.
+ */
+static U32 *COVER_map_at(COVER_map_t *map, U32 key) {
+  COVER_map_pair_t *pos = &map->data[COVER_map_index(map, key)];
+  if (pos->value == MAP_EMPTY_VALUE) {
+    pos->key = key;
+    pos->value = 0;
+  }
+  return &pos->value;
+}
+
+/**
+ * Deletes key from the map if present.
+ */
+static void COVER_map_remove(COVER_map_t *map, U32 key) {
+  U32 i = COVER_map_index(map, key);
+  COVER_map_pair_t *del = &map->data[i];
+  U32 shift = 1;
+  if (del->value == MAP_EMPTY_VALUE) {
+    return;
+  }
+  for (i = (i + 1) & map->sizeMask;; i = (i + 1) & map->sizeMask) {
+    COVER_map_pair_t *const pos = &map->data[i];
+    /* If the position is empty we are done */
+    if (pos->value == MAP_EMPTY_VALUE) {
+      del->value = MAP_EMPTY_VALUE;
+      return;
+    }
+    /* If pos can be moved to del do so */
+    if (((i - COVER_map_hash(map, pos->key)) & map->sizeMask) >= shift) {
+      del->key = pos->key;
+      del->value = pos->value;
+      del = pos;
+      shift = 1;
+    } else {
+      ++shift;
+    }
+  }
+}
+
+/**
+ * Destroyes a map that is inited with COVER_map_init().
+ */
+static void COVER_map_destroy(COVER_map_t *map) {
+  if (map->data) {
+    free(map->data);
+  }
+  map->data = NULL;
+  map->size = 0;
+}
+
+/*-*************************************
+* Context
+***************************************/
+
+typedef struct {
+  const BYTE *samples;
+  size_t *offsets;
+  const size_t *samplesSizes;
+  size_t nbSamples;
+  U32 *suffix;
+  size_t suffixSize;
+  U32 *freqs;
+  U32 *dmerAt;
+  unsigned d;
+} COVER_ctx_t;
+
+/* We need a global context for qsort... */
+static COVER_ctx_t *g_ctx = NULL;
+
+/*-*************************************
+*  Helper functions
+***************************************/
+
+/**
+ * Returns the sum of the sample sizes.
+ */
+static size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
+  size_t sum = 0;
+  size_t i;
+  for (i = 0; i < nbSamples; ++i) {
+    sum += samplesSizes[i];
+  }
+  return sum;
+}
+
+/**
+ * Returns -1 if the dmer at lp is less than the dmer at rp.
+ * Return 0 if the dmers at lp and rp are equal.
+ * Returns 1 if the dmer at lp is greater than the dmer at rp.
+ */
+static int COVER_cmp(COVER_ctx_t *ctx, const void *lp, const void *rp) {
+  U32 const lhs = *(U32 const *)lp;
+  U32 const rhs = *(U32 const *)rp;
+  return memcmp(ctx->samples + lhs, ctx->samples + rhs, ctx->d);
+}
+/**
+ * Faster version for d <= 8.
+ */
+static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) {
+  U64 const mask = (ctx->d == 8) ? (U64)-1 : (((U64)1 << (8 * ctx->d)) - 1);
+  U64 const lhs = MEM_readLE64(ctx->samples + *(U32 const *)lp) & mask;
+  U64 const rhs = MEM_readLE64(ctx->samples + *(U32 const *)rp) & mask;
+  if (lhs < rhs) {
+    return -1;
+  }
+  return (lhs > rhs);
+}
+
+/**
+ * Same as COVER_cmp() except ties are broken by pointer value
+ * NOTE: g_ctx must be set to call this function.  A global is required because
+ * qsort doesn't take an opaque pointer.
+ */
+static int COVER_strict_cmp(const void *lp, const void *rp) {
+  int result = COVER_cmp(g_ctx, lp, rp);
+  if (result == 0) {
+    result = lp < rp ? -1 : 1;
+  }
+  return result;
+}
+/**
+ * Faster version for d <= 8.
+ */
+static int COVER_strict_cmp8(const void *lp, const void *rp) {
+  int result = COVER_cmp8(g_ctx, lp, rp);
+  if (result == 0) {
+    result = lp < rp ? -1 : 1;
+  }
+  return result;
+}
+
+/**
+ * Returns the first pointer in [first, last) whose element does not compare
+ * less than value.  If no such element exists it returns last.
+ */
+static const size_t *COVER_lower_bound(const size_t *first, const size_t *last,
+                                       size_t value) {
+  size_t count = last - first;
+  while (count != 0) {
+    size_t step = count / 2;
+    const size_t *ptr = first;
+    ptr += step;
+    if (*ptr < value) {
+      first = ++ptr;
+      count -= step + 1;
+    } else {
+      count = step;
+    }
+  }
+  return first;
+}
+
+/**
+ * Generic groupBy function.
+ * Groups an array sorted by cmp into groups with equivalent values.
+ * Calls grp for each group.
+ */
+static void
+COVER_groupBy(const void *data, size_t count, size_t size, COVER_ctx_t *ctx,
+              int (*cmp)(COVER_ctx_t *, const void *, const void *),
+              void (*grp)(COVER_ctx_t *, const void *, const void *)) {
+  const BYTE *ptr = (const BYTE *)data;
+  size_t num = 0;
+  while (num < count) {
+    const BYTE *grpEnd = ptr + size;
+    ++num;
+    while (num < count && cmp(ctx, ptr, grpEnd) == 0) {
+      grpEnd += size;
+      ++num;
+    }
+    grp(ctx, ptr, grpEnd);
+    ptr = grpEnd;
+  }
+}
+
+/*-*************************************
+*  Cover functions
+***************************************/
+
+/**
+ * Called on each group of positions with the same dmer.
+ * Counts the frequency of each dmer and saves it in the suffix array.
+ * Fills `ctx->dmerAt`.
+ */
+static void COVER_group(COVER_ctx_t *ctx, const void *group,
+                        const void *groupEnd) {
+  /* The group consists of all the positions with the same first d bytes. */
+  const U32 *grpPtr = (const U32 *)group;
+  const U32 *grpEnd = (const U32 *)groupEnd;
+  /* The dmerId is how we will reference this dmer.
+   * This allows us to map the whole dmer space to a much smaller space, the
+   * size of the suffix array.
+   */
+  const U32 dmerId = (U32)(grpPtr - ctx->suffix);
+  /* Count the number of samples this dmer shows up in */
+  U32 freq = 0;
+  /* Details */
+  const size_t *curOffsetPtr = ctx->offsets;
+  const size_t *offsetsEnd = ctx->offsets + ctx->nbSamples;
+  /* Once *grpPtr >= curSampleEnd this occurrence of the dmer is in a
+   * different sample than the last.
+   */
+  size_t curSampleEnd = ctx->offsets[0];
+  for (; grpPtr != grpEnd; ++grpPtr) {
+    /* Save the dmerId for this position so we can get back to it. */
+    ctx->dmerAt[*grpPtr] = dmerId;
+    /* Dictionaries only help for the first reference to the dmer.
+     * After that zstd can reference the match from the previous reference.
+     * So only count each dmer once for each sample it is in.
+     */
+    if (*grpPtr < curSampleEnd) {
+      continue;
+    }
+    freq += 1;
+    /* Binary search to find the end of the sample *grpPtr is in.
+     * In the common case that grpPtr + 1 == grpEnd we can skip the binary
+     * search because the loop is over.
+     */
+    if (grpPtr + 1 != grpEnd) {
+      const size_t *sampleEndPtr =
+          COVER_lower_bound(curOffsetPtr, offsetsEnd, *grpPtr);
+      curSampleEnd = *sampleEndPtr;
+      curOffsetPtr = sampleEndPtr + 1;
+    }
+  }
+  /* At this point we are never going to look at this segment of the suffix
+   * array again.  We take advantage of this fact to save memory.
+   * We store the frequency of the dmer in the first position of the group,
+   * which is dmerId.
+   */
+  ctx->suffix[dmerId] = freq;
+}
+
+/**
+ * A segment is a range in the source as well as the score of the segment.
+ */
+typedef struct {
+  U32 begin;
+  U32 end;
+  U32 score;
+} COVER_segment_t;
+
+/**
+ * Selects the best segment in an epoch.
+ * Segments of are scored according to the function:
+ *
+ * Let F(d) be the frequency of dmer d.
+ * Let S_i be the dmer at position i of segment S which has length k.
+ *
+ *     Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1})
+ *
+ * Once the dmer d is in the dictionay we set F(d) = 0.
+ */
+static COVER_segment_t COVER_selectSegment(const COVER_ctx_t *ctx, U32 *freqs,
+                                           COVER_map_t *activeDmers, U32 begin,
+                                           U32 end,
+                                           ZDICT_cover_params_t parameters) {
+  /* Constants */
+  const U32 k = parameters.k;
+  const U32 d = parameters.d;
+  const U32 dmersInK = k - d + 1;
+  /* Try each segment (activeSegment) and save the best (bestSegment) */
+  COVER_segment_t bestSegment = {0, 0, 0};
+  COVER_segment_t activeSegment;
+  /* Reset the activeDmers in the segment */
+  COVER_map_clear(activeDmers);
+  /* The activeSegment starts at the beginning of the epoch. */
+  activeSegment.begin = begin;
+  activeSegment.end = begin;
+  activeSegment.score = 0;
+  /* Slide the activeSegment through the whole epoch.
+   * Save the best segment in bestSegment.
+   */
+  while (activeSegment.end < end) {
+    /* The dmerId for the dmer at the next position */
+    U32 newDmer = ctx->dmerAt[activeSegment.end];
+    /* The entry in activeDmers for this dmerId */
+    U32 *newDmerOcc = COVER_map_at(activeDmers, newDmer);
+    /* If the dmer isn't already present in the segment add its score. */
+    if (*newDmerOcc == 0) {
+      /* The paper suggest using the L-0.5 norm, but experiments show that it
+       * doesn't help.
+       */
+      activeSegment.score += freqs[newDmer];
+    }
+    /* Add the dmer to the segment */
+    activeSegment.end += 1;
+    *newDmerOcc += 1;
+
+    /* If the window is now too large, drop the first position */
+    if (activeSegment.end - activeSegment.begin == dmersInK + 1) {
+      U32 delDmer = ctx->dmerAt[activeSegment.begin];
+      U32 *delDmerOcc = COVER_map_at(activeDmers, delDmer);
+      activeSegment.begin += 1;
+      *delDmerOcc -= 1;
+      /* If this is the last occurence of the dmer, subtract its score */
+      if (*delDmerOcc == 0) {
+        COVER_map_remove(activeDmers, delDmer);
+        activeSegment.score -= freqs[delDmer];
+      }
+    }
+
+    /* If this segment is the best so far save it */
+    if (activeSegment.score > bestSegment.score) {
+      bestSegment = activeSegment;
+    }
+  }
+  {
+    /* Trim off the zero frequency head and tail from the segment. */
+    U32 newBegin = bestSegment.end;
+    U32 newEnd = bestSegment.begin;
+    U32 pos;
+    for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
+      U32 freq = freqs[ctx->dmerAt[pos]];
+      if (freq != 0) {
+        newBegin = MIN(newBegin, pos);
+        newEnd = pos + 1;
+      }
+    }
+    bestSegment.begin = newBegin;
+    bestSegment.end = newEnd;
+  }
+  {
+    /* Zero out the frequency of each dmer covered by the chosen segment. */
+    U32 pos;
+    for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
+      freqs[ctx->dmerAt[pos]] = 0;
+    }
+  }
+  return bestSegment;
+}
+
+/**
+ * Check the validity of the parameters.
+ * Returns non-zero if the parameters are valid and 0 otherwise.
+ */
+static int COVER_checkParameters(ZDICT_cover_params_t parameters,
+                                 size_t maxDictSize) {
+  /* k and d are required parameters */
+  if (parameters.d == 0 || parameters.k == 0) {
+    return 0;
+  }
+  /* k <= maxDictSize */
+  if (parameters.k > maxDictSize) {
+    return 0;
+  }
+  /* d <= k */
+  if (parameters.d > parameters.k) {
+    return 0;
+  }
+  return 1;
+}
+
+/**
+ * Clean up a context initialized with `COVER_ctx_init()`.
+ */
+static void COVER_ctx_destroy(COVER_ctx_t *ctx) {
+  if (!ctx) {
+    return;
+  }
+  if (ctx->suffix) {
+    free(ctx->suffix);
+    ctx->suffix = NULL;
+  }
+  if (ctx->freqs) {
+    free(ctx->freqs);
+    ctx->freqs = NULL;
+  }
+  if (ctx->dmerAt) {
+    free(ctx->dmerAt);
+    ctx->dmerAt = NULL;
+  }
+  if (ctx->offsets) {
+    free(ctx->offsets);
+    ctx->offsets = NULL;
+  }
+}
+
+/**
+ * Prepare a context for dictionary building.
+ * The context is only dependent on the parameter `d` and can used multiple
+ * times.
+ * Returns 1 on success or zero on error.
+ * The context must be destroyed with `COVER_ctx_destroy()`.
+ */
+static int COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
+                          const size_t *samplesSizes, unsigned nbSamples,
+                          unsigned d) {
+  const BYTE *const samples = (const BYTE *)samplesBuffer;
+  const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
+  /* Checks */
+  if (totalSamplesSize < MAX(d, sizeof(U64)) ||
+      totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) {
+    DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
+                 (U32)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20));
+    return 0;
+  }
+  /* Zero the context */
+  memset(ctx, 0, sizeof(*ctx));
+  DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbSamples,
+               (U32)totalSamplesSize);
+  ctx->samples = samples;
+  ctx->samplesSizes = samplesSizes;
+  ctx->nbSamples = nbSamples;
+  /* Partial suffix array */
+  ctx->suffixSize = totalSamplesSize - MAX(d, sizeof(U64)) + 1;
+  ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
+  /* Maps index to the dmerID */
+  ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
+  /* The offsets of each file */
+  ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t));
+  if (!ctx->suffix || !ctx->dmerAt || !ctx->offsets) {
+    DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n");
+    COVER_ctx_destroy(ctx);
+    return 0;
+  }
+  ctx->freqs = NULL;
+  ctx->d = d;
+
+  /* Fill offsets from the samlesSizes */
+  {
+    U32 i;
+    ctx->offsets[0] = 0;
+    for (i = 1; i <= nbSamples; ++i) {
+      ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1];
+    }
+  }
+  DISPLAYLEVEL(2, "Constructing partial suffix array\n");
+  {
+    /* suffix is a partial suffix array.
+     * It only sorts suffixes by their first parameters.d bytes.
+     * The sort is stable, so each dmer group is sorted by position in input.
+     */
+    U32 i;
+    for (i = 0; i < ctx->suffixSize; ++i) {
+      ctx->suffix[i] = i;
+    }
+    /* qsort doesn't take an opaque pointer, so pass as a global */
+    g_ctx = ctx;
+    qsort(ctx->suffix, ctx->suffixSize, sizeof(U32),
+          (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
+  }
+  DISPLAYLEVEL(2, "Computing frequencies\n");
+  /* For each dmer group (group of positions with the same first d bytes):
+   * 1. For each position we set dmerAt[position] = dmerID.  The dmerID is
+   *    (groupBeginPtr - suffix).  This allows us to go from position to
+   *    dmerID so we can look up values in freq.
+   * 2. We calculate how many samples the dmer occurs in and save it in
+   *    freqs[dmerId].
+   */
+  COVER_groupBy(ctx->suffix, ctx->suffixSize, sizeof(U32), ctx,
+                (ctx->d <= 8 ? &COVER_cmp8 : &COVER_cmp), &COVER_group);
+  ctx->freqs = ctx->suffix;
+  ctx->suffix = NULL;
+  return 1;
+}
+
+/**
+ * Given the prepared context build the dictionary.
+ */
+static size_t COVER_buildDictionary(const COVER_ctx_t *ctx, U32 *freqs,
+                                    COVER_map_t *activeDmers, void *dictBuffer,
+                                    size_t dictBufferCapacity,
+                                    ZDICT_cover_params_t parameters) {
+  BYTE *const dict = (BYTE *)dictBuffer;
+  size_t tail = dictBufferCapacity;
+  /* Divide the data up into epochs of equal size.
+   * We will select at least one segment from each epoch.
+   */
+  const U32 epochs = (U32)(dictBufferCapacity / parameters.k);
+  const U32 epochSize = (U32)(ctx->suffixSize / epochs);
+  size_t epoch;
+  DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n", epochs,
+               epochSize);
+  /* Loop through the epochs until there are no more segments or the dictionary
+   * is full.
+   */
+  for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs) {
+    const U32 epochBegin = (U32)(epoch * epochSize);
+    const U32 epochEnd = epochBegin + epochSize;
+    size_t segmentSize;
+    /* Select a segment */
+    COVER_segment_t segment = COVER_selectSegment(
+        ctx, freqs, activeDmers, epochBegin, epochEnd, parameters);
+    /* If the segment covers no dmers, then we are out of content */
+    if (segment.score == 0) {
+      break;
+    }
+    /* Trim the segment if necessary and if it is too small then we are done */
+    segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail);
+    if (segmentSize < parameters.d) {
+      break;
+    }
+    /* We fill the dictionary from the back to allow the best segments to be
+     * referenced with the smallest offsets.
+     */
+    tail -= segmentSize;
+    memcpy(dict + tail, ctx->samples + segment.begin, segmentSize);
+    DISPLAYUPDATE(
+        2, "\r%u%%       ",
+        (U32)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
+  }
+  DISPLAYLEVEL(2, "\r%79s\r", "");
+  return tail;
+}
+
+ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
+    void *dictBuffer, size_t dictBufferCapacity,
+    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+    ZDICT_cover_params_t parameters)
+{
+  BYTE* const dict = (BYTE*)dictBuffer;
+  COVER_ctx_t ctx;
+  COVER_map_t activeDmers;
+
+  /* Initialize global data */
+  g_displayLevel = parameters.zParams.notificationLevel;
+  /* Checks */
+  if (!COVER_checkParameters(parameters, dictBufferCapacity)) {
+    DISPLAYLEVEL(1, "Cover parameters incorrect\n");
+    return ERROR(GENERIC);
+  }
+  if (nbSamples == 0) {
+    DISPLAYLEVEL(1, "Cover must have at least one input file\n");
+    return ERROR(GENERIC);
+  }
+  if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
+    DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
+                 ZDICT_DICTSIZE_MIN);
+    return ERROR(dstSize_tooSmall);
+  }
+  /* Initialize context and activeDmers */
+  if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
+                      parameters.d)) {
+    return ERROR(GENERIC);
+  }
+  if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
+    DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
+    COVER_ctx_destroy(&ctx);
+    return ERROR(GENERIC);
+  }
+
+  DISPLAYLEVEL(2, "Building dictionary\n");
+  {
+    const size_t tail =
+        COVER_buildDictionary(&ctx, ctx.freqs, &activeDmers, dictBuffer,
+                              dictBufferCapacity, parameters);
+    const size_t dictionarySize = ZDICT_finalizeDictionary(
+        dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
+        samplesBuffer, samplesSizes, nbSamples, parameters.zParams);
+    if (!ZSTD_isError(dictionarySize)) {
+      DISPLAYLEVEL(2, "Constructed dictionary of size %u\n",
+                   (U32)dictionarySize);
+    }
+    COVER_ctx_destroy(&ctx);
+    COVER_map_destroy(&activeDmers);
+    return dictionarySize;
+  }
+}
+
+/**
+ * COVER_best_t is used for two purposes:
+ * 1. Synchronizing threads.
+ * 2. Saving the best parameters and dictionary.
+ *
+ * All of the methods except COVER_best_init() are thread safe if zstd is
+ * compiled with multithreaded support.
+ */
+typedef struct COVER_best_s {
+  ZSTD_pthread_mutex_t mutex;
+  ZSTD_pthread_cond_t cond;
+  size_t liveJobs;
+  void *dict;
+  size_t dictSize;
+  ZDICT_cover_params_t parameters;
+  size_t compressedSize;
+} COVER_best_t;
+
+/**
+ * Initialize the `COVER_best_t`.
+ */
+static void COVER_best_init(COVER_best_t *best) {
+  if (best==NULL) return; /* compatible with init on NULL */
+  (void)ZSTD_pthread_mutex_init(&best->mutex, NULL);
+  (void)ZSTD_pthread_cond_init(&best->cond, NULL);
+  best->liveJobs = 0;
+  best->dict = NULL;
+  best->dictSize = 0;
+  best->compressedSize = (size_t)-1;
+  memset(&best->parameters, 0, sizeof(best->parameters));
+}
+
+/**
+ * Wait until liveJobs == 0.
+ */
+static void COVER_best_wait(COVER_best_t *best) {
+  if (!best) {
+    return;
+  }
+  ZSTD_pthread_mutex_lock(&best->mutex);
+  while (best->liveJobs != 0) {
+    ZSTD_pthread_cond_wait(&best->cond, &best->mutex);
+  }
+  ZSTD_pthread_mutex_unlock(&best->mutex);
+}
+
+/**
+ * Call COVER_best_wait() and then destroy the COVER_best_t.
+ */
+static void COVER_best_destroy(COVER_best_t *best) {
+  if (!best) {
+    return;
+  }
+  COVER_best_wait(best);
+  if (best->dict) {
+    free(best->dict);
+  }
+  ZSTD_pthread_mutex_destroy(&best->mutex);
+  ZSTD_pthread_cond_destroy(&best->cond);
+}
+
+/**
+ * Called when a thread is about to be launched.
+ * Increments liveJobs.
+ */
+static void COVER_best_start(COVER_best_t *best) {
+  if (!best) {
+    return;
+  }
+  ZSTD_pthread_mutex_lock(&best->mutex);
+  ++best->liveJobs;
+  ZSTD_pthread_mutex_unlock(&best->mutex);
+}
+
+/**
+ * Called when a thread finishes executing, both on error or success.
+ * Decrements liveJobs and signals any waiting threads if liveJobs == 0.
+ * If this dictionary is the best so far save it and its parameters.
+ */
+static void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
+                              ZDICT_cover_params_t parameters, void *dict,
+                              size_t dictSize) {
+  if (!best) {
+    return;
+  }
+  {
+    size_t liveJobs;
+    ZSTD_pthread_mutex_lock(&best->mutex);
+    --best->liveJobs;
+    liveJobs = best->liveJobs;
+    /* If the new dictionary is better */
+    if (compressedSize < best->compressedSize) {
+      /* Allocate space if necessary */
+      if (!best->dict || best->dictSize < dictSize) {
+        if (best->dict) {
+          free(best->dict);
+        }
+        best->dict = malloc(dictSize);
+        if (!best->dict) {
+          best->compressedSize = ERROR(GENERIC);
+          best->dictSize = 0;
+          return;
+        }
+      }
+      /* Save the dictionary, parameters, and size */
+      memcpy(best->dict, dict, dictSize);
+      best->dictSize = dictSize;
+      best->parameters = parameters;
+      best->compressedSize = compressedSize;
+    }
+    ZSTD_pthread_mutex_unlock(&best->mutex);
+    if (liveJobs == 0) {
+      ZSTD_pthread_cond_broadcast(&best->cond);
+    }
+  }
+}
+
+/**
+ * Parameters for COVER_tryParameters().
+ */
+typedef struct COVER_tryParameters_data_s {
+  const COVER_ctx_t *ctx;
+  COVER_best_t *best;
+  size_t dictBufferCapacity;
+  ZDICT_cover_params_t parameters;
+} COVER_tryParameters_data_t;
+
+/**
+ * Tries a set of parameters and upates the COVER_best_t with the results.
+ * This function is thread safe if zstd is compiled with multithreaded support.
+ * It takes its parameters as an *OWNING* opaque pointer to support threading.
+ */
+static void COVER_tryParameters(void *opaque) {
+  /* Save parameters as local variables */
+  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t *)opaque;
+  const COVER_ctx_t *const ctx = data->ctx;
+  const ZDICT_cover_params_t parameters = data->parameters;
+  size_t dictBufferCapacity = data->dictBufferCapacity;
+  size_t totalCompressedSize = ERROR(GENERIC);
+  /* Allocate space for hash table, dict, and freqs */
+  COVER_map_t activeDmers;
+  BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
+  U32 *freqs = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
+  if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
+    DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
+    goto _cleanup;
+  }
+  if (!dict || !freqs) {
+    DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
+    goto _cleanup;
+  }
+  /* Copy the frequencies because we need to modify them */
+  memcpy(freqs, ctx->freqs, ctx->suffixSize * sizeof(U32));
+  /* Build the dictionary */
+  {
+    const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict,
+                                              dictBufferCapacity, parameters);
+    dictBufferCapacity = ZDICT_finalizeDictionary(
+        dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
+        ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbSamples,
+        parameters.zParams);
+    if (ZDICT_isError(dictBufferCapacity)) {
+      DISPLAYLEVEL(1, "Failed to finalize dictionary\n");
+      goto _cleanup;
+    }
+  }
+  /* Check total compressed size */
+  {
+    /* Pointers */
+    ZSTD_CCtx *cctx;
+    ZSTD_CDict *cdict;
+    void *dst;
+    /* Local variables */
+    size_t dstCapacity;
+    size_t i;
+    /* Allocate dst with enough space to compress the maximum sized sample */
+    {
+      size_t maxSampleSize = 0;
+      for (i = 0; i < ctx->nbSamples; ++i) {
+        maxSampleSize = MAX(ctx->samplesSizes[i], maxSampleSize);
+      }
+      dstCapacity = ZSTD_compressBound(maxSampleSize);
+      dst = malloc(dstCapacity);
+    }
+    /* Create the cctx and cdict */
+    cctx = ZSTD_createCCtx();
+    cdict = ZSTD_createCDict(dict, dictBufferCapacity,
+                             parameters.zParams.compressionLevel);
+    if (!dst || !cctx || !cdict) {
+      goto _compressCleanup;
+    }
+    /* Compress each sample and sum their sizes (or error) */
+    totalCompressedSize = dictBufferCapacity;
+    for (i = 0; i < ctx->nbSamples; ++i) {
+      const size_t size = ZSTD_compress_usingCDict(
+          cctx, dst, dstCapacity, ctx->samples + ctx->offsets[i],
+          ctx->samplesSizes[i], cdict);
+      if (ZSTD_isError(size)) {
+        totalCompressedSize = ERROR(GENERIC);
+        goto _compressCleanup;
+      }
+      totalCompressedSize += size;
+    }
+  _compressCleanup:
+    ZSTD_freeCCtx(cctx);
+    ZSTD_freeCDict(cdict);
+    if (dst) {
+      free(dst);
+    }
+  }
+
+_cleanup:
+  COVER_best_finish(data->best, totalCompressedSize, parameters, dict,
+                    dictBufferCapacity);
+  free(data);
+  COVER_map_destroy(&activeDmers);
+  if (dict) {
+    free(dict);
+  }
+  if (freqs) {
+    free(freqs);
+  }
+}
+
+ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
+    void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
+    const size_t *samplesSizes, unsigned nbSamples,
+    ZDICT_cover_params_t *parameters) {
+  /* constants */
+  const unsigned nbThreads = parameters->nbThreads;
+  const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
+  const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
+  const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
+  const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k;
+  const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps;
+  const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1);
+  const unsigned kIterations =
+      (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize);
+  /* Local variables */
+  const int displayLevel = parameters->zParams.notificationLevel;
+  unsigned iteration = 1;
+  unsigned d;
+  unsigned k;
+  COVER_best_t best;
+  POOL_ctx *pool = NULL;
+
+  /* Checks */
+  if (kMinK < kMaxD || kMaxK < kMinK) {
+    LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
+    return ERROR(GENERIC);
+  }
+  if (nbSamples == 0) {
+    DISPLAYLEVEL(1, "Cover must have at least one input file\n");
+    return ERROR(GENERIC);
+  }
+  if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
+    DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
+                 ZDICT_DICTSIZE_MIN);
+    return ERROR(dstSize_tooSmall);
+  }
+  if (nbThreads > 1) {
+    pool = POOL_create(nbThreads, 1);
+    if (!pool) {
+      return ERROR(memory_allocation);
+    }
+  }
+  /* Initialization */
+  COVER_best_init(&best);
+  /* Turn down global display level to clean up display at level 2 and below */
+  g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1;
+  /* Loop through d first because each new value needs a new context */
+  LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n",
+                    kIterations);
+  for (d = kMinD; d <= kMaxD; d += 2) {
+    /* Initialize the context for this value of d */
+    COVER_ctx_t ctx;
+    LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
+    if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d)) {
+      LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
+      COVER_best_destroy(&best);
+      POOL_free(pool);
+      return ERROR(GENERIC);
+    }
+    /* Loop through k reusing the same context */
+    for (k = kMinK; k <= kMaxK; k += kStepSize) {
+      /* Prepare the arguments */
+      COVER_tryParameters_data_t *data = (COVER_tryParameters_data_t *)malloc(
+          sizeof(COVER_tryParameters_data_t));
+      LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k);
+      if (!data) {
+        LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n");
+        COVER_best_destroy(&best);
+        COVER_ctx_destroy(&ctx);
+        POOL_free(pool);
+        return ERROR(GENERIC);
+      }
+      data->ctx = &ctx;
+      data->best = &best;
+      data->dictBufferCapacity = dictBufferCapacity;
+      data->parameters = *parameters;
+      data->parameters.k = k;
+      data->parameters.d = d;
+      data->parameters.steps = kSteps;
+      data->parameters.zParams.notificationLevel = g_displayLevel;
+      /* Check the parameters */
+      if (!COVER_checkParameters(data->parameters, dictBufferCapacity)) {
+        DISPLAYLEVEL(1, "Cover parameters incorrect\n");
+        free(data);
+        continue;
+      }
+      /* Call the function and pass ownership of data to it */
+      COVER_best_start(&best);
+      if (pool) {
+        POOL_add(pool, &COVER_tryParameters, data);
+      } else {
+        COVER_tryParameters(data);
+      }
+      /* Print status */
+      LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%%       ",
+                         (U32)((iteration * 100) / kIterations));
+      ++iteration;
+    }
+    COVER_best_wait(&best);
+    COVER_ctx_destroy(&ctx);
+  }
+  LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", "");
+  /* Fill the output buffer and parameters with output of the best parameters */
+  {
+    const size_t dictSize = best.dictSize;
+    if (ZSTD_isError(best.compressedSize)) {
+      const size_t compressedSize = best.compressedSize;
+      COVER_best_destroy(&best);
+      POOL_free(pool);
+      return compressedSize;
+    }
+    *parameters = best.parameters;
+    memcpy(dictBuffer, best.dict, dictSize);
+    COVER_best_destroy(&best);
+    POOL_free(pool);
+    return dictSize;
+  }
+}
diff --git a/vendor/github.com/DataDog/zstd/cpu.h b/vendor/github.com/DataDog/zstd/cpu.h
new file mode 100755
index 0000000..4eb48e3
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/cpu.h
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2018-present, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_COMMON_CPU_H
+#define ZSTD_COMMON_CPU_H
+
+/**
+ * Implementation taken from folly/CpuId.h
+ * https://github.com/facebook/folly/blob/master/folly/CpuId.h
+ */
+
+#include <string.h>
+
+#include "mem.h"
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+typedef struct {
+    U32 f1c;
+    U32 f1d;
+    U32 f7b;
+    U32 f7c;
+} ZSTD_cpuid_t;
+
+MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
+    U32 f1c = 0;
+    U32 f1d = 0;
+    U32 f7b = 0;
+    U32 f7c = 0;
+#ifdef _MSC_VER
+    int reg[4];
+    __cpuid((int*)reg, 0);
+    {
+        int const n = reg[0];
+        if (n >= 1) {
+            __cpuid((int*)reg, 1);
+            f1c = (U32)reg[2];
+            f1d = (U32)reg[3];
+        }
+        if (n >= 7) {
+            __cpuidex((int*)reg, 7, 0);
+            f7b = (U32)reg[1];
+            f7c = (U32)reg[2];
+        }
+    }
+#elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__)
+    /* The following block like the normal cpuid branch below, but gcc
+     * reserves ebx for use of its pic register so we must specially
+     * handle the save and restore to avoid clobbering the register
+     */
+    U32 n;
+    __asm__(
+        "pushl %%ebx\n\t"
+        "cpuid\n\t"
+        "popl %%ebx\n\t"
+        : "=a"(n)
+        : "a"(0)
+        : "ecx", "edx");
+    if (n >= 1) {
+      U32 f1a;
+      __asm__(
+          "pushl %%ebx\n\t"
+          "cpuid\n\t"
+          "popl %%ebx\n\t"
+          : "=a"(f1a), "=c"(f1c), "=d"(f1d)
+          : "a"(1)
+          :);
+    }
+    if (n >= 7) {
+      __asm__(
+          "pushl %%ebx\n\t"
+          "cpuid\n\t"
+          "movl %%ebx, %%eax\n\r"
+          "popl %%ebx"
+          : "=a"(f7b), "=c"(f7c)
+          : "a"(7), "c"(0)
+          : "edx");
+    }
+#elif defined(__x86_64__) || defined(_M_X64) || defined(__i386__)
+    U32 n;
+    __asm__("cpuid" : "=a"(n) : "a"(0) : "ebx", "ecx", "edx");
+    if (n >= 1) {
+      U32 f1a;
+      __asm__("cpuid" : "=a"(f1a), "=c"(f1c), "=d"(f1d) : "a"(1) : "ebx");
+    }
+    if (n >= 7) {
+      U32 f7a;
+      __asm__("cpuid"
+              : "=a"(f7a), "=b"(f7b), "=c"(f7c)
+              : "a"(7), "c"(0)
+              : "edx");
+    }
+#endif
+    {
+        ZSTD_cpuid_t cpuid;
+        cpuid.f1c = f1c;
+        cpuid.f1d = f1d;
+        cpuid.f7b = f7b;
+        cpuid.f7c = f7c;
+        return cpuid;
+    }
+}
+
+#define X(name, r, bit)                                                        \
+  MEM_STATIC int ZSTD_cpuid_##name(ZSTD_cpuid_t const cpuid) {                 \
+    return ((cpuid.r) & (1U << bit)) != 0;                                     \
+  }
+
+/* cpuid(1): Processor Info and Feature Bits. */
+#define C(name, bit) X(name, f1c, bit)
+  C(sse3, 0)
+  C(pclmuldq, 1)
+  C(dtes64, 2)
+  C(monitor, 3)
+  C(dscpl, 4)
+  C(vmx, 5)
+  C(smx, 6)
+  C(eist, 7)
+  C(tm2, 8)
+  C(ssse3, 9)
+  C(cnxtid, 10)
+  C(fma, 12)
+  C(cx16, 13)
+  C(xtpr, 14)
+  C(pdcm, 15)
+  C(pcid, 17)
+  C(dca, 18)
+  C(sse41, 19)
+  C(sse42, 20)
+  C(x2apic, 21)
+  C(movbe, 22)
+  C(popcnt, 23)
+  C(tscdeadline, 24)
+  C(aes, 25)
+  C(xsave, 26)
+  C(osxsave, 27)
+  C(avx, 28)
+  C(f16c, 29)
+  C(rdrand, 30)
+#undef C
+#define D(name, bit) X(name, f1d, bit)
+  D(fpu, 0)
+  D(vme, 1)
+  D(de, 2)
+  D(pse, 3)
+  D(tsc, 4)
+  D(msr, 5)
+  D(pae, 6)
+  D(mce, 7)
+  D(cx8, 8)
+  D(apic, 9)
+  D(sep, 11)
+  D(mtrr, 12)
+  D(pge, 13)
+  D(mca, 14)
+  D(cmov, 15)
+  D(pat, 16)
+  D(pse36, 17)
+  D(psn, 18)
+  D(clfsh, 19)
+  D(ds, 21)
+  D(acpi, 22)
+  D(mmx, 23)
+  D(fxsr, 24)
+  D(sse, 25)
+  D(sse2, 26)
+  D(ss, 27)
+  D(htt, 28)
+  D(tm, 29)
+  D(pbe, 31)
+#undef D
+
+/* cpuid(7): Extended Features. */
+#define B(name, bit) X(name, f7b, bit)
+  B(bmi1, 3)
+  B(hle, 4)
+  B(avx2, 5)
+  B(smep, 7)
+  B(bmi2, 8)
+  B(erms, 9)
+  B(invpcid, 10)
+  B(rtm, 11)
+  B(mpx, 14)
+  B(avx512f, 16)
+  B(avx512dq, 17)
+  B(rdseed, 18)
+  B(adx, 19)
+  B(smap, 20)
+  B(avx512ifma, 21)
+  B(pcommit, 22)
+  B(clflushopt, 23)
+  B(clwb, 24)
+  B(avx512pf, 26)
+  B(avx512er, 27)
+  B(avx512cd, 28)
+  B(sha, 29)
+  B(avx512bw, 30)
+  B(avx512vl, 31)
+#undef B
+#define C(name, bit) X(name, f7c, bit)
+  C(prefetchwt1, 0)
+  C(avx512vbmi, 1)
+#undef C
+
+#undef X
+
+#endif /* ZSTD_COMMON_CPU_H */
diff --git a/vendor/github.com/DataDog/zstd/divsufsort.c b/vendor/github.com/DataDog/zstd/divsufsort.c
new file mode 100644
index 0000000..60cceb0
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/divsufsort.c
@@ -0,0 +1,1913 @@
+/*
+ * divsufsort.c for libdivsufsort-lite
+ * Copyright (c) 2003-2008 Yuta Mori All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+/*- Compiler specifics -*/
+#ifdef __clang__
+#pragma clang diagnostic ignored "-Wshorten-64-to-32"
+#endif
+
+#if defined(_MSC_VER)
+#  pragma warning(disable : 4244)
+#  pragma warning(disable : 4127)    /* C4127 : Condition expression is constant */
+#endif
+
+
+/*- Dependencies -*/
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "divsufsort.h"
+
+/*- Constants -*/
+#if defined(INLINE)
+# undef INLINE
+#endif
+#if !defined(INLINE)
+# define INLINE __inline
+#endif
+#if defined(ALPHABET_SIZE) && (ALPHABET_SIZE < 1)
+# undef ALPHABET_SIZE
+#endif
+#if !defined(ALPHABET_SIZE)
+# define ALPHABET_SIZE (256)
+#endif
+#define BUCKET_A_SIZE (ALPHABET_SIZE)
+#define BUCKET_B_SIZE (ALPHABET_SIZE * ALPHABET_SIZE)
+#if defined(SS_INSERTIONSORT_THRESHOLD)
+# if SS_INSERTIONSORT_THRESHOLD < 1
+#  undef SS_INSERTIONSORT_THRESHOLD
+#  define SS_INSERTIONSORT_THRESHOLD (1)
+# endif
+#else
+# define SS_INSERTIONSORT_THRESHOLD (8)
+#endif
+#if defined(SS_BLOCKSIZE)
+# if SS_BLOCKSIZE < 0
+#  undef SS_BLOCKSIZE
+#  define SS_BLOCKSIZE (0)
+# elif 32768 <= SS_BLOCKSIZE
+#  undef SS_BLOCKSIZE
+#  define SS_BLOCKSIZE (32767)
+# endif
+#else
+# define SS_BLOCKSIZE (1024)
+#endif
+/* minstacksize = log(SS_BLOCKSIZE) / log(3) * 2 */
+#if SS_BLOCKSIZE == 0
+# define SS_MISORT_STACKSIZE (96)
+#elif SS_BLOCKSIZE <= 4096
+# define SS_MISORT_STACKSIZE (16)
+#else
+# define SS_MISORT_STACKSIZE (24)
+#endif
+#define SS_SMERGE_STACKSIZE (32)
+#define TR_INSERTIONSORT_THRESHOLD (8)
+#define TR_STACKSIZE (64)
+
+
+/*- Macros -*/
+#ifndef SWAP
+# define SWAP(_a, _b) do { t = (_a); (_a) = (_b); (_b) = t; } while(0)
+#endif /* SWAP */
+#ifndef MIN
+# define MIN(_a, _b) (((_a) < (_b)) ? (_a) : (_b))
+#endif /* MIN */
+#ifndef MAX
+# define MAX(_a, _b) (((_a) > (_b)) ? (_a) : (_b))
+#endif /* MAX */
+#define STACK_PUSH(_a, _b, _c, _d)\
+  do {\
+    assert(ssize < STACK_SIZE);\
+    stack[ssize].a = (_a), stack[ssize].b = (_b),\
+    stack[ssize].c = (_c), stack[ssize++].d = (_d);\
+  } while(0)
+#define STACK_PUSH5(_a, _b, _c, _d, _e)\
+  do {\
+    assert(ssize < STACK_SIZE);\
+    stack[ssize].a = (_a), stack[ssize].b = (_b),\
+    stack[ssize].c = (_c), stack[ssize].d = (_d), stack[ssize++].e = (_e);\
+  } while(0)
+#define STACK_POP(_a, _b, _c, _d)\
+  do {\
+    assert(0 <= ssize);\
+    if(ssize == 0) { return; }\
+    (_a) = stack[--ssize].a, (_b) = stack[ssize].b,\
+    (_c) = stack[ssize].c, (_d) = stack[ssize].d;\
+  } while(0)
+#define STACK_POP5(_a, _b, _c, _d, _e)\
+  do {\
+    assert(0 <= ssize);\
+    if(ssize == 0) { return; }\
+    (_a) = stack[--ssize].a, (_b) = stack[ssize].b,\
+    (_c) = stack[ssize].c, (_d) = stack[ssize].d, (_e) = stack[ssize].e;\
+  } while(0)
+#define BUCKET_A(_c0) bucket_A[(_c0)]
+#if ALPHABET_SIZE == 256
+#define BUCKET_B(_c0, _c1) (bucket_B[((_c1) << 8) | (_c0)])
+#define BUCKET_BSTAR(_c0, _c1) (bucket_B[((_c0) << 8) | (_c1)])
+#else
+#define BUCKET_B(_c0, _c1) (bucket_B[(_c1) * ALPHABET_SIZE + (_c0)])
+#define BUCKET_BSTAR(_c0, _c1) (bucket_B[(_c0) * ALPHABET_SIZE + (_c1)])
+#endif
+
+
+/*- Private Functions -*/
+
+static const int lg_table[256]= {
+ -1,0,1,1,2,2,2,2,3,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,
+  5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,
+  6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,
+  6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,
+  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
+  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
+  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
+  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
+};
+
+#if (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE)
+
+static INLINE
+int
+ss_ilg(int n) {
+#if SS_BLOCKSIZE == 0
+  return (n & 0xffff0000) ?
+          ((n & 0xff000000) ?
+            24 + lg_table[(n >> 24) & 0xff] :
+            16 + lg_table[(n >> 16) & 0xff]) :
+          ((n & 0x0000ff00) ?
+             8 + lg_table[(n >>  8) & 0xff] :
+             0 + lg_table[(n >>  0) & 0xff]);
+#elif SS_BLOCKSIZE < 256
+  return lg_table[n];
+#else
+  return (n & 0xff00) ?
+          8 + lg_table[(n >> 8) & 0xff] :
+          0 + lg_table[(n >> 0) & 0xff];
+#endif
+}
+
+#endif /* (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE) */
+
+#if SS_BLOCKSIZE != 0
+
+static const int sqq_table[256] = {
+  0,  16,  22,  27,  32,  35,  39,  42,  45,  48,  50,  53,  55,  57,  59,  61,
+ 64,  65,  67,  69,  71,  73,  75,  76,  78,  80,  81,  83,  84,  86,  87,  89,
+ 90,  91,  93,  94,  96,  97,  98,  99, 101, 102, 103, 104, 106, 107, 108, 109,
+110, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126,
+128, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142,
+143, 144, 144, 145, 146, 147, 148, 149, 150, 150, 151, 152, 153, 154, 155, 155,
+156, 157, 158, 159, 160, 160, 161, 162, 163, 163, 164, 165, 166, 167, 167, 168,
+169, 170, 170, 171, 172, 173, 173, 174, 175, 176, 176, 177, 178, 178, 179, 180,
+181, 181, 182, 183, 183, 184, 185, 185, 186, 187, 187, 188, 189, 189, 190, 191,
+192, 192, 193, 193, 194, 195, 195, 196, 197, 197, 198, 199, 199, 200, 201, 201,
+202, 203, 203, 204, 204, 205, 206, 206, 207, 208, 208, 209, 209, 210, 211, 211,
+212, 212, 213, 214, 214, 215, 215, 216, 217, 217, 218, 218, 219, 219, 220, 221,
+221, 222, 222, 223, 224, 224, 225, 225, 226, 226, 227, 227, 228, 229, 229, 230,
+230, 231, 231, 232, 232, 233, 234, 234, 235, 235, 236, 236, 237, 237, 238, 238,
+239, 240, 240, 241, 241, 242, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247,
+247, 248, 248, 249, 249, 250, 250, 251, 251, 252, 252, 253, 253, 254, 254, 255
+};
+
+static INLINE
+int
+ss_isqrt(int x) {
+  int y, e;
+
+  if(x >= (SS_BLOCKSIZE * SS_BLOCKSIZE)) { return SS_BLOCKSIZE; }
+  e = (x & 0xffff0000) ?
+        ((x & 0xff000000) ?
+          24 + lg_table[(x >> 24) & 0xff] :
+          16 + lg_table[(x >> 16) & 0xff]) :
+        ((x & 0x0000ff00) ?
+           8 + lg_table[(x >>  8) & 0xff] :
+           0 + lg_table[(x >>  0) & 0xff]);
+
+  if(e >= 16) {
+    y = sqq_table[x >> ((e - 6) - (e & 1))] << ((e >> 1) - 7);
+    if(e >= 24) { y = (y + 1 + x / y) >> 1; }
+    y = (y + 1 + x / y) >> 1;
+  } else if(e >= 8) {
+    y = (sqq_table[x >> ((e - 6) - (e & 1))] >> (7 - (e >> 1))) + 1;
+  } else {
+    return sqq_table[x] >> 4;
+  }
+
+  return (x < (y * y)) ? y - 1 : y;
+}
+
+#endif /* SS_BLOCKSIZE != 0 */
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Compares two suffixes. */
+static INLINE
+int
+ss_compare(const unsigned char *T,
+           const int *p1, const int *p2,
+           int depth) {
+  const unsigned char *U1, *U2, *U1n, *U2n;
+
+  for(U1 = T + depth + *p1,
+      U2 = T + depth + *p2,
+      U1n = T + *(p1 + 1) + 2,
+      U2n = T + *(p2 + 1) + 2;
+      (U1 < U1n) && (U2 < U2n) && (*U1 == *U2);
+      ++U1, ++U2) {
+  }
+
+  return U1 < U1n ?
+        (U2 < U2n ? *U1 - *U2 : 1) :
+        (U2 < U2n ? -1 : 0);
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+#if (SS_BLOCKSIZE != 1) && (SS_INSERTIONSORT_THRESHOLD != 1)
+
+/* Insertionsort for small size groups */
+static
+void
+ss_insertionsort(const unsigned char *T, const int *PA,
+                 int *first, int *last, int depth) {
+  int *i, *j;
+  int t;
+  int r;
+
+  for(i = last - 2; first <= i; --i) {
+    for(t = *i, j = i + 1; 0 < (r = ss_compare(T, PA + t, PA + *j, depth));) {
+      do { *(j - 1) = *j; } while((++j < last) && (*j < 0));
+      if(last <= j) { break; }
+    }
+    if(r == 0) { *j = ~*j; }
+    *(j - 1) = t;
+  }
+}
+
+#endif /* (SS_BLOCKSIZE != 1) && (SS_INSERTIONSORT_THRESHOLD != 1) */
+
+
+/*---------------------------------------------------------------------------*/
+
+#if (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE)
+
+static INLINE
+void
+ss_fixdown(const unsigned char *Td, const int *PA,
+           int *SA, int i, int size) {
+  int j, k;
+  int v;
+  int c, d, e;
+
+  for(v = SA[i], c = Td[PA[v]]; (j = 2 * i + 1) < size; SA[i] = SA[k], i = k) {
+    d = Td[PA[SA[k = j++]]];
+    if(d < (e = Td[PA[SA[j]]])) { k = j; d = e; }
+    if(d <= c) { break; }
+  }
+  SA[i] = v;
+}
+
+/* Simple top-down heapsort. */
+static
+void
+ss_heapsort(const unsigned char *Td, const int *PA, int *SA, int size) {
+  int i, m;
+  int t;
+
+  m = size;
+  if((size % 2) == 0) {
+    m--;
+    if(Td[PA[SA[m / 2]]] < Td[PA[SA[m]]]) { SWAP(SA[m], SA[m / 2]); }
+  }
+
+  for(i = m / 2 - 1; 0 <= i; --i) { ss_fixdown(Td, PA, SA, i, m); }
+  if((size % 2) == 0) { SWAP(SA[0], SA[m]); ss_fixdown(Td, PA, SA, 0, m); }
+  for(i = m - 1; 0 < i; --i) {
+    t = SA[0], SA[0] = SA[i];
+    ss_fixdown(Td, PA, SA, 0, i);
+    SA[i] = t;
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Returns the median of three elements. */
+static INLINE
+int *
+ss_median3(const unsigned char *Td, const int *PA,
+           int *v1, int *v2, int *v3) {
+  int *t;
+  if(Td[PA[*v1]] > Td[PA[*v2]]) { SWAP(v1, v2); }
+  if(Td[PA[*v2]] > Td[PA[*v3]]) {
+    if(Td[PA[*v1]] > Td[PA[*v3]]) { return v1; }
+    else { return v3; }
+  }
+  return v2;
+}
+
+/* Returns the median of five elements. */
+static INLINE
+int *
+ss_median5(const unsigned char *Td, const int *PA,
+           int *v1, int *v2, int *v3, int *v4, int *v5) {
+  int *t;
+  if(Td[PA[*v2]] > Td[PA[*v3]]) { SWAP(v2, v3); }
+  if(Td[PA[*v4]] > Td[PA[*v5]]) { SWAP(v4, v5); }
+  if(Td[PA[*v2]] > Td[PA[*v4]]) { SWAP(v2, v4); SWAP(v3, v5); }
+  if(Td[PA[*v1]] > Td[PA[*v3]]) { SWAP(v1, v3); }
+  if(Td[PA[*v1]] > Td[PA[*v4]]) { SWAP(v1, v4); SWAP(v3, v5); }
+  if(Td[PA[*v3]] > Td[PA[*v4]]) { return v4; }
+  return v3;
+}
+
+/* Returns the pivot element. */
+static INLINE
+int *
+ss_pivot(const unsigned char *Td, const int *PA, int *first, int *last) {
+  int *middle;
+  int t;
+
+  t = last - first;
+  middle = first + t / 2;
+
+  if(t <= 512) {
+    if(t <= 32) {
+      return ss_median3(Td, PA, first, middle, last - 1);
+    } else {
+      t >>= 2;
+      return ss_median5(Td, PA, first, first + t, middle, last - 1 - t, last - 1);
+    }
+  }
+  t >>= 3;
+  first  = ss_median3(Td, PA, first, first + t, first + (t << 1));
+  middle = ss_median3(Td, PA, middle - t, middle, middle + t);
+  last   = ss_median3(Td, PA, last - 1 - (t << 1), last - 1 - t, last - 1);
+  return ss_median3(Td, PA, first, middle, last);
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Binary partition for substrings. */
+static INLINE
+int *
+ss_partition(const int *PA,
+                    int *first, int *last, int depth) {
+  int *a, *b;
+  int t;
+  for(a = first - 1, b = last;;) {
+    for(; (++a < b) && ((PA[*a] + depth) >= (PA[*a + 1] + 1));) { *a = ~*a; }
+    for(; (a < --b) && ((PA[*b] + depth) <  (PA[*b + 1] + 1));) { }
+    if(b <= a) { break; }
+    t = ~*b;
+    *b = *a;
+    *a = t;
+  }
+  if(first < a) { *first = ~*first; }
+  return a;
+}
+
+/* Multikey introsort for medium size groups. */
+static
+void
+ss_mintrosort(const unsigned char *T, const int *PA,
+              int *first, int *last,
+              int depth) {
+#define STACK_SIZE SS_MISORT_STACKSIZE
+  struct { int *a, *b, c; int d; } stack[STACK_SIZE];
+  const unsigned char *Td;
+  int *a, *b, *c, *d, *e, *f;
+  int s, t;
+  int ssize;
+  int limit;
+  int v, x = 0;
+
+  for(ssize = 0, limit = ss_ilg(last - first);;) {
+
+    if((last - first) <= SS_INSERTIONSORT_THRESHOLD) {
+#if 1 < SS_INSERTIONSORT_THRESHOLD
+      if(1 < (last - first)) { ss_insertionsort(T, PA, first, last, depth); }
+#endif
+      STACK_POP(first, last, depth, limit);
+      continue;
+    }
+
+    Td = T + depth;
+    if(limit-- == 0) { ss_heapsort(Td, PA, first, last - first); }
+    if(limit < 0) {
+      for(a = first + 1, v = Td[PA[*first]]; a < last; ++a) {
+        if((x = Td[PA[*a]]) != v) {
+          if(1 < (a - first)) { break; }
+          v = x;
+          first = a;
+        }
+      }
+      if(Td[PA[*first] - 1] < v) {
+        first = ss_partition(PA, first, a, depth);
+      }
+      if((a - first) <= (last - a)) {
+        if(1 < (a - first)) {
+          STACK_PUSH(a, last, depth, -1);
+          last = a, depth += 1, limit = ss_ilg(a - first);
+        } else {
+          first = a, limit = -1;
+        }
+      } else {
+        if(1 < (last - a)) {
+          STACK_PUSH(first, a, depth + 1, ss_ilg(a - first));
+          first = a, limit = -1;
+        } else {
+          last = a, depth += 1, limit = ss_ilg(a - first);
+        }
+      }
+      continue;
+    }
+
+    /* choose pivot */
+    a = ss_pivot(Td, PA, first, last);
+    v = Td[PA[*a]];
+    SWAP(*first, *a);
+
+    /* partition */
+    for(b = first; (++b < last) && ((x = Td[PA[*b]]) == v);) { }
+    if(((a = b) < last) && (x < v)) {
+      for(; (++b < last) && ((x = Td[PA[*b]]) <= v);) {
+        if(x == v) { SWAP(*b, *a); ++a; }
+      }
+    }
+    for(c = last; (b < --c) && ((x = Td[PA[*c]]) == v);) { }
+    if((b < (d = c)) && (x > v)) {
+      for(; (b < --c) && ((x = Td[PA[*c]]) >= v);) {
+        if(x == v) { SWAP(*c, *d); --d; }
+      }
+    }
+    for(; b < c;) {
+      SWAP(*b, *c);
+      for(; (++b < c) && ((x = Td[PA[*b]]) <= v);) {
+        if(x == v) { SWAP(*b, *a); ++a; }
+      }
+      for(; (b < --c) && ((x = Td[PA[*c]]) >= v);) {
+        if(x == v) { SWAP(*c, *d); --d; }
+      }
+    }
+
+    if(a <= d) {
+      c = b - 1;
+
+      if((s = a - first) > (t = b - a)) { s = t; }
+      for(e = first, f = b - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
+      if((s = d - c) > (t = last - d - 1)) { s = t; }
+      for(e = b, f = last - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
+
+      a = first + (b - a), c = last - (d - c);
+      b = (v <= Td[PA[*a] - 1]) ? a : ss_partition(PA, a, c, depth);
+
+      if((a - first) <= (last - c)) {
+        if((last - c) <= (c - b)) {
+          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
+          STACK_PUSH(c, last, depth, limit);
+          last = a;
+        } else if((a - first) <= (c - b)) {
+          STACK_PUSH(c, last, depth, limit);
+          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
+          last = a;
+        } else {
+          STACK_PUSH(c, last, depth, limit);
+          STACK_PUSH(first, a, depth, limit);
+          first = b, last = c, depth += 1, limit = ss_ilg(c - b);
+        }
+      } else {
+        if((a - first) <= (c - b)) {
+          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
+          STACK_PUSH(first, a, depth, limit);
+          first = c;
+        } else if((last - c) <= (c - b)) {
+          STACK_PUSH(first, a, depth, limit);
+          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
+          first = c;
+        } else {
+          STACK_PUSH(first, a, depth, limit);
+          STACK_PUSH(c, last, depth, limit);
+          first = b, last = c, depth += 1, limit = ss_ilg(c - b);
+        }
+      }
+    } else {
+      limit += 1;
+      if(Td[PA[*first] - 1] < v) {
+        first = ss_partition(PA, first, last, depth);
+        limit = ss_ilg(last - first);
+      }
+      depth += 1;
+    }
+  }
+#undef STACK_SIZE
+}
+
+#endif /* (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE) */
+
+
+/*---------------------------------------------------------------------------*/
+
+#if SS_BLOCKSIZE != 0
+
+static INLINE
+void
+ss_blockswap(int *a, int *b, int n) {
+  int t;
+  for(; 0 < n; --n, ++a, ++b) {
+    t = *a, *a = *b, *b = t;
+  }
+}
+
+static INLINE
+void
+ss_rotate(int *first, int *middle, int *last) {
+  int *a, *b, t;
+  int l, r;
+  l = middle - first, r = last - middle;
+  for(; (0 < l) && (0 < r);) {
+    if(l == r) { ss_blockswap(first, middle, l); break; }
+    if(l < r) {
+      a = last - 1, b = middle - 1;
+      t = *a;
+      do {
+        *a-- = *b, *b-- = *a;
+        if(b < first) {
+          *a = t;
+          last = a;
+          if((r -= l + 1) <= l) { break; }
+          a -= 1, b = middle - 1;
+          t = *a;
+        }
+      } while(1);
+    } else {
+      a = first, b = middle;
+      t = *a;
+      do {
+        *a++ = *b, *b++ = *a;
+        if(last <= b) {
+          *a = t;
+          first = a + 1;
+          if((l -= r + 1) <= r) { break; }
+          a += 1, b = middle;
+          t = *a;
+        }
+      } while(1);
+    }
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+static
+void
+ss_inplacemerge(const unsigned char *T, const int *PA,
+                int *first, int *middle, int *last,
+                int depth) {
+  const int *p;
+  int *a, *b;
+  int len, half;
+  int q, r;
+  int x;
+
+  for(;;) {
+    if(*(last - 1) < 0) { x = 1; p = PA + ~*(last - 1); }
+    else                { x = 0; p = PA +  *(last - 1); }
+    for(a = first, len = middle - first, half = len >> 1, r = -1;
+        0 < len;
+        len = half, half >>= 1) {
+      b = a + half;
+      q = ss_compare(T, PA + ((0 <= *b) ? *b : ~*b), p, depth);
+      if(q < 0) {
+        a = b + 1;
+        half -= (len & 1) ^ 1;
+      } else {
+        r = q;
+      }
+    }
+    if(a < middle) {
+      if(r == 0) { *a = ~*a; }
+      ss_rotate(a, middle, last);
+      last -= middle - a;
+      middle = a;
+      if(first == middle) { break; }
+    }
+    --last;
+    if(x != 0) { while(*--last < 0) { } }
+    if(middle == last) { break; }
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Merge-forward with internal buffer. */
+static
+void
+ss_mergeforward(const unsigned char *T, const int *PA,
+                int *first, int *middle, int *last,
+                int *buf, int depth) {
+  int *a, *b, *c, *bufend;
+  int t;
+  int r;
+
+  bufend = buf + (middle - first) - 1;
+  ss_blockswap(buf, first, middle - first);
+
+  for(t = *(a = first), b = buf, c = middle;;) {
+    r = ss_compare(T, PA + *b, PA + *c, depth);
+    if(r < 0) {
+      do {
+        *a++ = *b;
+        if(bufend <= b) { *bufend = t; return; }
+        *b++ = *a;
+      } while(*b < 0);
+    } else if(r > 0) {
+      do {
+        *a++ = *c, *c++ = *a;
+        if(last <= c) {
+          while(b < bufend) { *a++ = *b, *b++ = *a; }
+          *a = *b, *b = t;
+          return;
+        }
+      } while(*c < 0);
+    } else {
+      *c = ~*c;
+      do {
+        *a++ = *b;
+        if(bufend <= b) { *bufend = t; return; }
+        *b++ = *a;
+      } while(*b < 0);
+
+      do {
+        *a++ = *c, *c++ = *a;
+        if(last <= c) {
+          while(b < bufend) { *a++ = *b, *b++ = *a; }
+          *a = *b, *b = t;
+          return;
+        }
+      } while(*c < 0);
+    }
+  }
+}
+
+/* Merge-backward with internal buffer. */
+static
+void
+ss_mergebackward(const unsigned char *T, const int *PA,
+                 int *first, int *middle, int *last,
+                 int *buf, int depth) {
+  const int *p1, *p2;
+  int *a, *b, *c, *bufend;
+  int t;
+  int r;
+  int x;
+
+  bufend = buf + (last - middle) - 1;
+  ss_blockswap(buf, middle, last - middle);
+
+  x = 0;
+  if(*bufend < 0)       { p1 = PA + ~*bufend; x |= 1; }
+  else                  { p1 = PA +  *bufend; }
+  if(*(middle - 1) < 0) { p2 = PA + ~*(middle - 1); x |= 2; }
+  else                  { p2 = PA +  *(middle - 1); }
+  for(t = *(a = last - 1), b = bufend, c = middle - 1;;) {
+    r = ss_compare(T, p1, p2, depth);
+    if(0 < r) {
+      if(x & 1) { do { *a-- = *b, *b-- = *a; } while(*b < 0); x ^= 1; }
+      *a-- = *b;
+      if(b <= buf) { *buf = t; break; }
+      *b-- = *a;
+      if(*b < 0) { p1 = PA + ~*b; x |= 1; }
+      else       { p1 = PA +  *b; }
+    } else if(r < 0) {
+      if(x & 2) { do { *a-- = *c, *c-- = *a; } while(*c < 0); x ^= 2; }
+      *a-- = *c, *c-- = *a;
+      if(c < first) {
+        while(buf < b) { *a-- = *b, *b-- = *a; }
+        *a = *b, *b = t;
+        break;
+      }
+      if(*c < 0) { p2 = PA + ~*c; x |= 2; }
+      else       { p2 = PA +  *c; }
+    } else {
+      if(x & 1) { do { *a-- = *b, *b-- = *a; } while(*b < 0); x ^= 1; }
+      *a-- = ~*b;
+      if(b <= buf) { *buf = t; break; }
+      *b-- = *a;
+      if(x & 2) { do { *a-- = *c, *c-- = *a; } while(*c < 0); x ^= 2; }
+      *a-- = *c, *c-- = *a;
+      if(c < first) {
+        while(buf < b) { *a-- = *b, *b-- = *a; }
+        *a = *b, *b = t;
+        break;
+      }
+      if(*b < 0) { p1 = PA + ~*b; x |= 1; }
+      else       { p1 = PA +  *b; }
+      if(*c < 0) { p2 = PA + ~*c; x |= 2; }
+      else       { p2 = PA +  *c; }
+    }
+  }
+}
+
+/* D&C based merge. */
+static
+void
+ss_swapmerge(const unsigned char *T, const int *PA,
+             int *first, int *middle, int *last,
+             int *buf, int bufsize, int depth) {
+#define STACK_SIZE SS_SMERGE_STACKSIZE
+#define GETIDX(a) ((0 <= (a)) ? (a) : (~(a)))
+#define MERGE_CHECK(a, b, c)\
+  do {\
+    if(((c) & 1) ||\
+       (((c) & 2) && (ss_compare(T, PA + GETIDX(*((a) - 1)), PA + *(a), depth) == 0))) {\
+      *(a) = ~*(a);\
+    }\
+    if(((c) & 4) && ((ss_compare(T, PA + GETIDX(*((b) - 1)), PA + *(b), depth) == 0))) {\
+      *(b) = ~*(b);\
+    }\
+  } while(0)
+  struct { int *a, *b, *c; int d; } stack[STACK_SIZE];
+  int *l, *r, *lm, *rm;
+  int m, len, half;
+  int ssize;
+  int check, next;
+
+  for(check = 0, ssize = 0;;) {
+    if((last - middle) <= bufsize) {
+      if((first < middle) && (middle < last)) {
+        ss_mergebackward(T, PA, first, middle, last, buf, depth);
+      }
+      MERGE_CHECK(first, last, check);
+      STACK_POP(first, middle, last, check);
+      continue;
+    }
+
+    if((middle - first) <= bufsize) {
+      if(first < middle) {
+        ss_mergeforward(T, PA, first, middle, last, buf, depth);
+      }
+      MERGE_CHECK(first, last, check);
+      STACK_POP(first, middle, last, check);
+      continue;
+    }
+
+    for(m = 0, len = MIN(middle - first, last - middle), half = len >> 1;
+        0 < len;
+        len = half, half >>= 1) {
+      if(ss_compare(T, PA + GETIDX(*(middle + m + half)),
+                       PA + GETIDX(*(middle - m - half - 1)), depth) < 0) {
+        m += half + 1;
+        half -= (len & 1) ^ 1;
+      }
+    }
+
+    if(0 < m) {
+      lm = middle - m, rm = middle + m;
+      ss_blockswap(lm, middle, m);
+      l = r = middle, next = 0;
+      if(rm < last) {
+        if(*rm < 0) {
+          *rm = ~*rm;
+          if(first < lm) { for(; *--l < 0;) { } next |= 4; }
+          next |= 1;
+        } else if(first < lm) {
+          for(; *r < 0; ++r) { }
+          next |= 2;
+        }
+      }
+
+      if((l - first) <= (last - r)) {
+        STACK_PUSH(r, rm, last, (next & 3) | (check & 4));
+        middle = lm, last = l, check = (check & 3) | (next & 4);
+      } else {
+        if((next & 2) && (r == middle)) { next ^= 6; }
+        STACK_PUSH(first, lm, l, (check & 3) | (next & 4));
+        first = r, middle = rm, check = (next & 3) | (check & 4);
+      }
+    } else {
+      if(ss_compare(T, PA + GETIDX(*(middle - 1)), PA + *middle, depth) == 0) {
+        *middle = ~*middle;
+      }
+      MERGE_CHECK(first, last, check);
+      STACK_POP(first, middle, last, check);
+    }
+  }
+#undef STACK_SIZE
+}
+
+#endif /* SS_BLOCKSIZE != 0 */
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Substring sort */
+static
+void
+sssort(const unsigned char *T, const int *PA,
+       int *first, int *last,
+       int *buf, int bufsize,
+       int depth, int n, int lastsuffix) {
+  int *a;
+#if SS_BLOCKSIZE != 0
+  int *b, *middle, *curbuf;
+  int j, k, curbufsize, limit;
+#endif
+  int i;
+
+  if(lastsuffix != 0) { ++first; }
+
+#if SS_BLOCKSIZE == 0
+  ss_mintrosort(T, PA, first, last, depth);
+#else
+  if((bufsize < SS_BLOCKSIZE) &&
+      (bufsize < (last - first)) &&
+      (bufsize < (limit = ss_isqrt(last - first)))) {
+    if(SS_BLOCKSIZE < limit) { limit = SS_BLOCKSIZE; }
+    buf = middle = last - limit, bufsize = limit;
+  } else {
+    middle = last, limit = 0;
+  }
+  for(a = first, i = 0; SS_BLOCKSIZE < (middle - a); a += SS_BLOCKSIZE, ++i) {
+#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
+    ss_mintrosort(T, PA, a, a + SS_BLOCKSIZE, depth);
+#elif 1 < SS_BLOCKSIZE
+    ss_insertionsort(T, PA, a, a + SS_BLOCKSIZE, depth);
+#endif
+    curbufsize = last - (a + SS_BLOCKSIZE);
+    curbuf = a + SS_BLOCKSIZE;
+    if(curbufsize <= bufsize) { curbufsize = bufsize, curbuf = buf; }
+    for(b = a, k = SS_BLOCKSIZE, j = i; j & 1; b -= k, k <<= 1, j >>= 1) {
+      ss_swapmerge(T, PA, b - k, b, b + k, curbuf, curbufsize, depth);
+    }
+  }
+#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
+  ss_mintrosort(T, PA, a, middle, depth);
+#elif 1 < SS_BLOCKSIZE
+  ss_insertionsort(T, PA, a, middle, depth);
+#endif
+  for(k = SS_BLOCKSIZE; i != 0; k <<= 1, i >>= 1) {
+    if(i & 1) {
+      ss_swapmerge(T, PA, a - k, a, middle, buf, bufsize, depth);
+      a -= k;
+    }
+  }
+  if(limit != 0) {
+#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
+    ss_mintrosort(T, PA, middle, last, depth);
+#elif 1 < SS_BLOCKSIZE
+    ss_insertionsort(T, PA, middle, last, depth);
+#endif
+    ss_inplacemerge(T, PA, first, middle, last, depth);
+  }
+#endif
+
+  if(lastsuffix != 0) {
+    /* Insert last type B* suffix. */
+    int PAi[2]; PAi[0] = PA[*(first - 1)], PAi[1] = n - 2;
+    for(a = first, i = *(first - 1);
+        (a < last) && ((*a < 0) || (0 < ss_compare(T, &(PAi[0]), PA + *a, depth)));
+        ++a) {
+      *(a - 1) = *a;
+    }
+    *(a - 1) = i;
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+static INLINE
+int
+tr_ilg(int n) {
+  return (n & 0xffff0000) ?
+          ((n & 0xff000000) ?
+            24 + lg_table[(n >> 24) & 0xff] :
+            16 + lg_table[(n >> 16) & 0xff]) :
+          ((n & 0x0000ff00) ?
+             8 + lg_table[(n >>  8) & 0xff] :
+             0 + lg_table[(n >>  0) & 0xff]);
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Simple insertionsort for small size groups. */
+static
+void
+tr_insertionsort(const int *ISAd, int *first, int *last) {
+  int *a, *b;
+  int t, r;
+
+  for(a = first + 1; a < last; ++a) {
+    for(t = *a, b = a - 1; 0 > (r = ISAd[t] - ISAd[*b]);) {
+      do { *(b + 1) = *b; } while((first <= --b) && (*b < 0));
+      if(b < first) { break; }
+    }
+    if(r == 0) { *b = ~*b; }
+    *(b + 1) = t;
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+static INLINE
+void
+tr_fixdown(const int *ISAd, int *SA, int i, int size) {
+  int j, k;
+  int v;
+  int c, d, e;
+
+  for(v = SA[i], c = ISAd[v]; (j = 2 * i + 1) < size; SA[i] = SA[k], i = k) {
+    d = ISAd[SA[k = j++]];
+    if(d < (e = ISAd[SA[j]])) { k = j; d = e; }
+    if(d <= c) { break; }
+  }
+  SA[i] = v;
+}
+
+/* Simple top-down heapsort. */
+static
+void
+tr_heapsort(const int *ISAd, int *SA, int size) {
+  int i, m;
+  int t;
+
+  m = size;
+  if((size % 2) == 0) {
+    m--;
+    if(ISAd[SA[m / 2]] < ISAd[SA[m]]) { SWAP(SA[m], SA[m / 2]); }
+  }
+
+  for(i = m / 2 - 1; 0 <= i; --i) { tr_fixdown(ISAd, SA, i, m); }
+  if((size % 2) == 0) { SWAP(SA[0], SA[m]); tr_fixdown(ISAd, SA, 0, m); }
+  for(i = m - 1; 0 < i; --i) {
+    t = SA[0], SA[0] = SA[i];
+    tr_fixdown(ISAd, SA, 0, i);
+    SA[i] = t;
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Returns the median of three elements. */
+static INLINE
+int *
+tr_median3(const int *ISAd, int *v1, int *v2, int *v3) {
+  int *t;
+  if(ISAd[*v1] > ISAd[*v2]) { SWAP(v1, v2); }
+  if(ISAd[*v2] > ISAd[*v3]) {
+    if(ISAd[*v1] > ISAd[*v3]) { return v1; }
+    else { return v3; }
+  }
+  return v2;
+}
+
+/* Returns the median of five elements. */
+static INLINE
+int *
+tr_median5(const int *ISAd,
+           int *v1, int *v2, int *v3, int *v4, int *v5) {
+  int *t;
+  if(ISAd[*v2] > ISAd[*v3]) { SWAP(v2, v3); }
+  if(ISAd[*v4] > ISAd[*v5]) { SWAP(v4, v5); }
+  if(ISAd[*v2] > ISAd[*v4]) { SWAP(v2, v4); SWAP(v3, v5); }
+  if(ISAd[*v1] > ISAd[*v3]) { SWAP(v1, v3); }
+  if(ISAd[*v1] > ISAd[*v4]) { SWAP(v1, v4); SWAP(v3, v5); }
+  if(ISAd[*v3] > ISAd[*v4]) { return v4; }
+  return v3;
+}
+
+/* Returns the pivot element. */
+static INLINE
+int *
+tr_pivot(const int *ISAd, int *first, int *last) {
+  int *middle;
+  int t;
+
+  t = last - first;
+  middle = first + t / 2;
+
+  if(t <= 512) {
+    if(t <= 32) {
+      return tr_median3(ISAd, first, middle, last - 1);
+    } else {
+      t >>= 2;
+      return tr_median5(ISAd, first, first + t, middle, last - 1 - t, last - 1);
+    }
+  }
+  t >>= 3;
+  first  = tr_median3(ISAd, first, first + t, first + (t << 1));
+  middle = tr_median3(ISAd, middle - t, middle, middle + t);
+  last   = tr_median3(ISAd, last - 1 - (t << 1), last - 1 - t, last - 1);
+  return tr_median3(ISAd, first, middle, last);
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+typedef struct _trbudget_t trbudget_t;
+struct _trbudget_t {
+  int chance;
+  int remain;
+  int incval;
+  int count;
+};
+
+static INLINE
+void
+trbudget_init(trbudget_t *budget, int chance, int incval) {
+  budget->chance = chance;
+  budget->remain = budget->incval = incval;
+}
+
+static INLINE
+int
+trbudget_check(trbudget_t *budget, int size) {
+  if(size <= budget->remain) { budget->remain -= size; return 1; }
+  if(budget->chance == 0) { budget->count += size; return 0; }
+  budget->remain += budget->incval - size;
+  budget->chance -= 1;
+  return 1;
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+static INLINE
+void
+tr_partition(const int *ISAd,
+             int *first, int *middle, int *last,
+             int **pa, int **pb, int v) {
+  int *a, *b, *c, *d, *e, *f;
+  int t, s;
+  int x = 0;
+
+  for(b = middle - 1; (++b < last) && ((x = ISAd[*b]) == v);) { }
+  if(((a = b) < last) && (x < v)) {
+    for(; (++b < last) && ((x = ISAd[*b]) <= v);) {
+      if(x == v) { SWAP(*b, *a); ++a; }
+    }
+  }
+  for(c = last; (b < --c) && ((x = ISAd[*c]) == v);) { }
+  if((b < (d = c)) && (x > v)) {
+    for(; (b < --c) && ((x = ISAd[*c]) >= v);) {
+      if(x == v) { SWAP(*c, *d); --d; }
+    }
+  }
+  for(; b < c;) {
+    SWAP(*b, *c);
+    for(; (++b < c) && ((x = ISAd[*b]) <= v);) {
+      if(x == v) { SWAP(*b, *a); ++a; }
+    }
+    for(; (b < --c) && ((x = ISAd[*c]) >= v);) {
+      if(x == v) { SWAP(*c, *d); --d; }
+    }
+  }
+
+  if(a <= d) {
+    c = b - 1;
+    if((s = a - first) > (t = b - a)) { s = t; }
+    for(e = first, f = b - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
+    if((s = d - c) > (t = last - d - 1)) { s = t; }
+    for(e = b, f = last - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
+    first += (b - a), last -= (d - c);
+  }
+  *pa = first, *pb = last;
+}
+
+static
+void
+tr_copy(int *ISA, const int *SA,
+        int *first, int *a, int *b, int *last,
+        int depth) {
+  /* sort suffixes of middle partition
+     by using sorted order of suffixes of left and right partition. */
+  int *c, *d, *e;
+  int s, v;
+
+  v = b - SA - 1;
+  for(c = first, d = a - 1; c <= d; ++c) {
+    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
+      *++d = s;
+      ISA[s] = d - SA;
+    }
+  }
+  for(c = last - 1, e = d + 1, d = b; e < d; --c) {
+    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
+      *--d = s;
+      ISA[s] = d - SA;
+    }
+  }
+}
+
+static
+void
+tr_partialcopy(int *ISA, const int *SA,
+               int *first, int *a, int *b, int *last,
+               int depth) {
+  int *c, *d, *e;
+  int s, v;
+  int rank, lastrank, newrank = -1;
+
+  v = b - SA - 1;
+  lastrank = -1;
+  for(c = first, d = a - 1; c <= d; ++c) {
+    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
+      *++d = s;
+      rank = ISA[s + depth];
+      if(lastrank != rank) { lastrank = rank; newrank = d - SA; }
+      ISA[s] = newrank;
+    }
+  }
+
+  lastrank = -1;
+  for(e = d; first <= e; --e) {
+    rank = ISA[*e];
+    if(lastrank != rank) { lastrank = rank; newrank = e - SA; }
+    if(newrank != rank) { ISA[*e] = newrank; }
+  }
+
+  lastrank = -1;
+  for(c = last - 1, e = d + 1, d = b; e < d; --c) {
+    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
+      *--d = s;
+      rank = ISA[s + depth];
+      if(lastrank != rank) { lastrank = rank; newrank = d - SA; }
+      ISA[s] = newrank;
+    }
+  }
+}
+
+static
+void
+tr_introsort(int *ISA, const int *ISAd,
+             int *SA, int *first, int *last,
+             trbudget_t *budget) {
+#define STACK_SIZE TR_STACKSIZE
+  struct { const int *a; int *b, *c; int d, e; }stack[STACK_SIZE];
+  int *a, *b, *c;
+  int t;
+  int v, x = 0;
+  int incr = ISAd - ISA;
+  int limit, next;
+  int ssize, trlink = -1;
+
+  for(ssize = 0, limit = tr_ilg(last - first);;) {
+
+    if(limit < 0) {
+      if(limit == -1) {
+        /* tandem repeat partition */
+        tr_partition(ISAd - incr, first, first, last, &a, &b, last - SA - 1);
+
+        /* update ranks */
+        if(a < last) {
+          for(c = first, v = a - SA - 1; c < a; ++c) { ISA[*c] = v; }
+        }
+        if(b < last) {
+          for(c = a, v = b - SA - 1; c < b; ++c) { ISA[*c] = v; }
+        }
+
+        /* push */
+        if(1 < (b - a)) {
+          STACK_PUSH5(NULL, a, b, 0, 0);
+          STACK_PUSH5(ISAd - incr, first, last, -2, trlink);
+          trlink = ssize - 2;
+        }
+        if((a - first) <= (last - b)) {
+          if(1 < (a - first)) {
+            STACK_PUSH5(ISAd, b, last, tr_ilg(last - b), trlink);
+            last = a, limit = tr_ilg(a - first);
+          } else if(1 < (last - b)) {
+            first = b, limit = tr_ilg(last - b);
+          } else {
+            STACK_POP5(ISAd, first, last, limit, trlink);
+          }
+        } else {
+          if(1 < (last - b)) {
+            STACK_PUSH5(ISAd, first, a, tr_ilg(a - first), trlink);
+            first = b, limit = tr_ilg(last - b);
+          } else if(1 < (a - first)) {
+            last = a, limit = tr_ilg(a - first);
+          } else {
+            STACK_POP5(ISAd, first, last, limit, trlink);
+          }
+        }
+      } else if(limit == -2) {
+        /* tandem repeat copy */
+        a = stack[--ssize].b, b = stack[ssize].c;
+        if(stack[ssize].d == 0) {
+          tr_copy(ISA, SA, first, a, b, last, ISAd - ISA);
+        } else {
+          if(0 <= trlink) { stack[trlink].d = -1; }
+          tr_partialcopy(ISA, SA, first, a, b, last, ISAd - ISA);
+        }
+        STACK_POP5(ISAd, first, last, limit, trlink);
+      } else {
+        /* sorted partition */
+        if(0 <= *first) {
+          a = first;
+          do { ISA[*a] = a - SA; } while((++a < last) && (0 <= *a));
+          first = a;
+        }
+        if(first < last) {
+          a = first; do { *a = ~*a; } while(*++a < 0);
+          next = (ISA[*a] != ISAd[*a]) ? tr_ilg(a - first + 1) : -1;
+          if(++a < last) { for(b = first, v = a - SA - 1; b < a; ++b) { ISA[*b] = v; } }
+
+          /* push */
+          if(trbudget_check(budget, a - first)) {
+            if((a - first) <= (last - a)) {
+              STACK_PUSH5(ISAd, a, last, -3, trlink);
+              ISAd += incr, last = a, limit = next;
+            } else {
+              if(1 < (last - a)) {
+                STACK_PUSH5(ISAd + incr, first, a, next, trlink);
+                first = a, limit = -3;
+              } else {
+                ISAd += incr, last = a, limit = next;
+              }
+            }
+          } else {
+            if(0 <= trlink) { stack[trlink].d = -1; }
+            if(1 < (last - a)) {
+              first = a, limit = -3;
+            } else {
+              STACK_POP5(ISAd, first, last, limit, trlink);
+            }
+          }
+        } else {
+          STACK_POP5(ISAd, first, last, limit, trlink);
+        }
+      }
+      continue;
+    }
+
+    if((last - first) <= TR_INSERTIONSORT_THRESHOLD) {
+      tr_insertionsort(ISAd, first, last);
+      limit = -3;
+      continue;
+    }
+
+    if(limit-- == 0) {
+      tr_heapsort(ISAd, first, last - first);
+      for(a = last - 1; first < a; a = b) {
+        for(x = ISAd[*a], b = a - 1; (first <= b) && (ISAd[*b] == x); --b) { *b = ~*b; }
+      }
+      limit = -3;
+      continue;
+    }
+
+    /* choose pivot */
+    a = tr_pivot(ISAd, first, last);
+    SWAP(*first, *a);
+    v = ISAd[*first];
+
+    /* partition */
+    tr_partition(ISAd, first, first + 1, last, &a, &b, v);
+    if((last - first) != (b - a)) {
+      next = (ISA[*a] != v) ? tr_ilg(b - a) : -1;
+
+      /* update ranks */
+      for(c = first, v = a - SA - 1; c < a; ++c) { ISA[*c] = v; }
+      if(b < last) { for(c = a, v = b - SA - 1; c < b; ++c) { ISA[*c] = v; } }
+
+      /* push */
+      if((1 < (b - a)) && (trbudget_check(budget, b - a))) {
+        if((a - first) <= (last - b)) {
+          if((last - b) <= (b - a)) {
+            if(1 < (a - first)) {
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              STACK_PUSH5(ISAd, b, last, limit, trlink);
+              last = a;
+            } else if(1 < (last - b)) {
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              first = b;
+            } else {
+              ISAd += incr, first = a, last = b, limit = next;
+            }
+          } else if((a - first) <= (b - a)) {
+            if(1 < (a - first)) {
+              STACK_PUSH5(ISAd, b, last, limit, trlink);
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              last = a;
+            } else {
+              STACK_PUSH5(ISAd, b, last, limit, trlink);
+              ISAd += incr, first = a, last = b, limit = next;
+            }
+          } else {
+            STACK_PUSH5(ISAd, b, last, limit, trlink);
+            STACK_PUSH5(ISAd, first, a, limit, trlink);
+            ISAd += incr, first = a, last = b, limit = next;
+          }
+        } else {
+          if((a - first) <= (b - a)) {
+            if(1 < (last - b)) {
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              STACK_PUSH5(ISAd, first, a, limit, trlink);
+              first = b;
+            } else if(1 < (a - first)) {
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              last = a;
+            } else {
+              ISAd += incr, first = a, last = b, limit = next;
+            }
+          } else if((last - b) <= (b - a)) {
+            if(1 < (last - b)) {
+              STACK_PUSH5(ISAd, first, a, limit, trlink);
+              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
+              first = b;
+            } else {
+              STACK_PUSH5(ISAd, first, a, limit, trlink);
+              ISAd += incr, first = a, last = b, limit = next;
+            }
+          } else {
+            STACK_PUSH5(ISAd, first, a, limit, trlink);
+            STACK_PUSH5(ISAd, b, last, limit, trlink);
+            ISAd += incr, first = a, last = b, limit = next;
+          }
+        }
+      } else {
+        if((1 < (b - a)) && (0 <= trlink)) { stack[trlink].d = -1; }
+        if((a - first) <= (last - b)) {
+          if(1 < (a - first)) {
+            STACK_PUSH5(ISAd, b, last, limit, trlink);
+            last = a;
+          } else if(1 < (last - b)) {
+            first = b;
+          } else {
+            STACK_POP5(ISAd, first, last, limit, trlink);
+          }
+        } else {
+          if(1 < (last - b)) {
+            STACK_PUSH5(ISAd, first, a, limit, trlink);
+            first = b;
+          } else if(1 < (a - first)) {
+            last = a;
+          } else {
+            STACK_POP5(ISAd, first, last, limit, trlink);
+          }
+        }
+      }
+    } else {
+      if(trbudget_check(budget, last - first)) {
+        limit = tr_ilg(last - first), ISAd += incr;
+      } else {
+        if(0 <= trlink) { stack[trlink].d = -1; }
+        STACK_POP5(ISAd, first, last, limit, trlink);
+      }
+    }
+  }
+#undef STACK_SIZE
+}
+
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Tandem repeat sort */
+static
+void
+trsort(int *ISA, int *SA, int n, int depth) {
+  int *ISAd;
+  int *first, *last;
+  trbudget_t budget;
+  int t, skip, unsorted;
+
+  trbudget_init(&budget, tr_ilg(n) * 2 / 3, n);
+/*  trbudget_init(&budget, tr_ilg(n) * 3 / 4, n); */
+  for(ISAd = ISA + depth; -n < *SA; ISAd += ISAd - ISA) {
+    first = SA;
+    skip = 0;
+    unsorted = 0;
+    do {
+      if((t = *first) < 0) { first -= t; skip += t; }
+      else {
+        if(skip != 0) { *(first + skip) = skip; skip = 0; }
+        last = SA + ISA[t] + 1;
+        if(1 < (last - first)) {
+          budget.count = 0;
+          tr_introsort(ISA, ISAd, SA, first, last, &budget);
+          if(budget.count != 0) { unsorted += budget.count; }
+          else { skip = first - last; }
+        } else if((last - first) == 1) {
+          skip = -1;
+        }
+        first = last;
+      }
+    } while(first < (SA + n));
+    if(skip != 0) { *(first + skip) = skip; }
+    if(unsorted == 0) { break; }
+  }
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/* Sorts suffixes of type B*. */
+static
+int
+sort_typeBstar(const unsigned char *T, int *SA,
+               int *bucket_A, int *bucket_B,
+               int n, int openMP) {
+  int *PAb, *ISAb, *buf;
+#ifdef LIBBSC_OPENMP
+  int *curbuf;
+  int l;
+#endif
+  int i, j, k, t, m, bufsize;
+  int c0, c1;
+#ifdef LIBBSC_OPENMP
+  int d0, d1;
+#endif
+  (void)openMP;
+
+  /* Initialize bucket arrays. */
+  for(i = 0; i < BUCKET_A_SIZE; ++i) { bucket_A[i] = 0; }
+  for(i = 0; i < BUCKET_B_SIZE; ++i) { bucket_B[i] = 0; }
+
+  /* Count the number of occurrences of the first one or two characters of each
+     type A, B and B* suffix. Moreover, store the beginning position of all
+     type B* suffixes into the array SA. */
+  for(i = n - 1, m = n, c0 = T[n - 1]; 0 <= i;) {
+    /* type A suffix. */
+    do { ++BUCKET_A(c1 = c0); } while((0 <= --i) && ((c0 = T[i]) >= c1));
+    if(0 <= i) {
+      /* type B* suffix. */
+      ++BUCKET_BSTAR(c0, c1);
+      SA[--m] = i;
+      /* type B suffix. */
+      for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) <= c1); --i, c1 = c0) {
+        ++BUCKET_B(c0, c1);
+      }
+    }
+  }
+  m = n - m;
+/*
+note:
+  A type B* suffix is lexicographically smaller than a type B suffix that
+  begins with the same first two characters.
+*/
+
+  /* Calculate the index of start/end point of each bucket. */
+  for(c0 = 0, i = 0, j = 0; c0 < ALPHABET_SIZE; ++c0) {
+    t = i + BUCKET_A(c0);
+    BUCKET_A(c0) = i + j; /* start point */
+    i = t + BUCKET_B(c0, c0);
+    for(c1 = c0 + 1; c1 < ALPHABET_SIZE; ++c1) {
+      j += BUCKET_BSTAR(c0, c1);
+      BUCKET_BSTAR(c0, c1) = j; /* end point */
+      i += BUCKET_B(c0, c1);
+    }
+  }
+
+  if(0 < m) {
+    /* Sort the type B* suffixes by their first two characters. */
+    PAb = SA + n - m; ISAb = SA + m;
+    for(i = m - 2; 0 <= i; --i) {
+      t = PAb[i], c0 = T[t], c1 = T[t + 1];
+      SA[--BUCKET_BSTAR(c0, c1)] = i;
+    }
+    t = PAb[m - 1], c0 = T[t], c1 = T[t + 1];
+    SA[--BUCKET_BSTAR(c0, c1)] = m - 1;
+
+    /* Sort the type B* substrings using sssort. */
+#ifdef LIBBSC_OPENMP
+    if (openMP)
+    {
+        buf = SA + m;
+        c0 = ALPHABET_SIZE - 2, c1 = ALPHABET_SIZE - 1, j = m;
+#pragma omp parallel default(shared) private(bufsize, curbuf, k, l, d0, d1)
+        {
+          bufsize = (n - (2 * m)) / omp_get_num_threads();
+          curbuf = buf + omp_get_thread_num() * bufsize;
+          k = 0;
+          for(;;) {
+            #pragma omp critical(sssort_lock)
+            {
+              if(0 < (l = j)) {
+                d0 = c0, d1 = c1;
+                do {
+                  k = BUCKET_BSTAR(d0, d1);
+                  if(--d1 <= d0) {
+                    d1 = ALPHABET_SIZE - 1;
+                    if(--d0 < 0) { break; }
+                  }
+                } while(((l - k) <= 1) && (0 < (l = k)));
+                c0 = d0, c1 = d1, j = k;
+              }
+            }
+            if(l == 0) { break; }
+            sssort(T, PAb, SA + k, SA + l,
+                   curbuf, bufsize, 2, n, *(SA + k) == (m - 1));
+          }
+        }
+    }
+    else
+    {
+        buf = SA + m, bufsize = n - (2 * m);
+        for(c0 = ALPHABET_SIZE - 2, j = m; 0 < j; --c0) {
+          for(c1 = ALPHABET_SIZE - 1; c0 < c1; j = i, --c1) {
+            i = BUCKET_BSTAR(c0, c1);
+            if(1 < (j - i)) {
+              sssort(T, PAb, SA + i, SA + j,
+                     buf, bufsize, 2, n, *(SA + i) == (m - 1));
+            }
+          }
+        }
+    }
+#else
+    buf = SA + m, bufsize = n - (2 * m);
+    for(c0 = ALPHABET_SIZE - 2, j = m; 0 < j; --c0) {
+      for(c1 = ALPHABET_SIZE - 1; c0 < c1; j = i, --c1) {
+        i = BUCKET_BSTAR(c0, c1);
+        if(1 < (j - i)) {
+          sssort(T, PAb, SA + i, SA + j,
+                 buf, bufsize, 2, n, *(SA + i) == (m - 1));
+        }
+      }
+    }
+#endif
+
+    /* Compute ranks of type B* substrings. */
+    for(i = m - 1; 0 <= i; --i) {
+      if(0 <= SA[i]) {
+        j = i;
+        do { ISAb[SA[i]] = i; } while((0 <= --i) && (0 <= SA[i]));
+        SA[i + 1] = i - j;
+        if(i <= 0) { break; }
+      }
+      j = i;
+      do { ISAb[SA[i] = ~SA[i]] = j; } while(SA[--i] < 0);
+      ISAb[SA[i]] = j;
+    }
+
+    /* Construct the inverse suffix array of type B* suffixes using trsort. */
+    trsort(ISAb, SA, m, 1);
+
+    /* Set the sorted order of tyoe B* suffixes. */
+    for(i = n - 1, j = m, c0 = T[n - 1]; 0 <= i;) {
+      for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) >= c1); --i, c1 = c0) { }
+      if(0 <= i) {
+        t = i;
+        for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) <= c1); --i, c1 = c0) { }
+        SA[ISAb[--j]] = ((t == 0) || (1 < (t - i))) ? t : ~t;
+      }
+    }
+
+    /* Calculate the index of start/end point of each bucket. */
+    BUCKET_B(ALPHABET_SIZE - 1, ALPHABET_SIZE - 1) = n; /* end point */
+    for(c0 = ALPHABET_SIZE - 2, k = m - 1; 0 <= c0; --c0) {
+      i = BUCKET_A(c0 + 1) - 1;
+      for(c1 = ALPHABET_SIZE - 1; c0 < c1; --c1) {
+        t = i - BUCKET_B(c0, c1);
+        BUCKET_B(c0, c1) = i; /* end point */
+
+        /* Move all type B* suffixes to the correct position. */
+        for(i = t, j = BUCKET_BSTAR(c0, c1);
+            j <= k;
+            --i, --k) { SA[i] = SA[k]; }
+      }
+      BUCKET_BSTAR(c0, c0 + 1) = i - BUCKET_B(c0, c0) + 1; /* start point */
+      BUCKET_B(c0, c0) = i; /* end point */
+    }
+  }
+
+  return m;
+}
+
+/* Constructs the suffix array by using the sorted order of type B* suffixes. */
+static
+void
+construct_SA(const unsigned char *T, int *SA,
+             int *bucket_A, int *bucket_B,
+             int n, int m) {
+  int *i, *j, *k;
+  int s;
+  int c0, c1, c2;
+
+  if(0 < m) {
+    /* Construct the sorted order of type B suffixes by using
+       the sorted order of type B* suffixes. */
+    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
+      /* Scan the suffix array from right to left. */
+      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
+          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
+          i <= j;
+          --j) {
+        if(0 < (s = *j)) {
+          assert(T[s] == c1);
+          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
+          assert(T[s - 1] <= T[s]);
+          *j = ~s;
+          c0 = T[--s];
+          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
+          if(c0 != c2) {
+            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
+            k = SA + BUCKET_B(c2 = c0, c1);
+          }
+          assert(k < j);
+          *k-- = s;
+        } else {
+          assert(((s == 0) && (T[s] == c1)) || (s < 0));
+          *j = ~s;
+        }
+      }
+    }
+  }
+
+  /* Construct the suffix array by using
+     the sorted order of type B suffixes. */
+  k = SA + BUCKET_A(c2 = T[n - 1]);
+  *k++ = (T[n - 2] < c2) ? ~(n - 1) : (n - 1);
+  /* Scan the suffix array from left to right. */
+  for(i = SA, j = SA + n; i < j; ++i) {
+    if(0 < (s = *i)) {
+      assert(T[s - 1] >= T[s]);
+      c0 = T[--s];
+      if((s == 0) || (T[s - 1] < c0)) { s = ~s; }
+      if(c0 != c2) {
+        BUCKET_A(c2) = k - SA;
+        k = SA + BUCKET_A(c2 = c0);
+      }
+      assert(i < k);
+      *k++ = s;
+    } else {
+      assert(s < 0);
+      *i = ~s;
+    }
+  }
+}
+
+/* Constructs the burrows-wheeler transformed string directly
+   by using the sorted order of type B* suffixes. */
+static
+int
+construct_BWT(const unsigned char *T, int *SA,
+              int *bucket_A, int *bucket_B,
+              int n, int m) {
+  int *i, *j, *k, *orig;
+  int s;
+  int c0, c1, c2;
+
+  if(0 < m) {
+    /* Construct the sorted order of type B suffixes by using
+       the sorted order of type B* suffixes. */
+    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
+      /* Scan the suffix array from right to left. */
+      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
+          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
+          i <= j;
+          --j) {
+        if(0 < (s = *j)) {
+          assert(T[s] == c1);
+          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
+          assert(T[s - 1] <= T[s]);
+          c0 = T[--s];
+          *j = ~((int)c0);
+          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
+          if(c0 != c2) {
+            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
+            k = SA + BUCKET_B(c2 = c0, c1);
+          }
+          assert(k < j);
+          *k-- = s;
+        } else if(s != 0) {
+          *j = ~s;
+#ifndef NDEBUG
+        } else {
+          assert(T[s] == c1);
+#endif
+        }
+      }
+    }
+  }
+
+  /* Construct the BWTed string by using
+     the sorted order of type B suffixes. */
+  k = SA + BUCKET_A(c2 = T[n - 1]);
+  *k++ = (T[n - 2] < c2) ? ~((int)T[n - 2]) : (n - 1);
+  /* Scan the suffix array from left to right. */
+  for(i = SA, j = SA + n, orig = SA; i < j; ++i) {
+    if(0 < (s = *i)) {
+      assert(T[s - 1] >= T[s]);
+      c0 = T[--s];
+      *i = c0;
+      if((0 < s) && (T[s - 1] < c0)) { s = ~((int)T[s - 1]); }
+      if(c0 != c2) {
+        BUCKET_A(c2) = k - SA;
+        k = SA + BUCKET_A(c2 = c0);
+      }
+      assert(i < k);
+      *k++ = s;
+    } else if(s != 0) {
+      *i = ~s;
+    } else {
+      orig = i;
+    }
+  }
+
+  return orig - SA;
+}
+
+/* Constructs the burrows-wheeler transformed string directly
+   by using the sorted order of type B* suffixes. */
+static
+int
+construct_BWT_indexes(const unsigned char *T, int *SA,
+                      int *bucket_A, int *bucket_B,
+                      int n, int m,
+                      unsigned char * num_indexes, int * indexes) {
+  int *i, *j, *k, *orig;
+  int s;
+  int c0, c1, c2;
+
+  int mod = n / 8;
+  {
+      mod |= mod >> 1;  mod |= mod >> 2;
+      mod |= mod >> 4;  mod |= mod >> 8;
+      mod |= mod >> 16; mod >>= 1;
+
+      *num_indexes = (unsigned char)((n - 1) / (mod + 1));
+  }
+
+  if(0 < m) {
+    /* Construct the sorted order of type B suffixes by using
+       the sorted order of type B* suffixes. */
+    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
+      /* Scan the suffix array from right to left. */
+      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
+          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
+          i <= j;
+          --j) {
+        if(0 < (s = *j)) {
+          assert(T[s] == c1);
+          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
+          assert(T[s - 1] <= T[s]);
+
+          if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = j - SA;
+
+          c0 = T[--s];
+          *j = ~((int)c0);
+          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
+          if(c0 != c2) {
+            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
+            k = SA + BUCKET_B(c2 = c0, c1);
+          }
+          assert(k < j);
+          *k-- = s;
+        } else if(s != 0) {
+          *j = ~s;
+#ifndef NDEBUG
+        } else {
+          assert(T[s] == c1);
+#endif
+        }
+      }
+    }
+  }
+
+  /* Construct the BWTed string by using
+     the sorted order of type B suffixes. */
+  k = SA + BUCKET_A(c2 = T[n - 1]);
+  if (T[n - 2] < c2) {
+    if (((n - 1) & mod) == 0) indexes[(n - 1) / (mod + 1) - 1] = k - SA;
+    *k++ = ~((int)T[n - 2]);
+  }
+  else {
+    *k++ = n - 1;
+  }
+
+  /* Scan the suffix array from left to right. */
+  for(i = SA, j = SA + n, orig = SA; i < j; ++i) {
+    if(0 < (s = *i)) {
+      assert(T[s - 1] >= T[s]);
+
+      if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = i - SA;
+
+      c0 = T[--s];
+      *i = c0;
+      if(c0 != c2) {
+        BUCKET_A(c2) = k - SA;
+        k = SA + BUCKET_A(c2 = c0);
+      }
+      assert(i < k);
+      if((0 < s) && (T[s - 1] < c0)) {
+          if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = k - SA;
+          *k++ = ~((int)T[s - 1]);
+      } else
+        *k++ = s;
+    } else if(s != 0) {
+      *i = ~s;
+    } else {
+      orig = i;
+    }
+  }
+
+  return orig - SA;
+}
+
+
+/*---------------------------------------------------------------------------*/
+
+/*- Function -*/
+
+int
+divsufsort(const unsigned char *T, int *SA, int n, int openMP) {
+  int *bucket_A, *bucket_B;
+  int m;
+  int err = 0;
+
+  /* Check arguments. */
+  if((T == NULL) || (SA == NULL) || (n < 0)) { return -1; }
+  else if(n == 0) { return 0; }
+  else if(n == 1) { SA[0] = 0; return 0; }
+  else if(n == 2) { m = (T[0] < T[1]); SA[m ^ 1] = 0, SA[m] = 1; return 0; }
+
+  bucket_A = (int *)malloc(BUCKET_A_SIZE * sizeof(int));
+  bucket_B = (int *)malloc(BUCKET_B_SIZE * sizeof(int));
+
+  /* Suffixsort. */
+  if((bucket_A != NULL) && (bucket_B != NULL)) {
+    m = sort_typeBstar(T, SA, bucket_A, bucket_B, n, openMP);
+    construct_SA(T, SA, bucket_A, bucket_B, n, m);
+  } else {
+    err = -2;
+  }
+
+  free(bucket_B);
+  free(bucket_A);
+
+  return err;
+}
+
+int
+divbwt(const unsigned char *T, unsigned char *U, int *A, int n, unsigned char * num_indexes, int * indexes, int openMP) {
+  int *B;
+  int *bucket_A, *bucket_B;
+  int m, pidx, i;
+
+  /* Check arguments. */
+  if((T == NULL) || (U == NULL) || (n < 0)) { return -1; }
+  else if(n <= 1) { if(n == 1) { U[0] = T[0]; } return n; }
+
+  if((B = A) == NULL) { B = (int *)malloc((size_t)(n + 1) * sizeof(int)); }
+  bucket_A = (int *)malloc(BUCKET_A_SIZE * sizeof(int));
+  bucket_B = (int *)malloc(BUCKET_B_SIZE * sizeof(int));
+
+  /* Burrows-Wheeler Transform. */
+  if((B != NULL) && (bucket_A != NULL) && (bucket_B != NULL)) {
+    m = sort_typeBstar(T, B, bucket_A, bucket_B, n, openMP);
+
+    if (num_indexes == NULL || indexes == NULL) {
+        pidx = construct_BWT(T, B, bucket_A, bucket_B, n, m);
+    } else {
+        pidx = construct_BWT_indexes(T, B, bucket_A, bucket_B, n, m, num_indexes, indexes);
+    }
+
+    /* Copy to output string. */
+    U[0] = T[n - 1];
+    for(i = 0; i < pidx; ++i) { U[i + 1] = (unsigned char)B[i]; }
+    for(i += 1; i < n; ++i) { U[i] = (unsigned char)B[i]; }
+    pidx += 1;
+  } else {
+    pidx = -2;
+  }
+
+  free(bucket_B);
+  free(bucket_A);
+  if(A == NULL) { free(B); }
+
+  return pidx;
+}
diff --git a/vendor/github.com/DataDog/zstd/divsufsort.h b/vendor/github.com/DataDog/zstd/divsufsort.h
new file mode 100644
index 0000000..5440994
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/divsufsort.h
@@ -0,0 +1,67 @@
+/*
+ * divsufsort.h for libdivsufsort-lite
+ * Copyright (c) 2003-2008 Yuta Mori All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef _DIVSUFSORT_H
+#define _DIVSUFSORT_H 1
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+
+/*- Prototypes -*/
+
+/**
+ * Constructs the suffix array of a given string.
+ * @param T [0..n-1] The input string.
+ * @param SA [0..n-1] The output array of suffixes.
+ * @param n The length of the given string.
+ * @param openMP enables OpenMP optimization.
+ * @return 0 if no error occurred, -1 or -2 otherwise.
+ */
+int
+divsufsort(const unsigned char *T, int *SA, int n, int openMP);
+
+/**
+ * Constructs the burrows-wheeler transformed string of a given string.
+ * @param T [0..n-1] The input string.
+ * @param U [0..n-1] The output string. (can be T)
+ * @param A [0..n-1] The temporary array. (can be NULL)
+ * @param n The length of the given string.
+ * @param num_indexes The length of secondary indexes array. (can be NULL)
+ * @param indexes The secondary indexes array. (can be NULL)
+ * @param openMP enables OpenMP optimization.
+ * @return The primary index if no error occurred, -1 or -2 otherwise.
+ */
+int
+divbwt(const unsigned char *T, unsigned char *U, int *A, int n, unsigned char * num_indexes, int * indexes, int openMP);
+
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif /* __cplusplus */
+
+#endif /* _DIVSUFSORT_H */
diff --git a/vendor/github.com/DataDog/zstd/entropy_common.c b/vendor/github.com/DataDog/zstd/entropy_common.c
new file mode 100644
index 0000000..b37a082
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/entropy_common.c
@@ -0,0 +1,221 @@
+/*
+   Common functions of New Generation Entropy library
+   Copyright (C) 2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+*************************************************************************** */
+
+/* *************************************
+*  Dependencies
+***************************************/
+#include "mem.h"
+#include "error_private.h"       /* ERR_*, ERROR */
+#define FSE_STATIC_LINKING_ONLY  /* FSE_MIN_TABLELOG */
+#include "fse.h"
+#define HUF_STATIC_LINKING_ONLY  /* HUF_TABLELOG_ABSOLUTEMAX */
+#include "huf.h"
+
+
+/*===   Version   ===*/
+unsigned FSE_versionNumber(void) { return FSE_VERSION_NUMBER; }
+
+
+/*===   Error Management   ===*/
+unsigned FSE_isError(size_t code) { return ERR_isError(code); }
+const char* FSE_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+unsigned HUF_isError(size_t code) { return ERR_isError(code); }
+const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/*-**************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) & (charnum<=*maxSVPtr)) {
+        if (previous0) {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF) {
+                n0 += 24;
+                if (ip < iend-5) {
+                    ip += 2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                } else {
+                    bitStream >>= 16;
+                    bitCount   += 16;
+            }   }
+            while ((bitStream & 3) == 3) {
+                n0 += 3;
+                bitStream >>= 2;
+                bitCount += 2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            } else {
+                bitStream >>= 2;
+        }   }
+        {   int const max = (2*threshold-1) - remaining;
+            int count;
+
+            if ((bitStream & (threshold-1)) < (U32)max) {
+                count = bitStream & (threshold-1);
+                bitCount += nbBits-1;
+            } else {
+                count = bitStream & (2*threshold-1);
+                if (count >= threshold) count -= max;
+                bitCount += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= count < 0 ? -count : count;   /* -1 means +1 */
+            normalizedCounter[charnum++] = (short)count;
+            previous0 = !count;
+            while (remaining < threshold) {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+            } else {
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+    }   }   /* while ((remaining>1) & (charnum<=*maxSVPtr)) */
+    if (remaining != 1) return ERROR(corruption_detected);
+    if (bitCount > 32) return ERROR(corruption_detected);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    return ip-istart;
+}
+
+
+/*! HUF_readStats() :
+    Read compact Huffman tree, saved by HUF_writeCTable().
+    `huffWeight` is destination buffer.
+    `rankStats` is assumed to be a table of at least HUF_TABLELOG_MAX U32.
+    @return : size read from `src` , or an error Code .
+    Note : Needed by HUF_readCTable() and HUF_readDTableX?() .
+*/
+size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    /* memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128) {  /* special header */
+        oSize = iSize - 127;
+        iSize = ((oSize+1)/2);
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        if (oSize >= hwSize) return ERROR(corruption_detected);
+        ip += 1;
+        {   U32 n;
+            for (n=0; n<oSize; n+=2) {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+    }   }   }
+    else  {   /* header compressed with FSE (normal case) */
+        FSE_DTable fseWorkspace[FSE_DTABLE_SIZE_U32(6)];  /* 6 is max possible tableLog for HUF header (maybe even 5, to be tested) */
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSE_decompress_wksp(huffWeight, hwSize-1, ip+1, iSize, fseWorkspace, 6);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSE_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
+    weightTotal = 0;
+    {   U32 n; for (n=0; n<oSize; n++) {
+            if (huffWeight[n] >= HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+            rankStats[huffWeight[n]]++;
+            weightTotal += (1 << huffWeight[n]) >> 1;
+    }   }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    {   U32 const tableLog = BIT_highbit32(weightTotal) + 1;
+        if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+        *tableLogPtr = tableLog;
+        /* determine last weight */
+        {   U32 const total = 1 << tableLog;
+            U32 const rest = total - weightTotal;
+            U32 const verif = 1 << BIT_highbit32(rest);
+            U32 const lastWeight = BIT_highbit32(rest) + 1;
+            if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+            huffWeight[oSize] = (BYTE)lastWeight;
+            rankStats[lastWeight]++;
+    }   }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    return iSize+1;
+}
diff --git a/vendor/github.com/DataDog/zstd/error_private.c b/vendor/github.com/DataDog/zstd/error_private.c
new file mode 100644
index 0000000..d004ee6
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/error_private.c
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* The purpose of this file is to have a single list of error strings embedded in binary */
+
+#include "error_private.h"
+
+const char* ERR_getErrorString(ERR_enum code)
+{
+    static const char* const notErrorCode = "Unspecified error code";
+    switch( code )
+    {
+    case PREFIX(no_error): return "No error detected";
+    case PREFIX(GENERIC):  return "Error (generic)";
+    case PREFIX(prefix_unknown): return "Unknown frame descriptor";
+    case PREFIX(version_unsupported): return "Version not supported";
+    case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
+    case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
+    case PREFIX(corruption_detected): return "Corrupted block detected";
+    case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
+    case PREFIX(parameter_unsupported): return "Unsupported parameter";
+    case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
+    case PREFIX(init_missing): return "Context should be init first";
+    case PREFIX(memory_allocation): return "Allocation error : not enough memory";
+    case PREFIX(workSpace_tooSmall): return "workSpace buffer is not large enough";
+    case PREFIX(stage_wrong): return "Operation not authorized at current processing stage";
+    case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
+    case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
+    case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
+    case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
+    case PREFIX(dictionary_wrong): return "Dictionary mismatch";
+    case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
+    case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
+    case PREFIX(srcSize_wrong): return "Src size is incorrect";
+        /* following error codes are not stable and may be removed or changed in a future version */
+    case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
+    case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
+    case PREFIX(maxCode):
+    default: return notErrorCode;
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/error_private.h b/vendor/github.com/DataDog/zstd/error_private.h
new file mode 100644
index 0000000..0d2fa7e
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/error_private.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* Note : this module is expected to remain private, do not expose it */
+
+#ifndef ERROR_H_MODULE
+#define ERROR_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************************
+*  Dependencies
+******************************************/
+#include <stddef.h>        /* size_t */
+#include "zstd_errors.h"  /* enum list */
+
+
+/* ****************************************
+*  Compiler-specific
+******************************************/
+#if defined(__GNUC__)
+#  define ERR_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define ERR_STATIC static inline
+#elif defined(_MSC_VER)
+#  define ERR_STATIC static __inline
+#else
+#  define ERR_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/*-****************************************
+*  Customization (error_public.h)
+******************************************/
+typedef ZSTD_ErrorCode ERR_enum;
+#define PREFIX(name) ZSTD_error_##name
+
+
+/*-****************************************
+*  Error codes handling
+******************************************/
+#undef ERROR   /* reported already defined on VS 2015 (Rich Geldreich) */
+#define ERROR(name) ZSTD_ERROR(name)
+#define ZSTD_ERROR(name) ((size_t)-PREFIX(name))
+
+ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }
+
+ERR_STATIC ERR_enum ERR_getErrorCode(size_t code) { if (!ERR_isError(code)) return (ERR_enum)0; return (ERR_enum) (0-code); }
+
+
+/*-****************************************
+*  Error Strings
+******************************************/
+
+const char* ERR_getErrorString(ERR_enum code);   /* error_private.c */
+
+ERR_STATIC const char* ERR_getErrorName(size_t code)
+{
+    return ERR_getErrorString(ERR_getErrorCode(code));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ERROR_H_MODULE */
diff --git a/vendor/github.com/DataDog/zstd/errors.go b/vendor/github.com/DataDog/zstd/errors.go
new file mode 100644
index 0000000..38db0d5
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/errors.go
@@ -0,0 +1,35 @@
+package zstd
+
+/*
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+*/
+import "C"
+
+// ErrorCode is an error returned by the zstd library.
+type ErrorCode int
+
+// Error returns the error string given by zstd
+func (e ErrorCode) Error() string {
+	return C.GoString(C.ZSTD_getErrorName(C.size_t(e)))
+}
+
+func cIsError(code int) bool {
+	return int(C.ZSTD_isError(C.size_t(code))) != 0
+}
+
+// getError returns an error for the return code, or nil if it's not an error
+func getError(code int) error {
+	if code < 0 && cIsError(code) {
+		return ErrorCode(code)
+	}
+	return nil
+}
+
+// IsDstSizeTooSmallError returns whether the error correspond to zstd standard sDstSizeTooSmall error
+func IsDstSizeTooSmallError(e error) bool {
+	if e != nil && e.Error() == "Destination buffer is too small" {
+		return true
+	}
+	return false
+}
diff --git a/vendor/github.com/DataDog/zstd/fse.h b/vendor/github.com/DataDog/zstd/fse.h
new file mode 100644
index 0000000..6a1d272
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/fse.h
@@ -0,0 +1,704 @@
+/* ******************************************************************
+   FSE : Finite State Entropy codec
+   Public Prototypes declaration
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef FSE_H
+#define FSE_H
+
+
+/*-*****************************************
+*  Dependencies
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+
+
+/*-*****************************************
+*  FSE_PUBLIC_API : control library symbols visibility
+******************************************/
+#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
+#  define FSE_PUBLIC_API __attribute__ ((visibility ("default")))
+#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
+#  define FSE_PUBLIC_API __declspec(dllexport)
+#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
+#  define FSE_PUBLIC_API __declspec(dllimport) /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define FSE_PUBLIC_API
+#endif
+
+/*------   Version   ------*/
+#define FSE_VERSION_MAJOR    0
+#define FSE_VERSION_MINOR    9
+#define FSE_VERSION_RELEASE  0
+
+#define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE
+#define FSE_QUOTE(str) #str
+#define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str)
+#define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION)
+
+#define FSE_VERSION_NUMBER  (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE)
+FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /**< library version number; to be used when checking dll version */
+
+/*-****************************************
+*  FSE simple functions
+******************************************/
+/*! FSE_compress() :
+    Compress content of buffer 'src', of size 'srcSize', into destination buffer 'dst'.
+    'dst' buffer must be already allocated. Compression runs faster is dstCapacity >= FSE_compressBound(srcSize).
+    @return : size of compressed data (<= dstCapacity).
+    Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
+                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression instead.
+                     if FSE_isError(return), compression failed (more details using FSE_getErrorName())
+*/
+FSE_PUBLIC_API size_t FSE_compress(void* dst, size_t dstCapacity,
+                             const void* src, size_t srcSize);
+
+/*! FSE_decompress():
+    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstCapacity'.
+    @return : size of regenerated data (<= maxDstSize),
+              or an error code, which can be tested using FSE_isError() .
+
+    ** Important ** : FSE_decompress() does not decompress non-compressible nor RLE data !!!
+    Why ? : making this distinction requires a header.
+    Header management is intentionally delegated to the user layer, which can better manage special cases.
+*/
+FSE_PUBLIC_API size_t FSE_decompress(void* dst,  size_t dstCapacity,
+                               const void* cSrc, size_t cSrcSize);
+
+
+/*-*****************************************
+*  Tool functions
+******************************************/
+FSE_PUBLIC_API size_t FSE_compressBound(size_t size);       /* maximum compressed size */
+
+/* Error Management */
+FSE_PUBLIC_API unsigned    FSE_isError(size_t code);        /* tells if a return value is an error code */
+FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
+
+
+/*-*****************************************
+*  FSE advanced functions
+******************************************/
+/*! FSE_compress2() :
+    Same as FSE_compress(), but allows the selection of 'maxSymbolValue' and 'tableLog'
+    Both parameters can be defined as '0' to mean : use default value
+    @return : size of compressed data
+    Special values : if return == 0, srcData is not compressible => Nothing is stored within cSrc !!!
+                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression.
+                     if FSE_isError(return), it's an error code.
+*/
+FSE_PUBLIC_API size_t FSE_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
+
+
+/*-*****************************************
+*  FSE detailed API
+******************************************/
+/*!
+FSE_compress() does the following:
+1. count symbol occurrence from source[] into table count[]
+2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)
+3. save normalized counters to memory buffer using writeNCount()
+4. build encoding table 'CTable' from normalized counters
+5. encode the data stream using encoding table 'CTable'
+
+FSE_decompress() does the following:
+1. read normalized counters with readNCount()
+2. build decoding table 'DTable' from normalized counters
+3. decode the data stream using decoding table 'DTable'
+
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and provide normalized distribution using external method.
+*/
+
+/* *** COMPRESSION *** */
+
+/*! FSE_count():
+    Provides the precise count of each byte within a table 'count'.
+    'count' is a table of unsigned int, of minimum size (*maxSymbolValuePtr+1).
+    *maxSymbolValuePtr will be updated if detected smaller than initial value.
+    @return : the count of the most frequent symbol (which is not identified).
+              if return == srcSize, there is only one symbol.
+              Can also return an error code, which can be tested with FSE_isError(). */
+FSE_PUBLIC_API size_t FSE_count(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+
+/*! FSE_optimalTableLog():
+    dynamically downsize 'tableLog' when conditions are met.
+    It saves CPU time, by using smaller tables, while preserving or even improving compression ratio.
+    @return : recommended tableLog (necessarily <= 'maxTableLog') */
+FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
+
+/*! FSE_normalizeCount():
+    normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
+    'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
+    @return : tableLog,
+              or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog, const unsigned* count, size_t srcSize, unsigned maxSymbolValue);
+
+/*! FSE_NCountWriteBound():
+    Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
+    Typically useful for allocation purpose. */
+FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSE_writeNCount():
+    Compactly save 'normalizedCounter' into 'buffer'.
+    @return : size of the compressed table,
+              or an errorCode, which can be tested using FSE_isError(). */
+FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+
+/*! Constructor and Destructor of FSE_CTable.
+    Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
+typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */
+FSE_PUBLIC_API FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog);
+FSE_PUBLIC_API void        FSE_freeCTable (FSE_CTable* ct);
+
+/*! FSE_buildCTable():
+    Builds `ct`, which must be already allocated, using FSE_createCTable().
+    @return : 0, or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSE_compress_usingCTable():
+    Compress `src` using `ct` into `dst` which must be already allocated.
+    @return : size of compressed data (<= `dstCapacity`),
+              or 0 if compressed data could not fit into `dst`,
+              or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct);
+
+/*!
+Tutorial :
+----------
+The first step is to count all symbols. FSE_count() does this job very fast.
+Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells.
+'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0]
+maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value)
+FSE_count() will return the number of occurrence of the most frequent symbol.
+This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
+
+The next step is to normalize the frequencies.
+FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'.
+It also guarantees a minimum of 1 to any Symbol with frequency >= 1.
+You can use 'tableLog'==0 to mean "use default tableLog value".
+If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(),
+which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default").
+
+The result of FSE_normalizeCount() will be saved into a table,
+called 'normalizedCounter', which is a table of signed short.
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells.
+The return value is tableLog if everything proceeded as expected.
+It is 0 if there is a single symbol within distribution.
+If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()).
+
+'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount().
+'buffer' must be already allocated.
+For guaranteed success, buffer size must be at least FSE_headerBound().
+The result of the function is the number of bytes written into 'buffer'.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small).
+
+'normalizedCounter' can then be used to create the compression table 'CTable'.
+The space required by 'CTable' must be already allocated, using FSE_createCTable().
+You can then use FSE_buildCTable() to fill 'CTable'.
+If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()).
+
+'CTable' can then be used to compress 'src', with FSE_compress_usingCTable().
+Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize'
+The function returns the size of compressed data (without header), necessarily <= `dstCapacity`.
+If it returns '0', compressed data could not fit into 'dst'.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
+*/
+
+
+/* *** DECOMPRESSION *** */
+
+/*! FSE_readNCount():
+    Read compactly saved 'normalizedCounter' from 'rBuffer'.
+    @return : size read from 'rBuffer',
+              or an errorCode, which can be tested using FSE_isError().
+              maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, const void* rBuffer, size_t rBuffSize);
+
+/*! Constructor and Destructor of FSE_DTable.
+    Note that its size depends on 'tableLog' */
+typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+FSE_PUBLIC_API FSE_DTable* FSE_createDTable(unsigned tableLog);
+FSE_PUBLIC_API void        FSE_freeDTable(FSE_DTable* dt);
+
+/*! FSE_buildDTable():
+    Builds 'dt', which must be already allocated, using FSE_createDTable().
+    return : 0, or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_buildDTable (FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSE_decompress_usingDTable():
+    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
+    into `dst` which must be already allocated.
+    @return : size of regenerated data (necessarily <= `dstCapacity`),
+              or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
+
+/*!
+Tutorial :
+----------
+(Note : these functions only decompress FSE-compressed blocks.
+ If block is uncompressed, use memcpy() instead
+ If block is a single repeated byte, use memset() instead )
+
+The first step is to obtain the normalized frequencies of symbols.
+This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
+In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
+or size the table to handle worst case situations (typically 256).
+FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
+The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.
+Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.
+This is performed by the function FSE_buildDTable().
+The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+`FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable().
+`cSrcSize` must be strictly correct, otherwise decompression will fail.
+FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
+If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)
+*/
+
+#endif  /* FSE_H */
+
+#if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
+#define FSE_H_FSE_STATIC_LINKING_ONLY
+
+/* *** Dependency *** */
+#include "bitstream.h"
+
+
+/* *****************************************
+*  Static allocation
+*******************************************/
+/* FSE buffer bounds */
+#define FSE_NCOUNTBOUND 512
+#define FSE_BLOCKBOUND(size) (size + (size>>7))
+#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+/* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
+#define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue)   (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
+#define FSE_DTABLE_SIZE(maxTableLog)                   (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable))
+
+
+/* *****************************************
+*  FSE advanced API
+*******************************************/
+/* FSE_count_wksp() :
+ * Same as FSE_count(), but using an externally provided scratch buffer.
+ * `workSpace` size must be table of >= `1024` unsigned
+ */
+size_t FSE_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
+                 const void* source, size_t sourceSize, unsigned* workSpace);
+
+/** FSE_countFast() :
+ *  same as FSE_count(), but blindly trusts that all byte values within src are <= *maxSymbolValuePtr
+ */
+size_t FSE_countFast(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+
+/* FSE_countFast_wksp() :
+ * Same as FSE_countFast(), but using an externally provided scratch buffer.
+ * `workSpace` must be a table of minimum `1024` unsigned
+ */
+size_t FSE_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* workSpace);
+
+/*! FSE_count_simple() :
+ * Same as FSE_countFast(), but does not use any additional memory (not even on stack).
+ * This function is unsafe, and will segfault if any value within `src` is `> *maxSymbolValuePtr` (presuming it's also the size of `count`).
+*/
+size_t FSE_count_simple(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+
+
+
+unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
+/**< same as FSE_optimalTableLog(), which used `minus==2` */
+
+/* FSE_compress_wksp() :
+ * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
+ * FSE_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
+ */
+#define FSE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue)   ( FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) + ((maxTableLog > 12) ? (1 << (maxTableLog - 2)) : 1024) )
+size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+
+size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits);
+/**< build a fake FSE_CTable, designed for a flat distribution, where each symbol uses nbBits */
+
+size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
+/**< build a fake FSE_CTable, designed to compress always the same symbolValue */
+
+/* FSE_buildCTable_wksp() :
+ * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
+ * `wkspSize` must be >= `(1<<tableLog)`.
+ */
+size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+
+size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
+/**< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
+
+size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
+/**< build a fake FSE_DTable, designed to always generate the same symbolValue */
+
+size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog);
+/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DTABLE_SIZE_U32(maxLog)` */
+
+typedef enum {
+   FSE_repeat_none,  /**< Cannot use the previous table */
+   FSE_repeat_check, /**< Can use the previous table but it must be checked */
+   FSE_repeat_valid  /**< Can use the previous table and it is asumed to be valid */
+ } FSE_repeat;
+
+/* *****************************************
+*  FSE symbol compression API
+*******************************************/
+/*!
+   This API consists of small unitary functions, which highly benefit from being inlined.
+   Hence their body are included in next section.
+*/
+typedef struct {
+    ptrdiff_t   value;
+    const void* stateTable;
+    const void* symbolTT;
+    unsigned    stateLog;
+} FSE_CState_t;
+
+static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct);
+
+static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol);
+
+static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr);
+
+/**<
+These functions are inner components of FSE_compress_usingCTable().
+They allow the creation of custom streams, mixing multiple tables and bit sources.
+
+A key property to keep in mind is that encoding and decoding are done **in reverse direction**.
+So the first symbol you will encode is the last you will decode, like a LIFO stack.
+
+You will need a few variables to track your CStream. They are :
+
+FSE_CTable    ct;         // Provided by FSE_buildCTable()
+BIT_CStream_t bitStream;  // bitStream tracking structure
+FSE_CState_t  state;      // State tracking structure (can have several)
+
+
+The first thing to do is to init bitStream and state.
+    size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize);
+    FSE_initCState(&state, ct);
+
+Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError();
+You can then encode your input data, byte after byte.
+FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time.
+Remember decoding will be done in reverse direction.
+    FSE_encodeByte(&bitStream, &state, symbol);
+
+At any time, you can also add any bit sequence.
+Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders
+    BIT_addBits(&bitStream, bitField, nbBits);
+
+The above methods don't commit data to memory, they just store it into local register, for speed.
+Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
+Writing data to memory is a manual operation, performed by the flushBits function.
+    BIT_flushBits(&bitStream);
+
+Your last FSE encoding operation shall be to flush your last state value(s).
+    FSE_flushState(&bitStream, &state);
+
+Finally, you must close the bitStream.
+The function returns the size of CStream in bytes.
+If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible)
+If there is an error, it returns an errorCode (which can be tested using FSE_isError()).
+    size_t size = BIT_closeCStream(&bitStream);
+*/
+
+
+/* *****************************************
+*  FSE symbol decompression API
+*******************************************/
+typedef struct {
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+
+static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
+
+static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
+
+/**<
+Let's now decompose FSE_decompress_usingDTable() into its unitary components.
+You will decode FSE-encoded symbols from the bitStream,
+and also any other bitFields you put in, **in reverse order**.
+
+You will need a few variables to track your bitStream. They are :
+
+BIT_DStream_t DStream;    // Stream context
+FSE_DState_t  DState;     // State context. Multiple ones are possible
+FSE_DTable*   DTablePtr;  // Decoding table, provided by FSE_buildDTable()
+
+The first thing to do is to init the bitStream.
+    errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize);
+
+You should then retrieve your initial state(s)
+(in reverse flushing order if you have several ones) :
+    errorCode = FSE_initDState(&DState, &DStream, DTablePtr);
+
+You can then decode your data, symbol after symbol.
+For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'.
+Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out).
+    unsigned char symbol = FSE_decodeSymbol(&DState, &DStream);
+
+You can retrieve any bitfield you eventually stored into the bitStream (in reverse order)
+Note : maximum allowed nbBits is 25, for 32-bits compatibility
+    size_t bitField = BIT_readBits(&DStream, nbBits);
+
+All above operations only read from local register (which size depends on size_t).
+Refueling the register from memory is manually performed by the reload method.
+    endSignal = FSE_reloadDStream(&DStream);
+
+BIT_reloadDStream() result tells if there is still some more data to read from DStream.
+BIT_DStream_unfinished : there is still some data left into the DStream.
+BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled.
+BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed.
+BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted.
+
+When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop,
+to properly detect the exact end of stream.
+After each decoded symbol, check if DStream is fully consumed using this simple test :
+    BIT_reloadDStream(&DStream) >= BIT_DStream_completed
+
+When it's done, verify decompression is fully completed, by checking both DStream and the relevant states.
+Checking if DStream has reached its end is performed by :
+    BIT_endOfDStream(&DStream);
+Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible.
+    FSE_endOfDState(&DState);
+*/
+
+
+/* *****************************************
+*  FSE unsafe API
+*******************************************/
+static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* *****************************************
+*  Implementation of inlined functions
+*******************************************/
+typedef struct {
+    int deltaFindState;
+    U32 deltaNbBits;
+} FSE_symbolCompressionTransform; /* total 8 bytes */
+
+MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct)
+{
+    const void* ptr = ct;
+    const U16* u16ptr = (const U16*) ptr;
+    const U32 tableLog = MEM_read16(ptr);
+    statePtr->value = (ptrdiff_t)1<<tableLog;
+    statePtr->stateTable = u16ptr+2;
+    statePtr->symbolTT = ((const U32*)ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1));
+    statePtr->stateLog = tableLog;
+}
+
+
+/*! FSE_initCState2() :
+*   Same as FSE_initCState(), but the first symbol to include (which will be the last to be read)
+*   uses the smallest state value possible, saving the cost of this symbol */
+MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol)
+{
+    FSE_initCState(statePtr, ct);
+    {   const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
+        const U16* stateTable = (const U16*)(statePtr->stateTable);
+        U32 nbBitsOut  = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16);
+        statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits;
+        statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
+    }
+}
+
+MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, U32 symbol)
+{
+    FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
+    const U16* const stateTable = (const U16*)(statePtr->stateTable);
+    U32 const nbBitsOut  = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
+    BIT_addBits(bitC, statePtr->value, nbBitsOut);
+    statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
+}
+
+MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
+{
+    BIT_addBits(bitC, statePtr->value, statePtr->stateLog);
+    BIT_flushBits(bitC);
+}
+
+
+/* ======    Decompression    ====== */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSE_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSE_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;
+    DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
+    BIT_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr)
+{
+    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    return DInfo.symbol;
+}
+
+MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    size_t const lowBits = BIT_readBits(bitD, nbBits);
+    DStatePtr->state = DInfo.newState + lowBits;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BIT_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+/*! FSE_decodeSymbolFast() :
+    unsafe, only works if no symbol has a probability > 50% */
+MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BIT_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/* **************************************************************
+*  Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#ifndef FSE_MAX_MEMORY_USAGE
+#  define FSE_MAX_MEMORY_USAGE 14
+#endif
+#ifndef FSE_DEFAULT_MEMORY_USAGE
+#  define FSE_DEFAULT_MEMORY_USAGE 13
+#endif
+
+/*!FSE_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#ifndef FSE_MAX_SYMBOL_VALUE
+#  define FSE_MAX_SYMBOL_VALUE 255
+#endif
+
+/* **************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+#define FSE_DECODE_TYPE FSE_decode_t
+
+
+#endif   /* !FSE_COMMONDEFS_ONLY */
+
+
+/* ***************************************************************
+*  Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+#  error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+#define FSE_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
+
+
+#endif /* FSE_STATIC_LINKING_ONLY */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/vendor/github.com/DataDog/zstd/fse_compress.c b/vendor/github.com/DataDog/zstd/fse_compress.c
new file mode 100644
index 0000000..cb8f1fa
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/fse_compress.c
@@ -0,0 +1,849 @@
+/* ******************************************************************
+   FSE : Finite State Entropy encoder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+#include "bitstream.h"
+#include "compiler.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#include "error_private.h"
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_isError ERR_isError
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+
+/* Function templates */
+
+/* FSE_buildCTable_wksp() :
+ * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
+ * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)`
+ * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements
+ */
+size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+    U32 const tableSize = 1 << tableLog;
+    U32 const tableMask = tableSize - 1;
+    void* const ptr = ct;
+    U16* const tableU16 = ( (U16*) ptr) + 2;
+    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
+    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
+    U32 const step = FSE_TABLESTEP(tableSize);
+    U32 cumul[FSE_MAX_SYMBOL_VALUE+2];
+
+    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)workSpace;
+    U32 highThreshold = tableSize-1;
+
+    /* CTable header */
+    if (((size_t)1 << tableLog) * sizeof(FSE_FUNCTION_TYPE) > wkspSize) return ERROR(tableLog_tooLarge);
+    tableU16[-2] = (U16) tableLog;
+    tableU16[-1] = (U16) maxSymbolValue;
+
+    /* For explanations on how to distribute symbol values over the table :
+    *  http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
+
+    /* symbol start positions */
+    {   U32 u;
+        cumul[0] = 0;
+        for (u=1; u<=maxSymbolValue+1; u++) {
+            if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
+                cumul[u] = cumul[u-1] + 1;
+                tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
+            } else {
+                cumul[u] = cumul[u-1] + normalizedCounter[u-1];
+        }   }
+        cumul[maxSymbolValue+1] = tableSize+1;
+    }
+
+    /* Spread symbols */
+    {   U32 position = 0;
+        U32 symbol;
+        for (symbol=0; symbol<=maxSymbolValue; symbol++) {
+            int nbOccurences;
+            for (nbOccurences=0; nbOccurences<normalizedCounter[symbol]; nbOccurences++) {
+                tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* Low proba area */
+        }   }
+
+        if (position!=0) return ERROR(GENERIC);   /* Must have gone through all positions */
+    }
+
+    /* Build table */
+    {   U32 u; for (u=0; u<tableSize; u++) {
+        FSE_FUNCTION_TYPE s = tableSymbol[u];   /* note : static analyzer may not understand tableSymbol is properly initialized */
+        tableU16[cumul[s]++] = (U16) (tableSize+u);   /* TableU16 : sorted by symbol order; gives next state value */
+    }   }
+
+    /* Build Symbol Transformation Table */
+    {   unsigned total = 0;
+        unsigned s;
+        for (s=0; s<=maxSymbolValue; s++) {
+            switch (normalizedCounter[s])
+            {
+            case  0: break;
+
+            case -1:
+            case  1:
+                symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
+                symbolTT[s].deltaFindState = total - 1;
+                total ++;
+                break;
+            default :
+                {
+                    U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
+                    U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
+                    symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
+                    symbolTT[s].deltaFindState = total - normalizedCounter[s];
+                    total +=  normalizedCounter[s];
+    }   }   }   }
+
+    return 0;
+}
+
+
+size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE];   /* memset() is not necessary, even if static analyzer complain about it */
+    return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol));
+}
+
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/*-**************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
+{
+    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3;
+    return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
+}
+
+static size_t FSE_writeNCount_generic (void* header, size_t headerBufferSize,
+                                       const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
+                                       unsigned writeIsSafe)
+{
+    BYTE* const ostart = (BYTE*) header;
+    BYTE* out = ostart;
+    BYTE* const oend = ostart + headerBufferSize;
+    int nbBits;
+    const int tableSize = 1 << tableLog;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    bitStream = 0;
+    bitCount  = 0;
+    /* Table Size */
+    bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount;
+    bitCount  += 4;
+
+    /* Init */
+    remaining = tableSize+1;   /* +1 for extra accuracy */
+    threshold = tableSize;
+    nbBits = tableLog+1;
+
+    while (remaining>1) {  /* stops at 1 */
+        if (previous0) {
+            unsigned start = charnum;
+            while (!normalizedCounter[charnum]) charnum++;
+            while (charnum >= start+24) {
+                start+=24;
+                bitStream += 0xFFFFU << bitCount;
+                if ((!writeIsSafe) && (out > oend-2)) return ERROR(dstSize_tooSmall);   /* Buffer overflow */
+                out[0] = (BYTE) bitStream;
+                out[1] = (BYTE)(bitStream>>8);
+                out+=2;
+                bitStream>>=16;
+            }
+            while (charnum >= start+3) {
+                start+=3;
+                bitStream += 3 << bitCount;
+                bitCount += 2;
+            }
+            bitStream += (charnum-start) << bitCount;
+            bitCount += 2;
+            if (bitCount>16) {
+                if ((!writeIsSafe) && (out > oend - 2)) return ERROR(dstSize_tooSmall);   /* Buffer overflow */
+                out[0] = (BYTE)bitStream;
+                out[1] = (BYTE)(bitStream>>8);
+                out += 2;
+                bitStream >>= 16;
+                bitCount -= 16;
+        }   }
+        {   int count = normalizedCounter[charnum++];
+            int const max = (2*threshold-1)-remaining;
+            remaining -= count < 0 ? -count : count;
+            count++;   /* +1 for extra accuracy */
+            if (count>=threshold) count += max;   /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */
+            bitStream += count << bitCount;
+            bitCount  += nbBits;
+            bitCount  -= (count<max);
+            previous0  = (count==1);
+            if (remaining<1) return ERROR(GENERIC);
+            while (remaining<threshold) { nbBits--; threshold>>=1; }
+        }
+        if (bitCount>16) {
+            if ((!writeIsSafe) && (out > oend - 2)) return ERROR(dstSize_tooSmall);   /* Buffer overflow */
+            out[0] = (BYTE)bitStream;
+            out[1] = (BYTE)(bitStream>>8);
+            out += 2;
+            bitStream >>= 16;
+            bitCount -= 16;
+    }   }
+
+    /* flush remaining bitStream */
+    if ((!writeIsSafe) && (out > oend - 2)) return ERROR(dstSize_tooSmall);   /* Buffer overflow */
+    out[0] = (BYTE)bitStream;
+    out[1] = (BYTE)(bitStream>>8);
+    out+= (bitCount+7) /8;
+
+    if (charnum > maxSymbolValue + 1) return ERROR(GENERIC);
+
+    return (out-ostart);
+}
+
+
+size_t FSE_writeNCount (void* buffer, size_t bufferSize, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported */
+    if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported */
+
+    if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog))
+        return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0);
+
+    return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1);
+}
+
+
+
+/*-**************************************************************
+*  Counting histogram
+****************************************************************/
+/*! FSE_count_simple
+    This function counts byte values within `src`, and store the histogram into table `count`.
+    It doesn't use any additional memory.
+    But this function is unsafe : it doesn't check that all values within `src` can fit into `count`.
+    For this reason, prefer using a table `count` with 256 elements.
+    @return : count of most numerous element.
+*/
+size_t FSE_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
+                        const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* const end = ip + srcSize;
+    unsigned maxSymbolValue = *maxSymbolValuePtr;
+    unsigned max=0;
+
+    memset(count, 0, (maxSymbolValue+1)*sizeof(*count));
+    if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; }
+
+    while (ip<end) {
+        assert(*ip <= maxSymbolValue);
+        count[*ip++]++;
+    }
+
+    while (!count[maxSymbolValue]) maxSymbolValue--;
+    *maxSymbolValuePtr = maxSymbolValue;
+
+    { U32 s; for (s=0; s<=maxSymbolValue; s++) if (count[s] > max) max = count[s]; }
+
+    return (size_t)max;
+}
+
+
+/* FSE_count_parallel_wksp() :
+ * Same as FSE_count_parallel(), but using an externally provided scratch buffer.
+ * `workSpace` size must be a minimum of `1024 * sizeof(unsigned)`.
+ * @return : largest histogram frequency, or an error code (notably when histogram would be larger than *maxSymbolValuePtr). */
+static size_t FSE_count_parallel_wksp(
+                                unsigned* count, unsigned* maxSymbolValuePtr,
+                                const void* source, size_t sourceSize,
+                                unsigned checkMax, unsigned* const workSpace)
+{
+    const BYTE* ip = (const BYTE*)source;
+    const BYTE* const iend = ip+sourceSize;
+    unsigned maxSymbolValue = *maxSymbolValuePtr;
+    unsigned max=0;
+    U32* const Counting1 = workSpace;
+    U32* const Counting2 = Counting1 + 256;
+    U32* const Counting3 = Counting2 + 256;
+    U32* const Counting4 = Counting3 + 256;
+
+    memset(workSpace, 0, 4*256*sizeof(unsigned));
+
+    /* safety checks */
+    if (!sourceSize) {
+        memset(count, 0, maxSymbolValue + 1);
+        *maxSymbolValuePtr = 0;
+        return 0;
+    }
+    if (!maxSymbolValue) maxSymbolValue = 255;            /* 0 == default */
+
+    /* by stripes of 16 bytes */
+    {   U32 cached = MEM_read32(ip); ip += 4;
+        while (ip < iend-15) {
+            U32 c = cached; cached = MEM_read32(ip); ip += 4;
+            Counting1[(BYTE) c     ]++;
+            Counting2[(BYTE)(c>>8) ]++;
+            Counting3[(BYTE)(c>>16)]++;
+            Counting4[       c>>24 ]++;
+            c = cached; cached = MEM_read32(ip); ip += 4;
+            Counting1[(BYTE) c     ]++;
+            Counting2[(BYTE)(c>>8) ]++;
+            Counting3[(BYTE)(c>>16)]++;
+            Counting4[       c>>24 ]++;
+            c = cached; cached = MEM_read32(ip); ip += 4;
+            Counting1[(BYTE) c     ]++;
+            Counting2[(BYTE)(c>>8) ]++;
+            Counting3[(BYTE)(c>>16)]++;
+            Counting4[       c>>24 ]++;
+            c = cached; cached = MEM_read32(ip); ip += 4;
+            Counting1[(BYTE) c     ]++;
+            Counting2[(BYTE)(c>>8) ]++;
+            Counting3[(BYTE)(c>>16)]++;
+            Counting4[       c>>24 ]++;
+        }
+        ip-=4;
+    }
+
+    /* finish last symbols */
+    while (ip<iend) Counting1[*ip++]++;
+
+    if (checkMax) {   /* verify stats will fit into destination table */
+        U32 s; for (s=255; s>maxSymbolValue; s--) {
+            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
+            if (Counting1[s]) return ERROR(maxSymbolValue_tooSmall);
+    }   }
+
+    {   U32 s;
+        if (maxSymbolValue > 255) maxSymbolValue = 255;
+        for (s=0; s<=maxSymbolValue; s++) {
+            count[s] = Counting1[s] + Counting2[s] + Counting3[s] + Counting4[s];
+            if (count[s] > max) max = count[s];
+    }   }
+
+    while (!count[maxSymbolValue]) maxSymbolValue--;
+    *maxSymbolValuePtr = maxSymbolValue;
+    return (size_t)max;
+}
+
+/* FSE_countFast_wksp() :
+ * Same as FSE_countFast(), but using an externally provided scratch buffer.
+ * `workSpace` size must be table of >= `1024` unsigned */
+size_t FSE_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
+                          const void* source, size_t sourceSize,
+                          unsigned* workSpace)
+{
+    if (sourceSize < 1500) /* heuristic threshold */
+        return FSE_count_simple(count, maxSymbolValuePtr, source, sourceSize);
+    return FSE_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, 0, workSpace);
+}
+
+/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
+size_t FSE_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
+                     const void* source, size_t sourceSize)
+{
+    unsigned tmpCounters[1024];
+    return FSE_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters);
+}
+
+/* FSE_count_wksp() :
+ * Same as FSE_count(), but using an externally provided scratch buffer.
+ * `workSpace` size must be table of >= `1024` unsigned */
+size_t FSE_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
+                 const void* source, size_t sourceSize, unsigned* workSpace)
+{
+    if (*maxSymbolValuePtr < 255)
+        return FSE_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, 1, workSpace);
+    *maxSymbolValuePtr = 255;
+    return FSE_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace);
+}
+
+size_t FSE_count(unsigned* count, unsigned* maxSymbolValuePtr,
+                 const void* src, size_t srcSize)
+{
+    unsigned tmpCounters[1024];
+    return FSE_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters);
+}
+
+
+
+/*-**************************************************************
+*  FSE Compression Code
+****************************************************************/
+/*! FSE_sizeof_CTable() :
+    FSE_CTable is a variable size structure which contains :
+    `U16 tableLog;`
+    `U16 maxSymbolValue;`
+    `U16 nextStateNumber[1 << tableLog];`                         // This size is variable
+    `FSE_symbolCompressionTransform symbolTT[maxSymbolValue+1];`  // This size is variable
+Allocation is manual (C standard does not support variable-size structures).
+*/
+size_t FSE_sizeof_CTable (unsigned maxSymbolValue, unsigned tableLog)
+{
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    return FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
+}
+
+FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog)
+{
+    size_t size;
+    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
+    size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
+    return (FSE_CTable*)malloc(size);
+}
+
+void FSE_freeCTable (FSE_CTable* ct) { free(ct); }
+
+/* provides the minimum logSize to safely represent a distribution */
+static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
+{
+    U32 minBitsSrc = BIT_highbit32((U32)(srcSize - 1)) + 1;
+    U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
+    U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
+    assert(srcSize > 1); /* Not supported, RLE should be used instead */
+    return minBits;
+}
+
+unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
+{
+    U32 maxBitsSrc = BIT_highbit32((U32)(srcSize - 1)) - minus;
+    U32 tableLog = maxTableLog;
+    U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
+    assert(srcSize > 1); /* Not supported, RLE should be used instead */
+    if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
+    if (maxBitsSrc < tableLog) tableLog = maxBitsSrc;   /* Accuracy can be reduced */
+    if (minBits > tableLog) tableLog = minBits;   /* Need a minimum to safely represent all symbol values */
+    if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG;
+    if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG;
+    return tableLog;
+}
+
+unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
+{
+    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
+}
+
+
+/* Secondary normalization method.
+   To be used when primary method fails. */
+
+static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue)
+{
+    short const NOT_YET_ASSIGNED = -2;
+    U32 s;
+    U32 distributed = 0;
+    U32 ToDistribute;
+
+    /* Init */
+    U32 const lowThreshold = (U32)(total >> tableLog);
+    U32 lowOne = (U32)((total * 3) >> (tableLog + 1));
+
+    for (s=0; s<=maxSymbolValue; s++) {
+        if (count[s] == 0) {
+            norm[s]=0;
+            continue;
+        }
+        if (count[s] <= lowThreshold) {
+            norm[s] = -1;
+            distributed++;
+            total -= count[s];
+            continue;
+        }
+        if (count[s] <= lowOne) {
+            norm[s] = 1;
+            distributed++;
+            total -= count[s];
+            continue;
+        }
+
+        norm[s]=NOT_YET_ASSIGNED;
+    }
+    ToDistribute = (1 << tableLog) - distributed;
+
+    if ((total / ToDistribute) > lowOne) {
+        /* risk of rounding to zero */
+        lowOne = (U32)((total * 3) / (ToDistribute * 2));
+        for (s=0; s<=maxSymbolValue; s++) {
+            if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) {
+                norm[s] = 1;
+                distributed++;
+                total -= count[s];
+                continue;
+        }   }
+        ToDistribute = (1 << tableLog) - distributed;
+    }
+
+    if (distributed == maxSymbolValue+1) {
+        /* all values are pretty poor;
+           probably incompressible data (should have already been detected);
+           find max, then give all remaining points to max */
+        U32 maxV = 0, maxC = 0;
+        for (s=0; s<=maxSymbolValue; s++)
+            if (count[s] > maxC) { maxV=s; maxC=count[s]; }
+        norm[maxV] += (short)ToDistribute;
+        return 0;
+    }
+
+    if (total == 0) {
+        /* all of the symbols were low enough for the lowOne or lowThreshold */
+        for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1))
+            if (norm[s] > 0) { ToDistribute--; norm[s]++; }
+        return 0;
+    }
+
+    {   U64 const vStepLog = 62 - tableLog;
+        U64 const mid = (1ULL << (vStepLog-1)) - 1;
+        U64 const rStep = ((((U64)1<<vStepLog) * ToDistribute) + mid) / total;   /* scale on remaining */
+        U64 tmpTotal = mid;
+        for (s=0; s<=maxSymbolValue; s++) {
+            if (norm[s]==NOT_YET_ASSIGNED) {
+                U64 const end = tmpTotal + (count[s] * rStep);
+                U32 const sStart = (U32)(tmpTotal >> vStepLog);
+                U32 const sEnd = (U32)(end >> vStepLog);
+                U32 const weight = sEnd - sStart;
+                if (weight < 1)
+                    return ERROR(GENERIC);
+                norm[s] = (short)weight;
+                tmpTotal = end;
+    }   }   }
+
+    return 0;
+}
+
+
+size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
+                           const unsigned* count, size_t total,
+                           unsigned maxSymbolValue)
+{
+    /* Sanity checks */
+    if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
+    if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported size */
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported size */
+    if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */
+
+    {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
+        U64 const scale = 62 - tableLog;
+        U64 const step = ((U64)1<<62) / total;   /* <== here, one division ! */
+        U64 const vStep = 1ULL<<(scale-20);
+        int stillToDistribute = 1<<tableLog;
+        unsigned s;
+        unsigned largest=0;
+        short largestP=0;
+        U32 lowThreshold = (U32)(total >> tableLog);
+
+        for (s=0; s<=maxSymbolValue; s++) {
+            if (count[s] == total) return 0;   /* rle special case */
+            if (count[s] == 0) { normalizedCounter[s]=0; continue; }
+            if (count[s] <= lowThreshold) {
+                normalizedCounter[s] = -1;
+                stillToDistribute--;
+            } else {
+                short proba = (short)((count[s]*step) >> scale);
+                if (proba<8) {
+                    U64 restToBeat = vStep * rtbTable[proba];
+                    proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat;
+                }
+                if (proba > largestP) { largestP=proba; largest=s; }
+                normalizedCounter[s] = proba;
+                stillToDistribute -= proba;
+        }   }
+        if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
+            /* corner case, need another normalization method */
+            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue);
+            if (FSE_isError(errorCode)) return errorCode;
+        }
+        else normalizedCounter[largest] += (short)stillToDistribute;
+    }
+
+#if 0
+    {   /* Print Table (debug) */
+        U32 s;
+        U32 nTotal = 0;
+        for (s=0; s<=maxSymbolValue; s++)
+            printf("%3i: %4i \n", s, normalizedCounter[s]);
+        for (s=0; s<=maxSymbolValue; s++)
+            nTotal += abs(normalizedCounter[s]);
+        if (nTotal != (1U<<tableLog))
+            printf("Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog);
+        getchar();
+    }
+#endif
+
+    return tableLog;
+}
+
+
+/* fake FSE_CTable, for raw (uncompressed) input */
+size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits)
+{
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    void* const ptr = ct;
+    U16* const tableU16 = ( (U16*) ptr) + 2;
+    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableSize>>1);   /* assumption : tableLog >= 1 */
+    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);             /* min size */
+
+    /* header */
+    tableU16[-2] = (U16) nbBits;
+    tableU16[-1] = (U16) maxSymbolValue;
+
+    /* Build table */
+    for (s=0; s<tableSize; s++)
+        tableU16[s] = (U16)(tableSize + s);
+
+    /* Build Symbol Transformation Table */
+    {   const U32 deltaNbBits = (nbBits << 16) - (1 << nbBits);
+        for (s=0; s<=maxSymbolValue; s++) {
+            symbolTT[s].deltaNbBits = deltaNbBits;
+            symbolTT[s].deltaFindState = s-1;
+    }   }
+
+    return 0;
+}
+
+/* fake FSE_CTable, for rle input (always same symbol) */
+size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
+{
+    void* ptr = ct;
+    U16* tableU16 = ( (U16*) ptr) + 2;
+    void* FSCTptr = (U32*)ptr + 2;
+    FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr;
+
+    /* header */
+    tableU16[-2] = (U16) 0;
+    tableU16[-1] = (U16) symbolValue;
+
+    /* Build table */
+    tableU16[0] = 0;
+    tableU16[1] = 0;   /* just in case */
+
+    /* Build Symbol Transformation Table */
+    symbolTT[symbolValue].deltaNbBits = 0;
+    symbolTT[symbolValue].deltaFindState = 0;
+
+    return 0;
+}
+
+
+static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize,
+                           const void* src, size_t srcSize,
+                           const FSE_CTable* ct, const unsigned fast)
+{
+    const BYTE* const istart = (const BYTE*) src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip=iend;
+
+    BIT_CStream_t bitC;
+    FSE_CState_t CState1, CState2;
+
+    /* init */
+    if (srcSize <= 2) return 0;
+    { size_t const initError = BIT_initCStream(&bitC, dst, dstSize);
+      if (FSE_isError(initError)) return 0; /* not enough space available to write a bitstream */ }
+
+#define FSE_FLUSHBITS(s)  (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))
+
+    if (srcSize & 1) {
+        FSE_initCState2(&CState1, ct, *--ip);
+        FSE_initCState2(&CState2, ct, *--ip);
+        FSE_encodeSymbol(&bitC, &CState1, *--ip);
+        FSE_FLUSHBITS(&bitC);
+    } else {
+        FSE_initCState2(&CState2, ct, *--ip);
+        FSE_initCState2(&CState1, ct, *--ip);
+    }
+
+    /* join to mod 4 */
+    srcSize -= 2;
+    if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) {  /* test bit 2 */
+        FSE_encodeSymbol(&bitC, &CState2, *--ip);
+        FSE_encodeSymbol(&bitC, &CState1, *--ip);
+        FSE_FLUSHBITS(&bitC);
+    }
+
+    /* 2 or 4 encoding per loop */
+    while ( ip>istart ) {
+
+        FSE_encodeSymbol(&bitC, &CState2, *--ip);
+
+        if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 )   /* this test must be static */
+            FSE_FLUSHBITS(&bitC);
+
+        FSE_encodeSymbol(&bitC, &CState1, *--ip);
+
+        if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) {  /* this test must be static */
+            FSE_encodeSymbol(&bitC, &CState2, *--ip);
+            FSE_encodeSymbol(&bitC, &CState1, *--ip);
+        }
+
+        FSE_FLUSHBITS(&bitC);
+    }
+
+    FSE_flushCState(&bitC, &CState2);
+    FSE_flushCState(&bitC, &CState1);
+    return BIT_closeCStream(&bitC);
+}
+
+size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
+                           const void* src, size_t srcSize,
+                           const FSE_CTable* ct)
+{
+    unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize));
+
+    if (fast)
+        return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1);
+    else
+        return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0);
+}
+
+
+size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
+
+#define CHECK_V_F(e, f) size_t const e = f; if (ERR_isError(e)) return e
+#define CHECK_F(f)   { CHECK_V_F(_var_err__, f); }
+
+/* FSE_compress_wksp() :
+ * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
+ * `wkspSize` size must be `(1<<tableLog)`.
+ */
+size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + dstSize;
+
+    U32   count[FSE_MAX_SYMBOL_VALUE+1];
+    S16   norm[FSE_MAX_SYMBOL_VALUE+1];
+    FSE_CTable* CTable = (FSE_CTable*)workSpace;
+    size_t const CTableSize = FSE_CTABLE_SIZE_U32(tableLog, maxSymbolValue);
+    void* scratchBuffer = (void*)(CTable + CTableSize);
+    size_t const scratchBufferSize = wkspSize - (CTableSize * sizeof(FSE_CTable));
+
+    /* init conditions */
+    if (wkspSize < FSE_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge);
+    if (srcSize <= 1) return 0;  /* Not compressible */
+    if (!maxSymbolValue) maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    if (!tableLog) tableLog = FSE_DEFAULT_TABLELOG;
+
+    /* Scan input and build symbol stats */
+    {   CHECK_V_F(maxCount, FSE_count_wksp(count, &maxSymbolValue, src, srcSize, (unsigned*)scratchBuffer) );
+        if (maxCount == srcSize) return 1;   /* only a single symbol in src : rle */
+        if (maxCount == 1) return 0;         /* each symbol present maximum once => not compressible */
+        if (maxCount < (srcSize >> 7)) return 0;   /* Heuristic : not compressible enough */
+    }
+
+    tableLog = FSE_optimalTableLog(tableLog, srcSize, maxSymbolValue);
+    CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue) );
+
+    /* Write table description header */
+    {   CHECK_V_F(nc_err, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) );
+        op += nc_err;
+    }
+
+    /* Compress */
+    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, scratchBufferSize) );
+    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, oend - op, src, srcSize, CTable) );
+        if (cSize == 0) return 0;   /* not enough space for compressed data */
+        op += cSize;
+    }
+
+    /* check compressibility */
+    if ( (size_t)(op-ostart) >= srcSize-1 ) return 0;
+
+    return op-ostart;
+}
+
+typedef struct {
+    FSE_CTable CTable_max[FSE_CTABLE_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
+    BYTE scratchBuffer[1 << FSE_MAX_TABLELOG];
+} fseWkspMax_t;
+
+size_t FSE_compress2 (void* dst, size_t dstCapacity, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog)
+{
+    fseWkspMax_t scratchBuffer;
+    FSE_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE));   /* compilation failures here means scratchBuffer is not large enough */
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    return FSE_compress_wksp(dst, dstCapacity, src, srcSize, maxSymbolValue, tableLog, &scratchBuffer, sizeof(scratchBuffer));
+}
+
+size_t FSE_compress (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    return FSE_compress2(dst, dstCapacity, src, srcSize, FSE_MAX_SYMBOL_VALUE, FSE_DEFAULT_TABLELOG);
+}
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
diff --git a/vendor/github.com/DataDog/zstd/fse_decompress.c b/vendor/github.com/DataDog/zstd/fse_decompress.c
new file mode 100644
index 0000000..4c66c3b
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/fse_decompress.c
@@ -0,0 +1,309 @@
+/* ******************************************************************
+   FSE : Finite State Entropy decoder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include "bitstream.h"
+#include "compiler.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#include "error_private.h"
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_isError ERR_isError
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+/* check and forward error code */
+#define CHECK_F(f) { size_t const e = f; if (FSE_isError(e)) return e; }
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+
+/* Function templates */
+FSE_DTable* FSE_createDTable (unsigned tableLog)
+{
+    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
+    return (FSE_DTable*)malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
+}
+
+void FSE_freeDTable (FSE_DTable* dt)
+{
+    free(dt);
+}
+
+size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
+    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+
+    U32 const maxSV1 = maxSymbolValue + 1;
+    U32 const tableSize = 1 << tableLog;
+    U32 highThreshold = tableSize-1;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    {   FSE_DTableHeader DTableH;
+        DTableH.tableLog = (U16)tableLog;
+        DTableH.fastMode = 1;
+        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
+            U32 s;
+            for (s=0; s<maxSV1; s++) {
+                if (normalizedCounter[s]==-1) {
+                    tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+                    symbolNext[s] = 1;
+                } else {
+                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+                    symbolNext[s] = normalizedCounter[s];
+        }   }   }
+        memcpy(dt, &DTableH, sizeof(DTableH));
+    }
+
+    /* Spread symbols */
+    {   U32 const tableMask = tableSize-1;
+        U32 const step = FSE_TABLESTEP(tableSize);
+        U32 s, position = 0;
+        for (s=0; s<maxSV1; s++) {
+            int i;
+            for (i=0; i<normalizedCounter[s]; i++) {
+                tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }   }
+        if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+    }
+
+    /* Build Decoding table */
+    {   U32 u;
+        for (u=0; u<tableSize; u++) {
+            FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
+            U32 const nextState = symbolNext[symbol]++;
+            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+    }   }
+
+    return 0;
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/*-*******************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSV1 = tableMask+1;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<maxSV1; s++) {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BIT_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+
+    /* Init */
+    CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+    while (1) {
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+        *op++ = FSE_GETSYMBOL(&state1);
+        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+            *op++ = FSE_GETSYMBOL(&state2);
+            break;
+        }
+
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+        *op++ = FSE_GETSYMBOL(&state2);
+        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+            *op++ = FSE_GETSYMBOL(&state1);
+            break;
+    }   }
+
+    return op-ostart;
+}
+
+
+size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSE_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
+    const U32 fastMode = DTableH->fastMode;
+
+    /* select fast mode (static) */
+    if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSE_MAX_SYMBOL_VALUE+1];
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+
+    /* normal FSE decoding mode */
+    size_t const NCountLength = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSE_isError(NCountLength)) return NCountLength;
+    //if (NCountLength >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size; supposed to be already checked in NCountLength, only remaining case : NCountLength==cSrcSize */
+    if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
+    ip += NCountLength;
+    cSrcSize -= NCountLength;
+
+    CHECK_F( FSE_buildDTable (workSpace, counting, maxSymbolValue, tableLog) );
+
+    return FSE_decompress_usingDTable (dst, dstCapacity, ip, cSrcSize, workSpace);   /* always return, even if it is an error code */
+}
+
+
+typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+
+size_t FSE_decompress(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize)
+{
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    return FSE_decompress_wksp(dst, dstCapacity, cSrc, cSrcSize, dt, FSE_MAX_TABLELOG);
+}
+
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
diff --git a/vendor/github.com/DataDog/zstd/huf.h b/vendor/github.com/DataDog/zstd/huf.h
new file mode 100644
index 0000000..b4645b4
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/huf.h
@@ -0,0 +1,327 @@
+/* ******************************************************************
+   Huffman coder, part of New Generation Entropy library
+   header file
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef HUF_H_298734234
+#define HUF_H_298734234
+
+/* *** Dependencies *** */
+#include <stddef.h>    /* size_t */
+
+
+/* *** library symbols visibility *** */
+/* Note : when linking with -fvisibility=hidden on gcc, or by default on Visual,
+ *        HUF symbols remain "private" (internal symbols for library only).
+ *        Set macro FSE_DLL_EXPORT to 1 if you want HUF symbols visible on DLL interface */
+#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
+#  define HUF_PUBLIC_API __attribute__ ((visibility ("default")))
+#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
+#  define HUF_PUBLIC_API __declspec(dllexport)
+#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
+#  define HUF_PUBLIC_API __declspec(dllimport)  /* not required, just to generate faster code (saves a function pointer load from IAT and an indirect jump) */
+#else
+#  define HUF_PUBLIC_API
+#endif
+
+
+/* ========================== */
+/* ***  simple functions  *** */
+/* ========================== */
+
+/** HUF_compress() :
+ *  Compress content from buffer 'src', of size 'srcSize', into buffer 'dst'.
+ * 'dst' buffer must be already allocated.
+ *  Compression runs faster if `dstCapacity` >= HUF_compressBound(srcSize).
+ * `srcSize` must be <= `HUF_BLOCKSIZE_MAX` == 128 KB.
+ * @return : size of compressed data (<= `dstCapacity`).
+ *  Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
+ *                   if HUF_isError(return), compression failed (more details using HUF_getErrorName())
+ */
+HUF_PUBLIC_API size_t HUF_compress(void* dst, size_t dstCapacity,
+                             const void* src, size_t srcSize);
+
+/** HUF_decompress() :
+ *  Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
+ *  into already allocated buffer 'dst', of minimum size 'dstSize'.
+ * `originalSize` : **must** be the ***exact*** size of original (uncompressed) data.
+ *  Note : in contrast with FSE, HUF_decompress can regenerate
+ *         RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
+ *         because it knows size to regenerate (originalSize).
+ * @return : size of regenerated data (== originalSize),
+ *           or an error code, which can be tested using HUF_isError()
+ */
+HUF_PUBLIC_API size_t HUF_decompress(void* dst,  size_t originalSize,
+                               const void* cSrc, size_t cSrcSize);
+
+
+/* ***   Tool functions *** */
+#define HUF_BLOCKSIZE_MAX (128 * 1024)                  /**< maximum input size for a single block compressed with HUF_compress */
+HUF_PUBLIC_API size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
+
+/* Error Management */
+HUF_PUBLIC_API unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
+HUF_PUBLIC_API const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
+
+
+/* ***   Advanced function   *** */
+
+/** HUF_compress2() :
+ *  Same as HUF_compress(), but offers control over `maxSymbolValue` and `tableLog`.
+ * `maxSymbolValue` must be <= HUF_SYMBOLVALUE_MAX .
+ * `tableLog` must be `<= HUF_TABLELOG_MAX` . */
+HUF_PUBLIC_API size_t HUF_compress2 (void* dst, size_t dstCapacity,
+                               const void* src, size_t srcSize,
+                               unsigned maxSymbolValue, unsigned tableLog);
+
+/** HUF_compress4X_wksp() :
+ *  Same as HUF_compress2(), but uses externally allocated `workSpace`.
+ * `workspace` must have minimum alignment of 4, and be at least as large as HUF_WORKSPACE_SIZE */
+#define HUF_WORKSPACE_SIZE (6 << 10)
+#define HUF_WORKSPACE_SIZE_U32 (HUF_WORKSPACE_SIZE / sizeof(U32))
+HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
+                                     const void* src, size_t srcSize,
+                                     unsigned maxSymbolValue, unsigned tableLog,
+                                     void* workSpace, size_t wkspSize);
+
+#endif   /* HUF_H_298734234 */
+
+/* ******************************************************************
+ *  WARNING !!
+ *  The following section contains advanced and experimental definitions
+ *  which shall never be used in the context of a dynamic library,
+ *  because they are not guaranteed to remain stable in the future.
+ *  Only consider them in association with static linking.
+ * *****************************************************************/
+#if defined(HUF_STATIC_LINKING_ONLY) && !defined(HUF_H_HUF_STATIC_LINKING_ONLY)
+#define HUF_H_HUF_STATIC_LINKING_ONLY
+
+/* *** Dependencies *** */
+#include "mem.h"   /* U32 */
+
+
+/* *** Constants *** */
+#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
+#define HUF_SYMBOLVALUE_MAX  255
+
+#define HUF_TABLELOG_ABSOLUTEMAX  15  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
+#  error "HUF_TABLELOG_MAX is too large !"
+#endif
+
+
+/* ****************************************
+*  Static allocation
+******************************************/
+/* HUF buffer bounds */
+#define HUF_CTABLEBOUND 129
+#define HUF_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true when incompressible is pre-filtered with fast heuristic */
+#define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* static allocation of HUF's Compression Table */
+#define HUF_CTABLE_SIZE_U32(maxSymbolValue)   ((maxSymbolValue)+1)   /* Use tables of U32, for proper alignment */
+#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_U32(maxSymbolValue) * sizeof(U32))
+#define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
+    U32 name##hb[HUF_CTABLE_SIZE_U32(maxSymbolValue)]; \
+    void* name##hv = &(name##hb); \
+    HUF_CElt* name = (HUF_CElt*)(name##hv)   /* no final ; */
+
+/* static allocation of HUF's DTable */
+typedef U32 HUF_DTable;
+#define HUF_DTABLE_SIZE(maxTableLog)   (1 + (1<<(maxTableLog)))
+#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        HUF_DTable DTable[HUF_DTABLE_SIZE((maxTableLog)-1)] = { ((U32)((maxTableLog)-1) * 0x01000001) }
+#define HUF_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        HUF_DTable DTable[HUF_DTABLE_SIZE(maxTableLog)] = { ((U32)(maxTableLog) * 0x01000001) }
+
+
+/* ****************************************
+*  Advanced decompression functions
+******************************************/
+size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+
+size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< decodes RLE and uncompressed */
+size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize); /**< considers RLE and uncompressed as errors */
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize); /**< considers RLE and uncompressed as errors */
+size_t HUF_decompress4X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
+size_t HUF_decompress4X4_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+size_t HUF_decompress4X4_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
+
+
+/* ****************************************
+ *  HUF detailed API
+ * ****************************************/
+
+/*! HUF_compress() does the following:
+ *  1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
+ *  2. (optional) refine tableLog using HUF_optimalTableLog()
+ *  3. build Huffman table from count using HUF_buildCTable()
+ *  4. save Huffman table to memory buffer using HUF_writeCTable()
+ *  5. encode the data stream using HUF_compress4X_usingCTable()
+ *
+ *  The following API allows targeting specific sub-functions for advanced tasks.
+ *  For example, it's possible to compress several blocks using the same 'CTable',
+ *  or to save and regenerate 'CTable' using external methods.
+ */
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
+typedef struct HUF_CElt_s HUF_CElt;   /* incomplete type */
+size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits);   /* @return : maxNbBits; CTable and count can overlap. In which case, CTable will overwrite count content */
+size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+
+typedef enum {
+   HUF_repeat_none,  /**< Cannot use the previous table */
+   HUF_repeat_check, /**< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
+   HUF_repeat_valid  /**< Can use the previous table and it is asumed to be valid */
+ } HUF_repeat;
+/** HUF_compress4X_repeat() :
+ *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
+ *  If it uses hufTable it does not modify hufTable or repeat.
+ *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
+ *  If preferRepeat then the old table will always be used if valid. */
+size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
+                       const void* src, size_t srcSize,
+                       unsigned maxSymbolValue, unsigned tableLog,
+                       void* workSpace, size_t wkspSize,    /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+
+/** HUF_buildCTable_wksp() :
+ *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
+ * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
+ */
+#define HUF_CTABLE_WORKSPACE_SIZE_U32 (2*HUF_SYMBOLVALUE_MAX +1 +1)
+#define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
+size_t HUF_buildCTable_wksp (HUF_CElt* tree, const U32* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize);
+
+/*! HUF_readStats() :
+ *  Read compact Huffman tree, saved by HUF_writeCTable().
+ * `huffWeight` is destination buffer.
+ * @return : size read from `src` , or an error Code .
+ *  Note : Needed by HUF_readCTable() and HUF_readDTableXn() . */
+size_t HUF_readStats(BYTE* huffWeight, size_t hwSize,
+                     U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize);
+
+/** HUF_readCTable() :
+ *  Loading a CTable saved with HUF_writeCTable() */
+size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+
+
+/*
+ * HUF_decompress() does the following:
+ * 1. select the decompression algorithm (X2, X4) based on pre-computed heuristics
+ * 2. build Huffman table from save, using HUF_readDTableX?()
+ * 3. decode 1 or 4 segments in parallel using HUF_decompress?X?_usingDTable()
+ */
+
+/** HUF_selectDecoder() :
+ *  Tells which decoder is likely to decode faster,
+ *  based on a set of pre-computed metrics.
+ * @return : 0==HUF_decompress4X2, 1==HUF_decompress4X4 .
+ *  Assumption : 0 < dstSize <= 128 KB */
+U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
+
+/**
+ *  The minimum workspace size for the `workSpace` used in
+ *  HUF_readDTableX2_wksp() and HUF_readDTableX4_wksp().
+ *
+ *  The space used depends on HUF_TABLELOG_MAX, ranging from ~1500 bytes when
+ *  HUF_TABLE_LOG_MAX=12 to ~1850 bytes when HUF_TABLE_LOG_MAX=15.
+ *  Buffer overflow errors may potentially occur if code modifications result in
+ *  a required workspace size greater than that specified in the following
+ *  macro.
+ */
+#define HUF_DECOMPRESS_WORKSPACE_SIZE (2 << 10)
+#define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
+
+size_t HUF_readDTableX2 (HUF_DTable* DTable, const void* src, size_t srcSize);
+size_t HUF_readDTableX2_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
+size_t HUF_readDTableX4 (HUF_DTable* DTable, const void* src, size_t srcSize);
+size_t HUF_readDTableX4_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
+
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+
+
+/* ====================== */
+/* single stream variants */
+/* ====================== */
+
+size_t HUF_compress1X (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
+size_t HUF_compress1X_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);  /**< `workSpace` must be a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+/** HUF_compress1X_repeat() :
+ *  Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
+ *  If it uses hufTable it does not modify hufTable or repeat.
+ *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
+ *  If preferRepeat then the old table will always be used if valid. */
+size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
+                       const void* src, size_t srcSize,
+                       unsigned maxSymbolValue, unsigned tableLog,
+                       void* workSpace, size_t wkspSize,   /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+
+size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUF_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+
+size_t HUF_decompress1X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+size_t HUF_decompress1X_DCtx_wksp (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);
+size_t HUF_decompress1X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
+size_t HUF_decompress1X4_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+size_t HUF_decompress1X4_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
+
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);   /**< automatic selection of sing or double symbol decoder, based on DTable */
+size_t HUF_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+
+/* BMI2 variants.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
+size_t HUF_decompress1X2_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
+size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+
+#endif /* HUF_STATIC_LINKING_ONLY */
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/vendor/github.com/DataDog/zstd/huf_compress.c b/vendor/github.com/DataDog/zstd/huf_compress.c
new file mode 100644
index 0000000..83230b4
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/huf_compress.c
@@ -0,0 +1,788 @@
+/* ******************************************************************
+   Huffman encoder, part of New Generation Entropy library
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+#include "bitstream.h"
+#include "compiler.h"
+#define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
+#include "fse.h"        /* header compression */
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"
+#include "error_private.h"
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define HUF_isError ERR_isError
+#define HUF_STATIC_ASSERT(c) { enum { HUF_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+#define CHECK_V_F(e, f) size_t const e = f; if (ERR_isError(e)) return e
+#define CHECK_F(f)   { CHECK_V_F(_var_err__, f); }
+
+
+/* **************************************************************
+*  Utils
+****************************************************************/
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
+{
+    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+}
+
+
+/* *******************************************************
+*  HUF : Huffman block compression
+*********************************************************/
+/* HUF_compressWeights() :
+ * Same as FSE_compress(), but dedicated to huff0's weights compression.
+ * The use case needs much less stack memory.
+ * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
+ */
+#define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
+size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weightTable, size_t wtSize)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + dstSize;
+
+    U32 maxSymbolValue = HUF_TABLELOG_MAX;
+    U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
+
+    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
+    BYTE scratchBuffer[1<<MAX_FSE_TABLELOG_FOR_HUFF_HEADER];
+
+    U32 count[HUF_TABLELOG_MAX+1];
+    S16 norm[HUF_TABLELOG_MAX+1];
+
+    /* init conditions */
+    if (wtSize <= 1) return 0;  /* Not compressible */
+
+    /* Scan input and build symbol stats */
+    {   CHECK_V_F(maxCount, FSE_count_simple(count, &maxSymbolValue, weightTable, wtSize) );
+        if (maxCount == wtSize) return 1;   /* only a single symbol in src : rle */
+        if (maxCount == 1) return 0;         /* each symbol present maximum once => not compressible */
+    }
+
+    tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
+    CHECK_F( FSE_normalizeCount(norm, tableLog, count, wtSize, maxSymbolValue) );
+
+    /* Write table description header */
+    {   CHECK_V_F(hSize, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) );
+        op += hSize;
+    }
+
+    /* Compress */
+    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, sizeof(scratchBuffer)) );
+    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, oend - op, weightTable, wtSize, CTable) );
+        if (cSize == 0) return 0;   /* not enough space for compressed data */
+        op += cSize;
+    }
+
+    return op-ostart;
+}
+
+
+struct HUF_CElt_s {
+  U16  val;
+  BYTE nbBits;
+};   /* typedef'd to HUF_CElt within "huf.h" */
+
+/*! HUF_writeCTable() :
+    `CTable` : Huffman tree to save, using huf representation.
+    @return : size of saved CTable */
+size_t HUF_writeCTable (void* dst, size_t maxDstSize,
+                        const HUF_CElt* CTable, U32 maxSymbolValue, U32 huffLog)
+{
+    BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
+    BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
+    BYTE* op = (BYTE*)dst;
+    U32 n;
+
+     /* check conditions */
+    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
+
+    /* convert to weight */
+    bitsToWeight[0] = 0;
+    for (n=1; n<huffLog+1; n++)
+        bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
+    for (n=0; n<maxSymbolValue; n++)
+        huffWeight[n] = bitsToWeight[CTable[n].nbBits];
+
+    /* attempt weights compression by FSE */
+    {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, huffWeight, maxSymbolValue) );
+        if ((hSize>1) & (hSize < maxSymbolValue/2)) {   /* FSE compressed */
+            op[0] = (BYTE)hSize;
+            return hSize+1;
+    }   }
+
+    /* write raw values as 4-bits (max : 15) */
+    if (maxSymbolValue > (256-128)) return ERROR(GENERIC);   /* should not happen : likely means source cannot be compressed */
+    if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall);   /* not enough space within dst buffer */
+    op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
+    huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
+    for (n=0; n<maxSymbolValue; n+=2)
+        op[(n/2)+1] = (BYTE)((huffWeight[n] << 4) + huffWeight[n+1]);
+    return ((maxSymbolValue+1)/2) + 1;
+}
+
+
+size_t HUF_readCTable (HUF_CElt* CTable, U32* maxSymbolValuePtr, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];   /* init not required, even though some static analyzer may complain */
+    U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    U32 nbSymbols = 0;
+
+    /* get symbol weights */
+    CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
+
+    /* check result */
+    if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
+    if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
+
+    /* Prepare base value per rank */
+    {   U32 n, nextRankStart = 0;
+        for (n=1; n<=tableLog; n++) {
+            U32 current = nextRankStart;
+            nextRankStart += (rankVal[n] << (n-1));
+            rankVal[n] = current;
+    }   }
+
+    /* fill nbBits */
+    {   U32 n; for (n=0; n<nbSymbols; n++) {
+            const U32 w = huffWeight[n];
+            CTable[n].nbBits = (BYTE)(tableLog + 1 - w);
+    }   }
+
+    /* fill val */
+    {   U16 nbPerRank[HUF_TABLELOG_MAX+2]  = {0};  /* support w=0=>n=tableLog+1 */
+        U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
+        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; }
+        /* determine stating value per rank */
+        valPerRank[tableLog+1] = 0;   /* for w==0 */
+        {   U16 min = 0;
+            U32 n; for (n=tableLog; n>0; n--) {  /* start at n=tablelog <-> w=1 */
+                valPerRank[n] = min;     /* get starting value within each rank */
+                min += nbPerRank[n];
+                min >>= 1;
+        }   }
+        /* assign value within rank, symbol order */
+        { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; }
+    }
+
+    *maxSymbolValuePtr = nbSymbols - 1;
+    return readSize;
+}
+
+
+typedef struct nodeElt_s {
+    U32 count;
+    U16 parent;
+    BYTE byte;
+    BYTE nbBits;
+} nodeElt;
+
+static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
+{
+    const U32 largestBits = huffNode[lastNonNull].nbBits;
+    if (largestBits <= maxNbBits) return largestBits;   /* early exit : no elt > maxNbBits */
+
+    /* there are several too large elements (at least >= 2) */
+    {   int totalCost = 0;
+        const U32 baseCost = 1 << (largestBits - maxNbBits);
+        U32 n = lastNonNull;
+
+        while (huffNode[n].nbBits > maxNbBits) {
+            totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
+            huffNode[n].nbBits = (BYTE)maxNbBits;
+            n --;
+        }  /* n stops at huffNode[n].nbBits <= maxNbBits */
+        while (huffNode[n].nbBits == maxNbBits) n--;   /* n end at index of smallest symbol using < maxNbBits */
+
+        /* renorm totalCost */
+        totalCost >>= (largestBits - maxNbBits);  /* note : totalCost is necessarily a multiple of baseCost */
+
+        /* repay normalized cost */
+        {   U32 const noSymbol = 0xF0F0F0F0;
+            U32 rankLast[HUF_TABLELOG_MAX+2];
+            int pos;
+
+            /* Get pos of last (smallest) symbol per rank */
+            memset(rankLast, 0xF0, sizeof(rankLast));
+            {   U32 currentNbBits = maxNbBits;
+                for (pos=n ; pos >= 0; pos--) {
+                    if (huffNode[pos].nbBits >= currentNbBits) continue;
+                    currentNbBits = huffNode[pos].nbBits;   /* < maxNbBits */
+                    rankLast[maxNbBits-currentNbBits] = pos;
+            }   }
+
+            while (totalCost > 0) {
+                U32 nBitsToDecrease = BIT_highbit32(totalCost) + 1;
+                for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
+                    U32 highPos = rankLast[nBitsToDecrease];
+                    U32 lowPos = rankLast[nBitsToDecrease-1];
+                    if (highPos == noSymbol) continue;
+                    if (lowPos == noSymbol) break;
+                    {   U32 const highTotal = huffNode[highPos].count;
+                        U32 const lowTotal = 2 * huffNode[lowPos].count;
+                        if (highTotal <= lowTotal) break;
+                }   }
+                /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
+                /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */
+                while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
+                    nBitsToDecrease ++;
+                totalCost -= 1 << (nBitsToDecrease-1);
+                if (rankLast[nBitsToDecrease-1] == noSymbol)
+                    rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];   /* this rank is no longer empty */
+                huffNode[rankLast[nBitsToDecrease]].nbBits ++;
+                if (rankLast[nBitsToDecrease] == 0)    /* special case, reached largest symbol */
+                    rankLast[nBitsToDecrease] = noSymbol;
+                else {
+                    rankLast[nBitsToDecrease]--;
+                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
+                        rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
+            }   }   /* while (totalCost > 0) */
+
+            while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
+                if (rankLast[1] == noSymbol) {  /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
+                    while (huffNode[n].nbBits == maxNbBits) n--;
+                    huffNode[n+1].nbBits--;
+                    rankLast[1] = n+1;
+                    totalCost++;
+                    continue;
+                }
+                huffNode[ rankLast[1] + 1 ].nbBits--;
+                rankLast[1]++;
+                totalCost ++;
+    }   }   }   /* there are several too large elements (at least >= 2) */
+
+    return maxNbBits;
+}
+
+
+typedef struct {
+    U32 base;
+    U32 current;
+} rankPos;
+
+static void HUF_sort(nodeElt* huffNode, const U32* count, U32 maxSymbolValue)
+{
+    rankPos rank[32];
+    U32 n;
+
+    memset(rank, 0, sizeof(rank));
+    for (n=0; n<=maxSymbolValue; n++) {
+        U32 r = BIT_highbit32(count[n] + 1);
+        rank[r].base ++;
+    }
+    for (n=30; n>0; n--) rank[n-1].base += rank[n].base;
+    for (n=0; n<32; n++) rank[n].current = rank[n].base;
+    for (n=0; n<=maxSymbolValue; n++) {
+        U32 const c = count[n];
+        U32 const r = BIT_highbit32(c+1) + 1;
+        U32 pos = rank[r].current++;
+        while ((pos > rank[r].base) && (c > huffNode[pos-1].count)) {
+            huffNode[pos] = huffNode[pos-1];
+            pos--;
+        }
+        huffNode[pos].count = c;
+        huffNode[pos].byte  = (BYTE)n;
+    }
+}
+
+
+/** HUF_buildCTable_wksp() :
+ *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
+ *  `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as a table of HUF_CTABLE_WORKSPACE_SIZE_U32 unsigned.
+ */
+#define STARTNODE (HUF_SYMBOLVALUE_MAX+1)
+typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
+size_t HUF_buildCTable_wksp (HUF_CElt* tree, const U32* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
+{
+    nodeElt* const huffNode0 = (nodeElt*)workSpace;
+    nodeElt* const huffNode = huffNode0+1;
+    U32 n, nonNullRank;
+    int lowS, lowN;
+    U16 nodeNb = STARTNODE;
+    U32 nodeRoot;
+
+    /* safety checks */
+    if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
+    if (wkspSize < sizeof(huffNodeTable)) return ERROR(workSpace_tooSmall);
+    if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
+    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
+    memset(huffNode0, 0, sizeof(huffNodeTable));
+
+    /* sort, decreasing order */
+    HUF_sort(huffNode, count, maxSymbolValue);
+
+    /* init for parents */
+    nonNullRank = maxSymbolValue;
+    while(huffNode[nonNullRank].count == 0) nonNullRank--;
+    lowS = nonNullRank; nodeRoot = nodeNb + lowS - 1; lowN = nodeNb;
+    huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count;
+    huffNode[lowS].parent = huffNode[lowS-1].parent = nodeNb;
+    nodeNb++; lowS-=2;
+    for (n=nodeNb; n<=nodeRoot; n++) huffNode[n].count = (U32)(1U<<30);
+    huffNode0[0].count = (U32)(1U<<31);  /* fake entry, strong barrier */
+
+    /* create parents */
+    while (nodeNb <= nodeRoot) {
+        U32 n1 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
+        U32 n2 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
+        huffNode[nodeNb].count = huffNode[n1].count + huffNode[n2].count;
+        huffNode[n1].parent = huffNode[n2].parent = nodeNb;
+        nodeNb++;
+    }
+
+    /* distribute weights (unlimited tree height) */
+    huffNode[nodeRoot].nbBits = 0;
+    for (n=nodeRoot-1; n>=STARTNODE; n--)
+        huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
+    for (n=0; n<=nonNullRank; n++)
+        huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
+
+    /* enforce maxTableLog */
+    maxNbBits = HUF_setMaxHeight(huffNode, nonNullRank, maxNbBits);
+
+    /* fill result into tree (val, nbBits) */
+    {   U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
+        U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
+        if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
+        for (n=0; n<=nonNullRank; n++)
+            nbPerRank[huffNode[n].nbBits]++;
+        /* determine stating value per rank */
+        {   U16 min = 0;
+            for (n=maxNbBits; n>0; n--) {
+                valPerRank[n] = min;      /* get starting value within each rank */
+                min += nbPerRank[n];
+                min >>= 1;
+        }   }
+        for (n=0; n<=maxSymbolValue; n++)
+            tree[huffNode[n].byte].nbBits = huffNode[n].nbBits;   /* push nbBits per symbol, symbol order */
+        for (n=0; n<=maxSymbolValue; n++)
+            tree[n].val = valPerRank[tree[n].nbBits]++;   /* assign value within rank, symbol order */
+    }
+
+    return maxNbBits;
+}
+
+/** HUF_buildCTable() :
+ * @return : maxNbBits
+ *  Note : count is used before tree is written, so they can safely overlap
+ */
+size_t HUF_buildCTable (HUF_CElt* tree, const U32* count, U32 maxSymbolValue, U32 maxNbBits)
+{
+    huffNodeTable nodeTable;
+    return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, nodeTable, sizeof(nodeTable));
+}
+
+static size_t HUF_estimateCompressedSize(HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
+{
+    size_t nbBits = 0;
+    int s;
+    for (s = 0; s <= (int)maxSymbolValue; ++s) {
+        nbBits += CTable[s].nbBits * count[s];
+    }
+    return nbBits >> 3;
+}
+
+static int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
+  int bad = 0;
+  int s;
+  for (s = 0; s <= (int)maxSymbolValue; ++s) {
+    bad |= (count[s] != 0) & (CTable[s].nbBits == 0);
+  }
+  return !bad;
+}
+
+size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
+
+FORCE_INLINE_TEMPLATE void
+HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
+{
+    BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
+}
+
+#define HUF_FLUSHBITS(s)  BIT_flushBits(s)
+
+#define HUF_FLUSHBITS_1(stream) \
+    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
+
+#define HUF_FLUSHBITS_2(stream) \
+    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream)
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
+                                   const void* src, size_t srcSize,
+                                   const HUF_CElt* CTable)
+{
+    const BYTE* ip = (const BYTE*) src;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = ostart + dstSize;
+    BYTE* op = ostart;
+    size_t n;
+    BIT_CStream_t bitC;
+
+    /* init */
+    if (dstSize < 8) return 0;   /* not enough space to compress */
+    { size_t const initErr = BIT_initCStream(&bitC, op, oend-op);
+      if (HUF_isError(initErr)) return 0; }
+
+    n = srcSize & ~3;  /* join to mod 4 */
+    switch (srcSize & 3)
+    {
+        case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
+                 HUF_FLUSHBITS_2(&bitC);
+		 /* fall-through */
+        case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
+                 HUF_FLUSHBITS_1(&bitC);
+		 /* fall-through */
+        case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
+                 HUF_FLUSHBITS(&bitC);
+		 /* fall-through */
+        case 0 : /* fall-through */
+        default: break;
+    }
+
+    for (; n>0; n-=4) {  /* note : n&3==0 at this stage */
+        HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
+        HUF_FLUSHBITS_1(&bitC);
+        HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
+        HUF_FLUSHBITS_2(&bitC);
+        HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
+        HUF_FLUSHBITS_1(&bitC);
+        HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
+        HUF_FLUSHBITS(&bitC);
+    }
+
+    return BIT_closeCStream(&bitC);
+}
+
+#if DYNAMIC_BMI2
+
+static TARGET_ATTRIBUTE("bmi2") size_t
+HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
+                                   const void* src, size_t srcSize,
+                                   const HUF_CElt* CTable)
+{
+    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
+}
+
+static size_t
+HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
+                                      const void* src, size_t srcSize,
+                                      const HUF_CElt* CTable)
+{
+    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
+}
+
+static size_t
+HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
+                              const void* src, size_t srcSize,
+                              const HUF_CElt* CTable, const int bmi2)
+{
+    if (bmi2) {
+        return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
+    }
+    return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
+}
+
+#else
+
+static size_t
+HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
+                              const void* src, size_t srcSize,
+                              const HUF_CElt* CTable, const int bmi2)
+{
+    (void)bmi2;
+    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
+}
+
+#endif
+
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+{
+    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+}
+
+
+static size_t
+HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
+                              const void* src, size_t srcSize,
+                              const HUF_CElt* CTable, int bmi2)
+{
+    size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
+    const BYTE* ip = (const BYTE*) src;
+    const BYTE* const iend = ip + srcSize;
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* const oend = ostart + dstSize;
+    BYTE* op = ostart;
+
+    if (dstSize < 6 + 1 + 1 + 1 + 8) return 0;   /* minimum space to compress successfully */
+    if (srcSize < 12) return 0;   /* no saving possible : too small input */
+    op += 6;   /* jumpTable */
+
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, oend-op, ip, segmentSize, CTable, bmi2) );
+        if (cSize==0) return 0;
+        assert(cSize <= 65535);
+        MEM_writeLE16(ostart, (U16)cSize);
+        op += cSize;
+    }
+
+    ip += segmentSize;
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, oend-op, ip, segmentSize, CTable, bmi2) );
+        if (cSize==0) return 0;
+        assert(cSize <= 65535);
+        MEM_writeLE16(ostart+2, (U16)cSize);
+        op += cSize;
+    }
+
+    ip += segmentSize;
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, oend-op, ip, segmentSize, CTable, bmi2) );
+        if (cSize==0) return 0;
+        assert(cSize <= 65535);
+        MEM_writeLE16(ostart+4, (U16)cSize);
+        op += cSize;
+    }
+
+    ip += segmentSize;
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, oend-op, ip, iend-ip, CTable, bmi2) );
+        if (cSize==0) return 0;
+        op += cSize;
+    }
+
+    return op-ostart;
+}
+
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+{
+    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+}
+
+
+static size_t HUF_compressCTable_internal(
+                BYTE* const ostart, BYTE* op, BYTE* const oend,
+                const void* src, size_t srcSize,
+                unsigned singleStream, const HUF_CElt* CTable, const int bmi2)
+{
+    size_t const cSize = singleStream ?
+                         HUF_compress1X_usingCTable_internal(op, oend - op, src, srcSize, CTable, bmi2) :
+                         HUF_compress4X_usingCTable_internal(op, oend - op, src, srcSize, CTable, bmi2);
+    if (HUF_isError(cSize)) { return cSize; }
+    if (cSize==0) { return 0; }   /* uncompressible */
+    op += cSize;
+    /* check compressibility */
+    if ((size_t)(op-ostart) >= srcSize-1) { return 0; }
+    return op-ostart;
+}
+
+typedef struct {
+    U32 count[HUF_SYMBOLVALUE_MAX + 1];
+    HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
+    huffNodeTable nodeTable;
+} HUF_compress_tables_t;
+
+/* HUF_compress_internal() :
+ * `workSpace` must a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
+static size_t HUF_compress_internal (
+                void* dst, size_t dstSize,
+                const void* src, size_t srcSize,
+                unsigned maxSymbolValue, unsigned huffLog,
+                unsigned singleStream,
+                void* workSpace, size_t wkspSize,
+                HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
+                const int bmi2)
+{
+    HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = ostart + dstSize;
+    BYTE* op = ostart;
+
+    /* checks & inits */
+    if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
+    if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall);
+    if (!srcSize) return 0;  /* Uncompressed */
+    if (!dstSize) return 0;  /* cannot fit anything within dst budget */
+    if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);   /* current block size limit */
+    if (huffLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
+    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
+    if (!maxSymbolValue) maxSymbolValue = HUF_SYMBOLVALUE_MAX;
+    if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
+
+    /* Heuristic : If old table is valid, use it for small inputs */
+    if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
+        return HUF_compressCTable_internal(ostart, op, oend,
+                                           src, srcSize,
+                                           singleStream, oldHufTable, bmi2);
+    }
+
+    /* Scan input and build symbol stats */
+    {   CHECK_V_F(largest, FSE_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, table->count) );
+        if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
+        if (largest <= (srcSize >> 7)+1) return 0;   /* heuristic : probably not compressible enough */
+    }
+
+    /* Check validity of previous table */
+    if ( repeat
+      && *repeat == HUF_repeat_check
+      && !HUF_validateCTable(oldHufTable, table->count, maxSymbolValue)) {
+        *repeat = HUF_repeat_none;
+    }
+    /* Heuristic : use existing table for small inputs */
+    if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
+        return HUF_compressCTable_internal(ostart, op, oend,
+                                           src, srcSize,
+                                           singleStream, oldHufTable, bmi2);
+    }
+
+    /* Build Huffman Tree */
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    {   CHECK_V_F(maxBits, HUF_buildCTable_wksp(table->CTable, table->count,
+                                                maxSymbolValue, huffLog,
+                                                table->nodeTable, sizeof(table->nodeTable)) );
+        huffLog = (U32)maxBits;
+        /* Zero unused symbols in CTable, so we can check it for validity */
+        memset(table->CTable + (maxSymbolValue + 1), 0,
+               sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt)));
+    }
+
+    /* Write table description header */
+    {   CHECK_V_F(hSize, HUF_writeCTable (op, dstSize, table->CTable, maxSymbolValue, huffLog) );
+        /* Check if using previous huffman table is beneficial */
+        if (repeat && *repeat != HUF_repeat_none) {
+            size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
+            size_t const newSize = HUF_estimateCompressedSize(table->CTable, table->count, maxSymbolValue);
+            if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
+                return HUF_compressCTable_internal(ostart, op, oend,
+                                                   src, srcSize,
+                                                   singleStream, oldHufTable, bmi2);
+        }   }
+
+        /* Use the new huffman table */
+        if (hSize + 12ul >= srcSize) { return 0; }
+        op += hSize;
+        if (repeat) { *repeat = HUF_repeat_none; }
+        if (oldHufTable)
+            memcpy(oldHufTable, table->CTable, sizeof(table->CTable));  /* Save new table */
+    }
+    return HUF_compressCTable_internal(ostart, op, oend,
+                                       src, srcSize,
+                                       singleStream, table->CTable, bmi2);
+}
+
+
+size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
+                      const void* src, size_t srcSize,
+                      unsigned maxSymbolValue, unsigned huffLog,
+                      void* workSpace, size_t wkspSize)
+{
+    return HUF_compress_internal(dst, dstSize, src, srcSize,
+                                 maxSymbolValue, huffLog, 1 /*single stream*/,
+                                 workSpace, wkspSize,
+                                 NULL, NULL, 0, 0 /*bmi2*/);
+}
+
+size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
+                      const void* src, size_t srcSize,
+                      unsigned maxSymbolValue, unsigned huffLog,
+                      void* workSpace, size_t wkspSize,
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
+{
+    return HUF_compress_internal(dst, dstSize, src, srcSize,
+                                 maxSymbolValue, huffLog, 1 /*single stream*/,
+                                 workSpace, wkspSize, hufTable,
+                                 repeat, preferRepeat, bmi2);
+}
+
+size_t HUF_compress1X (void* dst, size_t dstSize,
+                 const void* src, size_t srcSize,
+                 unsigned maxSymbolValue, unsigned huffLog)
+{
+    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
+    return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
+}
+
+/* HUF_compress4X_repeat():
+ * compress input using 4 streams.
+ * provide workspace to generate compression tables */
+size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
+                      const void* src, size_t srcSize,
+                      unsigned maxSymbolValue, unsigned huffLog,
+                      void* workSpace, size_t wkspSize)
+{
+    return HUF_compress_internal(dst, dstSize, src, srcSize,
+                                 maxSymbolValue, huffLog, 0 /*4 streams*/,
+                                 workSpace, wkspSize,
+                                 NULL, NULL, 0, 0 /*bmi2*/);
+}
+
+/* HUF_compress4X_repeat():
+ * compress input using 4 streams.
+ * re-use an existing huffman compression table */
+size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
+                      const void* src, size_t srcSize,
+                      unsigned maxSymbolValue, unsigned huffLog,
+                      void* workSpace, size_t wkspSize,
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
+{
+    return HUF_compress_internal(dst, dstSize, src, srcSize,
+                                 maxSymbolValue, huffLog, 0 /* 4 streams */,
+                                 workSpace, wkspSize,
+                                 hufTable, repeat, preferRepeat, bmi2);
+}
+
+size_t HUF_compress2 (void* dst, size_t dstSize,
+                const void* src, size_t srcSize,
+                unsigned maxSymbolValue, unsigned huffLog)
+{
+    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
+    return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
+}
+
+size_t HUF_compress (void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return HUF_compress2(dst, maxDstSize, src, srcSize, 255, HUF_TABLELOG_DEFAULT);
+}
diff --git a/vendor/github.com/DataDog/zstd/huf_decompress.c b/vendor/github.com/DataDog/zstd/huf_decompress.c
new file mode 100644
index 0000000..73f5c46
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/huf_decompress.c
@@ -0,0 +1,1096 @@
+/* ******************************************************************
+   Huffman decoder, part of New Generation Entropy library
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Dependencies
+****************************************************************/
+#include <string.h>     /* memcpy, memset */
+#include "bitstream.h"  /* BIT_* */
+#include "compiler.h"
+#include "fse.h"        /* header compression */
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"
+#include "error_private.h"
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define HUF_isError ERR_isError
+#define HUF_STATIC_ASSERT(c) { enum { HUF_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+#define CHECK_F(f) { size_t const err_ = (f); if (HUF_isError(err_)) return err_; }
+
+
+/* **************************************************************
+*  Byte alignment for workSpace management
+****************************************************************/
+#define HUF_ALIGN(x, a)         HUF_ALIGN_MASK((x), (a) - 1)
+#define HUF_ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
+
+
+/*-***************************/
+/*  generic DTableDesc       */
+/*-***************************/
+typedef struct { BYTE maxTableLog; BYTE tableType; BYTE tableLog; BYTE reserved; } DTableDesc;
+
+static DTableDesc HUF_getDTableDesc(const HUF_DTable* table)
+{
+    DTableDesc dtd;
+    memcpy(&dtd, table, sizeof(dtd));
+    return dtd;
+}
+
+
+/*-***************************/
+/*  single-symbol decoding   */
+/*-***************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUF_DEltX2;   /* single-symbol decoding */
+
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize)
+{
+    U32 tableLog = 0;
+    U32 nbSymbols = 0;
+    size_t iSize;
+    void* const dtPtr = DTable + 1;
+    HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
+
+    U32* rankVal;
+    BYTE* huffWeight;
+    size_t spaceUsed32 = 0;
+
+    rankVal = (U32 *)workSpace + spaceUsed32;
+    spaceUsed32 += HUF_TABLELOG_ABSOLUTEMAX + 1;
+    huffWeight = (BYTE *)((U32 *)workSpace + spaceUsed32);
+    spaceUsed32 += HUF_ALIGN(HUF_SYMBOLVALUE_MAX + 1, sizeof(U32)) >> 2;
+
+    if ((spaceUsed32 << 2) > wkspSize) return ERROR(tableLog_tooLarge);
+
+    HUF_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUF_DTable));
+    /* memset(huffWeight, 0, sizeof(huffWeight)); */   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* Table header */
+    {   DTableDesc dtd = HUF_getDTableDesc(DTable);
+        if (tableLog > (U32)(dtd.maxTableLog+1)) return ERROR(tableLog_tooLarge);   /* DTable too small, Huffman tree cannot fit in */
+        dtd.tableType = 0;
+        dtd.tableLog = (BYTE)tableLog;
+        memcpy(DTable, &dtd, sizeof(dtd));
+    }
+
+    /* Calculate starting value for each rank */
+    {   U32 n, nextRankStart = 0;
+        for (n=1; n<tableLog+1; n++) {
+            U32 const current = nextRankStart;
+            nextRankStart += (rankVal[n] << (n-1));
+            rankVal[n] = current;
+    }   }
+
+    /* fill DTable */
+    {   U32 n;
+        for (n=0; n<nbSymbols; n++) {
+            U32 const w = huffWeight[n];
+            U32 const length = (1 << w) >> 1;
+            U32 u;
+            HUF_DEltX2 D;
+            D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+            for (u = rankVal[w]; u < rankVal[w] + length; u++)
+                dt[u] = D;
+            rankVal[w] += length;
+    }   }
+
+    return iSize;
+}
+
+size_t HUF_readDTableX2(HUF_DTable* DTable, const void* src, size_t srcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_readDTableX2_wksp(DTable, src, srcSize,
+                                 workSpace, sizeof(workSpace));
+}
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX4;  /* double-symbols decoding */
+
+FORCE_INLINE_TEMPLATE BYTE
+HUF_decodeSymbolX2(BIT_DStream_t* Dstream, const HUF_DEltX2* dt, const U32 dtLog)
+{
+    size_t const val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+    BYTE const c = dt[val].byte;
+    BIT_skipBits(Dstream, dt[val].nbBits);
+    return c;
+}
+
+#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUF_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr)  \
+    if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+HINT_INLINE size_t
+HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-3)) {
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* [0-3] symbols remaining */
+    if (MEM_32bits())
+        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd))
+            HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, no need to reload */
+    while (p < pEnd)
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_decompress1X2_usingDTable_internal_body(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + dstSize;
+    const void* dtPtr = DTable + 1;
+    const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
+    BIT_DStream_t bitD;
+    DTableDesc const dtd = HUF_getDTableDesc(DTable);
+    U32 const dtLog = dtd.tableLog;
+
+    CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );
+
+    HUF_decodeStreamX2(op, &bitD, oend, dt, dtLog);
+
+    if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    return dstSize;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_decompress4X2_usingDTable_internal_body(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    /* Check */
+    if (cSrcSize < 10) return ERROR(corruption_detected);  /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable + 1;
+        const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        size_t const length1 = MEM_readLE16(istart);
+        size_t const length2 = MEM_readLE16(istart+2);
+        size_t const length3 = MEM_readLE16(istart+4);
+        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal = BIT_DStream_unfinished;
+        DTableDesc const dtd = HUF_getDTableDesc(DTable);
+        U32 const dtLog = dtd.tableLog;
+
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
+        CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
+        CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
+        CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
+
+        /* up to 16 symbols per loop (4 symbols per stream) in 64-bit mode */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        while ( (endSignal==BIT_DStream_unfinished) && (op4<(oend-3)) ) {
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+            BIT_reloadDStream(&bitD1);
+            BIT_reloadDStream(&bitD2);
+            BIT_reloadDStream(&bitD3);
+            BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        /* note : should not be necessary : op# advance in lock step, and we control op4.
+         *        but curiously, binary generated by gcc 7.2 & 7.3 with -mbmi2 runs faster when >=1 test is present */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+          if (!endCheck) return ERROR(corruption_detected); }
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE U32
+HUF_decodeSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BIT_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+FORCE_INLINE_TEMPLATE U32
+HUF_decodeLastSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BIT_skipBits(DStream, dt[val].nbBits);
+    else {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
+            BIT_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);
+    }   }
+    return 1;
+}
+
+#define HUF_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+HINT_INLINE size_t
+HUF_decodeStreamX4(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd,
+                const HUF_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-(sizeof(bitDPtr->bitContainer)-1))) {
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to end : up to 2 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p <= pEnd-2))
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUF_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_decompress1X4_usingDTable_internal_body(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    BIT_DStream_t bitD;
+
+    /* Init */
+    CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );
+
+    /* decode */
+    {   BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable+1;   /* force compiler to not use strict-aliasing */
+        const HUF_DEltX4* const dt = (const HUF_DEltX4*)dtPtr;
+        DTableDesc const dtd = HUF_getDTableDesc(DTable);
+        HUF_decodeStreamX4(ostart, &bitD, oend, dt, dtd.tableLog);
+    }
+
+    /* check */
+    if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    /* decoded size */
+    return dstSize;
+}
+
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_decompress4X4_usingDTable_internal_body(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable+1;
+        const HUF_DEltX4* const dt = (const HUF_DEltX4*)dtPtr;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        size_t const length1 = MEM_readLE16(istart);
+        size_t const length2 = MEM_readLE16(istart+2);
+        size_t const length3 = MEM_readLE16(istart+4);
+        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        size_t const segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+        DTableDesc const dtd = HUF_getDTableDesc(DTable);
+        U32 const dtLog = dtd.tableLog;
+
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
+        CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
+        CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
+        CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) & (op4<(oend-(sizeof(bitD4.bitContainer)-1))) ; ) {
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+          if (!endCheck) return ERROR(corruption_detected); }
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+typedef size_t (*HUF_decompress_usingDTable_t)(void *dst, size_t dstSize,
+                                               const void *cSrc,
+                                               size_t cSrcSize,
+                                               const HUF_DTable *DTable);
+#if DYNAMIC_BMI2
+
+#define X(fn)                                                               \
+                                                                            \
+    static size_t fn##_default(                                             \
+                  void* dst,  size_t dstSize,                               \
+            const void* cSrc, size_t cSrcSize,                              \
+            const HUF_DTable* DTable)                                       \
+    {                                                                       \
+        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
+    }                                                                       \
+                                                                            \
+    static TARGET_ATTRIBUTE("bmi2") size_t fn##_bmi2(                       \
+                  void* dst,  size_t dstSize,                               \
+            const void* cSrc, size_t cSrcSize,                              \
+            const HUF_DTable* DTable)                                       \
+    {                                                                       \
+        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
+    }                                                                       \
+                                                                            \
+    static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
+                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+    {                                                                       \
+        if (bmi2) {                                                         \
+            return fn##_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);         \
+        }                                                                   \
+        return fn##_default(dst, dstSize, cSrc, cSrcSize, DTable);          \
+    }
+
+#else
+
+#define X(fn)                                                               \
+    static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
+                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+    {                                                                       \
+        (void)bmi2;                                                         \
+        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
+    }
+
+#endif
+
+X(HUF_decompress1X2_usingDTable_internal)
+X(HUF_decompress4X2_usingDTable_internal)
+X(HUF_decompress1X4_usingDTable_internal)
+X(HUF_decompress4X4_usingDTable_internal)
+
+#undef X
+
+
+size_t HUF_decompress1X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    DTableDesc dtd = HUF_getDTableDesc(DTable);
+    if (dtd.tableType != 0) return ERROR(GENERIC);
+    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize, workSpace, wkspSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
+}
+
+
+size_t HUF_decompress1X2_DCtx(HUF_DTable* DCtx, void* dst, size_t dstSize,
+                              const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress1X2_DCtx_wksp(DCtx, dst, dstSize, cSrc, cSrcSize,
+                                       workSpace, sizeof(workSpace));
+}
+
+size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_TABLELOG_MAX);
+    return HUF_decompress1X2_DCtx (DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+size_t HUF_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    DTableDesc dtd = HUF_getDTableDesc(DTable);
+    if (dtd.tableType != 0) return ERROR(GENERIC);
+    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+static size_t HUF_decompress4X2_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize, int bmi2)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUF_readDTableX2_wksp (dctx, cSrc, cSrcSize,
+                                                workSpace, wkspSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+}
+
+size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize)
+{
+    return HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, 0);
+}
+
+
+size_t HUF_decompress4X2_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
+                                       workSpace, sizeof(workSpace));
+}
+size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_TABLELOG_MAX);
+    return HUF_decompress4X2_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+
+/* *************************/
+/* double-symbols decoding */
+/* *************************/
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+/* HUF_fillDTableX4Level2() :
+ * `rankValOrigin` must be a table of at least (HUF_TABLELOG_MAX + 1) U32 */
+static void HUF_fillDTableX4Level2(HUF_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUF_DEltX4 DElt;
+    U32 rankVal[HUF_TABLELOG_MAX + 1];
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1) {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    {   U32 s; for (s=0; s<sortedListSize; s++) {   /* note : sortedSymbols already skipped */
+            const U32 symbol = sortedSymbols[s].symbol;
+            const U32 weight = sortedSymbols[s].weight;
+            const U32 nbBits = nbBitsBaseline - weight;
+            const U32 length = 1 << (sizeLog-nbBits);
+            const U32 start = rankVal[weight];
+            U32 i = start;
+            const U32 end = start + length;
+
+            MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+            DElt.nbBits = (BYTE)(nbBits + consumed);
+            DElt.length = 2;
+            do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+            rankVal[weight] += length;
+    }   }
+}
+
+typedef U32 rankValCol_t[HUF_TABLELOG_MAX + 1];
+typedef rankValCol_t rankVal_t[HUF_TABLELOG_MAX];
+
+static void HUF_fillDTableX4(HUF_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUF_TABLELOG_MAX + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++) {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits) {   /* enough room for a second symbol */
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUF_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        } else {
+            HUF_DEltX4 DElt;
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits = (BYTE)(nbBits);
+            DElt.length = 1;
+            {   U32 const end = start + length;
+                U32 u;
+                for (u = start; u < end; u++) DTable[u] = DElt;
+        }   }
+        rankVal[weight] += length;
+    }
+}
+
+size_t HUF_readDTableX4_wksp(HUF_DTable* DTable, const void* src,
+                             size_t srcSize, void* workSpace,
+                             size_t wkspSize)
+{
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    DTableDesc dtd = HUF_getDTableDesc(DTable);
+    U32 const maxTableLog = dtd.maxTableLog;
+    size_t iSize;
+    void* dtPtr = DTable+1;   /* force compiler to avoid strict-aliasing */
+    HUF_DEltX4* const dt = (HUF_DEltX4*)dtPtr;
+    U32 *rankStart;
+
+    rankValCol_t* rankVal;
+    U32* rankStats;
+    U32* rankStart0;
+    sortedSymbol_t* sortedSymbol;
+    BYTE* weightList;
+    size_t spaceUsed32 = 0;
+
+    rankVal = (rankValCol_t *)((U32 *)workSpace + spaceUsed32);
+    spaceUsed32 += (sizeof(rankValCol_t) * HUF_TABLELOG_MAX) >> 2;
+    rankStats = (U32 *)workSpace + spaceUsed32;
+    spaceUsed32 += HUF_TABLELOG_MAX + 1;
+    rankStart0 = (U32 *)workSpace + spaceUsed32;
+    spaceUsed32 += HUF_TABLELOG_MAX + 2;
+    sortedSymbol = (sortedSymbol_t *)workSpace + (spaceUsed32 * sizeof(U32)) / sizeof(sortedSymbol_t);
+    spaceUsed32 += HUF_ALIGN(sizeof(sortedSymbol_t) * (HUF_SYMBOLVALUE_MAX + 1), sizeof(U32)) >> 2;
+    weightList = (BYTE *)((U32 *)workSpace + spaceUsed32);
+    spaceUsed32 += HUF_ALIGN(HUF_SYMBOLVALUE_MAX + 1, sizeof(U32)) >> 2;
+
+    if ((spaceUsed32 << 2) > wkspSize) return ERROR(tableLog_tooLarge);
+
+    rankStart = rankStart0 + 1;
+    memset(rankStats, 0, sizeof(U32) * (2 * HUF_TABLELOG_MAX + 2 + 1));
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX4) == sizeof(HUF_DTable));   /* if compiler fails here, assertion is wrong */
+    if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
+    /* memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(weightList, HUF_SYMBOLVALUE_MAX + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+
+    /* Get start index of each weight */
+    {   U32 w, nextRankStart = 0;
+        for (w=1; w<maxW+1; w++) {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {   U32 s;
+        for (s=0; s<nbSymbols; s++) {
+            U32 const w = weightList[s];
+            U32 const r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {   U32* const rankVal0 = rankVal[0];
+        {   int const rescale = (maxTableLog-tableLog) - 1;   /* tableLog <= maxTableLog */
+            U32 nextRankVal = 0;
+            U32 w;
+            for (w=1; w<maxW+1; w++) {
+                U32 current = nextRankVal;
+                nextRankVal += rankStats[w] << (w+rescale);
+                rankVal0[w] = current;
+        }   }
+        {   U32 const minBits = tableLog+1 - maxW;
+            U32 consumed;
+            for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
+                U32* const rankValPtr = rankVal[consumed];
+                U32 w;
+                for (w = 1; w < maxW+1; w++) {
+                    rankValPtr[w] = rankVal0[w] >> consumed;
+    }   }   }   }
+
+    HUF_fillDTableX4(dt, maxTableLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    dtd.tableLog = (BYTE)maxTableLog;
+    dtd.tableType = 1;
+    memcpy(DTable, &dtd, sizeof(dtd));
+    return iSize;
+}
+
+size_t HUF_readDTableX4(HUF_DTable* DTable, const void* src, size_t srcSize)
+{
+  U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+  return HUF_readDTableX4_wksp(DTable, src, srcSize,
+                               workSpace, sizeof(workSpace));
+}
+
+size_t HUF_decompress1X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    DTableDesc dtd = HUF_getDTableDesc(DTable);
+    if (dtd.tableType != 1) return ERROR(GENERIC);
+    return HUF_decompress1X4_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+size_t HUF_decompress1X4_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUF_readDTableX4_wksp(DCtx, cSrc, cSrcSize,
+                                               workSpace, wkspSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUF_decompress1X4_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
+}
+
+
+size_t HUF_decompress1X4_DCtx(HUF_DTable* DCtx, void* dst, size_t dstSize,
+                              const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress1X4_DCtx_wksp(DCtx, dst, dstSize, cSrc, cSrcSize,
+                                       workSpace, sizeof(workSpace));
+}
+
+size_t HUF_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX4(DTable, HUF_TABLELOG_MAX);
+    return HUF_decompress1X4_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+size_t HUF_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable)
+{
+    DTableDesc dtd = HUF_getDTableDesc(DTable);
+    if (dtd.tableType != 1) return ERROR(GENERIC);
+    return HUF_decompress4X4_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+static size_t HUF_decompress4X4_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize, int bmi2)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUF_readDTableX4_wksp(dctx, cSrc, cSrcSize,
+                                         workSpace, wkspSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUF_decompress4X4_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+}
+
+size_t HUF_decompress4X4_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
+                                   const void* cSrc, size_t cSrcSize,
+                                   void* workSpace, size_t wkspSize)
+{
+    return HUF_decompress4X4_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, /* bmi2 */ 0);
+}
+
+
+size_t HUF_decompress4X4_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize,
+                              const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress4X4_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
+                                       workSpace, sizeof(workSpace));
+}
+
+size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX4(DTable, HUF_TABLELOG_MAX);
+    return HUF_decompress4X4_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+
+/* ********************************/
+/* Generic decompression selector */
+/* ********************************/
+
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize,
+                                    const void* cSrc, size_t cSrcSize,
+                                    const HUF_DTable* DTable)
+{
+    DTableDesc const dtd = HUF_getDTableDesc(DTable);
+    return dtd.tableType ? HUF_decompress1X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
+                           HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize,
+                                    const void* cSrc, size_t cSrcSize,
+                                    const HUF_DTable* DTable)
+{
+    DTableDesc const dtd = HUF_getDTableDesc(DTable);
+    return dtd.tableType ? HUF_decompress4X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
+                           HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+}
+
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+/** HUF_selectDecoder() :
+ *  Tells which decoder is likely to decode faster,
+ *  based on a set of pre-computed metrics.
+ * @return : 0==HUF_decompress4X2, 1==HUF_decompress4X4 .
+ *  Assumption : 0 < dstSize <= 128 KB */
+U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize)
+{
+    assert(dstSize > 0);
+    assert(dstSize <= 128 KB);
+    /* decoder timing evaluation */
+    {   U32 const Q = (cSrcSize >= dstSize) ? 15 : (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 */
+        U32 const D256 = (U32)(dstSize >> 8);
+        U32 const DTime0 = algoTime[Q][0].tableTime + (algoTime[Q][0].decode256Time * D256);
+        U32 DTime1 = algoTime[Q][1].tableTime + (algoTime[Q][1].decode256Time * D256);
+        DTime1 += DTime1 >> 3;  /* advantage to algorithm using less memory, to reduce cache eviction */
+        return DTime1 < DTime0;
+}   }
+
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[2] = { HUF_decompress4X2, HUF_decompress4X4 };
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
+        return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+    }
+}
+
+size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUF_decompress4X4_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) :
+                        HUF_decompress4X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) ;
+    }
+}
+
+size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress4X_hufOnly_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
+                                         workSpace, sizeof(workSpace));
+}
+
+
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst,
+                                     size_t dstSize, const void* cSrc,
+                                     size_t cSrcSize, void* workSpace,
+                                     size_t wkspSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize == 0) return ERROR(corruption_detected);
+
+    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUF_decompress4X4_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize):
+                        HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
+    }
+}
+
+size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
+                                  const void* cSrc, size_t cSrcSize,
+                                  void* workSpace, size_t wkspSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUF_decompress1X4_DCtx_wksp(dctx, dst, dstSize, cSrc,
+                                cSrcSize, workSpace, wkspSize):
+                        HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
+                                cSrcSize, workSpace, wkspSize);
+    }
+}
+
+size_t HUF_decompress1X_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize,
+                             const void* cSrc, size_t cSrcSize)
+{
+    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    return HUF_decompress1X_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
+                                      workSpace, sizeof(workSpace));
+}
+
+
+size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+{
+    DTableDesc const dtd = HUF_getDTableDesc(DTable);
+    return dtd.tableType ? HUF_decompress1X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
+                           HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+}
+
+size_t HUF_decompress1X2_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+}
+
+size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+{
+    DTableDesc const dtd = HUF_getDTableDesc(DTable);
+    return dtd.tableType ? HUF_decompress4X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
+                           HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+}
+
+size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize == 0) return ERROR(corruption_detected);
+
+    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUF_decompress4X4_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2) :
+                        HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/mem.h b/vendor/github.com/DataDog/zstd/mem.h
new file mode 100644
index 0000000..47d2300
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/mem.h
@@ -0,0 +1,362 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-****************************************
+*  Dependencies
+******************************************/
+#include <stddef.h>     /* size_t, ptrdiff_t */
+#include <string.h>     /* memcpy */
+
+
+/*-****************************************
+*  Compiler specifics
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+#if defined(__GNUC__)
+#  define MEM_STATIC static __inline __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+/* code only tested on 32 and 64 bits systems */
+#define MEM_STATIC_ASSERT(c)   { enum { MEM_static_assert = 1/(int)(!!(c)) }; }
+MEM_STATIC void MEM_check(void) { MEM_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
+
+
+/*-**************************************************************
+*  Basic Types
+*****************************************************************/
+#if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef   uint8_t BYTE;
+  typedef  uint16_t U16;
+  typedef   int16_t S16;
+  typedef  uint32_t U32;
+  typedef   int32_t S32;
+  typedef  uint64_t U64;
+  typedef   int64_t S64;
+#else
+  typedef unsigned char      BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/*-**************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (i.e., not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets depending on alignment.
+ *            In some circumstances, it's the only known way to get the most performance (i.e. GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif defined(__INTEL_COMPILER) || defined(__GNUC__)
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(size_t)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard, by lying on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+MEM_STATIC size_t MEM_readST(const void* memPtr) { return *(const size_t*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(U32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+#if defined(_MSC_VER) || (defined(__INTEL_COMPILER) && defined(WIN32))
+    __pragma( pack(push, 1) )
+    typedef struct { U16 v; } unalign16;
+    typedef struct { U32 v; } unalign32;
+    typedef struct { U64 v; } unalign64;
+    typedef struct { size_t v; } unalignArch;
+    __pragma( pack(pop) )
+#else
+    typedef struct { U16 v; } __attribute__((packed)) unalign16;
+    typedef struct { U32 v; } __attribute__((packed)) unalign32;
+    typedef struct { U64 v; } __attribute__((packed)) unalign64;
+    typedef struct { size_t v; } __attribute__((packed)) unalignArch;
+#endif
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign16*)ptr)->v; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign32*)ptr)->v; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign64*)ptr)->v; }
+MEM_STATIC size_t MEM_readST(const void* ptr) { return ((const unalignArch*)ptr)->v; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign16*)memPtr)->v = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign32*)memPtr)->v = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC size_t MEM_readST(const void* memPtr)
+{
+    size_t val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write32(void* memPtr, U32 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write64(void* memPtr, U64 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif /* MEM_FORCE_MEMORY_ACCESS */
+
+MEM_STATIC U32 MEM_swap32(U32 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_ulong(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap32(in);
+#else
+    return  ((in << 24) & 0xff000000 ) |
+            ((in <<  8) & 0x00ff0000 ) |
+            ((in >>  8) & 0x0000ff00 ) |
+            ((in >> 24) & 0x000000ff );
+#endif
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_uint64(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap64(in);
+#else
+    return  ((in << 56) & 0xff00000000000000ULL) |
+            ((in << 40) & 0x00ff000000000000ULL) |
+            ((in << 24) & 0x0000ff0000000000ULL) |
+            ((in << 8)  & 0x000000ff00000000ULL) |
+            ((in >> 8)  & 0x00000000ff000000ULL) |
+            ((in >> 24) & 0x0000000000ff0000ULL) |
+            ((in >> 40) & 0x000000000000ff00ULL) |
+            ((in >> 56) & 0x00000000000000ffULL);
+#endif
+}
+
+MEM_STATIC size_t MEM_swapST(size_t in)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_swap32((U32)in);
+    else
+        return (size_t)MEM_swap64((U64)in);
+}
+
+/*=== Little endian r/w ===*/
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian()) {
+        MEM_write16(memPtr, val);
+    } else {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE24(const void* memPtr)
+{
+    return MEM_readLE16(memPtr) + (((const BYTE*)memPtr)[2] << 16);
+}
+
+MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
+{
+    MEM_writeLE16(memPtr, (U16)val);
+    ((BYTE*)memPtr)[2] = (BYTE)(val>>16);
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+        return MEM_swap32(MEM_read32(memPtr));
+}
+
+MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32)
+{
+    if (MEM_isLittleEndian())
+        MEM_write32(memPtr, val32);
+    else
+        MEM_write32(memPtr, MEM_swap32(val32));
+}
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+        return MEM_swap64(MEM_read64(memPtr));
+}
+
+MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64)
+{
+    if (MEM_isLittleEndian())
+        MEM_write64(memPtr, val64);
+    else
+        MEM_write64(memPtr, MEM_swap64(val64));
+}
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val)
+{
+    if (MEM_32bits())
+        MEM_writeLE32(memPtr, (U32)val);
+    else
+        MEM_writeLE64(memPtr, (U64)val);
+}
+
+/*=== Big endian r/w ===*/
+
+MEM_STATIC U32 MEM_readBE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_swap32(MEM_read32(memPtr));
+    else
+        return MEM_read32(memPtr);
+}
+
+MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32)
+{
+    if (MEM_isLittleEndian())
+        MEM_write32(memPtr, MEM_swap32(val32));
+    else
+        MEM_write32(memPtr, val32);
+}
+
+MEM_STATIC U64 MEM_readBE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_swap64(MEM_read64(memPtr));
+    else
+        return MEM_read64(memPtr);
+}
+
+MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64)
+{
+    if (MEM_isLittleEndian())
+        MEM_write64(memPtr, MEM_swap64(val64));
+    else
+        MEM_write64(memPtr, val64);
+}
+
+MEM_STATIC size_t MEM_readBEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readBE32(memPtr);
+    else
+        return (size_t)MEM_readBE64(memPtr);
+}
+
+MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val)
+{
+    if (MEM_32bits())
+        MEM_writeBE32(memPtr, (U32)val);
+    else
+        MEM_writeBE64(memPtr, (U64)val);
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
diff --git a/vendor/github.com/DataDog/zstd/pool.c b/vendor/github.com/DataDog/zstd/pool.c
new file mode 100644
index 0000000..773488b
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/pool.c
@@ -0,0 +1,283 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/* ======   Dependencies   ======= */
+#include <stddef.h>  /* size_t */
+#include "pool.h"
+#include "zstd_internal.h"  /* ZSTD_malloc, ZSTD_free */
+
+/* ======   Compiler specifics   ====== */
+#if defined(_MSC_VER)
+#  pragma warning(disable : 4204)        /* disable: C4204: non-constant aggregate initializer */
+#endif
+
+
+#ifdef ZSTD_MULTITHREAD
+
+#include "threading.h"   /* pthread adaptation */
+
+/* A job is a function and an opaque argument */
+typedef struct POOL_job_s {
+    POOL_function function;
+    void *opaque;
+} POOL_job;
+
+struct POOL_ctx_s {
+    ZSTD_customMem customMem;
+    /* Keep track of the threads */
+    ZSTD_pthread_t *threads;
+    size_t numThreads;
+
+    /* The queue is a circular buffer */
+    POOL_job *queue;
+    size_t queueHead;
+    size_t queueTail;
+    size_t queueSize;
+
+    /* The number of threads working on jobs */
+    size_t numThreadsBusy;
+    /* Indicates if the queue is empty */
+    int queueEmpty;
+
+    /* The mutex protects the queue */
+    ZSTD_pthread_mutex_t queueMutex;
+    /* Condition variable for pushers to wait on when the queue is full */
+    ZSTD_pthread_cond_t queuePushCond;
+    /* Condition variables for poppers to wait on when the queue is empty */
+    ZSTD_pthread_cond_t queuePopCond;
+    /* Indicates if the queue is shutting down */
+    int shutdown;
+};
+
+/* POOL_thread() :
+   Work thread for the thread pool.
+   Waits for jobs and executes them.
+   @returns : NULL on failure else non-null.
+*/
+static void* POOL_thread(void* opaque) {
+    POOL_ctx* const ctx = (POOL_ctx*)opaque;
+    if (!ctx) { return NULL; }
+    for (;;) {
+        /* Lock the mutex and wait for a non-empty queue or until shutdown */
+        ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+
+        while (ctx->queueEmpty && !ctx->shutdown) {
+            ZSTD_pthread_cond_wait(&ctx->queuePopCond, &ctx->queueMutex);
+        }
+        /* empty => shutting down: so stop */
+        if (ctx->queueEmpty) {
+            ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+            return opaque;
+        }
+        /* Pop a job off the queue */
+        {   POOL_job const job = ctx->queue[ctx->queueHead];
+            ctx->queueHead = (ctx->queueHead + 1) % ctx->queueSize;
+            ctx->numThreadsBusy++;
+            ctx->queueEmpty = ctx->queueHead == ctx->queueTail;
+            /* Unlock the mutex, signal a pusher, and run the job */
+            ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+            ZSTD_pthread_cond_signal(&ctx->queuePushCond);
+
+            job.function(job.opaque);
+
+            /* If the intended queue size was 0, signal after finishing job */
+            if (ctx->queueSize == 1) {
+                ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+                ctx->numThreadsBusy--;
+                ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+                ZSTD_pthread_cond_signal(&ctx->queuePushCond);
+        }   }
+    }  /* for (;;) */
+    /* Unreachable */
+}
+
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
+    return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
+}
+
+POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem) {
+    POOL_ctx* ctx;
+    /* Check the parameters */
+    if (!numThreads) { return NULL; }
+    /* Allocate the context and zero initialize */
+    ctx = (POOL_ctx*)ZSTD_calloc(sizeof(POOL_ctx), customMem);
+    if (!ctx) { return NULL; }
+    /* Initialize the job queue.
+     * It needs one extra space since one space is wasted to differentiate empty
+     * and full queues.
+     */
+    ctx->queueSize = queueSize + 1;
+    ctx->queue = (POOL_job*)ZSTD_malloc(ctx->queueSize * sizeof(POOL_job), customMem);
+    ctx->queueHead = 0;
+    ctx->queueTail = 0;
+    ctx->numThreadsBusy = 0;
+    ctx->queueEmpty = 1;
+    (void)ZSTD_pthread_mutex_init(&ctx->queueMutex, NULL);
+    (void)ZSTD_pthread_cond_init(&ctx->queuePushCond, NULL);
+    (void)ZSTD_pthread_cond_init(&ctx->queuePopCond, NULL);
+    ctx->shutdown = 0;
+    /* Allocate space for the thread handles */
+    ctx->threads = (ZSTD_pthread_t*)ZSTD_malloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
+    ctx->numThreads = 0;
+    ctx->customMem = customMem;
+    /* Check for errors */
+    if (!ctx->threads || !ctx->queue) { POOL_free(ctx); return NULL; }
+    /* Initialize the threads */
+    {   size_t i;
+        for (i = 0; i < numThreads; ++i) {
+            if (ZSTD_pthread_create(&ctx->threads[i], NULL, &POOL_thread, ctx)) {
+                ctx->numThreads = i;
+                POOL_free(ctx);
+                return NULL;
+        }   }
+        ctx->numThreads = numThreads;
+    }
+    return ctx;
+}
+
+/*! POOL_join() :
+    Shutdown the queue, wake any sleeping threads, and join all of the threads.
+*/
+static void POOL_join(POOL_ctx* ctx) {
+    /* Shut down the queue */
+    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+    ctx->shutdown = 1;
+    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+    /* Wake up sleeping threads */
+    ZSTD_pthread_cond_broadcast(&ctx->queuePushCond);
+    ZSTD_pthread_cond_broadcast(&ctx->queuePopCond);
+    /* Join all of the threads */
+    {   size_t i;
+        for (i = 0; i < ctx->numThreads; ++i) {
+            ZSTD_pthread_join(ctx->threads[i], NULL);
+    }   }
+}
+
+void POOL_free(POOL_ctx *ctx) {
+    if (!ctx) { return; }
+    POOL_join(ctx);
+    ZSTD_pthread_mutex_destroy(&ctx->queueMutex);
+    ZSTD_pthread_cond_destroy(&ctx->queuePushCond);
+    ZSTD_pthread_cond_destroy(&ctx->queuePopCond);
+    ZSTD_free(ctx->queue, ctx->customMem);
+    ZSTD_free(ctx->threads, ctx->customMem);
+    ZSTD_free(ctx, ctx->customMem);
+}
+
+size_t POOL_sizeof(POOL_ctx *ctx) {
+    if (ctx==NULL) return 0;  /* supports sizeof NULL */
+    return sizeof(*ctx)
+        + ctx->queueSize * sizeof(POOL_job)
+        + ctx->numThreads * sizeof(ZSTD_pthread_t);
+}
+
+/**
+ * Returns 1 if the queue is full and 0 otherwise.
+ *
+ * If the queueSize is 1 (the pool was created with an intended queueSize of 0),
+ * then a queue is empty if there is a thread free and no job is waiting.
+ */
+static int isQueueFull(POOL_ctx const* ctx) {
+    if (ctx->queueSize > 1) {
+        return ctx->queueHead == ((ctx->queueTail + 1) % ctx->queueSize);
+    } else {
+        return ctx->numThreadsBusy == ctx->numThreads ||
+               !ctx->queueEmpty;
+    }
+}
+
+
+static void POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
+{
+    POOL_job const job = {function, opaque};
+    assert(ctx != NULL);
+    if (ctx->shutdown) return;
+
+    ctx->queueEmpty = 0;
+    ctx->queue[ctx->queueTail] = job;
+    ctx->queueTail = (ctx->queueTail + 1) % ctx->queueSize;
+    ZSTD_pthread_cond_signal(&ctx->queuePopCond);
+}
+
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque)
+{
+    assert(ctx != NULL);
+    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+    /* Wait until there is space in the queue for the new job */
+    while (isQueueFull(ctx) && (!ctx->shutdown)) {
+        ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
+    }
+    POOL_add_internal(ctx, function, opaque);
+    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+}
+
+
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque)
+{
+    assert(ctx != NULL);
+    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+    if (isQueueFull(ctx)) {
+        ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+        return 0;
+    }
+    POOL_add_internal(ctx, function, opaque);
+    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+    return 1;
+}
+
+
+#else  /* ZSTD_MULTITHREAD  not defined */
+
+/* ========================== */
+/* No multi-threading support */
+/* ========================== */
+
+
+/* We don't need any data, but if it is empty, malloc() might return NULL. */
+struct POOL_ctx_s {
+    int dummy;
+};
+static POOL_ctx g_ctx;
+
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
+    return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
+}
+
+POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem) {
+    (void)numThreads;
+    (void)queueSize;
+    (void)customMem;
+    return &g_ctx;
+}
+
+void POOL_free(POOL_ctx* ctx) {
+    assert(!ctx || ctx == &g_ctx);
+    (void)ctx;
+}
+
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque) {
+    (void)ctx;
+    function(opaque);
+}
+
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque) {
+    (void)ctx;
+    function(opaque);
+    return 1;
+}
+
+size_t POOL_sizeof(POOL_ctx* ctx) {
+    if (ctx==NULL) return 0;  /* supports sizeof NULL */
+    assert(ctx == &g_ctx);
+    return sizeof(*ctx);
+}
+
+#endif  /* ZSTD_MULTITHREAD */
diff --git a/vendor/github.com/DataDog/zstd/pool.h b/vendor/github.com/DataDog/zstd/pool.h
new file mode 100644
index 0000000..a57e9b4
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/pool.h
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef POOL_H
+#define POOL_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+#include <stddef.h>   /* size_t */
+#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_customMem */
+#include "zstd.h"
+
+typedef struct POOL_ctx_s POOL_ctx;
+
+/*! POOL_create() :
+ *  Create a thread pool with at most `numThreads` threads.
+ * `numThreads` must be at least 1.
+ *  The maximum number of queued jobs before blocking is `queueSize`.
+ * @return : POOL_ctx pointer on success, else NULL.
+*/
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize);
+
+POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem);
+
+/*! POOL_free() :
+    Free a thread pool returned by POOL_create().
+*/
+void POOL_free(POOL_ctx* ctx);
+
+/*! POOL_sizeof() :
+    return memory usage of pool returned by POOL_create().
+*/
+size_t POOL_sizeof(POOL_ctx* ctx);
+
+/*! POOL_function :
+    The function type that can be added to a thread pool.
+*/
+typedef void (*POOL_function)(void*);
+/*! POOL_add_function :
+    The function type for a generic thread pool add function.
+*/
+typedef void (*POOL_add_function)(void*, POOL_function, void*);
+
+/*! POOL_add() :
+    Add the job `function(opaque)` to the thread pool. `ctx` must be valid.
+    Possibly blocks until there is room in the queue.
+    Note : The function may be executed asynchronously, so `opaque` must live until the function has been completed.
+*/
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque);
+
+
+/*! POOL_tryAdd() :
+    Add the job `function(opaque)` to the thread pool if a worker is available.
+    return immediately otherwise.
+   @return : 1 if successful, 0 if not.
+*/
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif
diff --git a/vendor/github.com/DataDog/zstd/threading.c b/vendor/github.com/DataDog/zstd/threading.c
new file mode 100644
index 0000000..8be8c8d
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/threading.c
@@ -0,0 +1,75 @@
+/**
+ * Copyright (c) 2016 Tino Reichardt
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ *
+ * You can contact the author at:
+ * - zstdmt source repository: https://github.com/mcmilk/zstdmt
+ */
+
+/**
+ * This file will hold wrapper for systems, which do not support pthreads
+ */
+
+/* create fake symbol to avoid empty trnaslation unit warning */
+int g_ZSTD_threading_useles_symbol;
+
+#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
+
+/**
+ * Windows minimalist Pthread Wrapper, based on :
+ * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
+ */
+
+
+/* ===  Dependencies  === */
+#include <process.h>
+#include <errno.h>
+#include "threading.h"
+
+
+/* ===  Implementation  === */
+
+static unsigned __stdcall worker(void *arg)
+{
+    ZSTD_pthread_t* const thread = (ZSTD_pthread_t*) arg;
+    thread->arg = thread->start_routine(thread->arg);
+    return 0;
+}
+
+int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
+            void* (*start_routine) (void*), void* arg)
+{
+    (void)unused;
+    thread->arg = arg;
+    thread->start_routine = start_routine;
+    thread->handle = (HANDLE) _beginthreadex(NULL, 0, worker, thread, 0, NULL);
+
+    if (!thread->handle)
+        return errno;
+    else
+        return 0;
+}
+
+int ZSTD_pthread_join(ZSTD_pthread_t thread, void **value_ptr)
+{
+    DWORD result;
+
+    if (!thread.handle) return 0;
+
+    result = WaitForSingleObject(thread.handle, INFINITE);
+    switch (result) {
+    case WAIT_OBJECT_0:
+        if (value_ptr) *value_ptr = thread.arg;
+        return 0;
+    case WAIT_ABANDONED:
+        return EINVAL;
+    default:
+        return GetLastError();
+    }
+}
+
+#endif   /* ZSTD_MULTITHREAD */
diff --git a/vendor/github.com/DataDog/zstd/threading.h b/vendor/github.com/DataDog/zstd/threading.h
new file mode 100644
index 0000000..d806c89
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/threading.h
@@ -0,0 +1,123 @@
+/**
+ * Copyright (c) 2016 Tino Reichardt
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ *
+ * You can contact the author at:
+ * - zstdmt source repository: https://github.com/mcmilk/zstdmt
+ */
+
+#ifndef THREADING_H_938743
+#define THREADING_H_938743
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
+
+/**
+ * Windows minimalist Pthread Wrapper, based on :
+ * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
+ */
+#ifdef WINVER
+#  undef WINVER
+#endif
+#define WINVER       0x0600
+
+#ifdef _WIN32_WINNT
+#  undef _WIN32_WINNT
+#endif
+#define _WIN32_WINNT 0x0600
+
+#ifndef WIN32_LEAN_AND_MEAN
+#  define WIN32_LEAN_AND_MEAN
+#endif
+
+#undef ERROR   /* reported already defined on VS 2015 (Rich Geldreich) */
+#include <windows.h>
+#undef ERROR
+#define ERROR(name) ZSTD_ERROR(name)
+
+
+/* mutex */
+#define ZSTD_pthread_mutex_t           CRITICAL_SECTION
+#define ZSTD_pthread_mutex_init(a, b)  ((void)(b), InitializeCriticalSection((a)), 0)
+#define ZSTD_pthread_mutex_destroy(a)  DeleteCriticalSection((a))
+#define ZSTD_pthread_mutex_lock(a)     EnterCriticalSection((a))
+#define ZSTD_pthread_mutex_unlock(a)   LeaveCriticalSection((a))
+
+/* condition variable */
+#define ZSTD_pthread_cond_t             CONDITION_VARIABLE
+#define ZSTD_pthread_cond_init(a, b)    ((void)(b), InitializeConditionVariable((a)), 0)
+#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
+#define ZSTD_pthread_cond_wait(a, b)    SleepConditionVariableCS((a), (b), INFINITE)
+#define ZSTD_pthread_cond_signal(a)     WakeConditionVariable((a))
+#define ZSTD_pthread_cond_broadcast(a)  WakeAllConditionVariable((a))
+
+/* ZSTD_pthread_create() and ZSTD_pthread_join() */
+typedef struct {
+    HANDLE handle;
+    void* (*start_routine)(void*);
+    void* arg;
+} ZSTD_pthread_t;
+
+int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
+                   void* (*start_routine) (void*), void* arg);
+
+int ZSTD_pthread_join(ZSTD_pthread_t thread, void** value_ptr);
+
+/**
+ * add here more wrappers as required
+ */
+
+
+#elif defined(ZSTD_MULTITHREAD)   /* posix assumed ; need a better detection method */
+/* ===   POSIX Systems   === */
+#  include <pthread.h>
+
+#define ZSTD_pthread_mutex_t            pthread_mutex_t
+#define ZSTD_pthread_mutex_init(a, b)   pthread_mutex_init((a), (b))
+#define ZSTD_pthread_mutex_destroy(a)   pthread_mutex_destroy((a))
+#define ZSTD_pthread_mutex_lock(a)      pthread_mutex_lock((a))
+#define ZSTD_pthread_mutex_unlock(a)    pthread_mutex_unlock((a))
+
+#define ZSTD_pthread_cond_t             pthread_cond_t
+#define ZSTD_pthread_cond_init(a, b)    pthread_cond_init((a), (b))
+#define ZSTD_pthread_cond_destroy(a)    pthread_cond_destroy((a))
+#define ZSTD_pthread_cond_wait(a, b)    pthread_cond_wait((a), (b))
+#define ZSTD_pthread_cond_signal(a)     pthread_cond_signal((a))
+#define ZSTD_pthread_cond_broadcast(a)  pthread_cond_broadcast((a))
+
+#define ZSTD_pthread_t                  pthread_t
+#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
+#define ZSTD_pthread_join(a, b)         pthread_join((a),(b))
+
+#else  /* ZSTD_MULTITHREAD not defined */
+/* No multithreading support */
+
+typedef int ZSTD_pthread_mutex_t;
+#define ZSTD_pthread_mutex_init(a, b)   ((void)(a), (void)(b), 0)
+#define ZSTD_pthread_mutex_destroy(a)   ((void)(a))
+#define ZSTD_pthread_mutex_lock(a)      ((void)(a))
+#define ZSTD_pthread_mutex_unlock(a)    ((void)(a))
+
+typedef int ZSTD_pthread_cond_t;
+#define ZSTD_pthread_cond_init(a, b)    ((void)(a), (void)(b), 0)
+#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
+#define ZSTD_pthread_cond_wait(a, b)    ((void)(a), (void)(b))
+#define ZSTD_pthread_cond_signal(a)     ((void)(a))
+#define ZSTD_pthread_cond_broadcast(a)  ((void)(a))
+
+/* do not use ZSTD_pthread_t */
+
+#endif /* ZSTD_MULTITHREAD */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* THREADING_H_938743 */
diff --git a/vendor/github.com/DataDog/zstd/travis_test_32.sh b/vendor/github.com/DataDog/zstd/travis_test_32.sh
new file mode 100755
index 0000000..d29c86c
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/travis_test_32.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+# Get utilities
+yum -y -q -e 0 install wget tar unzip gcc
+
+# Get Go
+wget -q https://dl.google.com/go/go1.11.1.linux-386.tar.gz
+tar -C /usr/local -xzf go1.11.1.linux-386.tar.gz
+export PATH=$PATH:/usr/local/go/bin
+
+# Get payload
+wget -q https://github.com/DataDog/zstd/files/2246767/mr.zip
+unzip mr.zip
+
+# Build and run tests
+cd zstd
+go build
+PAYLOAD=$(pwd)/mr go test -v
+PAYLOAD=$(pwd)/mr go test -bench .
diff --git a/vendor/github.com/DataDog/zstd/update.txt b/vendor/github.com/DataDog/zstd/update.txt
new file mode 100644
index 0000000..1de939f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/update.txt
@@ -0,0 +1,56 @@
+./lib/common/bitstream.h
+./lib/common/compiler.h
+./lib/compress/zstd_compress_internal.h
+./lib/compress/zstd_fast.h
+./lib/compress/zstd_double_fast.h
+./lib/compress/zstd_lazy.h
+./lib/compress/zstd_ldm.h
+./lib/dictBuilder/cover.c
+./lib/dictBuilder/divsufsort.c
+./lib/dictBuilder/divsufsort.h
+./lib/common/entropy_common.c
+./lib/common/error_private.c
+./lib/common/error_private.h
+./lib/compress/fse_compress.c
+./lib/common/fse_decompress.c
+./lib/common/fse.h
+./lib/compress/huf_compress.c
+./lib/decompress/huf_decompress.c
+./lib/common/huf.h
+./lib/common/mem.h
+./lib/common/pool.c
+./lib/common/pool.h
+./lib/common/threading.c
+./lib/common/threading.h
+./lib/common/xxhash.c
+./lib/common/xxhash.h
+./lib/deprecated/zbuff_common.c
+./lib/deprecated/zbuff_compress.c
+./lib/deprecated/zbuff_decompress.c
+./lib/deprecated/zbuff.h
+./lib/dictBuilder/zdict.c
+./lib/dictBuilder/zdict.h
+./lib/common/zstd_common.c
+./lib/compress/zstd_compress.c
+./lib/decompress/zstd_decompress.c
+./lib/common/zstd_errors.h
+./lib/zstd.h
+./lib/common/zstd_internal.h
+./lib/legacy/zstd_legacy.h
+./lib/compress/zstd_opt.c
+./lib/compress/zstd_opt.h
+./lib/legacy/zstd_v01.c
+./lib/legacy/zstd_v01.h
+./lib/legacy/zstd_v02.c
+./lib/legacy/zstd_v02.h
+./lib/legacy/zstd_v03.c
+./lib/legacy/zstd_v03.h
+./lib/legacy/zstd_v04.c
+./lib/legacy/zstd_v04.h
+./lib/legacy/zstd_v05.c
+./lib/legacy/zstd_v05.h
+./lib/legacy/zstd_v06.c
+./lib/legacy/zstd_v06.h
+./lib/legacy/zstd_v07.c
+./lib/legacy/zstd_v07.h
+
diff --git a/vendor/github.com/DataDog/zstd/xxhash.c b/vendor/github.com/DataDog/zstd/xxhash.c
new file mode 100644
index 0000000..9d9c0e9
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/xxhash.c
@@ -0,0 +1,875 @@
+/*
+*  xxHash - Fast Hash algorithm
+*  Copyright (C) 2012-2016, Yann Collet
+*
+*  BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+*
+*  Redistribution and use in source and binary forms, with or without
+*  modification, are permitted provided that the following conditions are
+*  met:
+*
+*  * Redistributions of source code must retain the above copyright
+*  notice, this list of conditions and the following disclaimer.
+*  * Redistributions in binary form must reproduce the above
+*  copyright notice, this list of conditions and the following disclaimer
+*  in the documentation and/or other materials provided with the
+*  distribution.
+*
+*  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+*  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+*  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+*  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+*  OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+*  SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+*  LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+*  DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+*  THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+*  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+*  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*
+*  You can contact the author at :
+*  - xxHash homepage: http://www.xxhash.com
+*  - xxHash source repository : https://github.com/Cyan4973/xxHash
+*/
+
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+/*!XXH_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method doesn't depend on compiler but violate C standard.
+ *            It can generate buggy code on targets which do not support unaligned memory accesses.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://stackoverflow.com/a/32095106/646947 for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define XXH_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+/*!XXH_ACCEPT_NULL_INPUT_POINTER :
+ * If the input pointer is a null pointer, xxHash default behavior is to trigger a memory access error, since it is a bad pointer.
+ * When this option is enabled, xxHash output for null input pointers will be the same as a null-length input.
+ * By default, this option is disabled. To enable it, uncomment below define :
+ */
+/* #define XXH_ACCEPT_NULL_INPUT_POINTER 1 */
+
+/*!XXH_FORCE_NATIVE_FORMAT :
+ * By default, xxHash library provides endian-independant Hash values, based on little-endian convention.
+ * Results are therefore identical for little-endian and big-endian CPU.
+ * This comes at a performance cost for big-endian CPU, since some swapping is required to emulate little-endian format.
+ * Should endian-independance be of no importance for your application, you may set the #define below to 1,
+ * to improve speed for Big-endian CPU.
+ * This option has no impact on Little_Endian CPU.
+ */
+#ifndef XXH_FORCE_NATIVE_FORMAT   /* can be defined externally */
+#  define XXH_FORCE_NATIVE_FORMAT 0
+#endif
+
+/*!XXH_FORCE_ALIGN_CHECK :
+ * This is a minor performance trick, only useful with lots of very small keys.
+ * It means : check for aligned/unaligned input.
+ * The check costs one initial branch per hash; set to 0 when the input data
+ * is guaranteed to be aligned.
+ */
+#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
+#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+/* Modify the local functions below should you wish to use some other memory routines */
+/* for malloc(), free() */
+#include <stdlib.h>
+static void* XXH_malloc(size_t s) { return malloc(s); }
+static void  XXH_free  (void* p)  { free(p); }
+/* for memcpy() */
+#include <string.h>
+static void* XXH_memcpy(void* dest, const void* src, size_t size) { return memcpy(dest,src,size); }
+
+#ifndef XXH_STATIC_LINKING_ONLY
+#  define XXH_STATIC_LINKING_ONLY
+#endif
+#include "xxhash.h"
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#if defined (__GNUC__) || defined(__cplusplus) || defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#  define INLINE_KEYWORD inline
+#else
+#  define INLINE_KEYWORD
+#endif
+
+#if defined(__GNUC__)
+#  define FORCE_INLINE_ATTR __attribute__((always_inline))
+#elif defined(_MSC_VER)
+#  define FORCE_INLINE_ATTR __forceinline
+#else
+#  define FORCE_INLINE_ATTR
+#endif
+
+#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR
+
+
+#ifdef _MSC_VER
+#  pragma warning(disable : 4127)      /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/* *************************************
+*  Basic Types
+***************************************/
+#ifndef MEM_MODULE
+# define MEM_MODULE
+# if !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint8_t  BYTE;
+    typedef uint16_t U16;
+    typedef uint32_t U32;
+    typedef  int32_t S32;
+    typedef uint64_t U64;
+#  else
+    typedef unsigned char      BYTE;
+    typedef unsigned short     U16;
+    typedef unsigned int       U32;
+    typedef   signed int       S32;
+    typedef unsigned long long U64;   /* if your compiler doesn't support unsigned long long, replace by another 64-bit type here. Note that xxhash.h will also need to be updated. */
+#  endif
+#endif
+
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static U32 XXH_read32(const void* memPtr) { return *(const U32*) memPtr; }
+static U64 XXH_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U32 u32; U64 u64; } __attribute__((packed)) unalign;
+
+static U32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+static U64 XXH_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+#else
+
+/* portable and safe solution. Generally efficient.
+ * see : http://stackoverflow.com/a/32095106/646947
+ */
+
+static U32 XXH_read32(const void* memPtr)
+{
+    U32 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+static U64 XXH_read64(const void* memPtr)
+{
+    U64 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+/* Note : although _rotl exists for minGW (GCC under windows), performance seems poor */
+#if defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) ((x << r) | (x >> (32 - r)))
+#  define XXH_rotl64(x,r) ((x << r) | (x >> (64 - r)))
+#endif
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#  define XXH_swap64 _byteswap_uint64
+#elif GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#  define XXH_swap64 __builtin_bswap64
+#else
+static U32 XXH_swap32 (U32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+static U64 XXH_swap64 (U64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* *************************************
+*  Architecture Macros
+***************************************/
+typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
+
+/* XXH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+    static const int g_one = 1;
+#   define XXH_CPU_LITTLE_ENDIAN   (*(const char*)(&g_one))
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment;
+
+FORCE_INLINE_TEMPLATE U32 XXH_readLE32_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return endian==XXH_littleEndian ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+    else
+        return endian==XXH_littleEndian ? *(const U32*)ptr : XXH_swap32(*(const U32*)ptr);
+}
+
+FORCE_INLINE_TEMPLATE U32 XXH_readLE32(const void* ptr, XXH_endianess endian)
+{
+    return XXH_readLE32_align(ptr, endian, XXH_unaligned);
+}
+
+static U32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+
+FORCE_INLINE_TEMPLATE U64 XXH_readLE64_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return endian==XXH_littleEndian ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+    else
+        return endian==XXH_littleEndian ? *(const U64*)ptr : XXH_swap64(*(const U64*)ptr);
+}
+
+FORCE_INLINE_TEMPLATE U64 XXH_readLE64(const void* ptr, XXH_endianess endian)
+{
+    return XXH_readLE64_align(ptr, endian, XXH_unaligned);
+}
+
+static U64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+
+
+/* *************************************
+*  Macros
+***************************************/
+#define XXH_STATIC_ASSERT(c)   { enum { XXH_static_assert = 1/(int)(!!(c)) }; }    /* use only *after* variable declarations */
+
+
+/* *************************************
+*  Constants
+***************************************/
+static const U32 PRIME32_1 = 2654435761U;
+static const U32 PRIME32_2 = 2246822519U;
+static const U32 PRIME32_3 = 3266489917U;
+static const U32 PRIME32_4 =  668265263U;
+static const U32 PRIME32_5 =  374761393U;
+
+static const U64 PRIME64_1 = 11400714785074694791ULL;
+static const U64 PRIME64_2 = 14029467366897019727ULL;
+static const U64 PRIME64_3 =  1609587929392839161ULL;
+static const U64 PRIME64_4 =  9650029242287828579ULL;
+static const U64 PRIME64_5 =  2870177450012600261ULL;
+
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* **************************
+*  Utils
+****************************/
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* restrict dstState, const XXH32_state_t* restrict srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* restrict dstState, const XXH64_state_t* restrict srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+
+/* ***************************
+*  Simple Hash Functions
+*****************************/
+
+static U32 XXH32_round(U32 seed, U32 input)
+{
+    seed += input * PRIME32_2;
+    seed  = XXH_rotl32(seed, 13);
+    seed *= PRIME32_1;
+    return seed;
+}
+
+FORCE_INLINE_TEMPLATE U32 XXH32_endian_align(const void* input, size_t len, U32 seed, XXH_endianess endian, XXH_alignment align)
+{
+    const BYTE* p = (const BYTE*)input;
+    const BYTE* bEnd = p + len;
+    U32 h32;
+#define XXH_get32bits(p) XXH_readLE32_align(p, endian, align)
+
+#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
+    if (p==NULL) {
+        len=0;
+        bEnd=p=(const BYTE*)(size_t)16;
+    }
+#endif
+
+    if (len>=16) {
+        const BYTE* const limit = bEnd - 16;
+        U32 v1 = seed + PRIME32_1 + PRIME32_2;
+        U32 v2 = seed + PRIME32_2;
+        U32 v3 = seed + 0;
+        U32 v4 = seed - PRIME32_1;
+
+        do {
+            v1 = XXH32_round(v1, XXH_get32bits(p)); p+=4;
+            v2 = XXH32_round(v2, XXH_get32bits(p)); p+=4;
+            v3 = XXH32_round(v3, XXH_get32bits(p)); p+=4;
+            v4 = XXH32_round(v4, XXH_get32bits(p)); p+=4;
+        } while (p<=limit);
+
+        h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+    } else {
+        h32  = seed + PRIME32_5;
+    }
+
+    h32 += (U32) len;
+
+    while (p+4<=bEnd) {
+        h32 += XXH_get32bits(p) * PRIME32_3;
+        h32  = XXH_rotl32(h32, 17) * PRIME32_4 ;
+        p+=4;
+    }
+
+    while (p<bEnd) {
+        h32 += (*p) * PRIME32_5;
+        h32 = XXH_rotl32(h32, 11) * PRIME32_1 ;
+        p++;
+    }
+
+    h32 ^= h32 >> 15;
+    h32 *= PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= PRIME32_3;
+    h32 ^= h32 >> 16;
+
+    return h32;
+}
+
+
+XXH_PUBLIC_API unsigned int XXH32 (const void* input, size_t len, unsigned int seed)
+{
+#if 0
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH32_CREATESTATE_STATIC(state);
+    XXH32_reset(state, seed);
+    XXH32_update(state, input, len);
+    return XXH32_digest(state);
+#else
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+                return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
+            else
+                return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
+    }   }
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
+    else
+        return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
+#endif
+}
+
+
+static U64 XXH64_round(U64 acc, U64 input)
+{
+    acc += input * PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= PRIME64_1;
+    return acc;
+}
+
+static U64 XXH64_mergeRound(U64 acc, U64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * PRIME64_1 + PRIME64_4;
+    return acc;
+}
+
+FORCE_INLINE_TEMPLATE U64 XXH64_endian_align(const void* input, size_t len, U64 seed, XXH_endianess endian, XXH_alignment align)
+{
+    const BYTE* p = (const BYTE*)input;
+    const BYTE* const bEnd = p + len;
+    U64 h64;
+#define XXH_get64bits(p) XXH_readLE64_align(p, endian, align)
+
+#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
+    if (p==NULL) {
+        len=0;
+        bEnd=p=(const BYTE*)(size_t)32;
+    }
+#endif
+
+    if (len>=32) {
+        const BYTE* const limit = bEnd - 32;
+        U64 v1 = seed + PRIME64_1 + PRIME64_2;
+        U64 v2 = seed + PRIME64_2;
+        U64 v3 = seed + 0;
+        U64 v4 = seed - PRIME64_1;
+
+        do {
+            v1 = XXH64_round(v1, XXH_get64bits(p)); p+=8;
+            v2 = XXH64_round(v2, XXH_get64bits(p)); p+=8;
+            v3 = XXH64_round(v3, XXH_get64bits(p)); p+=8;
+            v4 = XXH64_round(v4, XXH_get64bits(p)); p+=8;
+        } while (p<=limit);
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+
+    } else {
+        h64  = seed + PRIME64_5;
+    }
+
+    h64 += (U64) len;
+
+    while (p+8<=bEnd) {
+        U64 const k1 = XXH64_round(0, XXH_get64bits(p));
+        h64 ^= k1;
+        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
+        p+=8;
+    }
+
+    if (p+4<=bEnd) {
+        h64 ^= (U64)(XXH_get32bits(p)) * PRIME64_1;
+        h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
+        p+=4;
+    }
+
+    while (p<bEnd) {
+        h64 ^= (*p) * PRIME64_5;
+        h64 = XXH_rotl64(h64, 11) * PRIME64_1;
+        p++;
+    }
+
+    h64 ^= h64 >> 33;
+    h64 *= PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= PRIME64_3;
+    h64 ^= h64 >> 32;
+
+    return h64;
+}
+
+
+XXH_PUBLIC_API unsigned long long XXH64 (const void* input, size_t len, unsigned long long seed)
+{
+#if 0
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_CREATESTATE_STATIC(state);
+    XXH64_reset(state, seed);
+    XXH64_update(state, input, len);
+    return XXH64_digest(state);
+#else
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+                return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
+            else
+                return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
+    }   }
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
+    else
+        return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
+#endif
+}
+
+
+/* **************************************************
+*  Advanced Hash Functions
+****************************************************/
+
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+
+/*** Hash feed ***/
+
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, unsigned int seed)
+{
+    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state)-4);   /* do not write into reserved, for future removal */
+    state.v1 = seed + PRIME32_1 + PRIME32_2;
+    state.v2 = seed + PRIME32_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - PRIME32_1;
+    memcpy(statePtr, &state, sizeof(state));
+    return XXH_OK;
+}
+
+
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, unsigned long long seed)
+{
+    XXH64_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state)-8);   /* do not write into reserved, for future removal */
+    state.v1 = seed + PRIME64_1 + PRIME64_2;
+    state.v2 = seed + PRIME64_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - PRIME64_1;
+    memcpy(statePtr, &state, sizeof(state));
+    return XXH_OK;
+}
+
+
+FORCE_INLINE_TEMPLATE XXH_errorcode XXH32_update_endian (XXH32_state_t* state, const void* input, size_t len, XXH_endianess endian)
+{
+    const BYTE* p = (const BYTE*)input;
+    const BYTE* const bEnd = p + len;
+
+#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
+    if (input==NULL) return XXH_ERROR;
+#endif
+
+    state->total_len_32 += (unsigned)len;
+    state->large_len |= (len>=16) | (state->total_len_32>=16);
+
+    if (state->memsize + len < 16)  {   /* fill in tmp buffer */
+        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, len);
+        state->memsize += (unsigned)len;
+        return XXH_OK;
+    }
+
+    if (state->memsize) {   /* some data left from previous update */
+        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, 16-state->memsize);
+        {   const U32* p32 = state->mem32;
+            state->v1 = XXH32_round(state->v1, XXH_readLE32(p32, endian)); p32++;
+            state->v2 = XXH32_round(state->v2, XXH_readLE32(p32, endian)); p32++;
+            state->v3 = XXH32_round(state->v3, XXH_readLE32(p32, endian)); p32++;
+            state->v4 = XXH32_round(state->v4, XXH_readLE32(p32, endian)); p32++;
+        }
+        p += 16-state->memsize;
+        state->memsize = 0;
+    }
+
+    if (p <= bEnd-16) {
+        const BYTE* const limit = bEnd - 16;
+        U32 v1 = state->v1;
+        U32 v2 = state->v2;
+        U32 v3 = state->v3;
+        U32 v4 = state->v4;
+
+        do {
+            v1 = XXH32_round(v1, XXH_readLE32(p, endian)); p+=4;
+            v2 = XXH32_round(v2, XXH_readLE32(p, endian)); p+=4;
+            v3 = XXH32_round(v3, XXH_readLE32(p, endian)); p+=4;
+            v4 = XXH32_round(v4, XXH_readLE32(p, endian)); p+=4;
+        } while (p<=limit);
+
+        state->v1 = v1;
+        state->v2 = v2;
+        state->v3 = v3;
+        state->v4 = v4;
+    }
+
+    if (p < bEnd) {
+        XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+        state->memsize = (unsigned)(bEnd-p);
+    }
+
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* state_in, const void* input, size_t len)
+{
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH32_update_endian(state_in, input, len, XXH_littleEndian);
+    else
+        return XXH32_update_endian(state_in, input, len, XXH_bigEndian);
+}
+
+
+
+FORCE_INLINE_TEMPLATE U32 XXH32_digest_endian (const XXH32_state_t* state, XXH_endianess endian)
+{
+    const BYTE * p = (const BYTE*)state->mem32;
+    const BYTE* const bEnd = (const BYTE*)(state->mem32) + state->memsize;
+    U32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v1, 1) + XXH_rotl32(state->v2, 7) + XXH_rotl32(state->v3, 12) + XXH_rotl32(state->v4, 18);
+    } else {
+        h32 = state->v3 /* == seed */ + PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    while (p+4<=bEnd) {
+        h32 += XXH_readLE32(p, endian) * PRIME32_3;
+        h32  = XXH_rotl32(h32, 17) * PRIME32_4;
+        p+=4;
+    }
+
+    while (p<bEnd) {
+        h32 += (*p) * PRIME32_5;
+        h32  = XXH_rotl32(h32, 11) * PRIME32_1;
+        p++;
+    }
+
+    h32 ^= h32 >> 15;
+    h32 *= PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= PRIME32_3;
+    h32 ^= h32 >> 16;
+
+    return h32;
+}
+
+
+XXH_PUBLIC_API unsigned int XXH32_digest (const XXH32_state_t* state_in)
+{
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH32_digest_endian(state_in, XXH_littleEndian);
+    else
+        return XXH32_digest_endian(state_in, XXH_bigEndian);
+}
+
+
+
+/* **** XXH64 **** */
+
+FORCE_INLINE_TEMPLATE XXH_errorcode XXH64_update_endian (XXH64_state_t* state, const void* input, size_t len, XXH_endianess endian)
+{
+    const BYTE* p = (const BYTE*)input;
+    const BYTE* const bEnd = p + len;
+
+#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
+    if (input==NULL) return XXH_ERROR;
+#endif
+
+    state->total_len += len;
+
+    if (state->memsize + len < 32) {  /* fill in tmp buffer */
+        XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, len);
+        state->memsize += (U32)len;
+        return XXH_OK;
+    }
+
+    if (state->memsize) {   /* tmp buffer is full */
+        XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, 32-state->memsize);
+        state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0, endian));
+        state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1, endian));
+        state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2, endian));
+        state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3, endian));
+        p += 32-state->memsize;
+        state->memsize = 0;
+    }
+
+    if (p+32 <= bEnd) {
+        const BYTE* const limit = bEnd - 32;
+        U64 v1 = state->v1;
+        U64 v2 = state->v2;
+        U64 v3 = state->v3;
+        U64 v4 = state->v4;
+
+        do {
+            v1 = XXH64_round(v1, XXH_readLE64(p, endian)); p+=8;
+            v2 = XXH64_round(v2, XXH_readLE64(p, endian)); p+=8;
+            v3 = XXH64_round(v3, XXH_readLE64(p, endian)); p+=8;
+            v4 = XXH64_round(v4, XXH_readLE64(p, endian)); p+=8;
+        } while (p<=limit);
+
+        state->v1 = v1;
+        state->v2 = v2;
+        state->v3 = v3;
+        state->v4 = v4;
+    }
+
+    if (p < bEnd) {
+        XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+        state->memsize = (unsigned)(bEnd-p);
+    }
+
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* state_in, const void* input, size_t len)
+{
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH64_update_endian(state_in, input, len, XXH_littleEndian);
+    else
+        return XXH64_update_endian(state_in, input, len, XXH_bigEndian);
+}
+
+
+
+FORCE_INLINE_TEMPLATE U64 XXH64_digest_endian (const XXH64_state_t* state, XXH_endianess endian)
+{
+    const BYTE * p = (const BYTE*)state->mem64;
+    const BYTE* const bEnd = (const BYTE*)state->mem64 + state->memsize;
+    U64 h64;
+
+    if (state->total_len >= 32) {
+        U64 const v1 = state->v1;
+        U64 const v2 = state->v2;
+        U64 const v3 = state->v3;
+        U64 const v4 = state->v4;
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+    } else {
+        h64  = state->v3 + PRIME64_5;
+    }
+
+    h64 += (U64) state->total_len;
+
+    while (p+8<=bEnd) {
+        U64 const k1 = XXH64_round(0, XXH_readLE64(p, endian));
+        h64 ^= k1;
+        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
+        p+=8;
+    }
+
+    if (p+4<=bEnd) {
+        h64 ^= (U64)(XXH_readLE32(p, endian)) * PRIME64_1;
+        h64  = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
+        p+=4;
+    }
+
+    while (p<bEnd) {
+        h64 ^= (*p) * PRIME64_5;
+        h64  = XXH_rotl64(h64, 11) * PRIME64_1;
+        p++;
+    }
+
+    h64 ^= h64 >> 33;
+    h64 *= PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= PRIME64_3;
+    h64 ^= h64 >> 32;
+
+    return h64;
+}
+
+
+XXH_PUBLIC_API unsigned long long XXH64_digest (const XXH64_state_t* state_in)
+{
+    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
+
+    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
+        return XXH64_digest_endian(state_in, XXH_littleEndian);
+    else
+        return XXH64_digest_endian(state_in, XXH_bigEndian);
+}
+
+
+/* **************************
+*  Canonical representation
+****************************/
+
+/*! Default XXH result types are basic unsigned 32 and 64 bits.
+*   The canonical representation follows human-readable write convention, aka big-endian (large digits first).
+*   These functions allow transformation of hash result into and from its canonical format.
+*   This way, hash values can be written into a file or buffer, and remain comparable across different systems and programs.
+*/
+
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
diff --git a/vendor/github.com/DataDog/zstd/xxhash.h b/vendor/github.com/DataDog/zstd/xxhash.h
new file mode 100644
index 0000000..9bad1f5
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/xxhash.h
@@ -0,0 +1,305 @@
+/*
+   xxHash - Extremely Fast Hash algorithm
+   Header File
+   Copyright (C) 2012-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - xxHash source repository : https://github.com/Cyan4973/xxHash
+*/
+
+/* Notice extracted from xxHash homepage :
+
+xxHash is an extremely fast Hash algorithm, running at RAM speed limits.
+It also successfully passes all tests from the SMHasher suite.
+
+Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
+
+Name            Speed       Q.Score   Author
+xxHash          5.4 GB/s     10
+CrapWow         3.2 GB/s      2       Andrew
+MumurHash 3a    2.7 GB/s     10       Austin Appleby
+SpookyHash      2.0 GB/s     10       Bob Jenkins
+SBox            1.4 GB/s      9       Bret Mulvey
+Lookup3         1.2 GB/s      9       Bob Jenkins
+SuperFastHash   1.2 GB/s      1       Paul Hsieh
+CityHash64      1.05 GB/s    10       Pike & Alakuijala
+FNV             0.55 GB/s     5       Fowler, Noll, Vo
+CRC32           0.43 GB/s     9
+MD5-32          0.33 GB/s    10       Ronald L. Rivest
+SHA1-32         0.28 GB/s    10
+
+Q.Score is a measure of quality of the hash function.
+It depends on successfully passing SMHasher test set.
+10 is a perfect score.
+
+A 64-bits version, named XXH64, is available since r35.
+It offers much better speed, but for 64-bits applications only.
+Name     Speed on 64 bits    Speed on 32 bits
+XXH64       13.8 GB/s            1.9 GB/s
+XXH32        6.8 GB/s            6.0 GB/s
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+
+/* ****************************
+*  Definitions
+******************************/
+#include <stddef.h>   /* size_t */
+typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+
+
+/* ****************************
+*  API modifier
+******************************/
+/** XXH_PRIVATE_API
+*   This is useful if you want to include xxhash functions in `static` mode
+*   in order to inline them, and remove their symbol from the public list.
+*   Methodology :
+*     #define XXH_PRIVATE_API
+*     #include "xxhash.h"
+*   `xxhash.c` is automatically included.
+*   It's not useful to compile and link it as a separate module anymore.
+*/
+#ifdef XXH_PRIVATE_API
+#  ifndef XXH_STATIC_LINKING_ONLY
+#    define XXH_STATIC_LINKING_ONLY
+#  endif
+#  if defined(__GNUC__)
+#    define XXH_PUBLIC_API static __inline __attribute__((unused))
+#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#    define XXH_PUBLIC_API static inline
+#  elif defined(_MSC_VER)
+#    define XXH_PUBLIC_API static __inline
+#  else
+#    define XXH_PUBLIC_API static   /* this version may generate warnings for unused static functions; disable the relevant warning */
+#  endif
+#else
+#  define XXH_PUBLIC_API   /* do nothing */
+#endif /* XXH_PRIVATE_API */
+
+/*!XXH_NAMESPACE, aka Namespace Emulation :
+
+If you want to include _and expose_ xxHash functions from within your own library,
+but also want to avoid symbol collisions with another library which also includes xxHash,
+
+you can use XXH_NAMESPACE, to automatically prefix any public symbol from xxhash library
+with the value of XXH_NAMESPACE (so avoid to keep it NULL and avoid numeric values).
+
+Note that no change is required within the calling program as long as it includes `xxhash.h` :
+regular symbol name will be automatically translated by this header.
+*/
+#ifdef XXH_NAMESPACE
+#  define XXH_CAT(A,B) A##B
+#  define XXH_NAME2(A,B) XXH_CAT(A,B)
+#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
+#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
+#  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+#  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
+#  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
+#  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
+#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
+#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
+#  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
+#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
+#  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+#  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    6
+#define XXH_VERSION_RELEASE  2
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Simple Hash Functions
+******************************/
+typedef unsigned int       XXH32_hash_t;
+typedef unsigned long long XXH64_hash_t;
+
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, unsigned int seed);
+XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, unsigned long long seed);
+
+/*!
+XXH32() :
+    Calculate the 32-bits hash of sequence "length" bytes stored at memory address "input".
+    The memory between input & input+length must be valid (allocated and read-accessible).
+    "seed" can be used to alter the result predictably.
+    Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark) : 5.4 GB/s
+XXH64() :
+    Calculate the 64-bits hash of sequence of length "len" stored at memory address "input".
+    "seed" can be used to alter the result predictably.
+    This function runs 2x faster on 64-bits systems, but slower on 32-bits systems (see benchmark).
+*/
+
+
+/* ****************************
+*  Streaming Hash Functions
+******************************/
+typedef struct XXH32_state_s XXH32_state_t;   /* incomplete type */
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+
+/*! State allocation, compatible with dynamic libraries */
+
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+
+
+/* hash streaming */
+
+XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, unsigned int seed);
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
+
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, unsigned long long seed);
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+
+/*
+These functions generate the xxHash of an input provided in multiple segments.
+Note that, for small input, they are slower than single-call functions, due to state management.
+For small input, prefer `XXH32()` and `XXH64()` .
+
+XXH state must first be allocated, using XXH*_createState() .
+
+Start a new hash by initializing state with a seed, using XXH*_reset().
+
+Then, feed the hash state by calling XXH*_update() as many times as necessary.
+Obviously, input must be allocated and read accessible.
+The function returns an error code, with 0 meaning OK, and any other value meaning there is an error.
+
+Finally, a hash value can be produced anytime, by using XXH*_digest().
+This function returns the nn-bits hash as an int or long long.
+
+It's still possible to continue inserting input into the hash state after a digest,
+and generate some new hashes later on, by calling again XXH*_digest().
+
+When done, free XXH state space if it was allocated dynamically.
+*/
+
+
+/* **************************
+*  Utils
+****************************/
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* ! C99 */
+#  define restrict   /* disable restrict */
+#endif
+
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* restrict dst_state, const XXH32_state_t* restrict src_state);
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* restrict dst_state, const XXH64_state_t* restrict src_state);
+
+
+/* **************************
+*  Canonical representation
+****************************/
+/* Default result type for XXH functions are primitive unsigned 32 and 64 bits.
+*  The canonical representation uses human-readable write convention, aka big-endian (large digits first).
+*  These functions allow transformation of hash result into and from its canonical format.
+*  This way, hash values can be written into a file / memory, and remain comparable on different systems and programs.
+*/
+typedef struct { unsigned char digest[4]; } XXH32_canonical_t;
+typedef struct { unsigned char digest[8]; } XXH64_canonical_t;
+
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+/* ================================================================================================
+   This section contains definitions which are not guaranteed to remain stable.
+   They may change in future versions, becoming incompatible with a different version of the library.
+   They shall only be used with static linking.
+   Never use these definitions in association with dynamic linking !
+=================================================================================================== */
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXH_STATIC_H_3543687687345)
+#define XXH_STATIC_H_3543687687345
+
+/* These definitions are only meant to allow allocation of XXH state
+   statically, on stack, or in a struct for example.
+   Do not use members directly. */
+
+   struct XXH32_state_s {
+       unsigned total_len_32;
+       unsigned large_len;
+       unsigned v1;
+       unsigned v2;
+       unsigned v3;
+       unsigned v4;
+       unsigned mem32[4];   /* buffer defined as U32 for alignment */
+       unsigned memsize;
+       unsigned reserved;   /* never read nor write, will be removed in a future version */
+   };   /* typedef'd to XXH32_state_t */
+
+   struct XXH64_state_s {
+       unsigned long long total_len;
+       unsigned long long v1;
+       unsigned long long v2;
+       unsigned long long v3;
+       unsigned long long v4;
+       unsigned long long mem64[4];   /* buffer defined as U64 for alignment */
+       unsigned memsize;
+       unsigned reserved[2];          /* never read nor write, will be removed in a future version */
+   };   /* typedef'd to XXH64_state_t */
+
+
+#  ifdef XXH_PRIVATE_API
+#    include "xxhash.c"   /* include xxhash functions as `static`, for inlining */
+#  endif
+
+#endif /* XXH_STATIC_LINKING_ONLY && XXH_STATIC_H_3543687687345 */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/vendor/github.com/DataDog/zstd/zbuff.h b/vendor/github.com/DataDog/zstd/zbuff.h
new file mode 100644
index 0000000..a93115d
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zbuff.h
@@ -0,0 +1,213 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* ***************************************************************
+*  NOTES/WARNINGS
+******************************************************************/
+/* The streaming API defined here is deprecated.
+ * Consider migrating towards ZSTD_compressStream() API in `zstd.h`
+ * See 'lib/README.md'.
+ *****************************************************************/
+
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef ZSTD_BUFFERED_H_23987
+#define ZSTD_BUFFERED_H_23987
+
+/* *************************************
+*  Dependencies
+***************************************/
+#include <stddef.h>      /* size_t */
+#include "zstd.h"        /* ZSTD_CStream, ZSTD_DStream, ZSTDLIB_API */
+
+
+/* ***************************************************************
+*  Compiler specifics
+*****************************************************************/
+/* Deprecation warnings */
+/* Should these warnings be a problem,
+   it is generally possible to disable them,
+   typically with -Wno-deprecated-declarations for gcc
+   or _CRT_SECURE_NO_WARNINGS in Visual.
+   Otherwise, it's also possible to define ZBUFF_DISABLE_DEPRECATE_WARNINGS */
+#ifdef ZBUFF_DISABLE_DEPRECATE_WARNINGS
+#  define ZBUFF_DEPRECATED(message) ZSTDLIB_API  /* disable deprecation warnings */
+#else
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZBUFF_DEPRECATED(message) [[deprecated(message)]] ZSTDLIB_API
+#  elif (defined(__GNUC__) && (__GNUC__ >= 5)) || defined(__clang__)
+#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ >= 3)
+#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZBUFF_DEPRECATED for this compiler")
+#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API
+#  endif
+#endif /* ZBUFF_DISABLE_DEPRECATE_WARNINGS */
+
+
+/* *************************************
+*  Streaming functions
+***************************************/
+/* This is the easier "buffered" streaming API,
+*  using an internal buffer to lift all restrictions on user-provided buffers
+*  which can be any size, any place, for both input and output.
+*  ZBUFF and ZSTD are 100% interoperable,
+*  frames created by one can be decoded by the other one */
+
+typedef ZSTD_CStream ZBUFF_CCtx;
+ZBUFF_DEPRECATED("use ZSTD_createCStream") ZBUFF_CCtx* ZBUFF_createCCtx(void);
+ZBUFF_DEPRECATED("use ZSTD_freeCStream")   size_t      ZBUFF_freeCCtx(ZBUFF_CCtx* cctx);
+
+ZBUFF_DEPRECATED("use ZSTD_initCStream")           size_t ZBUFF_compressInit(ZBUFF_CCtx* cctx, int compressionLevel);
+ZBUFF_DEPRECATED("use ZSTD_initCStream_usingDict") size_t ZBUFF_compressInitDictionary(ZBUFF_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
+
+ZBUFF_DEPRECATED("use ZSTD_compressStream") size_t ZBUFF_compressContinue(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr, const void* src, size_t* srcSizePtr);
+ZBUFF_DEPRECATED("use ZSTD_flushStream")    size_t ZBUFF_compressFlush(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr);
+ZBUFF_DEPRECATED("use ZSTD_endStream")      size_t ZBUFF_compressEnd(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr);
+
+/*-*************************************************
+*  Streaming compression - howto
+*
+*  A ZBUFF_CCtx object is required to track streaming operation.
+*  Use ZBUFF_createCCtx() and ZBUFF_freeCCtx() to create/release resources.
+*  ZBUFF_CCtx objects can be reused multiple times.
+*
+*  Start by initializing ZBUF_CCtx.
+*  Use ZBUFF_compressInit() to start a new compression operation.
+*  Use ZBUFF_compressInitDictionary() for a compression which requires a dictionary.
+*
+*  Use ZBUFF_compressContinue() repetitively to consume input stream.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written within *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present again remaining data.
+*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each call, so save its content if it matters or change @dst .
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's just a hint, to improve latency)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  At any moment, it's possible to flush whatever data remains within buffer, using ZBUFF_compressFlush().
+*  The nb of bytes written into `dst` will be reported into *dstCapacityPtr.
+*  Note that the function cannot output more than *dstCapacityPtr,
+*  therefore, some content might still be left into internal buffer if *dstCapacityPtr is too small.
+*  @return : nb of bytes still present into internal buffer (0 if it's empty)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  ZBUFF_compressEnd() instructs to finish a frame.
+*  It will perform a flush and write frame epilogue.
+*  The epilogue is required for decoders to consider a frame completed.
+*  Similar to ZBUFF_compressFlush(), it may not be able to output the entire internal buffer content if *dstCapacityPtr is too small.
+*  In which case, call again ZBUFF_compressFlush() to complete the flush.
+*  @return : nb of bytes still present into internal buffer (0 if it's empty)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  Hint : _recommended buffer_ sizes (not compulsory) : ZBUFF_recommendedCInSize() / ZBUFF_recommendedCOutSize()
+*  input : ZBUFF_recommendedCInSize==128 KB block size is the internal unit, use this value to reduce intermediate stages (better latency)
+*  output : ZBUFF_recommendedCOutSize==ZSTD_compressBound(128 KB) + 3 + 3 : ensures it's always possible to write/flush/end a full block. Skip some buffering.
+*  By using both, it ensures that input will be entirely consumed, and output will always contain the result, reducing intermediate buffering.
+* **************************************************/
+
+
+typedef ZSTD_DStream ZBUFF_DCtx;
+ZBUFF_DEPRECATED("use ZSTD_createDStream") ZBUFF_DCtx* ZBUFF_createDCtx(void);
+ZBUFF_DEPRECATED("use ZSTD_freeDStream")   size_t      ZBUFF_freeDCtx(ZBUFF_DCtx* dctx);
+
+ZBUFF_DEPRECATED("use ZSTD_initDStream")           size_t ZBUFF_decompressInit(ZBUFF_DCtx* dctx);
+ZBUFF_DEPRECATED("use ZSTD_initDStream_usingDict") size_t ZBUFF_decompressInitDictionary(ZBUFF_DCtx* dctx, const void* dict, size_t dictSize);
+
+ZBUFF_DEPRECATED("use ZSTD_decompressStream") size_t ZBUFF_decompressContinue(ZBUFF_DCtx* dctx,
+                                            void* dst, size_t* dstCapacityPtr,
+                                      const void* src, size_t* srcSizePtr);
+
+/*-***************************************************************************
+*  Streaming decompression howto
+*
+*  A ZBUFF_DCtx object is required to track streaming operations.
+*  Use ZBUFF_createDCtx() and ZBUFF_freeDCtx() to create/release resources.
+*  Use ZBUFF_decompressInit() to start a new decompression operation,
+*   or ZBUFF_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFF_DCtx objects can be re-init multiple times.
+*
+*  Use ZBUFF_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change `dst`.
+*  @return : 0 when a frame is completely decoded and fully flushed,
+*            1 when there is still some data left within internal buffer to flush,
+*            >1 when more data is expected, with value being a suggested next input size (it's just a hint, which helps latency),
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFF_recommendedDInSize() and ZBUFF_recommendedDOutSize()
+*  output : ZBUFF_recommendedDOutSize== 128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFF_recommendedDInSize == 128KB + 3;
+*           just follow indications from ZBUFF_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+ZBUFF_DEPRECATED("use ZSTD_isError")      unsigned ZBUFF_isError(size_t errorCode);
+ZBUFF_DEPRECATED("use ZSTD_getErrorName") const char* ZBUFF_getErrorName(size_t errorCode);
+
+/** Functions below provide recommended buffer sizes for Compression or Decompression operations.
+*   These sizes are just hints, they tend to offer better latency */
+ZBUFF_DEPRECATED("use ZSTD_CStreamInSize")  size_t ZBUFF_recommendedCInSize(void);
+ZBUFF_DEPRECATED("use ZSTD_CStreamOutSize") size_t ZBUFF_recommendedCOutSize(void);
+ZBUFF_DEPRECATED("use ZSTD_DStreamInSize")  size_t ZBUFF_recommendedDInSize(void);
+ZBUFF_DEPRECATED("use ZSTD_DStreamOutSize") size_t ZBUFF_recommendedDOutSize(void);
+
+#endif  /* ZSTD_BUFFERED_H_23987 */
+
+
+#ifdef ZBUFF_STATIC_LINKING_ONLY
+#ifndef ZBUFF_STATIC_H_30298098432
+#define ZBUFF_STATIC_H_30298098432
+
+/* ====================================================================================
+ * The definitions in this section are considered experimental.
+ * They should never be used in association with a dynamic library, as they may change in the future.
+ * They are provided for advanced usages.
+ * Use them only in association with static linking.
+ * ==================================================================================== */
+
+/*--- Dependency ---*/
+#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_parameters, ZSTD_customMem */
+#include "zstd.h"
+
+
+/*--- Custom memory allocator ---*/
+/*! ZBUFF_createCCtx_advanced() :
+ *  Create a ZBUFF compression context using external alloc and free functions */
+ZBUFF_DEPRECATED("use ZSTD_createCStream_advanced") ZBUFF_CCtx* ZBUFF_createCCtx_advanced(ZSTD_customMem customMem);
+
+/*! ZBUFF_createDCtx_advanced() :
+ *  Create a ZBUFF decompression context using external alloc and free functions */
+ZBUFF_DEPRECATED("use ZSTD_createDStream_advanced") ZBUFF_DCtx* ZBUFF_createDCtx_advanced(ZSTD_customMem customMem);
+
+
+/*--- Advanced Streaming Initialization ---*/
+ZBUFF_DEPRECATED("use ZSTD_initDStream_usingDict") size_t ZBUFF_compressInit_advanced(ZBUFF_CCtx* zbc,
+                                               const void* dict, size_t dictSize,
+                                               ZSTD_parameters params, unsigned long long pledgedSrcSize);
+
+
+#endif    /* ZBUFF_STATIC_H_30298098432 */
+#endif    /* ZBUFF_STATIC_LINKING_ONLY */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/vendor/github.com/DataDog/zstd/zbuff_common.c b/vendor/github.com/DataDog/zstd/zbuff_common.c
new file mode 100644
index 0000000..661b9b0
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zbuff_common.c
@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include "error_private.h"
+#include "zbuff.h"
+
+/*-****************************************
+*  ZBUFF Error Management  (deprecated)
+******************************************/
+
+/*! ZBUFF_isError() :
+*   tells if a return value is an error code */
+unsigned ZBUFF_isError(size_t errorCode) { return ERR_isError(errorCode); }
+/*! ZBUFF_getErrorName() :
+*   provides error code string from function result (useful for debugging) */
+const char* ZBUFF_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
diff --git a/vendor/github.com/DataDog/zstd/zbuff_compress.c b/vendor/github.com/DataDog/zstd/zbuff_compress.c
new file mode 100644
index 0000000..f39c60d
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zbuff_compress.c
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+
+/* *************************************
+*  Dependencies
+***************************************/
+#define ZBUFF_STATIC_LINKING_ONLY
+#include "zbuff.h"
+
+
+/*-***********************************************************
+*  Streaming compression
+*
+*  A ZBUFF_CCtx object is required to track streaming operation.
+*  Use ZBUFF_createCCtx() and ZBUFF_freeCCtx() to create/release resources.
+*  Use ZBUFF_compressInit() to start a new compression operation.
+*  ZBUFF_CCtx objects can be reused multiple times.
+*
+*  Use ZBUFF_compressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to call again the function with remaining input.
+*  The content of dst will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters or change dst .
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to improve latency)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  ZBUFF_compressFlush() can be used to instruct ZBUFF to compress and output whatever remains within its buffer.
+*  Note that it will not output more than *dstCapacityPtr.
+*  Therefore, some content might still be left into its internal buffer if dst buffer is too small.
+*  @return : nb of bytes still present into internal buffer (0 if it's empty)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  ZBUFF_compressEnd() instructs to finish a frame.
+*  It will perform a flush and write frame epilogue.
+*  Similar to ZBUFF_compressFlush(), it may not be able to output the entire internal buffer content if *dstCapacityPtr is too small.
+*  @return : nb of bytes still present into internal buffer (0 if it's empty)
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory)
+*  input : ZSTD_BLOCKSIZE_MAX (128 KB), internal unit size, it improves latency to use this value.
+*  output : ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + ZBUFF_endFrameSize : ensures it's always possible to write/flush/end a full block at best speed.
+* ***********************************************************/
+
+ZBUFF_CCtx* ZBUFF_createCCtx(void)
+{
+    return ZSTD_createCStream();
+}
+
+ZBUFF_CCtx* ZBUFF_createCCtx_advanced(ZSTD_customMem customMem)
+{
+    return ZSTD_createCStream_advanced(customMem);
+}
+
+size_t ZBUFF_freeCCtx(ZBUFF_CCtx* zbc)
+{
+    return ZSTD_freeCStream(zbc);
+}
+
+
+/* ======   Initialization   ====== */
+
+size_t ZBUFF_compressInit_advanced(ZBUFF_CCtx* zbc,
+                                   const void* dict, size_t dictSize,
+                                   ZSTD_parameters params, unsigned long long pledgedSrcSize)
+{
+    if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;  /* preserve "0 == unknown" behavior */
+    return ZSTD_initCStream_advanced(zbc, dict, dictSize, params, pledgedSrcSize);
+}
+
+
+size_t ZBUFF_compressInitDictionary(ZBUFF_CCtx* zbc, const void* dict, size_t dictSize, int compressionLevel)
+{
+    return ZSTD_initCStream_usingDict(zbc, dict, dictSize, compressionLevel);
+}
+
+size_t ZBUFF_compressInit(ZBUFF_CCtx* zbc, int compressionLevel)
+{
+    return ZSTD_initCStream(zbc, compressionLevel);
+}
+
+/* ======   Compression   ====== */
+
+
+size_t ZBUFF_compressContinue(ZBUFF_CCtx* zbc,
+                              void* dst, size_t* dstCapacityPtr,
+                        const void* src, size_t* srcSizePtr)
+{
+    size_t result;
+    ZSTD_outBuffer outBuff;
+    ZSTD_inBuffer inBuff;
+    outBuff.dst = dst;
+    outBuff.pos = 0;
+    outBuff.size = *dstCapacityPtr;
+    inBuff.src = src;
+    inBuff.pos = 0;
+    inBuff.size = *srcSizePtr;
+    result = ZSTD_compressStream(zbc, &outBuff, &inBuff);
+    *dstCapacityPtr = outBuff.pos;
+    *srcSizePtr = inBuff.pos;
+    return result;
+}
+
+
+
+/* ======   Finalize   ====== */
+
+size_t ZBUFF_compressFlush(ZBUFF_CCtx* zbc, void* dst, size_t* dstCapacityPtr)
+{
+    size_t result;
+    ZSTD_outBuffer outBuff;
+    outBuff.dst = dst;
+    outBuff.pos = 0;
+    outBuff.size = *dstCapacityPtr;
+    result = ZSTD_flushStream(zbc, &outBuff);
+    *dstCapacityPtr = outBuff.pos;
+    return result;
+}
+
+
+size_t ZBUFF_compressEnd(ZBUFF_CCtx* zbc, void* dst, size_t* dstCapacityPtr)
+{
+    size_t result;
+    ZSTD_outBuffer outBuff;
+    outBuff.dst = dst;
+    outBuff.pos = 0;
+    outBuff.size = *dstCapacityPtr;
+    result = ZSTD_endStream(zbc, &outBuff);
+    *dstCapacityPtr = outBuff.pos;
+    return result;
+}
+
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+size_t ZBUFF_recommendedCInSize(void)  { return ZSTD_CStreamInSize(); }
+size_t ZBUFF_recommendedCOutSize(void) { return ZSTD_CStreamOutSize(); }
diff --git a/vendor/github.com/DataDog/zstd/zbuff_decompress.c b/vendor/github.com/DataDog/zstd/zbuff_decompress.c
new file mode 100644
index 0000000..923c22b
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zbuff_decompress.c
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+
+/* *************************************
+*  Dependencies
+***************************************/
+#define ZBUFF_STATIC_LINKING_ONLY
+#include "zbuff.h"
+
+
+ZBUFF_DCtx* ZBUFF_createDCtx(void)
+{
+    return ZSTD_createDStream();
+}
+
+ZBUFF_DCtx* ZBUFF_createDCtx_advanced(ZSTD_customMem customMem)
+{
+    return ZSTD_createDStream_advanced(customMem);
+}
+
+size_t ZBUFF_freeDCtx(ZBUFF_DCtx* zbd)
+{
+    return ZSTD_freeDStream(zbd);
+}
+
+
+/* *** Initialization *** */
+
+size_t ZBUFF_decompressInitDictionary(ZBUFF_DCtx* zbd, const void* dict, size_t dictSize)
+{
+    return ZSTD_initDStream_usingDict(zbd, dict, dictSize);
+}
+
+size_t ZBUFF_decompressInit(ZBUFF_DCtx* zbd)
+{
+    return ZSTD_initDStream(zbd);
+}
+
+
+/* *** Decompression *** */
+
+size_t ZBUFF_decompressContinue(ZBUFF_DCtx* zbd,
+                                void* dst, size_t* dstCapacityPtr,
+                          const void* src, size_t* srcSizePtr)
+{
+    ZSTD_outBuffer outBuff;
+    ZSTD_inBuffer inBuff;
+    size_t result;
+    outBuff.dst  = dst;
+    outBuff.pos  = 0;
+    outBuff.size = *dstCapacityPtr;
+    inBuff.src  = src;
+    inBuff.pos  = 0;
+    inBuff.size = *srcSizePtr;
+    result = ZSTD_decompressStream(zbd, &outBuff, &inBuff);
+    *dstCapacityPtr = outBuff.pos;
+    *srcSizePtr = inBuff.pos;
+    return result;
+}
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+size_t ZBUFF_recommendedDInSize(void)  { return ZSTD_DStreamInSize(); }
+size_t ZBUFF_recommendedDOutSize(void) { return ZSTD_DStreamOutSize(); }
diff --git a/vendor/github.com/DataDog/zstd/zdict.c b/vendor/github.com/DataDog/zstd/zdict.c
new file mode 100644
index 0000000..7d24e49
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zdict.c
@@ -0,0 +1,1108 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*-**************************************
+*  Tuning parameters
+****************************************/
+#define MINRATIO 4   /* minimum nb of apparition to be selected in dictionary */
+#define ZDICT_MAX_SAMPLES_SIZE (2000U << 20)
+#define ZDICT_MIN_SAMPLES_SIZE (ZDICT_CONTENTSIZE_MIN * MINRATIO)
+
+
+/*-**************************************
+*  Compiler Options
+****************************************/
+/* Unix Large Files support (>4GB) */
+#define _FILE_OFFSET_BITS 64
+#if (defined(__sun__) && (!defined(__LP64__)))   /* Sun Solaris 32-bits requires specific definitions */
+#  define _LARGEFILE_SOURCE
+#elif ! defined(__LP64__)                        /* No point defining Large file for 64 bit */
+#  define _LARGEFILE64_SOURCE
+#endif
+
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include <stdlib.h>        /* malloc, free */
+#include <string.h>        /* memset */
+#include <stdio.h>         /* fprintf, fopen, ftello64 */
+#include <time.h>          /* clock */
+
+#include "mem.h"           /* read */
+#include "fse.h"           /* FSE_normalizeCount, FSE_writeNCount */
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"           /* HUF_buildCTable, HUF_writeCTable */
+#include "zstd_internal.h" /* includes zstd.h */
+#include "xxhash.h"        /* XXH64 */
+#include "divsufsort.h"
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+#include "zdict.h"
+
+
+/*-*************************************
+*  Constants
+***************************************/
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define DICTLISTSIZE_DEFAULT 10000
+
+#define NOISELENGTH 32
+
+static const int g_compressionLevel_default = 3;
+static const U32 g_selectivity_default = 9;
+
+
+/*-*************************************
+*  Console display
+***************************************/
+#define DISPLAY(...)         { fprintf(stderr, __VA_ARGS__); fflush( stderr ); }
+#define DISPLAYLEVEL(l, ...) if (notificationLevel>=l) { DISPLAY(__VA_ARGS__); }    /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
+
+static clock_t ZDICT_clockSpan(clock_t nPrevious) { return clock() - nPrevious; }
+
+static void ZDICT_printHex(const void* ptr, size_t length)
+{
+    const BYTE* const b = (const BYTE*)ptr;
+    size_t u;
+    for (u=0; u<length; u++) {
+        BYTE c = b[u];
+        if (c<32 || c>126) c = '.';   /* non-printable char */
+        DISPLAY("%c", c);
+    }
+}
+
+
+/*-********************************************************
+*  Helper functions
+**********************************************************/
+unsigned ZDICT_isError(size_t errorCode) { return ERR_isError(errorCode); }
+
+const char* ZDICT_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
+
+unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize)
+{
+    if (dictSize < 8) return 0;
+    if (MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return 0;
+    return MEM_readLE32((const char*)dictBuffer + 4);
+}
+
+
+/*-********************************************************
+*  Dictionary training functions
+**********************************************************/
+static unsigned ZDICT_NbCommonBytes (size_t val)
+{
+    if (MEM_isLittleEndian()) {
+        if (MEM_64bits()) {
+#       if defined(_MSC_VER) && defined(_WIN64)
+            unsigned long r = 0;
+            _BitScanForward64( &r, (U64)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_ctzll((U64)val) >> 3);
+#       else
+            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2, 0, 3, 1, 3, 1, 4, 2, 7, 0, 2, 3, 6, 1, 5, 3, 5, 1, 3, 4, 4, 2, 5, 6, 7, 7, 0, 1, 2, 3, 3, 4, 6, 2, 6, 5, 5, 3, 4, 5, 6, 7, 1, 2, 4, 6, 4, 4, 5, 7, 2, 6, 5, 7, 6, 7, 7 };
+            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
+#       endif
+        } else { /* 32 bits */
+#       if defined(_MSC_VER)
+            unsigned long r=0;
+            _BitScanForward( &r, (U32)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_ctz((U32)val) >> 3);
+#       else
+            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0, 3, 2, 2, 1, 3, 2, 0, 1, 3, 3, 1, 2, 2, 2, 2, 0, 3, 1, 2, 0, 1, 0, 1, 1 };
+            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
+#       endif
+        }
+    } else {  /* Big Endian CPU */
+        if (MEM_64bits()) {
+#       if defined(_MSC_VER) && defined(_WIN64)
+            unsigned long r = 0;
+            _BitScanReverse64( &r, val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_clzll(val) >> 3);
+#       else
+            unsigned r;
+            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
+            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
+            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
+            r += (!val);
+            return r;
+#       endif
+        } else { /* 32 bits */
+#       if defined(_MSC_VER)
+            unsigned long r = 0;
+            _BitScanReverse( &r, (unsigned long)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_clz((U32)val) >> 3);
+#       else
+            unsigned r;
+            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
+            r += (!val);
+            return r;
+#       endif
+    }   }
+}
+
+
+/*! ZDICT_count() :
+    Count the nb of common bytes between 2 pointers.
+    Note : this function presumes end of buffer followed by noisy guard band.
+*/
+static size_t ZDICT_count(const void* pIn, const void* pMatch)
+{
+    const char* const pStart = (const char*)pIn;
+    for (;;) {
+        size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
+        if (!diff) {
+            pIn = (const char*)pIn+sizeof(size_t);
+            pMatch = (const char*)pMatch+sizeof(size_t);
+            continue;
+        }
+        pIn = (const char*)pIn+ZDICT_NbCommonBytes(diff);
+        return (size_t)((const char*)pIn - pStart);
+    }
+}
+
+
+typedef struct {
+    U32 pos;
+    U32 length;
+    U32 savings;
+} dictItem;
+
+static void ZDICT_initDictItem(dictItem* d)
+{
+    d->pos = 1;
+    d->length = 0;
+    d->savings = (U32)(-1);
+}
+
+
+#define LLIMIT 64          /* heuristic determined experimentally */
+#define MINMATCHLENGTH 7   /* heuristic determined experimentally */
+static dictItem ZDICT_analyzePos(
+                       BYTE* doneMarks,
+                       const int* suffix, U32 start,
+                       const void* buffer, U32 minRatio, U32 notificationLevel)
+{
+    U32 lengthList[LLIMIT] = {0};
+    U32 cumulLength[LLIMIT] = {0};
+    U32 savings[LLIMIT] = {0};
+    const BYTE* b = (const BYTE*)buffer;
+    size_t maxLength = LLIMIT;
+    size_t pos = suffix[start];
+    U32 end = start;
+    dictItem solution;
+
+    /* init */
+    memset(&solution, 0, sizeof(solution));
+    doneMarks[pos] = 1;
+
+    /* trivial repetition cases */
+    if ( (MEM_read16(b+pos+0) == MEM_read16(b+pos+2))
+       ||(MEM_read16(b+pos+1) == MEM_read16(b+pos+3))
+       ||(MEM_read16(b+pos+2) == MEM_read16(b+pos+4)) ) {
+        /* skip and mark segment */
+        U16 const pattern16 = MEM_read16(b+pos+4);
+        U32 u, patternEnd = 6;
+        while (MEM_read16(b+pos+patternEnd) == pattern16) patternEnd+=2 ;
+        if (b[pos+patternEnd] == b[pos+patternEnd-1]) patternEnd++;
+        for (u=1; u<patternEnd; u++)
+            doneMarks[pos+u] = 1;
+        return solution;
+    }
+
+    /* look forward */
+    {   size_t length;
+        do {
+            end++;
+            length = ZDICT_count(b + pos, b + suffix[end]);
+        } while (length >= MINMATCHLENGTH);
+    }
+
+    /* look backward */
+    {   size_t length;
+        do {
+            length = ZDICT_count(b + pos, b + *(suffix+start-1));
+            if (length >=MINMATCHLENGTH) start--;
+        } while(length >= MINMATCHLENGTH);
+    }
+
+    /* exit if not found a minimum nb of repetitions */
+    if (end-start < minRatio) {
+        U32 idx;
+        for(idx=start; idx<end; idx++)
+            doneMarks[suffix[idx]] = 1;
+        return solution;
+    }
+
+    {   int i;
+        U32 searchLength;
+        U32 refinedStart = start;
+        U32 refinedEnd = end;
+
+        DISPLAYLEVEL(4, "\n");
+        DISPLAYLEVEL(4, "found %3u matches of length >= %i at pos %7u  ", (U32)(end-start), MINMATCHLENGTH, (U32)pos);
+        DISPLAYLEVEL(4, "\n");
+
+        for (searchLength = MINMATCHLENGTH ; ; searchLength++) {
+            BYTE currentChar = 0;
+            U32 currentCount = 0;
+            U32 currentID = refinedStart;
+            U32 id;
+            U32 selectedCount = 0;
+            U32 selectedID = currentID;
+            for (id =refinedStart; id < refinedEnd; id++) {
+                if (b[suffix[id] + searchLength] != currentChar) {
+                    if (currentCount > selectedCount) {
+                        selectedCount = currentCount;
+                        selectedID = currentID;
+                    }
+                    currentID = id;
+                    currentChar = b[ suffix[id] + searchLength];
+                    currentCount = 0;
+                }
+                currentCount ++;
+            }
+            if (currentCount > selectedCount) {  /* for last */
+                selectedCount = currentCount;
+                selectedID = currentID;
+            }
+
+            if (selectedCount < minRatio)
+                break;
+            refinedStart = selectedID;
+            refinedEnd = refinedStart + selectedCount;
+        }
+
+        /* evaluate gain based on new ref */
+        start = refinedStart;
+        pos = suffix[refinedStart];
+        end = start;
+        memset(lengthList, 0, sizeof(lengthList));
+
+        /* look forward */
+        {   size_t length;
+            do {
+                end++;
+                length = ZDICT_count(b + pos, b + suffix[end]);
+                if (length >= LLIMIT) length = LLIMIT-1;
+                lengthList[length]++;
+            } while (length >=MINMATCHLENGTH);
+        }
+
+        /* look backward */
+        {   size_t length = MINMATCHLENGTH;
+            while ((length >= MINMATCHLENGTH) & (start > 0)) {
+                length = ZDICT_count(b + pos, b + suffix[start - 1]);
+                if (length >= LLIMIT) length = LLIMIT - 1;
+                lengthList[length]++;
+                if (length >= MINMATCHLENGTH) start--;
+            }
+        }
+
+        /* largest useful length */
+        memset(cumulLength, 0, sizeof(cumulLength));
+        cumulLength[maxLength-1] = lengthList[maxLength-1];
+        for (i=(int)(maxLength-2); i>=0; i--)
+            cumulLength[i] = cumulLength[i+1] + lengthList[i];
+
+        for (i=LLIMIT-1; i>=MINMATCHLENGTH; i--) if (cumulLength[i]>=minRatio) break;
+        maxLength = i;
+
+        /* reduce maxLength in case of final into repetitive data */
+        {   U32 l = (U32)maxLength;
+            BYTE const c = b[pos + maxLength-1];
+            while (b[pos+l-2]==c) l--;
+            maxLength = l;
+        }
+        if (maxLength < MINMATCHLENGTH) return solution;   /* skip : no long-enough solution */
+
+        /* calculate savings */
+        savings[5] = 0;
+        for (i=MINMATCHLENGTH; i<=(int)maxLength; i++)
+            savings[i] = savings[i-1] + (lengthList[i] * (i-3));
+
+        DISPLAYLEVEL(4, "Selected ref at position %u, of length %u : saves %u (ratio: %.2f)  \n",
+                     (U32)pos, (U32)maxLength, savings[maxLength], (double)savings[maxLength] / maxLength);
+
+        solution.pos = (U32)pos;
+        solution.length = (U32)maxLength;
+        solution.savings = savings[maxLength];
+
+        /* mark positions done */
+        {   U32 id;
+            for (id=start; id<end; id++) {
+                U32 p, pEnd, length;
+                U32 const testedPos = suffix[id];
+                if (testedPos == pos)
+                    length = solution.length;
+                else {
+                    length = (U32)ZDICT_count(b+pos, b+testedPos);
+                    if (length > solution.length) length = solution.length;
+                }
+                pEnd = (U32)(testedPos + length);
+                for (p=testedPos; p<pEnd; p++)
+                    doneMarks[p] = 1;
+    }   }   }
+
+    return solution;
+}
+
+
+static int isIncluded(const void* in, const void* container, size_t length)
+{
+    const char* const ip = (const char*) in;
+    const char* const into = (const char*) container;
+    size_t u;
+
+    for (u=0; u<length; u++) {  /* works because end of buffer is a noisy guard band */
+        if (ip[u] != into[u]) break;
+    }
+
+    return u==length;
+}
+
+/*! ZDICT_tryMerge() :
+    check if dictItem can be merged, do it if possible
+    @return : id of destination elt, 0 if not merged
+*/
+static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const void* buffer)
+{
+    const U32 tableSize = table->pos;
+    const U32 eltEnd = elt.pos + elt.length;
+    const char* const buf = (const char*) buffer;
+
+    /* tail overlap */
+    U32 u; for (u=1; u<tableSize; u++) {
+        if (u==eltNbToSkip) continue;
+        if ((table[u].pos > elt.pos) && (table[u].pos <= eltEnd)) {  /* overlap, existing > new */
+            /* append */
+            U32 const addedLength = table[u].pos - elt.pos;
+            table[u].length += addedLength;
+            table[u].pos = elt.pos;
+            table[u].savings += elt.savings * addedLength / elt.length;   /* rough approx */
+            table[u].savings += elt.length / 8;    /* rough approx bonus */
+            elt = table[u];
+            /* sort : improve rank */
+            while ((u>1) && (table[u-1].savings < elt.savings))
+            table[u] = table[u-1], u--;
+            table[u] = elt;
+            return u;
+    }   }
+
+    /* front overlap */
+    for (u=1; u<tableSize; u++) {
+        if (u==eltNbToSkip) continue;
+
+        if ((table[u].pos + table[u].length >= elt.pos) && (table[u].pos < elt.pos)) {  /* overlap, existing < new */
+            /* append */
+            int const addedLength = (int)eltEnd - (table[u].pos + table[u].length);
+            table[u].savings += elt.length / 8;    /* rough approx bonus */
+            if (addedLength > 0) {   /* otherwise, elt fully included into existing */
+                table[u].length += addedLength;
+                table[u].savings += elt.savings * addedLength / elt.length;   /* rough approx */
+            }
+            /* sort : improve rank */
+            elt = table[u];
+            while ((u>1) && (table[u-1].savings < elt.savings))
+                table[u] = table[u-1], u--;
+            table[u] = elt;
+            return u;
+        }
+
+        if (MEM_read64(buf + table[u].pos) == MEM_read64(buf + elt.pos + 1)) {
+            if (isIncluded(buf + table[u].pos, buf + elt.pos + 1, table[u].length)) {
+                size_t const addedLength = MAX( (int)elt.length - (int)table[u].length , 1 );
+                table[u].pos = elt.pos;
+                table[u].savings += (U32)(elt.savings * addedLength / elt.length);
+                table[u].length = MIN(elt.length, table[u].length + 1);
+                return u;
+            }
+        }
+    }
+
+    return 0;
+}
+
+
+static void ZDICT_removeDictItem(dictItem* table, U32 id)
+{
+    /* convention : table[0].pos stores nb of elts */
+    U32 const max = table[0].pos;
+    U32 u;
+    if (!id) return;   /* protection, should never happen */
+    for (u=id; u<max-1; u++)
+        table[u] = table[u+1];
+    table->pos--;
+}
+
+
+static void ZDICT_insertDictItem(dictItem* table, U32 maxSize, dictItem elt, const void* buffer)
+{
+    /* merge if possible */
+    U32 mergeId = ZDICT_tryMerge(table, elt, 0, buffer);
+    if (mergeId) {
+        U32 newMerge = 1;
+        while (newMerge) {
+            newMerge = ZDICT_tryMerge(table, table[mergeId], mergeId, buffer);
+            if (newMerge) ZDICT_removeDictItem(table, mergeId);
+            mergeId = newMerge;
+        }
+        return;
+    }
+
+    /* insert */
+    {   U32 current;
+        U32 nextElt = table->pos;
+        if (nextElt >= maxSize) nextElt = maxSize-1;
+        current = nextElt-1;
+        while (table[current].savings < elt.savings) {
+            table[current+1] = table[current];
+            current--;
+        }
+        table[current+1] = elt;
+        table->pos = nextElt+1;
+    }
+}
+
+
+static U32 ZDICT_dictSize(const dictItem* dictList)
+{
+    U32 u, dictSize = 0;
+    for (u=1; u<dictList[0].pos; u++)
+        dictSize += dictList[u].length;
+    return dictSize;
+}
+
+
+static size_t ZDICT_trainBuffer_legacy(dictItem* dictList, U32 dictListSize,
+                            const void* const buffer, size_t bufferSize,   /* buffer must end with noisy guard band */
+                            const size_t* fileSizes, unsigned nbFiles,
+                            U32 minRatio, U32 notificationLevel)
+{
+    int* const suffix0 = (int*)malloc((bufferSize+2)*sizeof(*suffix0));
+    int* const suffix = suffix0+1;
+    U32* reverseSuffix = (U32*)malloc((bufferSize)*sizeof(*reverseSuffix));
+    BYTE* doneMarks = (BYTE*)malloc((bufferSize+16)*sizeof(*doneMarks));   /* +16 for overflow security */
+    U32* filePos = (U32*)malloc(nbFiles * sizeof(*filePos));
+    size_t result = 0;
+    clock_t displayClock = 0;
+    clock_t const refreshRate = CLOCKS_PER_SEC * 3 / 10;
+
+#   define DISPLAYUPDATE(l, ...) if (notificationLevel>=l) { \
+            if (ZDICT_clockSpan(displayClock) > refreshRate)  \
+            { displayClock = clock(); DISPLAY(__VA_ARGS__); \
+            if (notificationLevel>=4) fflush(stderr); } }
+
+    /* init */
+    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
+    if (!suffix0 || !reverseSuffix || !doneMarks || !filePos) {
+        result = ERROR(memory_allocation);
+        goto _cleanup;
+    }
+    if (minRatio < MINRATIO) minRatio = MINRATIO;
+    memset(doneMarks, 0, bufferSize+16);
+
+    /* limit sample set size (divsufsort limitation)*/
+    if (bufferSize > ZDICT_MAX_SAMPLES_SIZE) DISPLAYLEVEL(3, "sample set too large : reduced to %u MB ...\n", (U32)(ZDICT_MAX_SAMPLES_SIZE>>20));
+    while (bufferSize > ZDICT_MAX_SAMPLES_SIZE) bufferSize -= fileSizes[--nbFiles];
+
+    /* sort */
+    DISPLAYLEVEL(2, "sorting %u files of total size %u MB ...\n", nbFiles, (U32)(bufferSize>>20));
+    {   int const divSuftSortResult = divsufsort((const unsigned char*)buffer, suffix, (int)bufferSize, 0);
+        if (divSuftSortResult != 0) { result = ERROR(GENERIC); goto _cleanup; }
+    }
+    suffix[bufferSize] = (int)bufferSize;   /* leads into noise */
+    suffix0[0] = (int)bufferSize;           /* leads into noise */
+    /* build reverse suffix sort */
+    {   size_t pos;
+        for (pos=0; pos < bufferSize; pos++)
+            reverseSuffix[suffix[pos]] = (U32)pos;
+        /* note filePos tracks borders between samples.
+           It's not used at this stage, but planned to become useful in a later update */
+        filePos[0] = 0;
+        for (pos=1; pos<nbFiles; pos++)
+            filePos[pos] = (U32)(filePos[pos-1] + fileSizes[pos-1]);
+    }
+
+    DISPLAYLEVEL(2, "finding patterns ... \n");
+    DISPLAYLEVEL(3, "minimum ratio : %u \n", minRatio);
+
+    {   U32 cursor; for (cursor=0; cursor < bufferSize; ) {
+            dictItem solution;
+            if (doneMarks[cursor]) { cursor++; continue; }
+            solution = ZDICT_analyzePos(doneMarks, suffix, reverseSuffix[cursor], buffer, minRatio, notificationLevel);
+            if (solution.length==0) { cursor++; continue; }
+            ZDICT_insertDictItem(dictList, dictListSize, solution, buffer);
+            cursor += solution.length;
+            DISPLAYUPDATE(2, "\r%4.2f %% \r", (double)cursor / bufferSize * 100);
+    }   }
+
+_cleanup:
+    free(suffix0);
+    free(reverseSuffix);
+    free(doneMarks);
+    free(filePos);
+    return result;
+}
+
+
+static void ZDICT_fillNoise(void* buffer, size_t length)
+{
+    unsigned const prime1 = 2654435761U;
+    unsigned const prime2 = 2246822519U;
+    unsigned acc = prime1;
+    size_t p=0;;
+    for (p=0; p<length; p++) {
+        acc *= prime2;
+        ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21);
+    }
+}
+
+
+typedef struct
+{
+    ZSTD_CCtx* ref;    /* contains reference to dictionary */
+    ZSTD_CCtx* zc;     /* working context */
+    void* workPlace;   /* must be ZSTD_BLOCKSIZE_MAX allocated */
+} EStats_ress_t;
+
+#define MAXREPOFFSET 1024
+
+static void ZDICT_countEStats(EStats_ress_t esr, ZSTD_parameters params,
+                              U32* countLit, U32* offsetcodeCount, U32* matchlengthCount, U32* litlengthCount, U32* repOffsets,
+                              const void* src, size_t srcSize,
+                              U32 notificationLevel)
+{
+    size_t const blockSizeMax = MIN (ZSTD_BLOCKSIZE_MAX, 1 << params.cParams.windowLog);
+    size_t cSize;
+
+    if (srcSize > blockSizeMax) srcSize = blockSizeMax;   /* protection vs large samples */
+    {   size_t const errorCode = ZSTD_copyCCtx(esr.zc, esr.ref, 0);
+        if (ZSTD_isError(errorCode)) { DISPLAYLEVEL(1, "warning : ZSTD_copyCCtx failed \n"); return; }
+    }
+    cSize = ZSTD_compressBlock(esr.zc, esr.workPlace, ZSTD_BLOCKSIZE_MAX, src, srcSize);
+    if (ZSTD_isError(cSize)) { DISPLAYLEVEL(3, "warning : could not compress sample size %u \n", (U32)srcSize); return; }
+
+    if (cSize) {  /* if == 0; block is not compressible */
+        const seqStore_t* const seqStorePtr = ZSTD_getSeqStore(esr.zc);
+
+        /* literals stats */
+        {   const BYTE* bytePtr;
+            for(bytePtr = seqStorePtr->litStart; bytePtr < seqStorePtr->lit; bytePtr++)
+                countLit[*bytePtr]++;
+        }
+
+        /* seqStats */
+        {   U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+            ZSTD_seqToCodes(seqStorePtr);
+
+            {   const BYTE* codePtr = seqStorePtr->ofCode;
+                U32 u;
+                for (u=0; u<nbSeq; u++) offsetcodeCount[codePtr[u]]++;
+            }
+
+            {   const BYTE* codePtr = seqStorePtr->mlCode;
+                U32 u;
+                for (u=0; u<nbSeq; u++) matchlengthCount[codePtr[u]]++;
+            }
+
+            {   const BYTE* codePtr = seqStorePtr->llCode;
+                U32 u;
+                for (u=0; u<nbSeq; u++) litlengthCount[codePtr[u]]++;
+            }
+
+            if (nbSeq >= 2) { /* rep offsets */
+                const seqDef* const seq = seqStorePtr->sequencesStart;
+                U32 offset1 = seq[0].offset - 3;
+                U32 offset2 = seq[1].offset - 3;
+                if (offset1 >= MAXREPOFFSET) offset1 = 0;
+                if (offset2 >= MAXREPOFFSET) offset2 = 0;
+                repOffsets[offset1] += 3;
+                repOffsets[offset2] += 1;
+    }   }   }
+}
+
+static size_t ZDICT_totalSampleSize(const size_t* fileSizes, unsigned nbFiles)
+{
+    size_t total=0;
+    unsigned u;
+    for (u=0; u<nbFiles; u++) total += fileSizes[u];
+    return total;
+}
+
+typedef struct { U32 offset; U32 count; } offsetCount_t;
+
+static void ZDICT_insertSortCount(offsetCount_t table[ZSTD_REP_NUM+1], U32 val, U32 count)
+{
+    U32 u;
+    table[ZSTD_REP_NUM].offset = val;
+    table[ZSTD_REP_NUM].count = count;
+    for (u=ZSTD_REP_NUM; u>0; u--) {
+        offsetCount_t tmp;
+        if (table[u-1].count >= table[u].count) break;
+        tmp = table[u-1];
+        table[u-1] = table[u];
+        table[u] = tmp;
+    }
+}
+
+/* ZDICT_flatLit() :
+ * rewrite `countLit` to contain a mostly flat but still compressible distribution of literals.
+ * necessary to avoid generating a non-compressible distribution that HUF_writeCTable() cannot encode.
+ */
+static void ZDICT_flatLit(U32* countLit)
+{
+    int u;
+    for (u=1; u<256; u++) countLit[u] = 2;
+    countLit[0]   = 4;
+    countLit[253] = 1;
+    countLit[254] = 1;
+}
+
+#define OFFCODE_MAX 30  /* only applicable to first block */
+static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
+                                   unsigned compressionLevel,
+                             const void*  srcBuffer, const size_t* fileSizes, unsigned nbFiles,
+                             const void* dictBuffer, size_t  dictBufferSize,
+                                   unsigned notificationLevel)
+{
+    U32 countLit[256];
+    HUF_CREATE_STATIC_CTABLE(hufTable, 255);
+    U32 offcodeCount[OFFCODE_MAX+1];
+    short offcodeNCount[OFFCODE_MAX+1];
+    U32 offcodeMax = ZSTD_highbit32((U32)(dictBufferSize + 128 KB));
+    U32 matchLengthCount[MaxML+1];
+    short matchLengthNCount[MaxML+1];
+    U32 litLengthCount[MaxLL+1];
+    short litLengthNCount[MaxLL+1];
+    U32 repOffset[MAXREPOFFSET];
+    offsetCount_t bestRepOffset[ZSTD_REP_NUM+1];
+    EStats_ress_t esr;
+    ZSTD_parameters params;
+    U32 u, huffLog = 11, Offlog = OffFSELog, mlLog = MLFSELog, llLog = LLFSELog, total;
+    size_t pos = 0, errorCode;
+    size_t eSize = 0;
+    size_t const totalSrcSize = ZDICT_totalSampleSize(fileSizes, nbFiles);
+    size_t const averageSampleSize = totalSrcSize / (nbFiles + !nbFiles);
+    BYTE* dstPtr = (BYTE*)dstBuffer;
+
+    /* init */
+    DEBUGLOG(4, "ZDICT_analyzeEntropy");
+    esr.ref = ZSTD_createCCtx();
+    esr.zc = ZSTD_createCCtx();
+    esr.workPlace = malloc(ZSTD_BLOCKSIZE_MAX);
+    if (!esr.ref || !esr.zc || !esr.workPlace) {
+        eSize = ERROR(memory_allocation);
+        DISPLAYLEVEL(1, "Not enough memory \n");
+        goto _cleanup;
+    }
+    if (offcodeMax>OFFCODE_MAX) { eSize = ERROR(dictionaryCreation_failed); goto _cleanup; }   /* too large dictionary */
+    for (u=0; u<256; u++) countLit[u] = 1;   /* any character must be described */
+    for (u=0; u<=offcodeMax; u++) offcodeCount[u] = 1;
+    for (u=0; u<=MaxML; u++) matchLengthCount[u] = 1;
+    for (u=0; u<=MaxLL; u++) litLengthCount[u] = 1;
+    memset(repOffset, 0, sizeof(repOffset));
+    repOffset[1] = repOffset[4] = repOffset[8] = 1;
+    memset(bestRepOffset, 0, sizeof(bestRepOffset));
+    if (compressionLevel<=0) compressionLevel = g_compressionLevel_default;
+    params = ZSTD_getParams(compressionLevel, averageSampleSize, dictBufferSize);
+    {   size_t const beginResult = ZSTD_compressBegin_advanced(esr.ref, dictBuffer, dictBufferSize, params, 0);
+        if (ZSTD_isError(beginResult)) {
+            DISPLAYLEVEL(1, "error : ZSTD_compressBegin_advanced() failed : %s \n", ZSTD_getErrorName(beginResult));
+            eSize = ERROR(GENERIC);
+            goto _cleanup;
+    }   }
+
+    /* collect stats on all samples */
+    for (u=0; u<nbFiles; u++) {
+        ZDICT_countEStats(esr, params,
+                          countLit, offcodeCount, matchLengthCount, litLengthCount, repOffset,
+                         (const char*)srcBuffer + pos, fileSizes[u],
+                          notificationLevel);
+        pos += fileSizes[u];
+    }
+
+    /* analyze, build stats, starting with literals */
+    {   size_t maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
+        if (HUF_isError(maxNbBits)) {
+            eSize = ERROR(GENERIC);
+            DISPLAYLEVEL(1, " HUF_buildCTable error \n");
+            goto _cleanup;
+        }
+        if (maxNbBits==8) {  /* not compressible : will fail on HUF_writeCTable() */
+            DISPLAYLEVEL(2, "warning : pathological dataset : literals are not compressible : samples are noisy or too regular \n");
+            ZDICT_flatLit(countLit);  /* replace distribution by a fake "mostly flat but still compressible" distribution, that HUF_writeCTable() can encode */
+            maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
+            assert(maxNbBits==9);
+        }
+        huffLog = (U32)maxNbBits;
+    }
+
+    /* looking for most common first offsets */
+    {   U32 offset;
+        for (offset=1; offset<MAXREPOFFSET; offset++)
+            ZDICT_insertSortCount(bestRepOffset, offset, repOffset[offset]);
+    }
+    /* note : the result of this phase should be used to better appreciate the impact on statistics */
+
+    total=0; for (u=0; u<=offcodeMax; u++) total+=offcodeCount[u];
+    errorCode = FSE_normalizeCount(offcodeNCount, Offlog, offcodeCount, total, offcodeMax);
+    if (FSE_isError(errorCode)) {
+        eSize = ERROR(GENERIC);
+        DISPLAYLEVEL(1, "FSE_normalizeCount error with offcodeCount \n");
+        goto _cleanup;
+    }
+    Offlog = (U32)errorCode;
+
+    total=0; for (u=0; u<=MaxML; u++) total+=matchLengthCount[u];
+    errorCode = FSE_normalizeCount(matchLengthNCount, mlLog, matchLengthCount, total, MaxML);
+    if (FSE_isError(errorCode)) {
+        eSize = ERROR(GENERIC);
+        DISPLAYLEVEL(1, "FSE_normalizeCount error with matchLengthCount \n");
+        goto _cleanup;
+    }
+    mlLog = (U32)errorCode;
+
+    total=0; for (u=0; u<=MaxLL; u++) total+=litLengthCount[u];
+    errorCode = FSE_normalizeCount(litLengthNCount, llLog, litLengthCount, total, MaxLL);
+    if (FSE_isError(errorCode)) {
+        eSize = ERROR(GENERIC);
+        DISPLAYLEVEL(1, "FSE_normalizeCount error with litLengthCount \n");
+        goto _cleanup;
+    }
+    llLog = (U32)errorCode;
+
+    /* write result to buffer */
+    {   size_t const hhSize = HUF_writeCTable(dstPtr, maxDstSize, hufTable, 255, huffLog);
+        if (HUF_isError(hhSize)) {
+            eSize = ERROR(GENERIC);
+            DISPLAYLEVEL(1, "HUF_writeCTable error \n");
+            goto _cleanup;
+        }
+        dstPtr += hhSize;
+        maxDstSize -= hhSize;
+        eSize += hhSize;
+    }
+
+    {   size_t const ohSize = FSE_writeNCount(dstPtr, maxDstSize, offcodeNCount, OFFCODE_MAX, Offlog);
+        if (FSE_isError(ohSize)) {
+            eSize = ERROR(GENERIC);
+            DISPLAYLEVEL(1, "FSE_writeNCount error with offcodeNCount \n");
+            goto _cleanup;
+        }
+        dstPtr += ohSize;
+        maxDstSize -= ohSize;
+        eSize += ohSize;
+    }
+
+    {   size_t const mhSize = FSE_writeNCount(dstPtr, maxDstSize, matchLengthNCount, MaxML, mlLog);
+        if (FSE_isError(mhSize)) {
+            eSize = ERROR(GENERIC);
+            DISPLAYLEVEL(1, "FSE_writeNCount error with matchLengthNCount \n");
+            goto _cleanup;
+        }
+        dstPtr += mhSize;
+        maxDstSize -= mhSize;
+        eSize += mhSize;
+    }
+
+    {   size_t const lhSize = FSE_writeNCount(dstPtr, maxDstSize, litLengthNCount, MaxLL, llLog);
+        if (FSE_isError(lhSize)) {
+            eSize = ERROR(GENERIC);
+            DISPLAYLEVEL(1, "FSE_writeNCount error with litlengthNCount \n");
+            goto _cleanup;
+        }
+        dstPtr += lhSize;
+        maxDstSize -= lhSize;
+        eSize += lhSize;
+    }
+
+    if (maxDstSize<12) {
+        eSize = ERROR(GENERIC);
+        DISPLAYLEVEL(1, "not enough space to write RepOffsets \n");
+        goto _cleanup;
+    }
+# if 0
+    MEM_writeLE32(dstPtr+0, bestRepOffset[0].offset);
+    MEM_writeLE32(dstPtr+4, bestRepOffset[1].offset);
+    MEM_writeLE32(dstPtr+8, bestRepOffset[2].offset);
+#else
+    /* at this stage, we don't use the result of "most common first offset",
+       as the impact of statistics is not properly evaluated */
+    MEM_writeLE32(dstPtr+0, repStartValue[0]);
+    MEM_writeLE32(dstPtr+4, repStartValue[1]);
+    MEM_writeLE32(dstPtr+8, repStartValue[2]);
+#endif
+    eSize += 12;
+
+_cleanup:
+    ZSTD_freeCCtx(esr.ref);
+    ZSTD_freeCCtx(esr.zc);
+    free(esr.workPlace);
+
+    return eSize;
+}
+
+
+
+size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
+                          const void* customDictContent, size_t dictContentSize,
+                          const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                          ZDICT_params_t params)
+{
+    size_t hSize;
+#define HBUFFSIZE 256   /* should prove large enough for all entropy headers */
+    BYTE header[HBUFFSIZE];
+    int const compressionLevel = (params.compressionLevel <= 0) ? g_compressionLevel_default : params.compressionLevel;
+    U32 const notificationLevel = params.notificationLevel;
+
+    /* check conditions */
+    DEBUGLOG(4, "ZDICT_finalizeDictionary");
+    if (dictBufferCapacity < dictContentSize) return ERROR(dstSize_tooSmall);
+    if (dictContentSize < ZDICT_CONTENTSIZE_MIN) return ERROR(srcSize_wrong);
+    if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) return ERROR(dstSize_tooSmall);
+
+    /* dictionary header */
+    MEM_writeLE32(header, ZSTD_MAGIC_DICTIONARY);
+    {   U64 const randomID = XXH64(customDictContent, dictContentSize, 0);
+        U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
+        U32 const dictID = params.dictID ? params.dictID : compliantID;
+        MEM_writeLE32(header+4, dictID);
+    }
+    hSize = 8;
+
+    /* entropy tables */
+    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
+    DISPLAYLEVEL(2, "statistics ... \n");
+    {   size_t const eSize = ZDICT_analyzeEntropy(header+hSize, HBUFFSIZE-hSize,
+                                  compressionLevel,
+                                  samplesBuffer, samplesSizes, nbSamples,
+                                  customDictContent, dictContentSize,
+                                  notificationLevel);
+        if (ZDICT_isError(eSize)) return eSize;
+        hSize += eSize;
+    }
+
+    /* copy elements in final buffer ; note : src and dst buffer can overlap */
+    if (hSize + dictContentSize > dictBufferCapacity) dictContentSize = dictBufferCapacity - hSize;
+    {   size_t const dictSize = hSize + dictContentSize;
+        char* dictEnd = (char*)dictBuffer + dictSize;
+        memmove(dictEnd - dictContentSize, customDictContent, dictContentSize);
+        memcpy(dictBuffer, header, hSize);
+        return dictSize;
+    }
+}
+
+
+size_t ZDICT_addEntropyTablesFromBuffer_advanced(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
+                                                 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                                                 ZDICT_params_t params)
+{
+    int const compressionLevel = (params.compressionLevel <= 0) ? g_compressionLevel_default : params.compressionLevel;
+    U32 const notificationLevel = params.notificationLevel;
+    size_t hSize = 8;
+
+    /* calculate entropy tables */
+    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
+    DISPLAYLEVEL(2, "statistics ... \n");
+    {   size_t const eSize = ZDICT_analyzeEntropy((char*)dictBuffer+hSize, dictBufferCapacity-hSize,
+                                  compressionLevel,
+                                  samplesBuffer, samplesSizes, nbSamples,
+                                  (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize,
+                                  notificationLevel);
+        if (ZDICT_isError(eSize)) return eSize;
+        hSize += eSize;
+    }
+
+    /* add dictionary header (after entropy tables) */
+    MEM_writeLE32(dictBuffer, ZSTD_MAGIC_DICTIONARY);
+    {   U64 const randomID = XXH64((char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize, 0);
+        U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
+        U32 const dictID = params.dictID ? params.dictID : compliantID;
+        MEM_writeLE32((char*)dictBuffer+4, dictID);
+    }
+
+    if (hSize + dictContentSize < dictBufferCapacity)
+        memmove((char*)dictBuffer + hSize, (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize);
+    return MIN(dictBufferCapacity, hSize+dictContentSize);
+}
+
+
+/*! ZDICT_trainFromBuffer_unsafe_legacy() :
+*   Warning : `samplesBuffer` must be followed by noisy guard band.
+*   @return : size of dictionary, or an error code which can be tested with ZDICT_isError()
+*/
+size_t ZDICT_trainFromBuffer_unsafe_legacy(
+                            void* dictBuffer, size_t maxDictSize,
+                            const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                            ZDICT_legacy_params_t params)
+{
+    U32 const dictListSize = MAX(MAX(DICTLISTSIZE_DEFAULT, nbSamples), (U32)(maxDictSize/16));
+    dictItem* const dictList = (dictItem*)malloc(dictListSize * sizeof(*dictList));
+    unsigned const selectivity = params.selectivityLevel == 0 ? g_selectivity_default : params.selectivityLevel;
+    unsigned const minRep = (selectivity > 30) ? MINRATIO : nbSamples >> selectivity;
+    size_t const targetDictSize = maxDictSize;
+    size_t const samplesBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
+    size_t dictSize = 0;
+    U32 const notificationLevel = params.zParams.notificationLevel;
+
+    /* checks */
+    if (!dictList) return ERROR(memory_allocation);
+    if (maxDictSize < ZDICT_DICTSIZE_MIN) { free(dictList); return ERROR(dstSize_tooSmall); }   /* requested dictionary size is too small */
+    if (samplesBuffSize < ZDICT_MIN_SAMPLES_SIZE) { free(dictList); return ERROR(dictionaryCreation_failed); }   /* not enough source to create dictionary */
+
+    /* init */
+    ZDICT_initDictItem(dictList);
+
+    /* build dictionary */
+    ZDICT_trainBuffer_legacy(dictList, dictListSize,
+                       samplesBuffer, samplesBuffSize,
+                       samplesSizes, nbSamples,
+                       minRep, notificationLevel);
+
+    /* display best matches */
+    if (params.zParams.notificationLevel>= 3) {
+        U32 const nb = MIN(25, dictList[0].pos);
+        U32 const dictContentSize = ZDICT_dictSize(dictList);
+        U32 u;
+        DISPLAYLEVEL(3, "\n %u segments found, of total size %u \n", dictList[0].pos-1, dictContentSize);
+        DISPLAYLEVEL(3, "list %u best segments \n", nb-1);
+        for (u=1; u<nb; u++) {
+            U32 const pos = dictList[u].pos;
+            U32 const length = dictList[u].length;
+            U32 const printedLength = MIN(40, length);
+            if ((pos > samplesBuffSize) || ((pos + length) > samplesBuffSize))
+                return ERROR(GENERIC);   /* should never happen */
+            DISPLAYLEVEL(3, "%3u:%3u bytes at pos %8u, savings %7u bytes |",
+                         u, length, pos, dictList[u].savings);
+            ZDICT_printHex((const char*)samplesBuffer+pos, printedLength);
+            DISPLAYLEVEL(3, "| \n");
+    }   }
+
+
+    /* create dictionary */
+    {   U32 dictContentSize = ZDICT_dictSize(dictList);
+        if (dictContentSize < ZDICT_CONTENTSIZE_MIN) { free(dictList); return ERROR(dictionaryCreation_failed); }   /* dictionary content too small */
+        if (dictContentSize < targetDictSize/4) {
+            DISPLAYLEVEL(2, "!  warning : selected content significantly smaller than requested (%u < %u) \n", dictContentSize, (U32)maxDictSize);
+            if (samplesBuffSize < 10 * targetDictSize)
+                DISPLAYLEVEL(2, "!  consider increasing the number of samples (total size : %u MB)\n", (U32)(samplesBuffSize>>20));
+            if (minRep > MINRATIO) {
+                DISPLAYLEVEL(2, "!  consider increasing selectivity to produce larger dictionary (-s%u) \n", selectivity+1);
+                DISPLAYLEVEL(2, "!  note : larger dictionaries are not necessarily better, test its efficiency on samples \n");
+            }
+        }
+
+        if ((dictContentSize > targetDictSize*3) && (nbSamples > 2*MINRATIO) && (selectivity>1)) {
+            U32 proposedSelectivity = selectivity-1;
+            while ((nbSamples >> proposedSelectivity) <= MINRATIO) { proposedSelectivity--; }
+            DISPLAYLEVEL(2, "!  note : calculated dictionary significantly larger than requested (%u > %u) \n", dictContentSize, (U32)maxDictSize);
+            DISPLAYLEVEL(2, "!  consider increasing dictionary size, or produce denser dictionary (-s%u) \n", proposedSelectivity);
+            DISPLAYLEVEL(2, "!  always test dictionary efficiency on real samples \n");
+        }
+
+        /* limit dictionary size */
+        {   U32 const max = dictList->pos;   /* convention : nb of useful elts within dictList */
+            U32 currentSize = 0;
+            U32 n; for (n=1; n<max; n++) {
+                currentSize += dictList[n].length;
+                if (currentSize > targetDictSize) { currentSize -= dictList[n].length; break; }
+            }
+            dictList->pos = n;
+            dictContentSize = currentSize;
+        }
+
+        /* build dict content */
+        {   U32 u;
+            BYTE* ptr = (BYTE*)dictBuffer + maxDictSize;
+            for (u=1; u<dictList->pos; u++) {
+                U32 l = dictList[u].length;
+                ptr -= l;
+                if (ptr<(BYTE*)dictBuffer) { free(dictList); return ERROR(GENERIC); }   /* should not happen */
+                memcpy(ptr, (const char*)samplesBuffer+dictList[u].pos, l);
+        }   }
+
+        dictSize = ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, maxDictSize,
+                                                             samplesBuffer, samplesSizes, nbSamples,
+                                                             params.zParams);
+    }
+
+    /* clean up */
+    free(dictList);
+    return dictSize;
+}
+
+
+/* ZDICT_trainFromBuffer_legacy() :
+ * issue : samplesBuffer need to be followed by a noisy guard band.
+ * work around : duplicate the buffer, and add the noise */
+size_t ZDICT_trainFromBuffer_legacy(void* dictBuffer, size_t dictBufferCapacity,
+                              const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                              ZDICT_legacy_params_t params)
+{
+    size_t result;
+    void* newBuff;
+    size_t const sBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
+    if (sBuffSize < ZDICT_MIN_SAMPLES_SIZE) return 0;   /* not enough content => no dictionary */
+
+    newBuff = malloc(sBuffSize + NOISELENGTH);
+    if (!newBuff) return ERROR(memory_allocation);
+
+    memcpy(newBuff, samplesBuffer, sBuffSize);
+    ZDICT_fillNoise((char*)newBuff + sBuffSize, NOISELENGTH);   /* guard band, for end of buffer condition */
+
+    result =
+        ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, dictBufferCapacity, newBuff,
+                                            samplesSizes, nbSamples, params);
+    free(newBuff);
+    return result;
+}
+
+
+size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
+                             const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
+{
+    ZDICT_cover_params_t params;
+    DEBUGLOG(3, "ZDICT_trainFromBuffer");
+    memset(&params, 0, sizeof(params));
+    params.d = 8;
+    params.steps = 4;
+    /* Default to level 6 since no compression level information is available */
+    params.zParams.compressionLevel = 6;
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=1)
+    params.zParams.notificationLevel = ZSTD_DEBUG;
+#endif
+    return ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, dictBufferCapacity,
+                                               samplesBuffer, samplesSizes, nbSamples,
+                                               &params);
+}
+
+size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
+                                  const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
+{
+    ZDICT_params_t params;
+    memset(&params, 0, sizeof(params));
+    return ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, dictBufferCapacity,
+                                                     samplesBuffer, samplesSizes, nbSamples,
+                                                     params);
+}
diff --git a/vendor/github.com/DataDog/zstd/zdict.h b/vendor/github.com/DataDog/zstd/zdict.h
new file mode 100644
index 0000000..ad459c2
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zdict.h
@@ -0,0 +1,212 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef DICTBUILDER_H_001
+#define DICTBUILDER_H_001
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*======  Dependencies  ======*/
+#include <stddef.h>  /* size_t */
+
+
+/* =====   ZDICTLIB_API : control library symbols visibility   ===== */
+#ifndef ZDICTLIB_VISIBILITY
+#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#    define ZDICTLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  else
+#    define ZDICTLIB_VISIBILITY
+#  endif
+#endif
+#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBILITY
+#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#  define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define ZDICTLIB_API ZDICTLIB_VISIBILITY
+#endif
+
+
+/*! ZDICT_trainFromBuffer():
+ *  Train a dictionary from an array of samples.
+ *  Redirect towards ZDICT_optimizeTrainFromBuffer_cover() single-threaded, with d=8 and steps=4.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *  Note: ZDICT_trainFromBuffer() requires about 9 bytes of memory for each input byte.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ */
+ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
+                                    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples);
+
+
+/*======   Helper functions   ======*/
+ZDICTLIB_API unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize);  /**< extracts dictID; @return zero if error (not a valid dictionary) */
+ZDICTLIB_API unsigned ZDICT_isError(size_t errorCode);
+ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);
+
+
+
+#ifdef ZDICT_STATIC_LINKING_ONLY
+
+/* ====================================================================================
+ * The definitions in this section are considered experimental.
+ * They should never be used with a dynamic library, as they may change in the future.
+ * They are provided for advanced usages.
+ * Use them only in association with static linking.
+ * ==================================================================================== */
+
+typedef struct {
+    int      compressionLevel;   /* optimize for a specific zstd compression level; 0 means default */
+    unsigned notificationLevel;  /* Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
+    unsigned dictID;             /* force dictID value; 0 means auto mode (32-bits random value) */
+} ZDICT_params_t;
+
+/*! ZDICT_cover_params_t:
+ *  k and d are the only required parameters.
+ *  For others, value 0 means default.
+ */
+typedef struct {
+    unsigned k;                  /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
+    unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
+    unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (32) : Higher means more parameters checked */
+    unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
+    ZDICT_params_t zParams;
+} ZDICT_cover_params_t;
+
+
+/*! ZDICT_trainFromBuffer_cover():
+ *  Train a dictionary from an array of samples using the COVER algorithm.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *  Note: ZDICT_trainFromBuffer_cover() requires about 9 bytes of memory for each input byte.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ */
+ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
+          void *dictBuffer, size_t dictBufferCapacity,
+    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+          ZDICT_cover_params_t parameters);
+
+/*! ZDICT_optimizeTrainFromBuffer_cover():
+ * The same requirements as above hold for all the parameters except `parameters`.
+ * This function tries many parameter combinations and picks the best parameters.
+ * `*parameters` is filled with the best parameters found,
+ * dictionary constructed with those parameters is stored in `dictBuffer`.
+ *
+ * All of the parameters d, k, steps are optional.
+ * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8, 10, 12, 14, 16}.
+ * if steps is zero it defaults to its default value.
+ * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [16, 2048].
+ *
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *           or an error code, which can be tested with ZDICT_isError().
+ *           On success `*parameters` contains the parameters selected.
+ * Note: ZDICT_optimizeTrainFromBuffer_cover() requires about 8 bytes of memory for each input byte and additionally another 5 bytes of memory for each byte of memory for each thread.
+ */
+ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
+          void* dictBuffer, size_t dictBufferCapacity,
+    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+          ZDICT_cover_params_t* parameters);
+
+/*! ZDICT_finalizeDictionary():
+ * Given a custom content as a basis for dictionary, and a set of samples,
+ * finalize dictionary by adding headers and statistics.
+ *
+ * Samples must be stored concatenated in a flat buffer `samplesBuffer`,
+ * supplied with an array of sizes `samplesSizes`, providing the size of each sample in order.
+ *
+ * dictContentSize must be >= ZDICT_CONTENTSIZE_MIN bytes.
+ * maxDictSize must be >= dictContentSize, and must be >= ZDICT_DICTSIZE_MIN bytes.
+ *
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`),
+ *           or an error code, which can be tested by ZDICT_isError().
+ * Note: ZDICT_finalizeDictionary() will push notifications into stderr if instructed to, using notificationLevel>0.
+ * Note 2: dictBuffer and dictContent can overlap
+ */
+#define ZDICT_CONTENTSIZE_MIN 128
+#define ZDICT_DICTSIZE_MIN    256
+ZDICTLIB_API size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
+                                const void* dictContent, size_t dictContentSize,
+                                const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                                ZDICT_params_t parameters);
+
+typedef struct {
+    unsigned selectivityLevel;   /* 0 means default; larger => select more => larger dictionary */
+    ZDICT_params_t zParams;
+} ZDICT_legacy_params_t;
+
+/*! ZDICT_trainFromBuffer_legacy():
+ *  Train a dictionary from an array of samples.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * `parameters` is optional and can be provided with values set to 0 to mean "default".
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ *  Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0.
+ */
+ZDICTLIB_API size_t ZDICT_trainFromBuffer_legacy(
+    void *dictBuffer, size_t dictBufferCapacity,
+    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+    ZDICT_legacy_params_t parameters);
+
+/* Deprecation warnings */
+/* It is generally possible to disable deprecation warnings from compiler,
+   for example with -Wno-deprecated-declarations for gcc
+   or _CRT_SECURE_NO_WARNINGS in Visual.
+   Otherwise, it's also possible to manually define ZDICT_DISABLE_DEPRECATE_WARNINGS */
+#ifdef ZDICT_DISABLE_DEPRECATE_WARNINGS
+#  define ZDICT_DEPRECATED(message) ZDICTLIB_API   /* disable deprecation warnings */
+#else
+#  define ZDICT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZDICT_DEPRECATED(message) [[deprecated(message)]] ZDICTLIB_API
+#  elif (ZDICT_GCC_VERSION >= 405) || defined(__clang__)
+#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated(message)))
+#  elif (ZDICT_GCC_VERSION >= 301)
+#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZDICT_DEPRECATED for this compiler")
+#    define ZDICT_DEPRECATED(message) ZDICTLIB_API
+#  endif
+#endif /* ZDICT_DISABLE_DEPRECATE_WARNINGS */
+
+ZDICT_DEPRECATED("use ZDICT_finalizeDictionary() instead")
+size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
+                                  const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples);
+
+
+#endif   /* ZDICT_STATIC_LINKING_ONLY */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* DICTBUILDER_H_001 */
diff --git a/vendor/github.com/DataDog/zstd/zstd.go b/vendor/github.com/DataDog/zstd/zstd.go
new file mode 100644
index 0000000..e9953d6
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd.go
@@ -0,0 +1,144 @@
+package zstd
+
+/*
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+#include "stdint.h"  // for uintptr_t
+
+// The following *_wrapper function are used for removing superflouos
+// memory allocations when calling the wrapped functions from Go code.
+// See https://github.com/golang/go/issues/24450 for details.
+
+static size_t ZSTD_compress_wrapper(uintptr_t dst, size_t maxDstSize, const uintptr_t src, size_t srcSize, int compressionLevel) {
+	return ZSTD_compress((void*)dst, maxDstSize, (const void*)src, srcSize, compressionLevel);
+}
+
+static size_t ZSTD_decompress_wrapper(uintptr_t dst, size_t maxDstSize, uintptr_t src, size_t srcSize) {
+	return ZSTD_decompress((void*)dst, maxDstSize, (const void *)src, srcSize);
+}
+
+*/
+import "C"
+import (
+	"bytes"
+	"errors"
+	"io/ioutil"
+	"unsafe"
+)
+
+// Defines best and standard values for zstd cli
+const (
+	BestSpeed          = 1
+	BestCompression    = 20
+	DefaultCompression = 5
+)
+
+var (
+	// ErrEmptySlice is returned when there is nothing to compress
+	ErrEmptySlice = errors.New("Bytes slice is empty")
+)
+
+// CompressBound returns the worst case size needed for a destination buffer,
+// which can be used to preallocate a destination buffer or select a previously
+// allocated buffer from a pool.
+// See zstd.h to mirror implementation of ZSTD_COMPRESSBOUND
+func CompressBound(srcSize int) int {
+	lowLimit := 128 << 10 // 128 kB
+	var margin int
+	if srcSize < lowLimit {
+		margin = (lowLimit - srcSize) >> 11
+	}
+	return srcSize + (srcSize >> 8) + margin
+}
+
+// cCompressBound is a cgo call to check the go implementation above against the c code.
+func cCompressBound(srcSize int) int {
+	return int(C.ZSTD_compressBound(C.size_t(srcSize)))
+}
+
+// Compress src into dst.  If you have a buffer to use, you can pass it to
+// prevent allocation.  If it is too small, or if nil is passed, a new buffer
+// will be allocated and returned.
+func Compress(dst, src []byte) ([]byte, error) {
+	return CompressLevel(dst, src, DefaultCompression)
+}
+
+// CompressLevel is the same as Compress but you can pass a compression level
+func CompressLevel(dst, src []byte, level int) ([]byte, error) {
+	bound := CompressBound(len(src))
+	if cap(dst) >= bound {
+		dst = dst[0:bound] // Reuse dst buffer
+	} else {
+		dst = make([]byte, bound)
+	}
+
+	srcPtr := C.uintptr_t(uintptr(0)) // Do not point anywhere, if src is empty
+	if len(src) > 0 {
+		srcPtr = C.uintptr_t(uintptr(unsafe.Pointer(&src[0])))
+	}
+
+	cWritten := C.ZSTD_compress_wrapper(
+		C.uintptr_t(uintptr(unsafe.Pointer(&dst[0]))),
+		C.size_t(len(dst)),
+		srcPtr,
+		C.size_t(len(src)),
+		C.int(level))
+
+	written := int(cWritten)
+	// Check if the return is an Error code
+	if err := getError(written); err != nil {
+		return nil, err
+	}
+	return dst[:written], nil
+}
+
+// Decompress src into dst.  If you have a buffer to use, you can pass it to
+// prevent allocation.  If it is too small, or if nil is passed, a new buffer
+// will be allocated and returned.
+func Decompress(dst, src []byte) ([]byte, error) {
+	if len(src) == 0 {
+		return []byte{}, ErrEmptySlice
+	}
+	decompress := func(dst, src []byte) ([]byte, error) {
+
+		cWritten := C.ZSTD_decompress_wrapper(
+			C.uintptr_t(uintptr(unsafe.Pointer(&dst[0]))),
+			C.size_t(len(dst)),
+			C.uintptr_t(uintptr(unsafe.Pointer(&src[0]))),
+			C.size_t(len(src)))
+
+		written := int(cWritten)
+		// Check error
+		if err := getError(written); err != nil {
+			return nil, err
+		}
+		return dst[:written], nil
+	}
+
+	if dst == nil {
+		// Attempt to use zStd to determine decompressed size (may result in error or 0)
+		size := int(C.size_t(C.ZSTD_getDecompressedSize(unsafe.Pointer(&src[0]), C.size_t(len(src)))))
+
+		if err := getError(size); err != nil {
+			return nil, err
+		}
+
+		if size > 0 {
+			dst = make([]byte, size)
+		} else {
+			dst = make([]byte, len(src)*3) // starting guess
+		}
+	}
+	for i := 0; i < 3; i++ { // 3 tries to allocate a bigger buffer
+		result, err := decompress(dst, src)
+		if !IsDstSizeTooSmallError(err) {
+			return result, err
+		}
+		dst = make([]byte, len(dst)*2) // Grow buffer by 2
+	}
+
+	// We failed getting a dst buffer of correct size, use stream API
+	r := NewReader(bytes.NewReader(src))
+	defer r.Close()
+	return ioutil.ReadAll(r)
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd.h b/vendor/github.com/DataDog/zstd/zstd.h
new file mode 100644
index 0000000..6405da6
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd.h
@@ -0,0 +1,1399 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef ZSTD_H_235446
+#define ZSTD_H_235446
+
+/* ======   Dependency   ======*/
+#include <stddef.h>   /* size_t */
+
+
+/* =====   ZSTDLIB_API : control library symbols visibility   ===== */
+#ifndef ZSTDLIB_VISIBILITY
+#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#    define ZSTDLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  else
+#    define ZSTDLIB_VISIBILITY
+#  endif
+#endif
+#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBILITY
+#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#  define ZSTDLIB_API __declspec(dllimport) ZSTDLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define ZSTDLIB_API ZSTDLIB_VISIBILITY
+#endif
+
+
+/*******************************************************************************************************
+  Introduction
+
+  zstd, short for Zstandard, is a fast lossless compression algorithm,
+  targeting real-time compression scenarios at zlib-level and better compression ratios.
+  The zstd compression library provides in-memory compression and decompression functions.
+  The library supports compression levels from 1 up to ZSTD_maxCLevel() which is currently 22.
+  Levels >= 20, labeled `--ultra`, should be used with caution, as they require more memory.
+  Compression can be done in:
+    - a single step (described as Simple API)
+    - a single step, reusing a context (described as Explicit context)
+    - unbounded multiple steps (described as Streaming compression)
+  The compression ratio achievable on small data can be highly improved using a dictionary in:
+    - a single step (described as Simple dictionary API)
+    - a single step, reusing a dictionary (described as Bulk-processing dictionary API)
+
+  Advanced experimental functions can be accessed using #define ZSTD_STATIC_LINKING_ONLY before including zstd.h.
+  Advanced experimental APIs shall never be used with a dynamic library.
+  They are not "stable", their definition may change in the future. Only static linking is allowed.
+*********************************************************************************************************/
+
+/*------   Version   ------*/
+#define ZSTD_VERSION_MAJOR    1
+#define ZSTD_VERSION_MINOR    3
+#define ZSTD_VERSION_RELEASE  4
+
+#define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
+ZSTDLIB_API unsigned ZSTD_versionNumber(void);   /**< useful to check dll version */
+
+#define ZSTD_LIB_VERSION ZSTD_VERSION_MAJOR.ZSTD_VERSION_MINOR.ZSTD_VERSION_RELEASE
+#define ZSTD_QUOTE(str) #str
+#define ZSTD_EXPAND_AND_QUOTE(str) ZSTD_QUOTE(str)
+#define ZSTD_VERSION_STRING ZSTD_EXPAND_AND_QUOTE(ZSTD_LIB_VERSION)
+ZSTDLIB_API const char* ZSTD_versionString(void);   /* added in v1.3.0 */
+
+
+/***************************************
+*  Simple API
+***************************************/
+/*! ZSTD_compress() :
+ *  Compresses `src` content as a single zstd compressed frame into already allocated `dst`.
+ *  Hint : compression runs faster if `dstCapacity` >=  `ZSTD_compressBound(srcSize)`.
+ *  @return : compressed size written into `dst` (<= `dstCapacity),
+ *            or an error code if it fails (which can be tested using ZSTD_isError()). */
+ZSTDLIB_API size_t ZSTD_compress( void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize,
+                                  int compressionLevel);
+
+/*! ZSTD_decompress() :
+ *  `compressedSize` : must be the _exact_ size of some number of compressed and/or skippable frames.
+ *  `dstCapacity` is an upper bound of originalSize to regenerate.
+ *  If user cannot imply a maximum upper bound, it's better to use streaming mode to decompress data.
+ *  @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
+ *            or an errorCode if it fails (which can be tested using ZSTD_isError()). */
+ZSTDLIB_API size_t ZSTD_decompress( void* dst, size_t dstCapacity,
+                              const void* src, size_t compressedSize);
+
+/*! ZSTD_getFrameContentSize() : added in v1.3.0
+ *  `src` should point to the start of a ZSTD encoded frame.
+ *  `srcSize` must be at least as large as the frame header.
+ *            hint : any size >= `ZSTD_frameHeaderSize_max` is large enough.
+ *  @return : - decompressed size of the frame in `src`, if known
+ *            - ZSTD_CONTENTSIZE_UNKNOWN if the size cannot be determined
+ *            - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small)
+ *   note 1 : a 0 return value means the frame is valid but "empty".
+ *   note 2 : decompressed size is an optional field, it may not be present, typically in streaming mode.
+ *            When `return==ZSTD_CONTENTSIZE_UNKNOWN`, data to decompress could be any size.
+ *            In which case, it's necessary to use streaming mode to decompress data.
+ *            Optionally, application can rely on some implicit limit,
+ *            as ZSTD_decompress() only needs an upper bound of decompressed size.
+ *            (For example, data could be necessarily cut into blocks <= 16 KB).
+ *   note 3 : decompressed size is always present when compression is done with ZSTD_compress()
+ *   note 4 : decompressed size can be very large (64-bits value),
+ *            potentially larger than what local system can handle as a single memory segment.
+ *            In which case, it's necessary to use streaming mode to decompress data.
+ *   note 5 : If source is untrusted, decompressed size could be wrong or intentionally modified.
+ *            Always ensure return value fits within application's authorized limits.
+ *            Each application can set its own limits.
+ *   note 6 : This function replaces ZSTD_getDecompressedSize() */
+#define ZSTD_CONTENTSIZE_UNKNOWN (0ULL - 1)
+#define ZSTD_CONTENTSIZE_ERROR   (0ULL - 2)
+ZSTDLIB_API unsigned long long ZSTD_getFrameContentSize(const void *src, size_t srcSize);
+
+/*! ZSTD_getDecompressedSize() :
+ *  NOTE: This function is now obsolete, in favor of ZSTD_getFrameContentSize().
+ *  Both functions work the same way, but ZSTD_getDecompressedSize() blends
+ *  "empty", "unknown" and "error" results to the same return value (0),
+ *  while ZSTD_getFrameContentSize() gives them separate return values.
+ * `src` is the start of a zstd compressed frame.
+ * @return : content size to be decompressed, as a 64-bits value _if known and not empty_, 0 otherwise. */
+ZSTDLIB_API unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
+
+
+/*======  Helper functions  ======*/
+#define ZSTD_COMPRESSBOUND(srcSize)   ((srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
+ZSTDLIB_API size_t      ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
+ZSTDLIB_API unsigned    ZSTD_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
+ZSTDLIB_API const char* ZSTD_getErrorName(size_t code);     /*!< provides readable string from an error code */
+ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compression level available */
+
+
+/***************************************
+*  Explicit context
+***************************************/
+/*= Compression context
+ *  When compressing many times,
+ *  it is recommended to allocate a context just once, and re-use it for each successive compression operation.
+ *  This will make workload friendlier for system's memory.
+ *  Use one context per thread for parallel execution in multi-threaded environments. */
+typedef struct ZSTD_CCtx_s ZSTD_CCtx;
+ZSTDLIB_API ZSTD_CCtx* ZSTD_createCCtx(void);
+ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);
+
+/*! ZSTD_compressCCtx() :
+ *  Same as ZSTD_compress(), requires an allocated ZSTD_CCtx (see ZSTD_createCCtx()). */
+ZSTDLIB_API size_t ZSTD_compressCCtx(ZSTD_CCtx* ctx,
+                                     void* dst, size_t dstCapacity,
+                               const void* src, size_t srcSize,
+                                     int compressionLevel);
+
+/*= Decompression context
+ *  When decompressing many times,
+ *  it is recommended to allocate a context only once,
+ *  and re-use it for each successive compression operation.
+ *  This will make workload friendlier for system's memory.
+ *  Use one context per thread for parallel execution. */
+typedef struct ZSTD_DCtx_s ZSTD_DCtx;
+ZSTDLIB_API ZSTD_DCtx* ZSTD_createDCtx(void);
+ZSTDLIB_API size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);
+
+/*! ZSTD_decompressDCtx() :
+ *  Same as ZSTD_decompress(), requires an allocated ZSTD_DCtx (see ZSTD_createDCtx()) */
+ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* ctx,
+                                       void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize);
+
+
+/**************************
+*  Simple dictionary API
+***************************/
+/*! ZSTD_compress_usingDict() :
+ *  Compression using a predefined Dictionary (see dictBuilder/zdict.h).
+ *  Note : This function loads the dictionary, resulting in significant startup delay.
+ *  Note : When `dict == NULL || dictSize < 8` no dictionary is used. */
+ZSTDLIB_API size_t ZSTD_compress_usingDict(ZSTD_CCtx* ctx,
+                                           void* dst, size_t dstCapacity,
+                                     const void* src, size_t srcSize,
+                                     const void* dict,size_t dictSize,
+                                           int compressionLevel);
+
+/*! ZSTD_decompress_usingDict() :
+ *  Decompression using a predefined Dictionary (see dictBuilder/zdict.h).
+ *  Dictionary must be identical to the one used during compression.
+ *  Note : This function loads the dictionary, resulting in significant startup delay.
+ *  Note : When `dict == NULL || dictSize < 8` no dictionary is used. */
+ZSTDLIB_API size_t ZSTD_decompress_usingDict(ZSTD_DCtx* dctx,
+                                             void* dst, size_t dstCapacity,
+                                       const void* src, size_t srcSize,
+                                       const void* dict,size_t dictSize);
+
+
+/**********************************
+ *  Bulk processing dictionary API
+ *********************************/
+typedef struct ZSTD_CDict_s ZSTD_CDict;
+
+/*! ZSTD_createCDict() :
+ *  When compressing multiple messages / blocks with the same dictionary, it's recommended to load it just once.
+ *  ZSTD_createCDict() will create a digested dictionary, ready to start future compression operations without startup delay.
+ *  ZSTD_CDict can be created once and shared by multiple threads concurrently, since its usage is read-only.
+ *  `dictBuffer` can be released after ZSTD_CDict creation, since its content is copied within CDict */
+ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict(const void* dictBuffer, size_t dictSize,
+                                         int compressionLevel);
+
+/*! ZSTD_freeCDict() :
+ *  Function frees memory allocated by ZSTD_createCDict(). */
+ZSTDLIB_API size_t      ZSTD_freeCDict(ZSTD_CDict* CDict);
+
+/*! ZSTD_compress_usingCDict() :
+ *  Compression using a digested Dictionary.
+ *  Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
+ *  Note that compression level is decided during dictionary creation.
+ *  Frame parameters are hardcoded (dictID=yes, contentSize=yes, checksum=no) */
+ZSTDLIB_API size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx,
+                                            void* dst, size_t dstCapacity,
+                                      const void* src, size_t srcSize,
+                                      const ZSTD_CDict* cdict);
+
+
+typedef struct ZSTD_DDict_s ZSTD_DDict;
+
+/*! ZSTD_createDDict() :
+ *  Create a digested dictionary, ready to start decompression operation without startup delay.
+ *  dictBuffer can be released after DDict creation, as its content is copied inside DDict */
+ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict(const void* dictBuffer, size_t dictSize);
+
+/*! ZSTD_freeDDict() :
+ *  Function frees memory allocated with ZSTD_createDDict() */
+ZSTDLIB_API size_t      ZSTD_freeDDict(ZSTD_DDict* ddict);
+
+/*! ZSTD_decompress_usingDDict() :
+ *  Decompression using a digested Dictionary.
+ *  Faster startup than ZSTD_decompress_usingDict(), recommended when same dictionary is used multiple times. */
+ZSTDLIB_API size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
+                                              void* dst, size_t dstCapacity,
+                                        const void* src, size_t srcSize,
+                                        const ZSTD_DDict* ddict);
+
+
+/****************************
+*  Streaming
+****************************/
+
+typedef struct ZSTD_inBuffer_s {
+  const void* src;    /**< start of input buffer */
+  size_t size;        /**< size of input buffer */
+  size_t pos;         /**< position where reading stopped. Will be updated. Necessarily 0 <= pos <= size */
+} ZSTD_inBuffer;
+
+typedef struct ZSTD_outBuffer_s {
+  void*  dst;         /**< start of output buffer */
+  size_t size;        /**< size of output buffer */
+  size_t pos;         /**< position where writing stopped. Will be updated. Necessarily 0 <= pos <= size */
+} ZSTD_outBuffer;
+
+
+
+/*-***********************************************************************
+*  Streaming compression - HowTo
+*
+*  A ZSTD_CStream object is required to track streaming operation.
+*  Use ZSTD_createCStream() and ZSTD_freeCStream() to create/release resources.
+*  ZSTD_CStream objects can be reused multiple times on consecutive compression operations.
+*  It is recommended to re-use ZSTD_CStream in situations where many streaming operations will be achieved consecutively,
+*  since it will play nicer with system's memory, by re-using already allocated memory.
+*  Use one separate ZSTD_CStream per thread for parallel execution.
+*
+*  Start a new compression by initializing ZSTD_CStream.
+*  Use ZSTD_initCStream() to start a new compression operation.
+*  Use ZSTD_initCStream_usingDict() or ZSTD_initCStream_usingCDict() for a compression which requires a dictionary (experimental section)
+*
+*  Use ZSTD_compressStream() repetitively to consume input stream.
+*  The function will automatically update both `pos` fields.
+*  Note that it may not consume the entire input, in which case `pos < size`,
+*  and it's up to the caller to present again remaining data.
+*  @return : a size hint, preferred nb of bytes to use as input for next function call
+*            or an error code, which can be tested using ZSTD_isError().
+*            Note 1 : it's just a hint, to help latency a little, any other value will work fine.
+*            Note 2 : size hint is guaranteed to be <= ZSTD_CStreamInSize()
+*
+*  At any moment, it's possible to flush whatever data remains within internal buffer, using ZSTD_flushStream().
+*  `output->pos` will be updated.
+*  Note that some content might still be left within internal buffer if `output->size` is too small.
+*  @return : nb of bytes still present within internal buffer (0 if it's empty)
+*            or an error code, which can be tested using ZSTD_isError().
+*
+*  ZSTD_endStream() instructs to finish a frame.
+*  It will perform a flush and write frame epilogue.
+*  The epilogue is required for decoders to consider a frame completed.
+*  ZSTD_endStream() may not be able to flush full data if `output->size` is too small.
+*  In which case, call again ZSTD_endStream() to complete the flush.
+*  @return : 0 if frame fully completed and fully flushed,
+             or >0 if some data is still present within internal buffer
+                  (value is minimum size estimation for remaining data to flush, but it could be more)
+*            or an error code, which can be tested using ZSTD_isError().
+*
+* *******************************************************************/
+
+typedef ZSTD_CCtx ZSTD_CStream;  /**< CCtx and CStream are now effectively same object (>= v1.3.0) */
+                                 /* Continue to distinguish them for compatibility with versions <= v1.2.0 */
+/*===== ZSTD_CStream management functions =====*/
+ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream(void);
+ZSTDLIB_API size_t ZSTD_freeCStream(ZSTD_CStream* zcs);
+
+/*===== Streaming compression functions =====*/
+ZSTDLIB_API size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel);
+ZSTDLIB_API size_t ZSTD_compressStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
+ZSTDLIB_API size_t ZSTD_flushStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output);
+ZSTDLIB_API size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output);
+
+ZSTDLIB_API size_t ZSTD_CStreamInSize(void);    /**< recommended size for input buffer */
+ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output buffer. Guarantee to successfully flush at least one complete compressed block in all circumstances. */
+
+
+
+/*-***************************************************************************
+*  Streaming decompression - HowTo
+*
+*  A ZSTD_DStream object is required to track streaming operations.
+*  Use ZSTD_createDStream() and ZSTD_freeDStream() to create/release resources.
+*  ZSTD_DStream objects can be re-used multiple times.
+*
+*  Use ZSTD_initDStream() to start a new decompression operation,
+*   or ZSTD_initDStream_usingDict() if decompression requires a dictionary.
+*   @return : recommended first input size
+*
+*  Use ZSTD_decompressStream() repetitively to consume your input.
+*  The function will update both `pos` fields.
+*  If `input.pos < input.size`, some input has not been consumed.
+*  It's up to the caller to present again remaining data.
+*  If `output.pos < output.size`, decoder has flushed everything it could.
+*  @return : 0 when a frame is completely decoded and fully flushed,
+*            an error code, which can be tested using ZSTD_isError(),
+*            any other value > 0, which means there is still some decoding to do to complete current frame.
+*            The return value is a suggested next input size (a hint to improve latency) that will never load more than the current frame.
+* *******************************************************************************/
+
+typedef ZSTD_DCtx ZSTD_DStream;  /**< DCtx and DStream are now effectively same object (>= v1.3.0) */
+                                 /* For compatibility with versions <= v1.2.0, continue to consider them separated. */
+/*===== ZSTD_DStream management functions =====*/
+ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream(void);
+ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);
+
+/*===== Streaming decompression functions =====*/
+ZSTDLIB_API size_t ZSTD_initDStream(ZSTD_DStream* zds);
+ZSTDLIB_API size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
+
+ZSTDLIB_API size_t ZSTD_DStreamInSize(void);    /*!< recommended size for input buffer */
+ZSTDLIB_API size_t ZSTD_DStreamOutSize(void);   /*!< recommended size for output buffer. Guarantee to successfully flush at least one complete block in all circumstances. */
+
+#endif  /* ZSTD_H_235446 */
+
+
+
+/****************************************************************************************
+ * START OF ADVANCED AND EXPERIMENTAL FUNCTIONS
+ * The definitions in this section are considered experimental.
+ * They should never be used with a dynamic library, as prototypes may change in the future.
+ * They are provided for advanced scenarios.
+ * Use them only in association with static linking.
+ * ***************************************************************************************/
+
+#if defined(ZSTD_STATIC_LINKING_ONLY) && !defined(ZSTD_H_ZSTD_STATIC_LINKING_ONLY)
+#define ZSTD_H_ZSTD_STATIC_LINKING_ONLY
+
+/* --- Constants ---*/
+#define ZSTD_MAGICNUMBER            0xFD2FB528   /* >= v0.8.0 */
+#define ZSTD_MAGIC_SKIPPABLE_START  0x184D2A50U
+#define ZSTD_MAGIC_DICTIONARY       0xEC30A437   /* >= v0.7.0 */
+
+#define ZSTD_WINDOWLOG_MAX_32   30
+#define ZSTD_WINDOWLOG_MAX_64   31
+#define ZSTD_WINDOWLOG_MAX    ((unsigned)(sizeof(size_t) == 4 ? ZSTD_WINDOWLOG_MAX_32 : ZSTD_WINDOWLOG_MAX_64))
+#define ZSTD_WINDOWLOG_MIN      10
+#define ZSTD_HASHLOG_MAX      ((ZSTD_WINDOWLOG_MAX < 30) ? ZSTD_WINDOWLOG_MAX : 30)
+#define ZSTD_HASHLOG_MIN         6
+#define ZSTD_CHAINLOG_MAX_32    29
+#define ZSTD_CHAINLOG_MAX_64    30
+#define ZSTD_CHAINLOG_MAX     ((unsigned)(sizeof(size_t) == 4 ? ZSTD_CHAINLOG_MAX_32 : ZSTD_CHAINLOG_MAX_64))
+#define ZSTD_CHAINLOG_MIN       ZSTD_HASHLOG_MIN
+#define ZSTD_HASHLOG3_MAX       17
+#define ZSTD_SEARCHLOG_MAX     (ZSTD_WINDOWLOG_MAX-1)
+#define ZSTD_SEARCHLOG_MIN       1
+#define ZSTD_SEARCHLENGTH_MAX    7   /* only for ZSTD_fast, other strategies are limited to 6 */
+#define ZSTD_SEARCHLENGTH_MIN    3   /* only for ZSTD_btopt, other strategies are limited to 4 */
+#define ZSTD_TARGETLENGTH_MIN    1   /* only used by btopt, btultra and btfast */
+#define ZSTD_LDM_MINMATCH_MIN    4
+#define ZSTD_LDM_MINMATCH_MAX 4096
+#define ZSTD_LDM_BUCKETSIZELOG_MAX 8
+
+#define ZSTD_FRAMEHEADERSIZE_PREFIX 5   /* minimum input size to know frame header size */
+#define ZSTD_FRAMEHEADERSIZE_MIN    6
+#define ZSTD_FRAMEHEADERSIZE_MAX   18   /* for static allocation */
+static const size_t ZSTD_frameHeaderSize_prefix = ZSTD_FRAMEHEADERSIZE_PREFIX;
+static const size_t ZSTD_frameHeaderSize_min = ZSTD_FRAMEHEADERSIZE_MIN;
+static const size_t ZSTD_frameHeaderSize_max = ZSTD_FRAMEHEADERSIZE_MAX;
+static const size_t ZSTD_skippableHeaderSize = 8;  /* magic number + skippable frame length */
+
+
+/*--- Advanced types ---*/
+typedef enum { ZSTD_fast=1, ZSTD_dfast, ZSTD_greedy, ZSTD_lazy, ZSTD_lazy2,
+               ZSTD_btlazy2, ZSTD_btopt, ZSTD_btultra } ZSTD_strategy;   /* from faster to stronger */
+
+typedef struct {
+    unsigned windowLog;      /**< largest match distance : larger == more compression, more memory needed during decompression */
+    unsigned chainLog;       /**< fully searched segment : larger == more compression, slower, more memory (useless for fast) */
+    unsigned hashLog;        /**< dispatch table : larger == faster, more memory */
+    unsigned searchLog;      /**< nb of searches : larger == more compression, slower */
+    unsigned searchLength;   /**< match length searched : larger == faster decompression, sometimes less compression */
+    unsigned targetLength;   /**< acceptable match size for optimal parser (only) : larger == more compression, slower */
+    ZSTD_strategy strategy;
+} ZSTD_compressionParameters;
+
+typedef struct {
+    unsigned contentSizeFlag; /**< 1: content size will be in frame header (when known) */
+    unsigned checksumFlag;    /**< 1: generate a 32-bits checksum at end of frame, for error detection */
+    unsigned noDictIDFlag;    /**< 1: no dictID will be saved into frame header (if dictionary compression) */
+} ZSTD_frameParameters;
+
+typedef struct {
+    ZSTD_compressionParameters cParams;
+    ZSTD_frameParameters fParams;
+} ZSTD_parameters;
+
+typedef struct ZSTD_CCtx_params_s ZSTD_CCtx_params;
+
+typedef enum {
+    ZSTD_dct_auto=0,      /* dictionary is "full" when starting with ZSTD_MAGIC_DICTIONARY, otherwise it is "rawContent" */
+    ZSTD_dct_rawContent,  /* ensures dictionary is always loaded as rawContent, even if it starts with ZSTD_MAGIC_DICTIONARY */
+    ZSTD_dct_fullDict     /* refuses to load a dictionary if it does not respect Zstandard's specification */
+} ZSTD_dictContentType_e;
+
+typedef enum {
+    ZSTD_dlm_byCopy = 0, /**< Copy dictionary content internally */
+    ZSTD_dlm_byRef,      /**< Reference dictionary content -- the dictionary buffer must outlive its users. */
+} ZSTD_dictLoadMethod_e;
+
+
+
+/***************************************
+*  Frame size functions
+***************************************/
+
+/*! ZSTD_findFrameCompressedSize() :
+ *  `src` should point to the start of a ZSTD encoded frame or skippable frame
+ *  `srcSize` must be >= first frame size
+ *  @return : the compressed size of the first frame starting at `src`,
+ *            suitable to pass to `ZSTD_decompress` or similar,
+ *            or an error code if input is invalid */
+ZSTDLIB_API size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize);
+
+/*! ZSTD_findDecompressedSize() :
+ *  `src` should point the start of a series of ZSTD encoded and/or skippable frames
+ *  `srcSize` must be the _exact_ size of this series
+ *       (i.e. there should be a frame boundary exactly at `srcSize` bytes after `src`)
+ *  @return : - decompressed size of all data in all successive frames
+ *            - if the decompressed size cannot be determined: ZSTD_CONTENTSIZE_UNKNOWN
+ *            - if an error occurred: ZSTD_CONTENTSIZE_ERROR
+ *
+ *   note 1 : decompressed size is an optional field, that may not be present, especially in streaming mode.
+ *            When `return==ZSTD_CONTENTSIZE_UNKNOWN`, data to decompress could be any size.
+ *            In which case, it's necessary to use streaming mode to decompress data.
+ *   note 2 : decompressed size is always present when compression is done with ZSTD_compress()
+ *   note 3 : decompressed size can be very large (64-bits value),
+ *            potentially larger than what local system can handle as a single memory segment.
+ *            In which case, it's necessary to use streaming mode to decompress data.
+ *   note 4 : If source is untrusted, decompressed size could be wrong or intentionally modified.
+ *            Always ensure result fits within application's authorized limits.
+ *            Each application can set its own limits.
+ *   note 5 : ZSTD_findDecompressedSize handles multiple frames, and so it must traverse the input to
+ *            read each contained frame header.  This is fast as most of the data is skipped,
+ *            however it does mean that all frame data must be present and valid. */
+ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize);
+
+/*! ZSTD_frameHeaderSize() :
+*   `src` should point to the start of a ZSTD frame
+*   `srcSize` must be >= ZSTD_frameHeaderSize_prefix.
+*   @return : size of the Frame Header */
+ZSTDLIB_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
+
+
+/***************************************
+*  Memory management
+***************************************/
+
+/*! ZSTD_sizeof_*() :
+ *  These functions give the current memory usage of selected object.
+ *  Object memory usage can evolve when re-used. */
+ZSTDLIB_API size_t ZSTD_sizeof_CCtx(const ZSTD_CCtx* cctx);
+ZSTDLIB_API size_t ZSTD_sizeof_DCtx(const ZSTD_DCtx* dctx);
+ZSTDLIB_API size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs);
+ZSTDLIB_API size_t ZSTD_sizeof_DStream(const ZSTD_DStream* zds);
+ZSTDLIB_API size_t ZSTD_sizeof_CDict(const ZSTD_CDict* cdict);
+ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
+
+/*! ZSTD_estimate*() :
+ *  These functions make it possible to estimate memory usage
+ *  of a future {D,C}Ctx, before its creation.
+ *  ZSTD_estimateCCtxSize() will provide a budget large enough for any compression level up to selected one.
+ *  It will also consider src size to be arbitrarily "large", which is worst case.
+ *  If srcSize is known to always be small, ZSTD_estimateCCtxSize_usingCParams() can provide a tighter estimation.
+ *  ZSTD_estimateCCtxSize_usingCParams() can be used in tandem with ZSTD_getCParams() to create cParams from compressionLevel.
+ *  ZSTD_estimateCCtxSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbWorkers is >= 1.
+ *  Note : CCtx size estimation is only correct for single-threaded compression. */
+ZSTDLIB_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
+ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
+
+/*! ZSTD_estimateCStreamSize() :
+ *  ZSTD_estimateCStreamSize() will provide a budget large enough for any compression level up to selected one.
+ *  It will also consider src size to be arbitrarily "large", which is worst case.
+ *  If srcSize is known to always be small, ZSTD_estimateCStreamSize_usingCParams() can provide a tighter estimation.
+ *  ZSTD_estimateCStreamSize_usingCParams() can be used in tandem with ZSTD_getCParams() to create cParams from compressionLevel.
+ *  ZSTD_estimateCStreamSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParam_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_p_nbWorkers is >= 1.
+ *  Note : CStream size estimation is only correct for single-threaded compression.
+ *  ZSTD_DStream memory budget depends on window Size.
+ *  This information can be passed manually, using ZSTD_estimateDStreamSize,
+ *  or deducted from a valid frame Header, using ZSTD_estimateDStreamSize_fromFrame();
+ *  Note : if streaming is init with function ZSTD_init?Stream_usingDict(),
+ *         an internal ?Dict will be created, which additional size is not estimated here.
+ *         In this case, get total size by adding ZSTD_estimate?DictSize */
+ZSTDLIB_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
+ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
+ZSTDLIB_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
+
+/*! ZSTD_estimate?DictSize() :
+ *  ZSTD_estimateCDictSize() will bet that src size is relatively "small", and content is copied, like ZSTD_createCDict().
+ *  ZSTD_estimateCDictSize_advanced() makes it possible to control compression parameters precisely, like ZSTD_createCDict_advanced().
+ *  Note : dictionaries created by reference (`ZSTD_dlm_byRef`) are logically smaller.
+ */
+ZSTDLIB_API size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel);
+ZSTDLIB_API size_t ZSTD_estimateCDictSize_advanced(size_t dictSize, ZSTD_compressionParameters cParams, ZSTD_dictLoadMethod_e dictLoadMethod);
+ZSTDLIB_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod);
+
+/*! ZSTD_initStatic*() :
+ *  Initialize an object using a pre-allocated fixed-size buffer.
+ *  workspace: The memory area to emplace the object into.
+ *             Provided pointer *must be 8-bytes aligned*.
+ *             Buffer must outlive object.
+ *  workspaceSize: Use ZSTD_estimate*Size() to determine
+ *                 how large workspace must be to support target scenario.
+ * @return : pointer to object (same address as workspace, just different type),
+ *           or NULL if error (size too small, incorrect alignment, etc.)
+ *  Note : zstd will never resize nor malloc() when using a static buffer.
+ *         If the object requires more memory than available,
+ *         zstd will just error out (typically ZSTD_error_memory_allocation).
+ *  Note 2 : there is no corresponding "free" function.
+ *           Since workspace is allocated externally, it must be freed externally too.
+ *  Note 3 : cParams : use ZSTD_getCParams() to convert a compression level
+ *           into its associated cParams.
+ *  Limitation 1 : currently not compatible with internal dictionary creation, triggered by
+ *                 ZSTD_CCtx_loadDictionary(), ZSTD_initCStream_usingDict() or ZSTD_initDStream_usingDict().
+ *  Limitation 2 : static cctx currently not compatible with multi-threading.
+ *  Limitation 3 : static dctx is incompatible with legacy support.
+ */
+ZSTDLIB_API ZSTD_CCtx*    ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_API ZSTD_CStream* ZSTD_initStaticCStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticCCtx() */
+
+ZSTDLIB_API ZSTD_DCtx*    ZSTD_initStaticDCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_API ZSTD_DStream* ZSTD_initStaticDStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticDCtx() */
+
+ZSTDLIB_API const ZSTD_CDict* ZSTD_initStaticCDict(
+                                        void* workspace, size_t workspaceSize,
+                                        const void* dict, size_t dictSize,
+                                        ZSTD_dictLoadMethod_e dictLoadMethod,
+                                        ZSTD_dictContentType_e dictContentType,
+                                        ZSTD_compressionParameters cParams);
+
+ZSTDLIB_API const ZSTD_DDict* ZSTD_initStaticDDict(
+                                        void* workspace, size_t workspaceSize,
+                                        const void* dict, size_t dictSize,
+                                        ZSTD_dictLoadMethod_e dictLoadMethod,
+                                        ZSTD_dictContentType_e dictContentType);
+
+/*! Custom memory allocation :
+ *  These prototypes make it possible to pass your own allocation/free functions.
+ *  ZSTD_customMem is provided at creation time, using ZSTD_create*_advanced() variants listed below.
+ *  All allocation/free operations will be completed using these custom variants instead of regular <stdlib.h> ones.
+ */
+typedef void* (*ZSTD_allocFunction) (void* opaque, size_t size);
+typedef void  (*ZSTD_freeFunction) (void* opaque, void* address);
+typedef struct { ZSTD_allocFunction customAlloc; ZSTD_freeFunction customFree; void* opaque; } ZSTD_customMem;
+static ZSTD_customMem const ZSTD_defaultCMem = { NULL, NULL, NULL };  /**< this constant defers to stdlib's functions */
+
+ZSTDLIB_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
+ZSTDLIB_API ZSTD_DCtx*    ZSTD_createDCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem);
+
+ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictSize,
+                                                  ZSTD_dictLoadMethod_e dictLoadMethod,
+                                                  ZSTD_dictContentType_e dictContentType,
+                                                  ZSTD_compressionParameters cParams,
+                                                  ZSTD_customMem customMem);
+
+ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
+                                                  ZSTD_dictLoadMethod_e dictLoadMethod,
+                                                  ZSTD_dictContentType_e dictContentType,
+                                                  ZSTD_customMem customMem);
+
+
+
+/***************************************
+*  Advanced compression functions
+***************************************/
+
+/*! ZSTD_createCDict_byReference() :
+ *  Create a digested dictionary for compression
+ *  Dictionary content is simply referenced, and therefore stays in dictBuffer.
+ *  It is important that dictBuffer outlives CDict, it must remain read accessible throughout the lifetime of CDict */
+ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_byReference(const void* dictBuffer, size_t dictSize, int compressionLevel);
+
+/*! ZSTD_getCParams() :
+*   @return ZSTD_compressionParameters structure for a selected compression level and estimated srcSize.
+*   `estimatedSrcSize` value is optional, select 0 if not known */
+ZSTDLIB_API ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+
+/*! ZSTD_getParams() :
+*   same as ZSTD_getCParams(), but @return a full `ZSTD_parameters` object instead of sub-component `ZSTD_compressionParameters`.
+*   All fields of `ZSTD_frameParameters` are set to default : contentSize=1, checksum=0, noDictID=0 */
+ZSTDLIB_API ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+
+/*! ZSTD_checkCParams() :
+*   Ensure param values remain within authorized range */
+ZSTDLIB_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
+
+/*! ZSTD_adjustCParams() :
+ *  optimize params for a given `srcSize` and `dictSize`.
+ *  both values are optional, select `0` if unknown. */
+ZSTDLIB_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
+
+/*! ZSTD_compress_advanced() :
+*   Same as ZSTD_compress_usingDict(), with fine-tune control over each compression parameter */
+ZSTDLIB_API size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
+                                  void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize,
+                            const void* dict,size_t dictSize,
+                                  ZSTD_parameters params);
+
+/*! ZSTD_compress_usingCDict_advanced() :
+*   Same as ZSTD_compress_usingCDict(), with fine-tune control over frame parameters */
+ZSTDLIB_API size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+                                  void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize,
+                            const ZSTD_CDict* cdict, ZSTD_frameParameters fParams);
+
+
+/*--- Advanced decompression functions ---*/
+
+/*! ZSTD_isFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier.
+ *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
+ *  Note 2 : Legacy Frame Identifiers are considered valid only if Legacy Support is enabled.
+ *  Note 3 : Skippable Frame Identifiers are considered valid. */
+ZSTDLIB_API unsigned ZSTD_isFrame(const void* buffer, size_t size);
+
+/*! ZSTD_createDDict_byReference() :
+ *  Create a digested dictionary, ready to start decompression operation without startup delay.
+ *  Dictionary content is referenced, and therefore stays in dictBuffer.
+ *  It is important that dictBuffer outlives DDict,
+ *  it must remain read accessible throughout the lifetime of DDict */
+ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize);
+
+
+/*! ZSTD_getDictID_fromDict() :
+ *  Provides the dictID stored within dictionary.
+ *  if @return == 0, the dictionary is not conformant with Zstandard specification.
+ *  It can still be loaded, but as a content-only dictionary. */
+ZSTDLIB_API unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize);
+
+/*! ZSTD_getDictID_fromDDict() :
+ *  Provides the dictID of the dictionary loaded into `ddict`.
+ *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
+ *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
+ZSTDLIB_API unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict);
+
+/*! ZSTD_getDictID_fromFrame() :
+ *  Provides the dictID required to decompressed the frame stored within `src`.
+ *  If @return == 0, the dictID could not be decoded.
+ *  This could for one of the following reasons :
+ *  - The frame does not require a dictionary to be decoded (most common case).
+ *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden information.
+ *    Note : this use case also happens when using a non-conformant dictionary.
+ *  - `srcSize` is too small, and as a result, the frame header could not be decoded (only possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`).
+ *  - This is not a Zstandard frame.
+ *  When identifying the exact failure cause, it's possible to use ZSTD_getFrameHeader(), which will provide a more precise error code. */
+ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
+
+
+/********************************************************************
+*  Advanced streaming functions
+********************************************************************/
+
+/*=====   Advanced Streaming compression functions  =====*/
+ZSTDLIB_API size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs, int compressionLevel, unsigned long long pledgedSrcSize);   /**< pledgedSrcSize must be correct. If it is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN. Note that, for compatibility with older programs, "0" also disables frame content size field. It may be enabled in the future. */
+ZSTDLIB_API size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs, const void* dict, size_t dictSize, int compressionLevel); /**< creates of an internal CDict (incompatible with static CCtx), except if dict == NULL or dictSize < 8, in which case no dict is used. Note: dict is loaded with ZSTD_dm_auto (treated as a full zstd dictionary if it begins with ZSTD_MAGIC_DICTIONARY, else as raw content) and ZSTD_dlm_byCopy.*/
+ZSTDLIB_API size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs, const void* dict, size_t dictSize,
+                                             ZSTD_parameters params, unsigned long long pledgedSrcSize);  /**< pledgedSrcSize must be correct. If srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN. dict is loaded with ZSTD_dm_auto and ZSTD_dlm_byCopy. */
+ZSTDLIB_API size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);  /**< note : cdict will just be referenced, and must outlive compression session */
+ZSTDLIB_API size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs, const ZSTD_CDict* cdict, ZSTD_frameParameters fParams, unsigned long long pledgedSrcSize);  /**< same as ZSTD_initCStream_usingCDict(), with control over frame parameters. pledgedSrcSize must be correct. If srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN. */
+
+/*! ZSTD_resetCStream() :
+ *  start a new compression job, using same parameters from previous job.
+ *  This is typically useful to skip dictionary loading stage, since it will re-use it in-place..
+ *  Note that zcs must be init at least once before using ZSTD_resetCStream().
+ *  If pledgedSrcSize is not known at reset time, use macro ZSTD_CONTENTSIZE_UNKNOWN.
+ *  If pledgedSrcSize > 0, its value must be correct, as it will be written in header, and controlled at the end.
+ *  For the time being, pledgedSrcSize==0 is interpreted as "srcSize unknown" for compatibility with older programs,
+ *  but it will change to mean "empty" in future version, so use macro ZSTD_CONTENTSIZE_UNKNOWN instead.
+ * @return : 0, or an error code (which can be tested using ZSTD_isError()) */
+ZSTDLIB_API size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
+
+
+typedef struct {
+    unsigned long long ingested;
+    unsigned long long consumed;
+    unsigned long long produced;
+} ZSTD_frameProgression;
+
+/* ZSTD_getFrameProgression():
+ * tells how much data has been ingested (read from input)
+ * consumed (input actually compressed) and produced (output) for current frame.
+ * Therefore, (ingested - consumed) is amount of input data buffered internally, not yet compressed.
+ * Can report progression inside worker threads (multi-threading and non-blocking mode).
+ */
+ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx);
+
+
+
+/*=====   Advanced Streaming decompression functions  =====*/
+typedef enum { DStream_p_maxWindowSize } ZSTD_DStreamParameter_e;
+ZSTDLIB_API size_t ZSTD_setDStreamParameter(ZSTD_DStream* zds, ZSTD_DStreamParameter_e paramType, unsigned paramValue);   /* obsolete : this API will be removed in a future version */
+ZSTDLIB_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize); /**< note: no dictionary will be used if dict == NULL or dictSize < 8 */
+ZSTDLIB_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);  /**< note : ddict is referenced, it must outlive decompression session */
+ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);  /**< re-use decompression parameters from previous init; saves dictionary loading */
+
+
+/*********************************************************************
+*  Buffer-less and synchronous inner streaming functions
+*
+*  This is an advanced API, giving full control over buffer management, for users which need direct control over memory.
+*  But it's also a complex one, with several restrictions, documented below.
+*  Prefer normal streaming API for an easier experience.
+********************************************************************* */
+
+/**
+  Buffer-less streaming compression (synchronous mode)
+
+  A ZSTD_CCtx object is required to track streaming operations.
+  Use ZSTD_createCCtx() / ZSTD_freeCCtx() to manage resource.
+  ZSTD_CCtx object can be re-used multiple times within successive compression operations.
+
+  Start by initializing a context.
+  Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression,
+  or ZSTD_compressBegin_advanced(), for finer parameter control.
+  It's also possible to duplicate a reference context which has already been initialized, using ZSTD_copyCCtx()
+
+  Then, consume your input using ZSTD_compressContinue().
+  There are some important considerations to keep in mind when using this advanced function :
+  - ZSTD_compressContinue() has no internal buffer. It uses externally provided buffers only.
+  - Interface is synchronous : input is consumed entirely and produces 1+ compressed blocks.
+  - Caller must ensure there is enough space in `dst` to store compressed data under worst case scenario.
+    Worst case evaluation is provided by ZSTD_compressBound().
+    ZSTD_compressContinue() doesn't guarantee recover after a failed compression.
+  - ZSTD_compressContinue() presumes prior input ***is still accessible and unmodified*** (up to maximum distance size, see WindowLog).
+    It remembers all previous contiguous blocks, plus one separated memory segment (which can itself consists of multiple contiguous blocks)
+  - ZSTD_compressContinue() detects that prior input has been overwritten when `src` buffer overlaps.
+    In which case, it will "discard" the relevant memory section from its history.
+
+  Finish a frame with ZSTD_compressEnd(), which will write the last block(s) and optional checksum.
+  It's possible to use srcSize==0, in which case, it will write a final empty block to end the frame.
+  Without last block mark, frames are considered unfinished (hence corrupted) by compliant decoders.
+
+  `ZSTD_CCtx` object can be re-used (ZSTD_compressBegin()) to compress again.
+*/
+
+/*=====   Buffer-less streaming compression functions  =====*/
+ZSTDLIB_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
+ZSTDLIB_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
+ZSTDLIB_API size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
+ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTDLIB_API size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
+
+ZSTDLIB_API size_t ZSTD_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+/*-
+  Buffer-less streaming decompression (synchronous mode)
+
+  A ZSTD_DCtx object is required to track streaming operations.
+  Use ZSTD_createDCtx() / ZSTD_freeDCtx() to manage it.
+  A ZSTD_DCtx object can be re-used multiple times.
+
+  First typical operation is to retrieve frame parameters, using ZSTD_getFrameHeader().
+  Frame header is extracted from the beginning of compressed frame, so providing only the frame's beginning is enough.
+  Data fragment must be large enough to ensure successful decoding.
+ `ZSTD_frameHeaderSize_max` bytes is guaranteed to always be large enough.
+  @result : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
+           >0 : `srcSize` is too small, please provide at least @result bytes on next attempt.
+           errorCode, which can be tested using ZSTD_isError().
+
+  It fills a ZSTD_frameHeader structure with important information to correctly decode the frame,
+  such as the dictionary ID, content size, or maximum back-reference distance (`windowSize`).
+  Note that these values could be wrong, either because of data corruption, or because a 3rd party deliberately spoofs false information.
+  As a consequence, check that values remain within valid application range.
+  For example, do not allocate memory blindly, check that `windowSize` is within expectation.
+  Each application can set its own limits, depending on local restrictions.
+  For extended interoperability, it is recommended to support `windowSize` of at least 8 MB.
+
+  ZSTD_decompressContinue() needs previous data blocks during decompression, up to `windowSize` bytes.
+  ZSTD_decompressContinue() is very sensitive to contiguity,
+  if 2 blocks don't follow each other, make sure that either the compressor breaks contiguity at the same place,
+  or that previous contiguous segment is large enough to properly handle maximum back-reference distance.
+  There are multiple ways to guarantee this condition.
+
+  The most memory efficient way is to use a round buffer of sufficient size.
+  Sufficient size is determined by invoking ZSTD_decodingBufferSize_min(),
+  which can @return an error code if required value is too large for current system (in 32-bits mode).
+  In a round buffer methodology, ZSTD_decompressContinue() decompresses each block next to previous one,
+  up to the moment there is not enough room left in the buffer to guarantee decoding another full block,
+  which maximum size is provided in `ZSTD_frameHeader` structure, field `blockSizeMax`.
+  At which point, decoding can resume from the beginning of the buffer.
+  Note that already decoded data stored in the buffer should be flushed before being overwritten.
+
+  There are alternatives possible, for example using two or more buffers of size `windowSize` each, though they consume more memory.
+
+  Finally, if you control the compression process, you can also ignore all buffer size rules,
+  as long as the encoder and decoder progress in "lock-step",
+  aka use exactly the same buffer sizes, break contiguity at the same place, etc.
+
+  Once buffers are setup, start decompression, with ZSTD_decompressBegin().
+  If decompression requires a dictionary, use ZSTD_decompressBegin_usingDict() or ZSTD_decompressBegin_usingDDict().
+
+  Then use ZSTD_nextSrcSizeToDecompress() and ZSTD_decompressContinue() alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how many bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() requires this _exact_ amount of bytes, or it will fail.
+
+ @result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
+  It can be zero : it just means ZSTD_decompressContinue() has decoded some metadata item.
+  It can also be an error code, which can be tested with ZSTD_isError().
+
+  A frame is fully decoded when ZSTD_nextSrcSizeToDecompress() returns zero.
+  Context can then be reset to start a new decompression.
+
+  Note : it's possible to know if next input to present is a header or a block, using ZSTD_nextInputType().
+  This information is not required to properly decode a frame.
+
+  == Special case : skippable frames ==
+
+  Skippable frames allow integration of user-defined data into a flow of concatenated frames.
+  Skippable frames will be ignored (skipped) by decompressor.
+  The format of skippable frames is as follows :
+  a) Skippable frame ID - 4 Bytes, Little endian format, any value from 0x184D2A50 to 0x184D2A5F
+  b) Frame Size - 4 Bytes, Little endian format, unsigned 32-bits
+  c) Frame Content - any content (User Data) of length equal to Frame Size
+  For skippable frames ZSTD_getFrameHeader() returns zfhPtr->frameType==ZSTD_skippableFrame.
+  For skippable frames ZSTD_decompressContinue() always returns 0 : it only skips the content.
+*/
+
+/*=====   Buffer-less streaming decompression functions  =====*/
+typedef enum { ZSTD_frame, ZSTD_skippableFrame } ZSTD_frameType_e;
+typedef struct {
+    unsigned long long frameContentSize; /* if == ZSTD_CONTENTSIZE_UNKNOWN, it means this field is not available. 0 means "empty" */
+    unsigned long long windowSize;       /* can be very large, up to <= frameContentSize */
+    unsigned blockSizeMax;
+    ZSTD_frameType_e frameType;          /* if == ZSTD_skippableFrame, frameContentSize is the size of skippable content */
+    unsigned headerSize;
+    unsigned dictID;
+    unsigned checksumFlag;
+} ZSTD_frameHeader;
+ZSTDLIB_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
+ZSTDLIB_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /**< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
+
+ZSTDLIB_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
+ZSTDLIB_API size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_API size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
+
+ZSTDLIB_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
+ZSTDLIB_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+/* misc */
+ZSTDLIB_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
+typedef enum { ZSTDnit_frameHeader, ZSTDnit_blockHeader, ZSTDnit_block, ZSTDnit_lastBlock, ZSTDnit_checksum, ZSTDnit_skippableFrame } ZSTD_nextInputType_e;
+ZSTDLIB_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
+
+
+
+/* ============================================ */
+/**       New advanced API (experimental)       */
+/* ============================================ */
+
+/* notes on API design :
+ *   In this proposal, parameters are pushed one by one into an existing context,
+ *   and then applied on all subsequent compression jobs.
+ *   When no parameter is ever provided, CCtx is created with compression level ZSTD_CLEVEL_DEFAULT.
+ *
+ *   This API is intended to replace all others advanced / experimental API entry points.
+ *   But it stands a reasonable chance to become "stable", after a reasonable testing period.
+ */
+
+/* note on naming convention :
+ *   Initially, the API favored names like ZSTD_setCCtxParameter() .
+ *   In this proposal, convention is changed towards ZSTD_CCtx_setParameter() .
+ *   The main driver is that it identifies more clearly the target object type.
+ *   It feels clearer when considering multiple targets :
+ *   ZSTD_CDict_setParameter() (rather than ZSTD_setCDictParameter())
+ *   ZSTD_CCtxParams_setParameter()  (rather than ZSTD_setCCtxParamsParameter() )
+ *   etc...
+ */
+
+/* note on enum design :
+ * All enum will be pinned to explicit values before reaching "stable API" status */
+
+typedef enum {
+    /* Opened question : should we have a format ZSTD_f_auto ?
+     * Today, it would mean exactly the same as ZSTD_f_zstd1.
+     * But, in the future, should several formats become supported,
+     * on the compression side, it would mean "default format".
+     * On the decompression side, it would mean "automatic format detection",
+     * so that ZSTD_f_zstd1 would mean "accept *only* zstd frames".
+     * Since meaning is a little different, another option could be to define different enums for compression and decompression.
+     * This question could be kept for later, when there are actually multiple formats to support,
+     * but there is also the question of pinning enum values, and pinning value `0` is especially important */
+    ZSTD_f_zstd1 = 0,        /* zstd frame format, specified in zstd_compression_format.md (default) */
+    ZSTD_f_zstd1_magicless,  /* Variant of zstd frame format, without initial 4-bytes magic number.
+                              * Useful to save 4 bytes per generated frame.
+                              * Decoder cannot recognise automatically this format, requiring instructions. */
+} ZSTD_format_e;
+
+typedef enum {
+    /* compression format */
+    ZSTD_p_format = 10,      /* See ZSTD_format_e enum definition.
+                              * Cast selected format as unsigned for ZSTD_CCtx_setParameter() compatibility. */
+
+    /* compression parameters */
+    ZSTD_p_compressionLevel=100, /* Update all compression parameters according to pre-defined cLevel table
+                              * Default level is ZSTD_CLEVEL_DEFAULT==3.
+                              * Special: value 0 means "do not change cLevel".
+                              * Note 1 : it's possible to pass a negative compression level by casting it to unsigned type.
+                              * Note 2 : setting a level sets all default values of other compression parameters.
+                              * Note 3 : setting compressionLevel automatically updates ZSTD_p_compressLiterals. */
+    ZSTD_p_windowLog,        /* Maximum allowed back-reference distance, expressed as power of 2.
+                              * Must be clamped between ZSTD_WINDOWLOG_MIN and ZSTD_WINDOWLOG_MAX.
+                              * Special: value 0 means "use default windowLog".
+                              * Note: Using a window size greater than ZSTD_MAXWINDOWSIZE_DEFAULT (default: 2^27)
+                              *       requires explicitly allowing such window size during decompression stage. */
+    ZSTD_p_hashLog,          /* Size of the probe table, as a power of 2.
+                              * Resulting table size is (1 << (hashLog+2)).
+                              * Must be clamped between ZSTD_HASHLOG_MIN and ZSTD_HASHLOG_MAX.
+                              * Larger tables improve compression ratio of strategies <= dFast,
+                              * and improve speed of strategies > dFast.
+                              * Special: value 0 means "use default hashLog". */
+    ZSTD_p_chainLog,         /* Size of the full-search table, as a power of 2.
+                              * Resulting table size is (1 << (chainLog+2)).
+                              * Larger tables result in better and slower compression.
+                              * This parameter is useless when using "fast" strategy.
+                              * Special: value 0 means "use default chainLog". */
+    ZSTD_p_searchLog,        /* Number of search attempts, as a power of 2.
+                              * More attempts result in better and slower compression.
+                              * This parameter is useless when using "fast" and "dFast" strategies.
+                              * Special: value 0 means "use default searchLog". */
+    ZSTD_p_minMatch,         /* Minimum size of searched matches (note : repCode matches can be smaller).
+                              * Larger values make faster compression and decompression, but decrease ratio.
+                              * Must be clamped between ZSTD_SEARCHLENGTH_MIN and ZSTD_SEARCHLENGTH_MAX.
+                              * Note that currently, for all strategies < btopt, effective minimum is 4.
+                              *                    , for all strategies > fast, effective maximum is 6.
+                              * Special: value 0 means "use default minMatchLength". */
+    ZSTD_p_targetLength,     /* Impact of this field depends on strategy.
+                              * For strategies btopt & btultra:
+                              *     Length of Match considered "good enough" to stop search.
+                              *     Larger values make compression stronger, and slower.
+                              * For strategy fast:
+                              *     Distance between match sampling.
+                              *     Larger values make compression faster, and weaker.
+                              * Special: value 0 means "use default targetLength". */
+    ZSTD_p_compressionStrategy, /* See ZSTD_strategy enum definition.
+                              * Cast selected strategy as unsigned for ZSTD_CCtx_setParameter() compatibility.
+                              * The higher the value of selected strategy, the more complex it is,
+                              * resulting in stronger and slower compression.
+                              * Special: value 0 means "use default strategy". */
+
+    ZSTD_p_enableLongDistanceMatching=160, /* Enable long distance matching.
+                                         * This parameter is designed to improve compression ratio
+                                         * for large inputs, by finding large matches at long distance.
+                                         * It increases memory usage and window size.
+                                         * Note: enabling this parameter increases ZSTD_p_windowLog to 128 MB
+                                         * except when expressly set to a different value. */
+    ZSTD_p_ldmHashLog,       /* Size of the table for long distance matching, as a power of 2.
+                              * Larger values increase memory usage and compression ratio,
+                              * but decrease compression speed.
+                              * Must be clamped between ZSTD_HASHLOG_MIN and ZSTD_HASHLOG_MAX
+                              * default: windowlog - 7.
+                              * Special: value 0 means "automatically determine hashlog". */
+    ZSTD_p_ldmMinMatch,      /* Minimum match size for long distance matcher.
+                              * Larger/too small values usually decrease compression ratio.
+                              * Must be clamped between ZSTD_LDM_MINMATCH_MIN and ZSTD_LDM_MINMATCH_MAX.
+                              * Special: value 0 means "use default value" (default: 64). */
+    ZSTD_p_ldmBucketSizeLog, /* Log size of each bucket in the LDM hash table for collision resolution.
+                              * Larger values improve collision resolution but decrease compression speed.
+                              * The maximum value is ZSTD_LDM_BUCKETSIZELOG_MAX .
+                              * Special: value 0 means "use default value" (default: 3). */
+    ZSTD_p_ldmHashEveryLog,  /* Frequency of inserting/looking up entries in the LDM hash table.
+                              * Must be clamped between 0 and (ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN).
+                              * Default is MAX(0, (windowLog - ldmHashLog)), optimizing hash table usage.
+                              * Larger values improve compression speed.
+                              * Deviating far from default value will likely result in a compression ratio decrease.
+                              * Special: value 0 means "automatically determine hashEveryLog". */
+
+    /* frame parameters */
+    ZSTD_p_contentSizeFlag=200, /* Content size will be written into frame header _whenever known_ (default:1)
+                              * Content size must be known at the beginning of compression,
+                              * it is provided using ZSTD_CCtx_setPledgedSrcSize() */
+    ZSTD_p_checksumFlag,     /* A 32-bits checksum of content is written at end of frame (default:0) */
+    ZSTD_p_dictIDFlag,       /* When applicable, dictionary's ID is written into frame header (default:1) */
+
+    /* multi-threading parameters */
+    /* These parameters are only useful if multi-threading is enabled (ZSTD_MULTITHREAD).
+     * They return an error otherwise. */
+    ZSTD_p_nbWorkers=400,    /* Select how many threads will be spawned to compress in parallel.
+                              * When nbWorkers >= 1, triggers asynchronous mode :
+                              * ZSTD_compress_generic() consumes some input, flush some output if possible, and immediately gives back control to caller,
+                              * while compression work is performed in parallel, within worker threads.
+                              * (note : a strong exception to this rule is when first invocation sets ZSTD_e_end : it becomes a blocking call).
+                              * More workers improve speed, but also increase memory usage.
+                              * Default value is `0`, aka "single-threaded mode" : no worker is spawned, compression is performed inside Caller's thread, all invocations are blocking */
+    ZSTD_p_jobSize,          /* Size of a compression job. This value is enforced only in non-blocking mode.
+                              * Each compression job is completed in parallel, so this value indirectly controls the nb of active threads.
+                              * 0 means default, which is dynamically determined based on compression parameters.
+                              * Job size must be a minimum of overlapSize, or 1 MB, whichever is largest.
+                              * The minimum size is automatically and transparently enforced */
+    ZSTD_p_overlapSizeLog,   /* Size of previous input reloaded at the beginning of each job.
+                              * 0 => no overlap, 6(default) => use 1/8th of windowSize, >=9 => use full windowSize */
+
+    /* =================================================================== */
+    /* experimental parameters - no stability guaranteed                   */
+    /* =================================================================== */
+
+    ZSTD_p_compressLiterals=1000, /* control huffman compression of literals (enabled) by default.
+                              * disabling it improves speed and decreases compression ratio by a large amount.
+                              * note : this setting is automatically updated when changing compression level.
+                              *        positive compression levels set ZSTD_p_compressLiterals to 1.
+                              *        negative compression levels set ZSTD_p_compressLiterals to 0. */
+
+    ZSTD_p_forceMaxWindow=1100, /* Force back-reference distances to remain < windowSize,
+                              * even when referencing into Dictionary content (default:0) */
+
+} ZSTD_cParameter;
+
+
+/*! ZSTD_CCtx_setParameter() :
+ *  Set one compression parameter, selected by enum ZSTD_cParameter.
+ *  Setting a parameter is generally only possible during frame initialization (before starting compression),
+ *  except for a few exceptions which can be updated during compression: compressionLevel, hashLog, chainLog, searchLog, minMatch, targetLength and strategy.
+ *  Note : when `value` is an enum, cast it to unsigned for proper type checking.
+ *  @result : informational value (typically, value being set clamped correctly),
+ *            or an error code (which can be tested with ZSTD_isError()). */
+ZSTDLIB_API size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, unsigned value);
+
+/*! ZSTD_CCtx_setPledgedSrcSize() :
+ *  Total input data size to be compressed as a single frame.
+ *  This value will be controlled at the end, and result in error if not respected.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Note 1 : 0 means zero, empty.
+ *           In order to mean "unknown content size", pass constant ZSTD_CONTENTSIZE_UNKNOWN.
+ *           ZSTD_CONTENTSIZE_UNKNOWN is default value for any new compression job.
+ *  Note 2 : If all data is provided and consumed in a single round,
+ *           this value is overriden by srcSize instead. */
+ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize);
+
+/*! ZSTD_CCtx_loadDictionary() :
+ *  Create an internal CDict from `dict` buffer.
+ *  Decompression will have to use same dictionary.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Special: Adding a NULL (or 0-size) dictionary invalidates previous dictionary,
+ *           meaning "return to no-dictionary mode".
+ *  Note 1 : Dictionary will be used for all future compression jobs.
+ *           To return to "no-dictionary" situation, load a NULL dictionary
+ *  Note 2 : Loading a dictionary involves building tables, which are dependent on compression parameters.
+ *           For this reason, compression parameters cannot be changed anymore after loading a dictionary.
+ *           It's also a CPU consuming operation, with non-negligible impact on latency.
+ *  Note 3 :`dict` content will be copied internally.
+ *           Use ZSTD_CCtx_loadDictionary_byReference() to reference dictionary content instead.
+ *           In such a case, dictionary buffer must outlive its users.
+ *  Note 4 : Use ZSTD_CCtx_loadDictionary_advanced()
+ *           to precisely select how dictionary content must be interpreted. */
+ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
+ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
+ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+
+
+/*! ZSTD_CCtx_refCDict() :
+ *  Reference a prepared dictionary, to be used for all next compression jobs.
+ *  Note that compression parameters are enforced from within CDict,
+ *  and supercede any compression parameter previously set within CCtx.
+ *  The dictionary will remain valid for future compression jobs using same CCtx.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Special : adding a NULL CDict means "return to no-dictionary mode".
+ *  Note 1 : Currently, only one dictionary can be managed.
+ *           Adding a new dictionary effectively "discards" any previous one.
+ *  Note 2 : CDict is just referenced, its lifetime must outlive CCtx. */
+ZSTDLIB_API size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
+
+/*! ZSTD_CCtx_refPrefix() :
+ *  Reference a prefix (single-usage dictionary) for next compression job.
+ *  Decompression need same prefix to properly regenerate data.
+ *  Prefix is **only used once**. Tables are discarded at end of compression job.
+ *  Subsequent compression jobs will be done without prefix (if none is explicitly referenced).
+ *  If there is a need to use same prefix multiple times, consider embedding it into a ZSTD_CDict instead.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Special: Adding any prefix (including NULL) invalidates any previous prefix or dictionary
+ *  Note 1 : Prefix buffer is referenced. It must outlive compression job.
+ *  Note 2 : Referencing a prefix involves building tables, which are dependent on compression parameters.
+ *           It's a CPU consuming operation, with non-negligible impact on latency.
+ *  Note 3 : By default, the prefix is treated as raw content (ZSTD_dm_rawContent).
+ *           Use ZSTD_CCtx_refPrefix_advanced() to alter dictMode. */
+ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize);
+ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+
+/*! ZSTD_CCtx_reset() :
+ *  Return a CCtx to clean state.
+ *  Useful after an error, or to interrupt an ongoing compression job and start a new one.
+ *  Any internal data not yet flushed is cancelled.
+ *  Dictionary (if any) is dropped.
+ *  All parameters are back to default values.
+ *  It's possible to modify compression parameters after a reset.
+ */
+ZSTDLIB_API void ZSTD_CCtx_reset(ZSTD_CCtx* cctx);
+
+
+
+typedef enum {
+    ZSTD_e_continue=0, /* collect more data, encoder decides when to output compressed result, for optimal conditions */
+    ZSTD_e_flush,      /* flush any data provided so far - frame will continue, future data can still reference previous data for better compression */
+    ZSTD_e_end         /* flush any remaining data and close current frame. Any additional data starts a new frame. */
+} ZSTD_EndDirective;
+
+/*! ZSTD_compress_generic() :
+ *  Behave about the same as ZSTD_compressStream. To note :
+ *  - Compression parameters are pushed into CCtx before starting compression, using ZSTD_CCtx_setParameter()
+ *  - Compression parameters cannot be changed once compression is started.
+ *  - outpot->pos must be <= dstCapacity, input->pos must be <= srcSize
+ *  - outpot->pos and input->pos will be updated. They are guaranteed to remain below their respective limit.
+ *  - In single-thread mode (default), function is blocking : it completed its job before returning to caller.
+ *  - In multi-thread mode, function is non-blocking : it just acquires a copy of input, and distribute job to internal worker threads,
+ *                                                     and then immediately returns, just indicating that there is some data remaining to be flushed.
+ *                                                     The function nonetheless guarantees forward progress : it will return only after it reads or write at least 1+ byte.
+ *  - Exception : in multi-threading mode, if the first call requests a ZSTD_e_end directive, it is blocking : it will complete compression before giving back control to caller.
+ *  - @return provides a minimum amount of data remaining to be flushed from internal buffers
+ *            or an error code, which can be tested using ZSTD_isError().
+ *            if @return != 0, flush is not fully completed, there is still some data left within internal buffers.
+ *            This is useful for ZSTD_e_flush, since in this case more flushes are necessary to empty all buffers.
+ *            For ZSTD_e_end, @return == 0 when internal buffers are fully flushed and frame is completed.
+ *  - after a ZSTD_e_end directive, if internal buffer is not fully flushed (@return != 0),
+ *            only ZSTD_e_end or ZSTD_e_flush operations are allowed.
+ *            Before starting a new compression job, or changing compression parameters,
+ *            it is required to fully flush internal buffers.
+ */
+ZSTDLIB_API size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
+                                          ZSTD_outBuffer* output,
+                                          ZSTD_inBuffer* input,
+                                          ZSTD_EndDirective endOp);
+
+
+/*! ZSTD_compress_generic_simpleArgs() :
+ *  Same as ZSTD_compress_generic(),
+ *  but using only integral types as arguments.
+ *  Argument list is larger than ZSTD_{in,out}Buffer,
+ *  but can be helpful for binders from dynamic languages
+ *  which have troubles handling structures containing memory pointers.
+ */
+ZSTDLIB_API size_t ZSTD_compress_generic_simpleArgs (
+                            ZSTD_CCtx* cctx,
+                            void* dst, size_t dstCapacity, size_t* dstPos,
+                      const void* src, size_t srcSize, size_t* srcPos,
+                            ZSTD_EndDirective endOp);
+
+
+/*! ZSTD_CCtx_params :
+ *  Quick howto :
+ *  - ZSTD_createCCtxParams() : Create a ZSTD_CCtx_params structure
+ *  - ZSTD_CCtxParam_setParameter() : Push parameters one by one into
+ *                                    an existing ZSTD_CCtx_params structure.
+ *                                    This is similar to
+ *                                    ZSTD_CCtx_setParameter().
+ *  - ZSTD_CCtx_setParametersUsingCCtxParams() : Apply parameters to
+ *                                    an existing CCtx.
+ *                                    These parameters will be applied to
+ *                                    all subsequent compression jobs.
+ *  - ZSTD_compress_generic() : Do compression using the CCtx.
+ *  - ZSTD_freeCCtxParams() : Free the memory.
+ *
+ *  This can be used with ZSTD_estimateCCtxSize_advanced_usingCCtxParams()
+ *  for static allocation for single-threaded compression.
+ */
+ZSTDLIB_API ZSTD_CCtx_params* ZSTD_createCCtxParams(void);
+ZSTDLIB_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);
+
+
+/*! ZSTD_CCtxParams_reset() :
+ *  Reset params to default values.
+ */
+ZSTDLIB_API size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params);
+
+/*! ZSTD_CCtxParams_init() :
+ *  Initializes the compression parameters of cctxParams according to
+ *  compression level. All other parameters are reset to their default values.
+ */
+ZSTDLIB_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel);
+
+/*! ZSTD_CCtxParams_init_advanced() :
+ *  Initializes the compression and frame parameters of cctxParams according to
+ *  params. All other parameters are reset to their default values.
+ */
+ZSTDLIB_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
+
+
+/*! ZSTD_CCtxParam_setParameter() :
+ *  Similar to ZSTD_CCtx_setParameter.
+ *  Set one compression parameter, selected by enum ZSTD_cParameter.
+ *  Parameters must be applied to a ZSTD_CCtx using ZSTD_CCtx_setParametersUsingCCtxParams().
+ *  Note : when `value` is an enum, cast it to unsigned for proper type checking.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ */
+ZSTDLIB_API size_t ZSTD_CCtxParam_setParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, unsigned value);
+
+/*! ZSTD_CCtx_setParametersUsingCCtxParams() :
+ *  Apply a set of ZSTD_CCtx_params to the compression context.
+ *  This can be done even after compression is started,
+ *    if nbWorkers==0, this will have no impact until a new compression is started.
+ *    if nbWorkers>=1, new parameters will be picked up at next job,
+ *       with a few restrictions (windowLog, pledgedSrcSize, nbWorkers, jobSize, and overlapLog are not updated).
+ */
+ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
+        ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params);
+
+
+/*===   Advanced parameters for decompression API  ===*/
+
+/* The following parameters must be set after creating a ZSTD_DCtx* (or ZSTD_DStream*) object,
+ * but before starting decompression of a frame.
+ */
+
+/*! ZSTD_DCtx_loadDictionary() :
+ *  Create an internal DDict from dict buffer,
+ *  to be used to decompress next frames.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Special : Adding a NULL (or 0-size) dictionary invalidates any previous dictionary,
+ *            meaning "return to no-dictionary mode".
+ *  Note 1 : `dict` content will be copied internally.
+ *            Use ZSTD_DCtx_loadDictionary_byReference()
+ *            to reference dictionary content instead.
+ *            In which case, the dictionary buffer must outlive its users.
+ *  Note 2 : Loading a dictionary involves building tables,
+ *           which has a non-negligible impact on CPU usage and latency.
+ *  Note 3 : Use ZSTD_DCtx_loadDictionary_advanced() to select
+ *           how dictionary content will be interpreted and loaded.
+ */
+ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+
+
+/*! ZSTD_DCtx_refDDict() :
+ *  Reference a prepared dictionary, to be used to decompress next frames.
+ *  The dictionary remains active for decompression of future frames using same DCtx.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Note 1 : Currently, only one dictionary can be managed.
+ *           Referencing a new dictionary effectively "discards" any previous one.
+ *  Special : adding a NULL DDict means "return to no-dictionary mode".
+ *  Note 2 : DDict is just referenced, its lifetime must outlive its usage from DCtx.
+ */
+ZSTDLIB_API size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
+
+
+/*! ZSTD_DCtx_refPrefix() :
+ *  Reference a prefix (single-usage dictionary) for next compression job.
+ *  Prefix is **only used once**. It must be explicitly referenced before each frame.
+ *  If there is a need to use same prefix multiple times, consider embedding it into a ZSTD_DDict instead.
+ * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Note 1 : Adding any prefix (including NULL) invalidates any previously set prefix or dictionary
+ *  Note 2 : Prefix buffer is referenced. It must outlive compression job.
+ *  Note 3 : By default, the prefix is treated as raw content (ZSTD_dm_rawContent).
+ *           Use ZSTD_CCtx_refPrefix_advanced() to alter dictMode.
+ *  Note 4 : Referencing a raw content prefix has almost no cpu nor memory cost.
+ */
+ZSTDLIB_API size_t ZSTD_DCtx_refPrefix(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize);
+ZSTDLIB_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+
+
+/*! ZSTD_DCtx_setMaxWindowSize() :
+ *  Refuses allocating internal buffers for frames requiring a window size larger than provided limit.
+ *  This is useful to prevent a decoder context from reserving too much memory for itself (potential attack scenario).
+ *  This parameter is only useful in streaming mode, since no internal buffer is allocated in direct mode.
+ *  By default, a decompression context accepts all window sizes <= (1 << ZSTD_WINDOWLOG_MAX)
+ * @return : 0, or an error code (which can be tested using ZSTD_isError()).
+ */
+ZSTDLIB_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize);
+
+
+/*! ZSTD_DCtx_setFormat() :
+ *  Instruct the decoder context about what kind of data to decode next.
+ *  This instruction is mandatory to decode data without a fully-formed header,
+ *  such ZSTD_f_zstd1_magicless for example.
+ * @return : 0, or an error code (which can be tested using ZSTD_isError()).
+ */
+ZSTDLIB_API size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
+
+
+/*! ZSTD_decompress_generic() :
+ *  Behave the same as ZSTD_decompressStream.
+ *  Decompression parameters cannot be changed once decompression is started.
+ * @return : an error code, which can be tested using ZSTD_isError()
+ *           if >0, a hint, nb of expected input bytes for next invocation.
+ *           `0` means : a frame has just been fully decoded and flushed.
+ */
+ZSTDLIB_API size_t ZSTD_decompress_generic(ZSTD_DCtx* dctx,
+                                           ZSTD_outBuffer* output,
+                                           ZSTD_inBuffer* input);
+
+
+/*! ZSTD_decompress_generic_simpleArgs() :
+ *  Same as ZSTD_decompress_generic(),
+ *  but using only integral types as arguments.
+ *  Argument list is larger than ZSTD_{in,out}Buffer,
+ *  but can be helpful for binders from dynamic languages
+ *  which have troubles handling structures containing memory pointers.
+ */
+ZSTDLIB_API size_t ZSTD_decompress_generic_simpleArgs (
+                            ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity, size_t* dstPos,
+                      const void* src, size_t srcSize, size_t* srcPos);
+
+
+/*! ZSTD_DCtx_reset() :
+ *  Return a DCtx to clean state.
+ *  If a decompression was ongoing, any internal data not yet flushed is cancelled.
+ *  All parameters are back to default values, including sticky ones.
+ *  Dictionary (if any) is dropped.
+ *  Parameters can be modified again after a reset.
+ */
+ZSTDLIB_API void ZSTD_DCtx_reset(ZSTD_DCtx* dctx);
+
+
+
+/* ============================ */
+/**       Block level API       */
+/* ============================ */
+
+/*!
+    Block functions produce and decode raw zstd blocks, without frame metadata.
+    Frame metadata cost is typically ~18 bytes, which can be non-negligible for very small blocks (< 100 bytes).
+    User will have to take in charge required information to regenerate data, such as compressed and content sizes.
+
+    A few rules to respect :
+    - Compressing and decompressing require a context structure
+      + Use ZSTD_createCCtx() and ZSTD_createDCtx()
+    - It is necessary to init context before starting
+      + compression : any ZSTD_compressBegin*() variant, including with dictionary
+      + decompression : any ZSTD_decompressBegin*() variant, including with dictionary
+      + copyCCtx() and copyDCtx() can be used too
+    - Block size is limited, it must be <= ZSTD_getBlockSize() <= ZSTD_BLOCKSIZE_MAX == 128 KB
+      + If input is larger than a block size, it's necessary to split input data into multiple blocks
+      + For inputs larger than a single block size, consider using the regular ZSTD_compress() instead.
+        Frame metadata is not that costly, and quickly becomes negligible as source size grows larger.
+    - When a block is considered not compressible enough, ZSTD_compressBlock() result will be zero.
+      In which case, nothing is produced into `dst`.
+      + User must test for such outcome and deal directly with uncompressed data
+      + ZSTD_decompressBlock() doesn't accept uncompressed data as input !!!
+      + In case of multiple successive blocks, should some of them be uncompressed,
+        decoder must be informed of their existence in order to follow proper history.
+        Use ZSTD_insertBlock() for such a case.
+*/
+
+#define ZSTD_BLOCKSIZELOG_MAX 17
+#define ZSTD_BLOCKSIZE_MAX   (1<<ZSTD_BLOCKSIZELOG_MAX)   /* define, for static allocation */
+/*=====   Raw zstd block functions  =====*/
+ZSTDLIB_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
+ZSTDLIB_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_API size_t ZSTD_insertBlock    (ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize);  /**< insert uncompressed block into `dctx` history. Useful for multi-blocks decompression. */
+
+
+#endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/vendor/github.com/DataDog/zstd/zstd_common.c b/vendor/github.com/DataDog/zstd/zstd_common.c
new file mode 100644
index 0000000..bccc948
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_common.c
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include <stdlib.h>      /* malloc, calloc, free */
+#include <string.h>      /* memset */
+#include "error_private.h"
+#include "zstd_internal.h"
+
+
+/*-****************************************
+*  Version
+******************************************/
+unsigned ZSTD_versionNumber(void) { return ZSTD_VERSION_NUMBER; }
+
+const char* ZSTD_versionString(void) { return ZSTD_VERSION_STRING; }
+
+
+/*-****************************************
+*  ZSTD Error Management
+******************************************/
+/*! ZSTD_isError() :
+ *  tells if a return value is an error code */
+unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }
+
+/*! ZSTD_getErrorName() :
+ *  provides error code string from function result (useful for debugging) */
+const char* ZSTD_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+/*! ZSTD_getError() :
+ *  convert a `size_t` function result into a proper ZSTD_errorCode enum */
+ZSTD_ErrorCode ZSTD_getErrorCode(size_t code) { return ERR_getErrorCode(code); }
+
+/*! ZSTD_getErrorString() :
+ *  provides error code string from enum */
+const char* ZSTD_getErrorString(ZSTD_ErrorCode code) { return ERR_getErrorString(code); }
+
+/*! g_debuglog_enable :
+ *  turn on/off debug traces (global switch) */
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG >= 2)
+int g_debuglog_enable = 1;
+#endif
+
+
+/*=**************************************************************
+*  Custom allocator
+****************************************************************/
+void* ZSTD_malloc(size_t size, ZSTD_customMem customMem)
+{
+    if (customMem.customAlloc)
+        return customMem.customAlloc(customMem.opaque, size);
+    return malloc(size);
+}
+
+void* ZSTD_calloc(size_t size, ZSTD_customMem customMem)
+{
+    if (customMem.customAlloc) {
+        /* calloc implemented as malloc+memset;
+         * not as efficient as calloc, but next best guess for custom malloc */
+        void* const ptr = customMem.customAlloc(customMem.opaque, size);
+        memset(ptr, 0, size);
+        return ptr;
+    }
+    return calloc(1, size);
+}
+
+void ZSTD_free(void* ptr, ZSTD_customMem customMem)
+{
+    if (ptr!=NULL) {
+        if (customMem.customFree)
+            customMem.customFree(customMem.opaque, ptr);
+        else
+            free(ptr);
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_compress.c b/vendor/github.com/DataDog/zstd/zstd_compress.c
new file mode 100644
index 0000000..2aa26da
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_compress.c
@@ -0,0 +1,3449 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*-*************************************
+*  Tuning parameters
+***************************************/
+#ifndef ZSTD_CLEVEL_DEFAULT
+#  define ZSTD_CLEVEL_DEFAULT 3
+#endif
+
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include <string.h>         /* memset */
+#include "cpu.h"
+#include "mem.h"
+#define FSE_STATIC_LINKING_ONLY   /* FSE_encodeSymbol */
+#include "fse.h"
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"
+#include "zstd_compress_internal.h"
+#include "zstd_fast.h"
+#include "zstd_double_fast.h"
+#include "zstd_lazy.h"
+#include "zstd_opt.h"
+#include "zstd_ldm.h"
+
+
+/*-*************************************
+*  Helper functions
+***************************************/
+size_t ZSTD_compressBound(size_t srcSize) {
+    return ZSTD_COMPRESSBOUND(srcSize);
+}
+
+
+/*-*************************************
+*  Context memory management
+***************************************/
+struct ZSTD_CDict_s {
+    void* dictBuffer;
+    const void* dictContent;
+    size_t dictContentSize;
+    void* workspace;
+    size_t workspaceSize;
+    ZSTD_matchState_t matchState;
+    ZSTD_compressedBlockState_t cBlockState;
+    ZSTD_compressionParameters cParams;
+    ZSTD_customMem customMem;
+    U32 dictID;
+};  /* typedef'd to ZSTD_CDict within "zstd.h" */
+
+ZSTD_CCtx* ZSTD_createCCtx(void)
+{
+    return ZSTD_createCCtx_advanced(ZSTD_defaultCMem);
+}
+
+ZSTD_CCtx* ZSTD_createCCtx_advanced(ZSTD_customMem customMem)
+{
+    ZSTD_STATIC_ASSERT(zcss_init==0);
+    ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN==(0ULL - 1));
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+    {   ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_calloc(sizeof(ZSTD_CCtx), customMem);
+        if (!cctx) return NULL;
+        cctx->customMem = customMem;
+        cctx->requestedParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;
+        cctx->requestedParams.fParams.contentSizeFlag = 1;
+        cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+        return cctx;
+    }
+}
+
+ZSTD_CCtx* ZSTD_initStaticCCtx(void *workspace, size_t workspaceSize)
+{
+    ZSTD_CCtx* const cctx = (ZSTD_CCtx*) workspace;
+    if (workspaceSize <= sizeof(ZSTD_CCtx)) return NULL;  /* minimum size */
+    if ((size_t)workspace & 7) return NULL;  /* must be 8-aligned */
+    memset(workspace, 0, workspaceSize);   /* may be a bit generous, could memset be smaller ? */
+    cctx->staticSize = workspaceSize;
+    cctx->workSpace = (void*)(cctx+1);
+    cctx->workSpaceSize = workspaceSize - sizeof(ZSTD_CCtx);
+
+    /* statically sized space. entropyWorkspace never moves (but prev/next block swap places) */
+    if (cctx->workSpaceSize < HUF_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t)) return NULL;
+    assert(((size_t)cctx->workSpace & (sizeof(void*)-1)) == 0);   /* ensure correct alignment */
+    cctx->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)cctx->workSpace;
+    cctx->blockState.nextCBlock = cctx->blockState.prevCBlock + 1;
+    {
+        void* const ptr = cctx->blockState.nextCBlock + 1;
+        cctx->entropyWorkspace = (U32*)ptr;
+    }
+    cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+    return cctx;
+}
+
+size_t ZSTD_freeCCtx(ZSTD_CCtx* cctx)
+{
+    if (cctx==NULL) return 0;   /* support free on NULL */
+    if (cctx->staticSize) return ERROR(memory_allocation);   /* not compatible with static CCtx */
+    ZSTD_free(cctx->workSpace, cctx->customMem); cctx->workSpace = NULL;
+    ZSTD_freeCDict(cctx->cdictLocal); cctx->cdictLocal = NULL;
+#ifdef ZSTD_MULTITHREAD
+    ZSTDMT_freeCCtx(cctx->mtctx); cctx->mtctx = NULL;
+#endif
+    ZSTD_free(cctx, cctx->customMem);
+    return 0;   /* reserved as a potential error code in the future */
+}
+
+
+static size_t ZSTD_sizeof_mtctx(const ZSTD_CCtx* cctx)
+{
+#ifdef ZSTD_MULTITHREAD
+    return ZSTDMT_sizeof_CCtx(cctx->mtctx);
+#else
+    (void) cctx;
+    return 0;
+#endif
+}
+
+
+size_t ZSTD_sizeof_CCtx(const ZSTD_CCtx* cctx)
+{
+    if (cctx==NULL) return 0;   /* support sizeof on NULL */
+    return sizeof(*cctx) + cctx->workSpaceSize
+           + ZSTD_sizeof_CDict(cctx->cdictLocal)
+           + ZSTD_sizeof_mtctx(cctx);
+}
+
+size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs)
+{
+    return ZSTD_sizeof_CCtx(zcs);  /* same object */
+}
+
+/* private API call, for dictBuilder only */
+const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }
+
+ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize)
+{
+    ZSTD_compressionParameters cParams = ZSTD_getCParams(CCtxParams->compressionLevel, srcSizeHint, dictSize);
+    if (CCtxParams->ldmParams.enableLdm) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG;
+    if (CCtxParams->cParams.windowLog) cParams.windowLog = CCtxParams->cParams.windowLog;
+    if (CCtxParams->cParams.hashLog) cParams.hashLog = CCtxParams->cParams.hashLog;
+    if (CCtxParams->cParams.chainLog) cParams.chainLog = CCtxParams->cParams.chainLog;
+    if (CCtxParams->cParams.searchLog) cParams.searchLog = CCtxParams->cParams.searchLog;
+    if (CCtxParams->cParams.searchLength) cParams.searchLength = CCtxParams->cParams.searchLength;
+    if (CCtxParams->cParams.targetLength) cParams.targetLength = CCtxParams->cParams.targetLength;
+    if (CCtxParams->cParams.strategy) cParams.strategy = CCtxParams->cParams.strategy;
+    return cParams;
+}
+
+static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
+        ZSTD_compressionParameters cParams)
+{
+    ZSTD_CCtx_params cctxParams;
+    memset(&cctxParams, 0, sizeof(cctxParams));
+    cctxParams.cParams = cParams;
+    cctxParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;  /* should not matter, as all cParams are presumed properly defined */
+    assert(!ZSTD_checkCParams(cParams));
+    cctxParams.fParams.contentSizeFlag = 1;
+    return cctxParams;
+}
+
+static ZSTD_CCtx_params* ZSTD_createCCtxParams_advanced(
+        ZSTD_customMem customMem)
+{
+    ZSTD_CCtx_params* params;
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+    params = (ZSTD_CCtx_params*)ZSTD_calloc(
+            sizeof(ZSTD_CCtx_params), customMem);
+    if (!params) { return NULL; }
+    params->customMem = customMem;
+    params->compressionLevel = ZSTD_CLEVEL_DEFAULT;
+    params->fParams.contentSizeFlag = 1;
+    return params;
+}
+
+ZSTD_CCtx_params* ZSTD_createCCtxParams(void)
+{
+    return ZSTD_createCCtxParams_advanced(ZSTD_defaultCMem);
+}
+
+size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params)
+{
+    if (params == NULL) { return 0; }
+    ZSTD_free(params, params->customMem);
+    return 0;
+}
+
+size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params)
+{
+    return ZSTD_CCtxParams_init(params, ZSTD_CLEVEL_DEFAULT);
+}
+
+size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel) {
+    if (!cctxParams) { return ERROR(GENERIC); }
+    memset(cctxParams, 0, sizeof(*cctxParams));
+    cctxParams->compressionLevel = compressionLevel;
+    cctxParams->fParams.contentSizeFlag = 1;
+    return 0;
+}
+
+size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params)
+{
+    if (!cctxParams) { return ERROR(GENERIC); }
+    CHECK_F( ZSTD_checkCParams(params.cParams) );
+    memset(cctxParams, 0, sizeof(*cctxParams));
+    cctxParams->cParams = params.cParams;
+    cctxParams->fParams = params.fParams;
+    cctxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
+    assert(!ZSTD_checkCParams(params.cParams));
+    return 0;
+}
+
+/* ZSTD_assignParamsToCCtxParams() :
+ * params is presumed valid at this stage */
+static ZSTD_CCtx_params ZSTD_assignParamsToCCtxParams(
+        ZSTD_CCtx_params cctxParams, ZSTD_parameters params)
+{
+    ZSTD_CCtx_params ret = cctxParams;
+    ret.cParams = params.cParams;
+    ret.fParams = params.fParams;
+    ret.compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
+    assert(!ZSTD_checkCParams(params.cParams));
+    return ret;
+}
+
+#define CLAMPCHECK(val,min,max) {            \
+    if (((val)<(min)) | ((val)>(max))) {     \
+        return ERROR(parameter_outOfBound);  \
+}   }
+
+
+static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
+{
+    switch(param)
+    {
+    case ZSTD_p_compressionLevel:
+    case ZSTD_p_hashLog:
+    case ZSTD_p_chainLog:
+    case ZSTD_p_searchLog:
+    case ZSTD_p_minMatch:
+    case ZSTD_p_targetLength:
+    case ZSTD_p_compressionStrategy:
+    case ZSTD_p_compressLiterals:
+        return 1;
+
+    case ZSTD_p_format:
+    case ZSTD_p_windowLog:
+    case ZSTD_p_contentSizeFlag:
+    case ZSTD_p_checksumFlag:
+    case ZSTD_p_dictIDFlag:
+    case ZSTD_p_forceMaxWindow :
+    case ZSTD_p_nbWorkers:
+    case ZSTD_p_jobSize:
+    case ZSTD_p_overlapSizeLog:
+    case ZSTD_p_enableLongDistanceMatching:
+    case ZSTD_p_ldmHashLog:
+    case ZSTD_p_ldmMinMatch:
+    case ZSTD_p_ldmBucketSizeLog:
+    case ZSTD_p_ldmHashEveryLog:
+    default:
+        return 0;
+    }
+}
+
+size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, unsigned value)
+{
+    DEBUGLOG(4, "ZSTD_CCtx_setParameter (%u, %u)", (U32)param, value);
+    if (cctx->streamStage != zcss_init) {
+        if (ZSTD_isUpdateAuthorized(param)) {
+            cctx->cParamsChanged = 1;
+        } else {
+            return ERROR(stage_wrong);
+    }   }
+
+    switch(param)
+    {
+    case ZSTD_p_format :
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_compressionLevel:
+        if (cctx->cdict) return ERROR(stage_wrong);
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_windowLog:
+    case ZSTD_p_hashLog:
+    case ZSTD_p_chainLog:
+    case ZSTD_p_searchLog:
+    case ZSTD_p_minMatch:
+    case ZSTD_p_targetLength:
+    case ZSTD_p_compressionStrategy:
+        if (cctx->cdict) return ERROR(stage_wrong);
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_compressLiterals:
+    case ZSTD_p_contentSizeFlag:
+    case ZSTD_p_checksumFlag:
+    case ZSTD_p_dictIDFlag:
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_forceMaxWindow :  /* Force back-references to remain < windowSize,
+                                   * even when referencing into Dictionary content.
+                                   * default : 0 when using a CDict, 1 when using a Prefix */
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_nbWorkers:
+        if ((value>0) && cctx->staticSize) {
+            return ERROR(parameter_unsupported);  /* MT not compatible with static alloc */
+        }
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_jobSize:
+    case ZSTD_p_overlapSizeLog:
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    case ZSTD_p_enableLongDistanceMatching:
+    case ZSTD_p_ldmHashLog:
+    case ZSTD_p_ldmMinMatch:
+    case ZSTD_p_ldmBucketSizeLog:
+    case ZSTD_p_ldmHashEveryLog:
+        if (cctx->cdict) return ERROR(stage_wrong);
+        return ZSTD_CCtxParam_setParameter(&cctx->requestedParams, param, value);
+
+    default: return ERROR(parameter_unsupported);
+    }
+}
+
+size_t ZSTD_CCtxParam_setParameter(
+        ZSTD_CCtx_params* CCtxParams, ZSTD_cParameter param, unsigned value)
+{
+    DEBUGLOG(4, "ZSTD_CCtxParam_setParameter (%u, %u)", (U32)param, value);
+    switch(param)
+    {
+    case ZSTD_p_format :
+        if (value > (unsigned)ZSTD_f_zstd1_magicless)
+            return ERROR(parameter_unsupported);
+        CCtxParams->format = (ZSTD_format_e)value;
+        return (size_t)CCtxParams->format;
+
+    case ZSTD_p_compressionLevel : {
+        int cLevel = (int)value;  /* cast expected to restore negative sign */
+        if (cLevel > ZSTD_maxCLevel()) cLevel = ZSTD_maxCLevel();
+        if (cLevel) {  /* 0 : does not change current level */
+            CCtxParams->disableLiteralCompression = (cLevel<0);  /* negative levels disable huffman */
+            CCtxParams->compressionLevel = cLevel;
+        }
+        if (CCtxParams->compressionLevel >= 0) return CCtxParams->compressionLevel;
+        return 0;  /* return type (size_t) cannot represent negative values */
+    }
+
+    case ZSTD_p_windowLog :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, ZSTD_WINDOWLOG_MIN, ZSTD_WINDOWLOG_MAX);
+        CCtxParams->cParams.windowLog = value;
+        return CCtxParams->cParams.windowLog;
+
+    case ZSTD_p_hashLog :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, ZSTD_HASHLOG_MIN, ZSTD_HASHLOG_MAX);
+        CCtxParams->cParams.hashLog = value;
+        return CCtxParams->cParams.hashLog;
+
+    case ZSTD_p_chainLog :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, ZSTD_CHAINLOG_MIN, ZSTD_CHAINLOG_MAX);
+        CCtxParams->cParams.chainLog = value;
+        return CCtxParams->cParams.chainLog;
+
+    case ZSTD_p_searchLog :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, ZSTD_SEARCHLOG_MIN, ZSTD_SEARCHLOG_MAX);
+        CCtxParams->cParams.searchLog = value;
+        return value;
+
+    case ZSTD_p_minMatch :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, ZSTD_SEARCHLENGTH_MIN, ZSTD_SEARCHLENGTH_MAX);
+        CCtxParams->cParams.searchLength = value;
+        return CCtxParams->cParams.searchLength;
+
+    case ZSTD_p_targetLength :
+        /* all values are valid. 0 => use default */
+        CCtxParams->cParams.targetLength = value;
+        return CCtxParams->cParams.targetLength;
+
+    case ZSTD_p_compressionStrategy :
+        if (value>0)   /* 0 => use default */
+            CLAMPCHECK(value, (unsigned)ZSTD_fast, (unsigned)ZSTD_btultra);
+        CCtxParams->cParams.strategy = (ZSTD_strategy)value;
+        return (size_t)CCtxParams->cParams.strategy;
+
+    case ZSTD_p_compressLiterals:
+        CCtxParams->disableLiteralCompression = !value;
+        return !CCtxParams->disableLiteralCompression;
+
+    case ZSTD_p_contentSizeFlag :
+        /* Content size written in frame header _when known_ (default:1) */
+        DEBUGLOG(4, "set content size flag = %u", (value>0));
+        CCtxParams->fParams.contentSizeFlag = value > 0;
+        return CCtxParams->fParams.contentSizeFlag;
+
+    case ZSTD_p_checksumFlag :
+        /* A 32-bits content checksum will be calculated and written at end of frame (default:0) */
+        CCtxParams->fParams.checksumFlag = value > 0;
+        return CCtxParams->fParams.checksumFlag;
+
+    case ZSTD_p_dictIDFlag : /* When applicable, dictionary's dictID is provided in frame header (default:1) */
+        DEBUGLOG(4, "set dictIDFlag = %u", (value>0));
+        CCtxParams->fParams.noDictIDFlag = !value;
+        return !CCtxParams->fParams.noDictIDFlag;
+
+    case ZSTD_p_forceMaxWindow :
+        CCtxParams->forceWindow = (value > 0);
+        return CCtxParams->forceWindow;
+
+    case ZSTD_p_nbWorkers :
+#ifndef ZSTD_MULTITHREAD
+        if (value>0) return ERROR(parameter_unsupported);
+        return 0;
+#else
+        return ZSTDMT_CCtxParam_setNbWorkers(CCtxParams, value);
+#endif
+
+    case ZSTD_p_jobSize :
+#ifndef ZSTD_MULTITHREAD
+        return ERROR(parameter_unsupported);
+#else
+        return ZSTDMT_CCtxParam_setMTCtxParameter(CCtxParams, ZSTDMT_p_jobSize, value);
+#endif
+
+    case ZSTD_p_overlapSizeLog :
+#ifndef ZSTD_MULTITHREAD
+        return ERROR(parameter_unsupported);
+#else
+        return ZSTDMT_CCtxParam_setMTCtxParameter(CCtxParams, ZSTDMT_p_overlapSectionLog, value);
+#endif
+
+    case ZSTD_p_enableLongDistanceMatching :
+        CCtxParams->ldmParams.enableLdm = (value>0);
+        return CCtxParams->ldmParams.enableLdm;
+
+    case ZSTD_p_ldmHashLog :
+        if (value>0)   /* 0 ==> auto */
+            CLAMPCHECK(value, ZSTD_HASHLOG_MIN, ZSTD_HASHLOG_MAX);
+        CCtxParams->ldmParams.hashLog = value;
+        return CCtxParams->ldmParams.hashLog;
+
+    case ZSTD_p_ldmMinMatch :
+        if (value>0)   /* 0 ==> default */
+            CLAMPCHECK(value, ZSTD_LDM_MINMATCH_MIN, ZSTD_LDM_MINMATCH_MAX);
+        CCtxParams->ldmParams.minMatchLength = value;
+        return CCtxParams->ldmParams.minMatchLength;
+
+    case ZSTD_p_ldmBucketSizeLog :
+        if (value > ZSTD_LDM_BUCKETSIZELOG_MAX)
+            return ERROR(parameter_outOfBound);
+        CCtxParams->ldmParams.bucketSizeLog = value;
+        return CCtxParams->ldmParams.bucketSizeLog;
+
+    case ZSTD_p_ldmHashEveryLog :
+        if (value > ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN)
+            return ERROR(parameter_outOfBound);
+        CCtxParams->ldmParams.hashEveryLog = value;
+        return CCtxParams->ldmParams.hashEveryLog;
+
+    default: return ERROR(parameter_unsupported);
+    }
+}
+
+/** ZSTD_CCtx_setParametersUsingCCtxParams() :
+ *  just applies `params` into `cctx`
+ *  no action is performed, parameters are merely stored.
+ *  If ZSTDMT is enabled, parameters are pushed to cctx->mtctx.
+ *    This is possible even if a compression is ongoing.
+ *    In which case, new parameters will be applied on the fly, starting with next compression job.
+ */
+size_t ZSTD_CCtx_setParametersUsingCCtxParams(
+        ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params)
+{
+    if (cctx->streamStage != zcss_init) return ERROR(stage_wrong);
+    if (cctx->cdict) return ERROR(stage_wrong);
+
+    cctx->requestedParams = *params;
+    return 0;
+}
+
+ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %u bytes", (U32)pledgedSrcSize);
+    if (cctx->streamStage != zcss_init) return ERROR(stage_wrong);
+    cctx->pledgedSrcSizePlusOne = pledgedSrcSize+1;
+    return 0;
+}
+
+size_t ZSTD_CCtx_loadDictionary_advanced(
+        ZSTD_CCtx* cctx, const void* dict, size_t dictSize,
+        ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType)
+{
+    if (cctx->streamStage != zcss_init) return ERROR(stage_wrong);
+    if (cctx->staticSize) return ERROR(memory_allocation);  /* no malloc for static CCtx */
+    DEBUGLOG(4, "ZSTD_CCtx_loadDictionary_advanced (size: %u)", (U32)dictSize);
+    ZSTD_freeCDict(cctx->cdictLocal);  /* in case one already exists */
+    if (dict==NULL || dictSize==0) {   /* no dictionary mode */
+        cctx->cdictLocal = NULL;
+        cctx->cdict = NULL;
+    } else {
+        ZSTD_compressionParameters const cParams =
+                ZSTD_getCParamsFromCCtxParams(&cctx->requestedParams, cctx->pledgedSrcSizePlusOne-1, dictSize);
+        cctx->cdictLocal = ZSTD_createCDict_advanced(
+                                dict, dictSize,
+                                dictLoadMethod, dictContentType,
+                                cParams, cctx->customMem);
+        cctx->cdict = cctx->cdictLocal;
+        if (cctx->cdictLocal == NULL)
+            return ERROR(memory_allocation);
+    }
+    return 0;
+}
+
+ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference(
+      ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
+{
+    return ZSTD_CCtx_loadDictionary_advanced(
+            cctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto);
+}
+
+ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
+{
+    return ZSTD_CCtx_loadDictionary_advanced(
+            cctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto);
+}
+
+
+size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
+{
+    if (cctx->streamStage != zcss_init) return ERROR(stage_wrong);
+    cctx->cdict = cdict;
+    memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));  /* exclusive */
+    return 0;
+}
+
+size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize)
+{
+    return ZSTD_CCtx_refPrefix_advanced(cctx, prefix, prefixSize, ZSTD_dct_rawContent);
+}
+
+size_t ZSTD_CCtx_refPrefix_advanced(
+        ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType)
+{
+    if (cctx->streamStage != zcss_init) return ERROR(stage_wrong);
+    cctx->cdict = NULL;   /* prefix discards any prior cdict */
+    cctx->prefixDict.dict = prefix;
+    cctx->prefixDict.dictSize = prefixSize;
+    cctx->prefixDict.dictContentType = dictContentType;
+    return 0;
+}
+
+static void ZSTD_startNewCompression(ZSTD_CCtx* cctx)
+{
+    cctx->streamStage = zcss_init;
+    cctx->pledgedSrcSizePlusOne = 0;
+}
+
+/*! ZSTD_CCtx_reset() :
+ *  Also dumps dictionary */
+void ZSTD_CCtx_reset(ZSTD_CCtx* cctx)
+{
+    ZSTD_startNewCompression(cctx);
+    cctx->cdict = NULL;
+}
+
+/** ZSTD_checkCParams() :
+    control CParam values remain within authorized range.
+    @return : 0, or an error code if one value is beyond authorized range */
+size_t ZSTD_checkCParams(ZSTD_compressionParameters cParams)
+{
+    CLAMPCHECK(cParams.windowLog, ZSTD_WINDOWLOG_MIN, ZSTD_WINDOWLOG_MAX);
+    CLAMPCHECK(cParams.chainLog, ZSTD_CHAINLOG_MIN, ZSTD_CHAINLOG_MAX);
+    CLAMPCHECK(cParams.hashLog, ZSTD_HASHLOG_MIN, ZSTD_HASHLOG_MAX);
+    CLAMPCHECK(cParams.searchLog, ZSTD_SEARCHLOG_MIN, ZSTD_SEARCHLOG_MAX);
+    CLAMPCHECK(cParams.searchLength, ZSTD_SEARCHLENGTH_MIN, ZSTD_SEARCHLENGTH_MAX);
+    if ((U32)(cParams.targetLength) < ZSTD_TARGETLENGTH_MIN)
+        return ERROR(parameter_unsupported);
+    if ((U32)(cParams.strategy) > (U32)ZSTD_btultra)
+        return ERROR(parameter_unsupported);
+    return 0;
+}
+
+/** ZSTD_clampCParams() :
+ *  make CParam values within valid range.
+ *  @return : valid CParams */
+static ZSTD_compressionParameters ZSTD_clampCParams(ZSTD_compressionParameters cParams)
+{
+#   define CLAMP(val,min,max) {      \
+        if (val<min) val=min;        \
+        else if (val>max) val=max;   \
+    }
+    CLAMP(cParams.windowLog, ZSTD_WINDOWLOG_MIN, ZSTD_WINDOWLOG_MAX);
+    CLAMP(cParams.chainLog, ZSTD_CHAINLOG_MIN, ZSTD_CHAINLOG_MAX);
+    CLAMP(cParams.hashLog, ZSTD_HASHLOG_MIN, ZSTD_HASHLOG_MAX);
+    CLAMP(cParams.searchLog, ZSTD_SEARCHLOG_MIN, ZSTD_SEARCHLOG_MAX);
+    CLAMP(cParams.searchLength, ZSTD_SEARCHLENGTH_MIN, ZSTD_SEARCHLENGTH_MAX);
+    if ((U32)(cParams.targetLength) < ZSTD_TARGETLENGTH_MIN) cParams.targetLength = ZSTD_TARGETLENGTH_MIN;
+    if ((U32)(cParams.strategy) > (U32)ZSTD_btultra) cParams.strategy = ZSTD_btultra;
+    return cParams;
+}
+
+/** ZSTD_cycleLog() :
+ *  condition for correct operation : hashLog > 1 */
+static U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat)
+{
+    U32 const btScale = ((U32)strat >= (U32)ZSTD_btlazy2);
+    return hashLog - btScale;
+}
+
+/** ZSTD_adjustCParams_internal() :
+    optimize `cPar` for a given input (`srcSize` and `dictSize`).
+    mostly downsizing to reduce memory consumption and initialization latency.
+    Both `srcSize` and `dictSize` are optional (use 0 if unknown).
+    Note : cPar is considered validated at this stage. Use ZSTD_checkCParams() to ensure that condition. */
+ZSTD_compressionParameters ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize)
+{
+    static const U64 minSrcSize = 513; /* (1<<9) + 1 */
+    static const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
+    assert(ZSTD_checkCParams(cPar)==0);
+
+    if (dictSize && (srcSize+1<2) /* srcSize unknown */ )
+        srcSize = minSrcSize;  /* presumed small when there is a dictionary */
+    else if (srcSize == 0)
+        srcSize = ZSTD_CONTENTSIZE_UNKNOWN;  /* 0 == unknown : presumed large */
+
+    /* resize windowLog if input is small enough, to use less memory */
+    if ( (srcSize < maxWindowResize)
+      && (dictSize < maxWindowResize) )  {
+        U32 const tSize = (U32)(srcSize + dictSize);
+        static U32 const hashSizeMin = 1 << ZSTD_HASHLOG_MIN;
+        U32 const srcLog = (tSize < hashSizeMin) ? ZSTD_HASHLOG_MIN :
+                            ZSTD_highbit32(tSize-1) + 1;
+        if (cPar.windowLog > srcLog) cPar.windowLog = srcLog;
+    }
+    if (cPar.hashLog > cPar.windowLog) cPar.hashLog = cPar.windowLog;
+    {   U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy);
+        if (cycleLog > cPar.windowLog)
+            cPar.chainLog -= (cycleLog - cPar.windowLog);
+    }
+
+    if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN)
+        cPar.windowLog = ZSTD_WINDOWLOG_ABSOLUTEMIN;  /* required for frame header */
+
+    return cPar;
+}
+
+ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize)
+{
+    cPar = ZSTD_clampCParams(cPar);
+    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize);
+}
+
+static size_t ZSTD_sizeof_matchState(ZSTD_compressionParameters const* cParams, const U32 forCCtx)
+{
+    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    size_t const hSize = ((size_t)1) << cParams->hashLog;
+    U32    const hashLog3 = (forCCtx && cParams->searchLength==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
+    size_t const h3Size = ((size_t)1) << hashLog3;
+    size_t const tableSpace = (chainSize + hSize + h3Size) * sizeof(U32);
+    size_t const optPotentialSpace = ((MaxML+1) + (MaxLL+1) + (MaxOff+1) + (1<<Litbits)) * sizeof(U32)
+                          + (ZSTD_OPT_NUM+1) * (sizeof(ZSTD_match_t)+sizeof(ZSTD_optimal_t));
+    size_t const optSpace = (forCCtx && ((cParams->strategy == ZSTD_btopt) ||
+                                         (cParams->strategy == ZSTD_btultra)))
+                                ? optPotentialSpace
+                                : 0;
+    DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u",
+                (U32)chainSize, (U32)hSize, (U32)h3Size);
+    return tableSpace + optSpace;
+}
+
+size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
+{
+    /* Estimate CCtx size is supported for single-threaded compression only. */
+    if (params->nbWorkers > 0) { return ERROR(GENERIC); }
+    {   ZSTD_compressionParameters const cParams =
+                ZSTD_getCParamsFromCCtxParams(params, 0, 0);
+        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog);
+        U32    const divider = (cParams.searchLength==3) ? 3 : 4;
+        size_t const maxNbSeq = blockSize / divider;
+        size_t const tokenSpace = blockSize + 11*maxNbSeq;
+        size_t const entropySpace = HUF_WORKSPACE_SIZE;
+        size_t const blockStateSpace = 2 * sizeof(ZSTD_compressedBlockState_t);
+        size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 1);
+
+        size_t const ldmSpace = ZSTD_ldm_getTableSize(params->ldmParams);
+        size_t const ldmSeqSpace = ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize) * sizeof(rawSeq);
+
+        size_t const neededSpace = entropySpace + blockStateSpace + tokenSpace +
+                                   matchStateSize + ldmSpace + ldmSeqSpace;
+
+        DEBUGLOG(5, "sizeof(ZSTD_CCtx) : %u", (U32)sizeof(ZSTD_CCtx));
+        DEBUGLOG(5, "estimate workSpace : %u", (U32)neededSpace);
+        return sizeof(ZSTD_CCtx) + neededSpace;
+    }
+}
+
+size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
+{
+    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
+    return ZSTD_estimateCCtxSize_usingCCtxParams(&params);
+}
+
+static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel)
+{
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams(compressionLevel, 0, 0);
+    return ZSTD_estimateCCtxSize_usingCParams(cParams);
+}
+
+size_t ZSTD_estimateCCtxSize(int compressionLevel)
+{
+    int level;
+    size_t memBudget = 0;
+    for (level=1; level<=compressionLevel; level++) {
+        size_t const newMB = ZSTD_estimateCCtxSize_internal(level);
+        if (newMB > memBudget) memBudget = newMB;
+    }
+    return memBudget;
+}
+
+size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
+{
+    if (params->nbWorkers > 0) { return ERROR(GENERIC); }
+    {   size_t const CCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(params);
+        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << params->cParams.windowLog);
+        size_t const inBuffSize = ((size_t)1 << params->cParams.windowLog) + blockSize;
+        size_t const outBuffSize = ZSTD_compressBound(blockSize) + 1;
+        size_t const streamingSize = inBuffSize + outBuffSize;
+
+        return CCtxSize + streamingSize;
+    }
+}
+
+size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams)
+{
+    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
+    return ZSTD_estimateCStreamSize_usingCCtxParams(&params);
+}
+
+static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel) {
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams(compressionLevel, 0, 0);
+    return ZSTD_estimateCStreamSize_usingCParams(cParams);
+}
+
+size_t ZSTD_estimateCStreamSize(int compressionLevel) {
+    int level;
+    size_t memBudget = 0;
+    for (level=1; level<=compressionLevel; level++) {
+        size_t const newMB = ZSTD_estimateCStreamSize_internal(level);
+        if (newMB > memBudget) memBudget = newMB;
+    }
+    return memBudget;
+}
+
+/* ZSTD_getFrameProgression():
+ * tells how much data has been consumed (input) and produced (output) for current frame.
+ * able to count progression inside worker threads (non-blocking mode).
+ */
+ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx)
+{
+#ifdef ZSTD_MULTITHREAD
+    if (cctx->appliedParams.nbWorkers > 0) {
+        return ZSTDMT_getFrameProgression(cctx->mtctx);
+    }
+#endif
+    {   ZSTD_frameProgression fp;
+        size_t const buffered = (cctx->inBuff == NULL) ? 0 :
+                                cctx->inBuffPos - cctx->inToCompress;
+        if (buffered) assert(cctx->inBuffPos >= cctx->inToCompress);
+        assert(buffered <= ZSTD_BLOCKSIZE_MAX);
+        fp.ingested = cctx->consumedSrcSize + buffered;
+        fp.consumed = cctx->consumedSrcSize;
+        fp.produced = cctx->producedCSize;
+        return fp;
+}   }
+
+
+static U32 ZSTD_equivalentCParams(ZSTD_compressionParameters cParams1,
+                                  ZSTD_compressionParameters cParams2)
+{
+    return (cParams1.hashLog  == cParams2.hashLog)
+         & (cParams1.chainLog == cParams2.chainLog)
+         & (cParams1.strategy == cParams2.strategy)   /* opt parser space */
+         & ((cParams1.searchLength==3) == (cParams2.searchLength==3));  /* hashlog3 space */
+}
+
+/** The parameters are equivalent if ldm is not enabled in both sets or
+ *  all the parameters are equivalent. */
+static U32 ZSTD_equivalentLdmParams(ldmParams_t ldmParams1,
+                                    ldmParams_t ldmParams2)
+{
+    return (!ldmParams1.enableLdm && !ldmParams2.enableLdm) ||
+           (ldmParams1.enableLdm == ldmParams2.enableLdm &&
+            ldmParams1.hashLog == ldmParams2.hashLog &&
+            ldmParams1.bucketSizeLog == ldmParams2.bucketSizeLog &&
+            ldmParams1.minMatchLength == ldmParams2.minMatchLength &&
+            ldmParams1.hashEveryLog == ldmParams2.hashEveryLog);
+}
+
+typedef enum { ZSTDb_not_buffered, ZSTDb_buffered } ZSTD_buffered_policy_e;
+
+/* ZSTD_sufficientBuff() :
+ * check internal buffers exist for streaming if buffPol == ZSTDb_buffered .
+ * Note : they are assumed to be correctly sized if ZSTD_equivalentCParams()==1 */
+static U32 ZSTD_sufficientBuff(size_t bufferSize1, size_t blockSize1,
+                            ZSTD_buffered_policy_e buffPol2,
+                            ZSTD_compressionParameters cParams2,
+                            U64 pledgedSrcSize)
+{
+    size_t const windowSize2 = MAX(1, (size_t)MIN(((U64)1 << cParams2.windowLog), pledgedSrcSize));
+    size_t const blockSize2 = MIN(ZSTD_BLOCKSIZE_MAX, windowSize2);
+    size_t const neededBufferSize2 = (buffPol2==ZSTDb_buffered) ? windowSize2 + blockSize2 : 0;
+    DEBUGLOG(4, "ZSTD_sufficientBuff: is windowSize2=%u <= wlog1=%u",
+                (U32)windowSize2, cParams2.windowLog);
+    DEBUGLOG(4, "ZSTD_sufficientBuff: is blockSize2=%u <= blockSize1=%u",
+                (U32)blockSize2, (U32)blockSize1);
+    return (blockSize2 <= blockSize1) /* seqStore space depends on blockSize */
+         & (neededBufferSize2 <= bufferSize1);
+}
+
+/** Equivalence for resetCCtx purposes */
+static U32 ZSTD_equivalentParams(ZSTD_CCtx_params params1,
+                                 ZSTD_CCtx_params params2,
+                                 size_t buffSize1, size_t blockSize1,
+                                 ZSTD_buffered_policy_e buffPol2,
+                                 U64 pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_equivalentParams: pledgedSrcSize=%u", (U32)pledgedSrcSize);
+    return ZSTD_equivalentCParams(params1.cParams, params2.cParams) &&
+           ZSTD_equivalentLdmParams(params1.ldmParams, params2.ldmParams) &&
+           ZSTD_sufficientBuff(buffSize1, blockSize1, buffPol2, params2.cParams, pledgedSrcSize);
+}
+
+static void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs)
+{
+    int i;
+    for (i = 0; i < ZSTD_REP_NUM; ++i)
+        bs->rep[i] = repStartValue[i];
+    bs->entropy.hufCTable_repeatMode = HUF_repeat_none;
+    bs->entropy.offcode_repeatMode = FSE_repeat_none;
+    bs->entropy.matchlength_repeatMode = FSE_repeat_none;
+    bs->entropy.litlength_repeatMode = FSE_repeat_none;
+}
+
+/*! ZSTD_invalidateMatchState()
+ * Invalidate all the matches in the match finder tables.
+ * Requires nextSrc and base to be set (can be NULL).
+ */
+static void ZSTD_invalidateMatchState(ZSTD_matchState_t* ms)
+{
+    ZSTD_window_clear(&ms->window);
+
+    ms->nextToUpdate = ms->window.dictLimit + 1;
+    ms->loadedDictEnd = 0;
+    ms->opt.litLengthSum = 0;  /* force reset of btopt stats */
+}
+
+/*! ZSTD_continueCCtx() :
+ *  reuse CCtx without reset (note : requires no dictionary) */
+static size_t ZSTD_continueCCtx(ZSTD_CCtx* cctx, ZSTD_CCtx_params params, U64 pledgedSrcSize)
+{
+    size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params.cParams.windowLog), pledgedSrcSize));
+    size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
+    DEBUGLOG(4, "ZSTD_continueCCtx: re-use context in place");
+
+    cctx->blockSize = blockSize;   /* previous block size could be different even for same windowLog, due to pledgedSrcSize */
+    cctx->appliedParams = params;
+    cctx->pledgedSrcSizePlusOne = pledgedSrcSize+1;
+    cctx->consumedSrcSize = 0;
+    cctx->producedCSize = 0;
+    if (pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN)
+        cctx->appliedParams.fParams.contentSizeFlag = 0;
+    DEBUGLOG(4, "pledged content size : %u ; flag : %u",
+        (U32)pledgedSrcSize, cctx->appliedParams.fParams.contentSizeFlag);
+    cctx->stage = ZSTDcs_init;
+    cctx->dictID = 0;
+    if (params.ldmParams.enableLdm)
+        ZSTD_window_clear(&cctx->ldmState.window);
+    ZSTD_referenceExternalSequences(cctx, NULL, 0);
+    ZSTD_invalidateMatchState(&cctx->blockState.matchState);
+    ZSTD_reset_compressedBlockState(cctx->blockState.prevCBlock);
+    XXH64_reset(&cctx->xxhState, 0);
+    return 0;
+}
+
+typedef enum { ZSTDcrp_continue, ZSTDcrp_noMemset } ZSTD_compResetPolicy_e;
+
+static void* ZSTD_reset_matchState(ZSTD_matchState_t* ms, void* ptr, ZSTD_compressionParameters const* cParams, ZSTD_compResetPolicy_e const crp, U32 const forCCtx)
+{
+    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    size_t const hSize = ((size_t)1) << cParams->hashLog;
+    U32    const hashLog3 = (forCCtx && cParams->searchLength==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
+    size_t const h3Size = ((size_t)1) << hashLog3;
+    size_t const tableSpace = (chainSize + hSize + h3Size) * sizeof(U32);
+
+    assert(((size_t)ptr & 3) == 0);
+
+    ms->hashLog3 = hashLog3;
+    memset(&ms->window, 0, sizeof(ms->window));
+    ZSTD_invalidateMatchState(ms);
+
+    /* opt parser space */
+    if (forCCtx && ((cParams->strategy == ZSTD_btopt) | (cParams->strategy == ZSTD_btultra))) {
+        DEBUGLOG(4, "reserving optimal parser space");
+        ms->opt.litFreq = (U32*)ptr;
+        ms->opt.litLengthFreq = ms->opt.litFreq + (1<<Litbits);
+        ms->opt.matchLengthFreq = ms->opt.litLengthFreq + (MaxLL+1);
+        ms->opt.offCodeFreq = ms->opt.matchLengthFreq + (MaxML+1);
+        ptr = ms->opt.offCodeFreq + (MaxOff+1);
+        ms->opt.matchTable = (ZSTD_match_t*)ptr;
+        ptr = ms->opt.matchTable + ZSTD_OPT_NUM+1;
+        ms->opt.priceTable = (ZSTD_optimal_t*)ptr;
+        ptr = ms->opt.priceTable + ZSTD_OPT_NUM+1;
+    }
+
+    /* table Space */
+    DEBUGLOG(4, "reset table : %u", crp!=ZSTDcrp_noMemset);
+    assert(((size_t)ptr & 3) == 0);  /* ensure ptr is properly aligned */
+    if (crp!=ZSTDcrp_noMemset) memset(ptr, 0, tableSpace);   /* reset tables only */
+    ms->hashTable = (U32*)(ptr);
+    ms->chainTable = ms->hashTable + hSize;
+    ms->hashTable3 = ms->chainTable + chainSize;
+    ptr = ms->hashTable3 + h3Size;
+
+    assert(((size_t)ptr & 3) == 0);
+    return ptr;
+}
+
+/*! ZSTD_resetCCtx_internal() :
+    note : `params` are assumed fully validated at this stage */
+static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
+                                      ZSTD_CCtx_params params, U64 pledgedSrcSize,
+                                      ZSTD_compResetPolicy_e const crp,
+                                      ZSTD_buffered_policy_e const zbuff)
+{
+    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u",
+                (U32)pledgedSrcSize, params.cParams.windowLog);
+    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+
+    if (crp == ZSTDcrp_continue) {
+        if (ZSTD_equivalentParams(zc->appliedParams, params,
+                                zc->inBuffSize, zc->blockSize,
+                                zbuff, pledgedSrcSize)) {
+            DEBUGLOG(4, "ZSTD_equivalentParams()==1 -> continue mode (wLog1=%u, blockSize1=%u)",
+                        zc->appliedParams.cParams.windowLog, (U32)zc->blockSize);
+            return ZSTD_continueCCtx(zc, params, pledgedSrcSize);
+    }   }
+    DEBUGLOG(4, "ZSTD_equivalentParams()==0 -> reset CCtx");
+
+    if (params.ldmParams.enableLdm) {
+        /* Adjust long distance matching parameters */
+        params.ldmParams.windowLog = params.cParams.windowLog;
+        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
+        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
+        assert(params.ldmParams.hashEveryLog < 32);
+        zc->ldmState.hashPower =
+                ZSTD_ldm_getHashPower(params.ldmParams.minMatchLength);
+    }
+
+    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params.cParams.windowLog), pledgedSrcSize));
+        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
+        U32    const divider = (params.cParams.searchLength==3) ? 3 : 4;
+        size_t const maxNbSeq = blockSize / divider;
+        size_t const tokenSpace = blockSize + 11*maxNbSeq;
+        size_t const buffOutSize = (zbuff==ZSTDb_buffered) ? ZSTD_compressBound(blockSize)+1 : 0;
+        size_t const buffInSize = (zbuff==ZSTDb_buffered) ? windowSize + blockSize : 0;
+        size_t const matchStateSize = ZSTD_sizeof_matchState(&params.cParams, /* forCCtx */ 1);
+        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params.ldmParams, blockSize);
+        void* ptr;
+
+        /* Check if workSpace is large enough, alloc a new one if needed */
+        {   size_t const entropySpace = HUF_WORKSPACE_SIZE;
+            size_t const blockStateSpace = 2 * sizeof(ZSTD_compressedBlockState_t);
+            size_t const bufferSpace = buffInSize + buffOutSize;
+            size_t const ldmSpace = ZSTD_ldm_getTableSize(params.ldmParams);
+            size_t const ldmSeqSpace = maxNbLdmSeq * sizeof(rawSeq);
+
+            size_t const neededSpace = entropySpace + blockStateSpace + ldmSpace +
+                                       ldmSeqSpace + matchStateSize + tokenSpace +
+                                       bufferSpace;
+            DEBUGLOG(4, "Need %uKB workspace, including %uKB for match state, and %uKB for buffers",
+                        (U32)(neededSpace>>10), (U32)(matchStateSize>>10), (U32)(bufferSpace>>10));
+            DEBUGLOG(4, "windowSize: %u - blockSize: %u", (U32)windowSize, (U32)blockSize);
+
+            if (zc->workSpaceSize < neededSpace) {  /* too small : resize */
+                DEBUGLOG(4, "Need to update workSpaceSize from %uK to %uK",
+                            (unsigned)(zc->workSpaceSize>>10),
+                            (unsigned)(neededSpace>>10));
+                /* static cctx : no resize, error out */
+                if (zc->staticSize) return ERROR(memory_allocation);
+
+                zc->workSpaceSize = 0;
+                ZSTD_free(zc->workSpace, zc->customMem);
+                zc->workSpace = ZSTD_malloc(neededSpace, zc->customMem);
+                if (zc->workSpace == NULL) return ERROR(memory_allocation);
+                zc->workSpaceSize = neededSpace;
+                ptr = zc->workSpace;
+
+                /* Statically sized space. entropyWorkspace never moves (but prev/next block swap places) */
+                assert(((size_t)zc->workSpace & 3) == 0);   /* ensure correct alignment */
+                assert(zc->workSpaceSize >= 2 * sizeof(ZSTD_compressedBlockState_t));
+                zc->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)zc->workSpace;
+                zc->blockState.nextCBlock = zc->blockState.prevCBlock + 1;
+                ptr = zc->blockState.nextCBlock + 1;
+                zc->entropyWorkspace = (U32*)ptr;
+        }   }
+
+        /* init params */
+        zc->appliedParams = params;
+        zc->pledgedSrcSizePlusOne = pledgedSrcSize+1;
+        zc->consumedSrcSize = 0;
+        zc->producedCSize = 0;
+        if (pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN)
+            zc->appliedParams.fParams.contentSizeFlag = 0;
+        DEBUGLOG(4, "pledged content size : %u ; flag : %u",
+            (U32)pledgedSrcSize, zc->appliedParams.fParams.contentSizeFlag);
+        zc->blockSize = blockSize;
+
+        XXH64_reset(&zc->xxhState, 0);
+        zc->stage = ZSTDcs_init;
+        zc->dictID = 0;
+
+        ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock);
+
+        ptr = zc->entropyWorkspace + HUF_WORKSPACE_SIZE_U32;
+
+        /* ldm hash table */
+        /* initialize bucketOffsets table later for pointer alignment */
+        if (params.ldmParams.enableLdm) {
+            size_t const ldmHSize = ((size_t)1) << params.ldmParams.hashLog;
+            memset(ptr, 0, ldmHSize * sizeof(ldmEntry_t));
+            assert(((size_t)ptr & 3) == 0); /* ensure ptr is properly aligned */
+            zc->ldmState.hashTable = (ldmEntry_t*)ptr;
+            ptr = zc->ldmState.hashTable + ldmHSize;
+            zc->ldmSequences = (rawSeq*)ptr;
+            ptr = zc->ldmSequences + maxNbLdmSeq;
+            zc->maxNbLdmSequences = maxNbLdmSeq;
+
+            memset(&zc->ldmState.window, 0, sizeof(zc->ldmState.window));
+        }
+        assert(((size_t)ptr & 3) == 0); /* ensure ptr is properly aligned */
+
+        ptr = ZSTD_reset_matchState(&zc->blockState.matchState, ptr, &params.cParams, crp, /* forCCtx */ 1);
+
+        /* sequences storage */
+        zc->seqStore.sequencesStart = (seqDef*)ptr;
+        ptr = zc->seqStore.sequencesStart + maxNbSeq;
+        zc->seqStore.llCode = (BYTE*) ptr;
+        zc->seqStore.mlCode = zc->seqStore.llCode + maxNbSeq;
+        zc->seqStore.ofCode = zc->seqStore.mlCode + maxNbSeq;
+        zc->seqStore.litStart = zc->seqStore.ofCode + maxNbSeq;
+        ptr = zc->seqStore.litStart + blockSize;
+
+        /* ldm bucketOffsets table */
+        if (params.ldmParams.enableLdm) {
+            size_t const ldmBucketSize =
+                  ((size_t)1) << (params.ldmParams.hashLog -
+                                  params.ldmParams.bucketSizeLog);
+            memset(ptr, 0, ldmBucketSize);
+            zc->ldmState.bucketOffsets = (BYTE*)ptr;
+            ptr = zc->ldmState.bucketOffsets + ldmBucketSize;
+            ZSTD_window_clear(&zc->ldmState.window);
+        }
+        ZSTD_referenceExternalSequences(zc, NULL, 0);
+
+        /* buffers */
+        zc->inBuffSize = buffInSize;
+        zc->inBuff = (char*)ptr;
+        zc->outBuffSize = buffOutSize;
+        zc->outBuff = zc->inBuff + buffInSize;
+
+        return 0;
+    }
+}
+
+/* ZSTD_invalidateRepCodes() :
+ * ensures next compression will not use repcodes from previous block.
+ * Note : only works with regular variant;
+ *        do not use with extDict variant ! */
+void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx) {
+    int i;
+    for (i=0; i<ZSTD_REP_NUM; i++) cctx->blockState.prevCBlock->rep[i] = 0;
+    assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
+}
+
+static size_t ZSTD_resetCCtx_usingCDict(ZSTD_CCtx* cctx,
+                            const ZSTD_CDict* cdict,
+                            unsigned windowLog,
+                            ZSTD_frameParameters fParams,
+                            U64 pledgedSrcSize,
+                            ZSTD_buffered_policy_e zbuff)
+{
+    {   ZSTD_CCtx_params params = cctx->requestedParams;
+        /* Copy only compression parameters related to tables. */
+        params.cParams = cdict->cParams;
+        if (windowLog) params.cParams.windowLog = windowLog;
+        params.fParams = fParams;
+        ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+                                ZSTDcrp_noMemset, zbuff);
+        assert(cctx->appliedParams.cParams.strategy == cdict->cParams.strategy);
+        assert(cctx->appliedParams.cParams.hashLog == cdict->cParams.hashLog);
+        assert(cctx->appliedParams.cParams.chainLog == cdict->cParams.chainLog);
+    }
+
+    /* copy tables */
+    {   size_t const chainSize = (cdict->cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << cdict->cParams.chainLog);
+        size_t const hSize =  (size_t)1 << cdict->cParams.hashLog;
+        size_t const tableSpace = (chainSize + hSize) * sizeof(U32);
+        assert((U32*)cctx->blockState.matchState.chainTable == (U32*)cctx->blockState.matchState.hashTable + hSize);  /* chainTable must follow hashTable */
+        assert((U32*)cctx->blockState.matchState.hashTable3 == (U32*)cctx->blockState.matchState.chainTable + chainSize);
+        assert((U32*)cdict->matchState.chainTable == (U32*)cdict->matchState.hashTable + hSize);  /* chainTable must follow hashTable */
+        assert((U32*)cdict->matchState.hashTable3 == (U32*)cdict->matchState.chainTable + chainSize);
+        memcpy(cctx->blockState.matchState.hashTable, cdict->matchState.hashTable, tableSpace);   /* presumes all tables follow each other */
+    }
+    /* Zero the hashTable3, since the cdict never fills it */
+    {   size_t const h3Size = (size_t)1 << cctx->blockState.matchState.hashLog3;
+        assert(cdict->matchState.hashLog3 == 0);
+        memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32));
+    }
+
+    /* copy dictionary offsets */
+    {
+        ZSTD_matchState_t const* srcMatchState = &cdict->matchState;
+        ZSTD_matchState_t* dstMatchState = &cctx->blockState.matchState;
+        dstMatchState->window       = srcMatchState->window;
+        dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
+        dstMatchState->nextToUpdate3= srcMatchState->nextToUpdate3;
+        dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
+    }
+    cctx->dictID = cdict->dictID;
+
+    /* copy block state */
+    memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
+
+    return 0;
+}
+
+/*! ZSTD_copyCCtx_internal() :
+ *  Duplicate an existing context `srcCCtx` into another one `dstCCtx`.
+ *  Only works during stage ZSTDcs_init (i.e. after creation, but before first call to ZSTD_compressContinue()).
+ *  The "context", in this case, refers to the hash and chain tables,
+ *  entropy tables, and dictionary references.
+ * `windowLog` value is enforced if != 0, otherwise value is copied from srcCCtx.
+ * @return : 0, or an error code */
+static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
+                            const ZSTD_CCtx* srcCCtx,
+                            ZSTD_frameParameters fParams,
+                            U64 pledgedSrcSize,
+                            ZSTD_buffered_policy_e zbuff)
+{
+    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
+    if (srcCCtx->stage!=ZSTDcs_init) return ERROR(stage_wrong);
+
+    memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem));
+    {   ZSTD_CCtx_params params = dstCCtx->requestedParams;
+        /* Copy only compression parameters related to tables. */
+        params.cParams = srcCCtx->appliedParams.cParams;
+        params.fParams = fParams;
+        ZSTD_resetCCtx_internal(dstCCtx, params, pledgedSrcSize,
+                                ZSTDcrp_noMemset, zbuff);
+        assert(dstCCtx->appliedParams.cParams.windowLog == srcCCtx->appliedParams.cParams.windowLog);
+        assert(dstCCtx->appliedParams.cParams.strategy == srcCCtx->appliedParams.cParams.strategy);
+        assert(dstCCtx->appliedParams.cParams.hashLog == srcCCtx->appliedParams.cParams.hashLog);
+        assert(dstCCtx->appliedParams.cParams.chainLog == srcCCtx->appliedParams.cParams.chainLog);
+        assert(dstCCtx->blockState.matchState.hashLog3 == srcCCtx->blockState.matchState.hashLog3);
+    }
+
+    /* copy tables */
+    {   size_t const chainSize = (srcCCtx->appliedParams.cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog);
+        size_t const hSize =  (size_t)1 << srcCCtx->appliedParams.cParams.hashLog;
+        size_t const h3Size = (size_t)1 << srcCCtx->blockState.matchState.hashLog3;
+        size_t const tableSpace = (chainSize + hSize + h3Size) * sizeof(U32);
+        assert((U32*)dstCCtx->blockState.matchState.chainTable == (U32*)dstCCtx->blockState.matchState.hashTable + hSize);  /* chainTable must follow hashTable */
+        assert((U32*)dstCCtx->blockState.matchState.hashTable3 == (U32*)dstCCtx->blockState.matchState.chainTable + chainSize);
+        memcpy(dstCCtx->blockState.matchState.hashTable, srcCCtx->blockState.matchState.hashTable, tableSpace);   /* presumes all tables follow each other */
+    }
+
+    /* copy dictionary offsets */
+    {
+        ZSTD_matchState_t const* srcMatchState = &srcCCtx->blockState.matchState;
+        ZSTD_matchState_t* dstMatchState = &dstCCtx->blockState.matchState;
+        dstMatchState->window       = srcMatchState->window;
+        dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
+        dstMatchState->nextToUpdate3= srcMatchState->nextToUpdate3;
+        dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
+    }
+    dstCCtx->dictID = srcCCtx->dictID;
+
+    /* copy block state */
+    memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock));
+
+    return 0;
+}
+
+/*! ZSTD_copyCCtx() :
+ *  Duplicate an existing context `srcCCtx` into another one `dstCCtx`.
+ *  Only works during stage ZSTDcs_init (i.e. after creation, but before first call to ZSTD_compressContinue()).
+ *  pledgedSrcSize==0 means "unknown".
+*   @return : 0, or an error code */
+size_t ZSTD_copyCCtx(ZSTD_CCtx* dstCCtx, const ZSTD_CCtx* srcCCtx, unsigned long long pledgedSrcSize)
+{
+    ZSTD_frameParameters fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
+    ZSTD_buffered_policy_e const zbuff = (ZSTD_buffered_policy_e)(srcCCtx->inBuffSize>0);
+    ZSTD_STATIC_ASSERT((U32)ZSTDb_buffered==1);
+    if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    fParams.contentSizeFlag = (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN);
+
+    return ZSTD_copyCCtx_internal(dstCCtx, srcCCtx,
+                                fParams, pledgedSrcSize,
+                                zbuff);
+}
+
+
+#define ZSTD_ROWSIZE 16
+/*! ZSTD_reduceTable() :
+ *  reduce table indexes by `reducerValue`, or squash to zero.
+ *  PreserveMark preserves "unsorted mark" for btlazy2 strategy.
+ *  It must be set to a clear 0/1 value, to remove branch during inlining.
+ *  Presume table size is a multiple of ZSTD_ROWSIZE
+ *  to help auto-vectorization */
+FORCE_INLINE_TEMPLATE void
+ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerValue, int const preserveMark)
+{
+    int const nbRows = (int)size / ZSTD_ROWSIZE;
+    int cellNb = 0;
+    int rowNb;
+    assert((size & (ZSTD_ROWSIZE-1)) == 0);  /* multiple of ZSTD_ROWSIZE */
+    assert(size < (1U<<31));   /* can be casted to int */
+    for (rowNb=0 ; rowNb < nbRows ; rowNb++) {
+        int column;
+        for (column=0; column<ZSTD_ROWSIZE; column++) {
+            if (preserveMark) {
+                U32 const adder = (table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) ? reducerValue : 0;
+                table[cellNb] += adder;
+            }
+            if (table[cellNb] < reducerValue) table[cellNb] = 0;
+            else table[cellNb] -= reducerValue;
+            cellNb++;
+    }   }
+}
+
+static void ZSTD_reduceTable(U32* const table, U32 const size, U32 const reducerValue)
+{
+    ZSTD_reduceTable_internal(table, size, reducerValue, 0);
+}
+
+static void ZSTD_reduceTable_btlazy2(U32* const table, U32 const size, U32 const reducerValue)
+{
+    ZSTD_reduceTable_internal(table, size, reducerValue, 1);
+}
+
+/*! ZSTD_reduceIndex() :
+*   rescale all indexes to avoid future overflow (indexes are U32) */
+static void ZSTD_reduceIndex (ZSTD_CCtx* zc, const U32 reducerValue)
+{
+    ZSTD_matchState_t* const ms = &zc->blockState.matchState;
+    {   U32 const hSize = (U32)1 << zc->appliedParams.cParams.hashLog;
+        ZSTD_reduceTable(ms->hashTable, hSize, reducerValue);
+    }
+
+    if (zc->appliedParams.cParams.strategy != ZSTD_fast) {
+        U32 const chainSize = (U32)1 << zc->appliedParams.cParams.chainLog;
+        if (zc->appliedParams.cParams.strategy == ZSTD_btlazy2)
+            ZSTD_reduceTable_btlazy2(ms->chainTable, chainSize, reducerValue);
+        else
+            ZSTD_reduceTable(ms->chainTable, chainSize, reducerValue);
+    }
+
+    if (ms->hashLog3) {
+        U32 const h3Size = (U32)1 << ms->hashLog3;
+        ZSTD_reduceTable(ms->hashTable3, h3Size, reducerValue);
+    }
+}
+
+
+/*-*******************************************************
+*  Block entropic compression
+*********************************************************/
+
+/* See doc/zstd_compression_format.md for detailed format description */
+
+size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    if (srcSize + ZSTD_blockHeaderSize > dstCapacity) return ERROR(dstSize_tooSmall);
+    memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
+    MEM_writeLE24(dst, (U32)(srcSize << 2) + (U32)bt_raw);
+    return ZSTD_blockHeaderSize+srcSize;
+}
+
+
+static size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    BYTE* const ostart = (BYTE* const)dst;
+    U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
+
+    if (srcSize + flSize > dstCapacity) return ERROR(dstSize_tooSmall);
+
+    switch(flSize)
+    {
+        case 1: /* 2 - 1 - 5 */
+            ostart[0] = (BYTE)((U32)set_basic + (srcSize<<3));
+            break;
+        case 2: /* 2 - 2 - 12 */
+            MEM_writeLE16(ostart, (U16)((U32)set_basic + (1<<2) + (srcSize<<4)));
+            break;
+        case 3: /* 2 - 2 - 20 */
+            MEM_writeLE32(ostart, (U32)((U32)set_basic + (3<<2) + (srcSize<<4)));
+            break;
+        default:   /* not necessary : flSize is {1,2,3} */
+            assert(0);
+    }
+
+    memcpy(ostart + flSize, src, srcSize);
+    return srcSize + flSize;
+}
+
+static size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    BYTE* const ostart = (BYTE* const)dst;
+    U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
+
+    (void)dstCapacity;  /* dstCapacity already guaranteed to be >=4, hence large enough */
+
+    switch(flSize)
+    {
+        case 1: /* 2 - 1 - 5 */
+            ostart[0] = (BYTE)((U32)set_rle + (srcSize<<3));
+            break;
+        case 2: /* 2 - 2 - 12 */
+            MEM_writeLE16(ostart, (U16)((U32)set_rle + (1<<2) + (srcSize<<4)));
+            break;
+        case 3: /* 2 - 2 - 20 */
+            MEM_writeLE32(ostart, (U32)((U32)set_rle + (3<<2) + (srcSize<<4)));
+            break;
+        default:   /* not necessary : flSize is {1,2,3} */
+            assert(0);
+    }
+
+    ostart[flSize] = *(const BYTE*)src;
+    return flSize+1;
+}
+
+
+static size_t ZSTD_minGain(size_t srcSize) { return (srcSize >> 6) + 2; }
+
+static size_t ZSTD_compressLiterals (ZSTD_entropyCTables_t const* prevEntropy,
+                                     ZSTD_entropyCTables_t* nextEntropy,
+                                     ZSTD_strategy strategy, int disableLiteralCompression,
+                                     void* dst, size_t dstCapacity,
+                               const void* src, size_t srcSize,
+                                     U32* workspace, const int bmi2)
+{
+    size_t const minGain = ZSTD_minGain(srcSize);
+    size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
+    BYTE*  const ostart = (BYTE*)dst;
+    U32 singleStream = srcSize < 256;
+    symbolEncodingType_e hType = set_compressed;
+    size_t cLitSize;
+
+    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i)",
+                disableLiteralCompression);
+
+    /* Prepare nextEntropy assuming reusing the existing table */
+    nextEntropy->hufCTable_repeatMode = prevEntropy->hufCTable_repeatMode;
+    memcpy(nextEntropy->hufCTable, prevEntropy->hufCTable,
+           sizeof(prevEntropy->hufCTable));
+
+    if (disableLiteralCompression)
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+
+    /* small ? don't even attempt compression (speed opt) */
+#   define COMPRESS_LITERALS_SIZE_MIN 63
+    {   size_t const minLitSize = (prevEntropy->hufCTable_repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
+        if (srcSize <= minLitSize) return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+    }
+
+    if (dstCapacity < lhSize+1) return ERROR(dstSize_tooSmall);   /* not enough space for compression */
+    {   HUF_repeat repeat = prevEntropy->hufCTable_repeatMode;
+        int const preferRepeat = strategy < ZSTD_lazy ? srcSize <= 1024 : 0;
+        if (repeat == HUF_repeat_valid && lhSize == 3) singleStream = 1;
+        cLitSize = singleStream ? HUF_compress1X_repeat(ostart+lhSize, dstCapacity-lhSize, src, srcSize, 255, 11,
+                                      workspace, HUF_WORKSPACE_SIZE, (HUF_CElt*)nextEntropy->hufCTable, &repeat, preferRepeat, bmi2)
+                                : HUF_compress4X_repeat(ostart+lhSize, dstCapacity-lhSize, src, srcSize, 255, 11,
+                                      workspace, HUF_WORKSPACE_SIZE, (HUF_CElt*)nextEntropy->hufCTable, &repeat, preferRepeat, bmi2);
+        if (repeat != HUF_repeat_none) {
+            /* reused the existing table */
+            hType = set_repeat;
+        }
+    }
+
+    if ((cLitSize==0) | (cLitSize >= srcSize - minGain) | ERR_isError(cLitSize)) {
+        memcpy(nextEntropy->hufCTable, prevEntropy->hufCTable, sizeof(prevEntropy->hufCTable));
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+    }
+    if (cLitSize==1) {
+        memcpy(nextEntropy->hufCTable, prevEntropy->hufCTable, sizeof(prevEntropy->hufCTable));
+        return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
+    }
+
+    if (hType == set_compressed) {
+        /* using a newly constructed table */
+        nextEntropy->hufCTable_repeatMode = HUF_repeat_check;
+    }
+
+    /* Build header */
+    switch(lhSize)
+    {
+    case 3: /* 2 - 2 - 10 - 10 */
+        {   U32 const lhc = hType + ((!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
+            MEM_writeLE24(ostart, lhc);
+            break;
+        }
+    case 4: /* 2 - 2 - 14 - 14 */
+        {   U32 const lhc = hType + (2 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<18);
+            MEM_writeLE32(ostart, lhc);
+            break;
+        }
+    case 5: /* 2 - 2 - 18 - 18 */
+        {   U32 const lhc = hType + (3 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<22);
+            MEM_writeLE32(ostart, lhc);
+            ostart[4] = (BYTE)(cLitSize >> 10);
+            break;
+        }
+    default:  /* not possible : lhSize is {3,4,5} */
+        assert(0);
+    }
+    return lhSize+cLitSize;
+}
+
+
+void ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
+{
+    const seqDef* const sequences = seqStorePtr->sequencesStart;
+    BYTE* const llCodeTable = seqStorePtr->llCode;
+    BYTE* const ofCodeTable = seqStorePtr->ofCode;
+    BYTE* const mlCodeTable = seqStorePtr->mlCode;
+    U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    U32 u;
+    for (u=0; u<nbSeq; u++) {
+        U32 const llv = sequences[u].litLength;
+        U32 const mlv = sequences[u].matchLength;
+        llCodeTable[u] = (BYTE)ZSTD_LLcode(llv);
+        ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offset);
+        mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv);
+    }
+    if (seqStorePtr->longLengthID==1)
+        llCodeTable[seqStorePtr->longLengthPos] = MaxLL;
+    if (seqStorePtr->longLengthID==2)
+        mlCodeTable[seqStorePtr->longLengthPos] = MaxML;
+}
+
+typedef enum {
+    ZSTD_defaultDisallowed = 0,
+    ZSTD_defaultAllowed = 1
+} ZSTD_defaultPolicy_e;
+
+MEM_STATIC
+symbolEncodingType_e ZSTD_selectEncodingType(
+        FSE_repeat* repeatMode, size_t const mostFrequent, size_t nbSeq,
+        U32 defaultNormLog, ZSTD_defaultPolicy_e const isDefaultAllowed)
+{
+#define MIN_SEQ_FOR_DYNAMIC_FSE   64
+#define MAX_SEQ_FOR_STATIC_FSE  1000
+    ZSTD_STATIC_ASSERT(ZSTD_defaultDisallowed == 0 && ZSTD_defaultAllowed != 0);
+    if ((mostFrequent == nbSeq) && (!isDefaultAllowed || nbSeq > 2)) {
+        DEBUGLOG(5, "Selected set_rle");
+        /* Prefer set_basic over set_rle when there are 2 or less symbols,
+         * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
+         * If basic encoding isn't possible, always choose RLE.
+         */
+        *repeatMode = FSE_repeat_check;
+        return set_rle;
+    }
+    if ( isDefaultAllowed
+      && (*repeatMode == FSE_repeat_valid) && (nbSeq < MAX_SEQ_FOR_STATIC_FSE)) {
+        DEBUGLOG(5, "Selected set_repeat");
+        return set_repeat;
+    }
+    if ( isDefaultAllowed
+      && ((nbSeq < MIN_SEQ_FOR_DYNAMIC_FSE) || (mostFrequent < (nbSeq >> (defaultNormLog-1)))) ) {
+        DEBUGLOG(5, "Selected set_basic");
+        /* The format allows default tables to be repeated, but it isn't useful.
+         * When using simple heuristics to select encoding type, we don't want
+         * to confuse these tables with dictionaries. When running more careful
+         * analysis, we don't need to waste time checking both repeating tables
+         * and default tables.
+         */
+        *repeatMode = FSE_repeat_none;
+        return set_basic;
+    }
+    DEBUGLOG(5, "Selected set_compressed");
+    *repeatMode = FSE_repeat_check;
+    return set_compressed;
+}
+
+MEM_STATIC
+size_t ZSTD_buildCTable(void* dst, size_t dstCapacity,
+        FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
+        U32* count, U32 max,
+        BYTE const* codeTable, size_t nbSeq,
+        S16 const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
+        FSE_CTable const* prevCTable, size_t prevCTableSize,
+        void* workspace, size_t workspaceSize)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE const* const oend = op + dstCapacity;
+
+    switch (type) {
+    case set_rle:
+        *op = codeTable[0];
+        CHECK_F(FSE_buildCTable_rle(nextCTable, (BYTE)max));
+        return 1;
+    case set_repeat:
+        memcpy(nextCTable, prevCTable, prevCTableSize);
+        return 0;
+    case set_basic:
+        CHECK_F(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, workspace, workspaceSize));  /* note : could be pre-calculated */
+        return 0;
+    case set_compressed: {
+        S16 norm[MaxSeq + 1];
+        size_t nbSeq_1 = nbSeq;
+        const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
+        if (count[codeTable[nbSeq-1]] > 1) {
+            count[codeTable[nbSeq-1]]--;
+            nbSeq_1--;
+        }
+        assert(nbSeq_1 > 1);
+        CHECK_F(FSE_normalizeCount(norm, tableLog, count, nbSeq_1, max));
+        {   size_t const NCountSize = FSE_writeNCount(op, oend - op, norm, max, tableLog);   /* overflow protected */
+            if (FSE_isError(NCountSize)) return NCountSize;
+            CHECK_F(FSE_buildCTable_wksp(nextCTable, norm, max, tableLog, workspace, workspaceSize));
+            return NCountSize;
+        }
+    }
+    default: return assert(0), ERROR(GENERIC);
+    }
+}
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_encodeSequences_body(
+            void* dst, size_t dstCapacity,
+            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
+            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
+            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
+            seqDef const* sequences, size_t nbSeq, int longOffsets)
+{
+    BIT_CStream_t blockStream;
+    FSE_CState_t  stateMatchLength;
+    FSE_CState_t  stateOffsetBits;
+    FSE_CState_t  stateLitLength;
+
+    CHECK_E(BIT_initCStream(&blockStream, dst, dstCapacity), dstSize_tooSmall); /* not enough space remaining */
+
+    /* first symbols */
+    FSE_initCState2(&stateMatchLength, CTable_MatchLength, mlCodeTable[nbSeq-1]);
+    FSE_initCState2(&stateOffsetBits,  CTable_OffsetBits,  ofCodeTable[nbSeq-1]);
+    FSE_initCState2(&stateLitLength,   CTable_LitLength,   llCodeTable[nbSeq-1]);
+    BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]);
+    if (MEM_32bits()) BIT_flushBits(&blockStream);
+    BIT_addBits(&blockStream, sequences[nbSeq-1].matchLength, ML_bits[mlCodeTable[nbSeq-1]]);
+    if (MEM_32bits()) BIT_flushBits(&blockStream);
+    if (longOffsets) {
+        U32 const ofBits = ofCodeTable[nbSeq-1];
+        int const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
+        if (extraBits) {
+            BIT_addBits(&blockStream, sequences[nbSeq-1].offset, extraBits);
+            BIT_flushBits(&blockStream);
+        }
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offset >> extraBits,
+                    ofBits - extraBits);
+    } else {
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offset, ofCodeTable[nbSeq-1]);
+    }
+    BIT_flushBits(&blockStream);
+
+    {   size_t n;
+        for (n=nbSeq-2 ; n<nbSeq ; n--) {      /* intentional underflow */
+            BYTE const llCode = llCodeTable[n];
+            BYTE const ofCode = ofCodeTable[n];
+            BYTE const mlCode = mlCodeTable[n];
+            U32  const llBits = LL_bits[llCode];
+            U32  const ofBits = ofCode;
+            U32  const mlBits = ML_bits[mlCode];
+            DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u",
+                        sequences[n].litLength,
+                        sequences[n].matchLength + MINMATCH,
+                        sequences[n].offset);
+                                                                            /* 32b*/  /* 64b*/
+                                                                            /* (7)*/  /* (7)*/
+            FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode);       /* 15 */  /* 15 */
+            FSE_encodeSymbol(&blockStream, &stateMatchLength, mlCode);      /* 24 */  /* 24 */
+            if (MEM_32bits()) BIT_flushBits(&blockStream);                  /* (7)*/
+            FSE_encodeSymbol(&blockStream, &stateLitLength, llCode);        /* 16 */  /* 33 */
+            if (MEM_32bits() || (ofBits+mlBits+llBits >= 64-7-(LLFSELog+MLFSELog+OffFSELog)))
+                BIT_flushBits(&blockStream);                                /* (7)*/
+            BIT_addBits(&blockStream, sequences[n].litLength, llBits);
+            if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream);
+            BIT_addBits(&blockStream, sequences[n].matchLength, mlBits);
+            if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream);
+            if (longOffsets) {
+                int const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
+                if (extraBits) {
+                    BIT_addBits(&blockStream, sequences[n].offset, extraBits);
+                    BIT_flushBits(&blockStream);                            /* (7)*/
+                }
+                BIT_addBits(&blockStream, sequences[n].offset >> extraBits,
+                            ofBits - extraBits);                            /* 31 */
+            } else {
+                BIT_addBits(&blockStream, sequences[n].offset, ofBits);     /* 31 */
+            }
+            BIT_flushBits(&blockStream);                                    /* (7)*/
+    }   }
+
+    DEBUGLOG(6, "ZSTD_encodeSequences: flushing ML state with %u bits", stateMatchLength.stateLog);
+    FSE_flushCState(&blockStream, &stateMatchLength);
+    DEBUGLOG(6, "ZSTD_encodeSequences: flushing Off state with %u bits", stateOffsetBits.stateLog);
+    FSE_flushCState(&blockStream, &stateOffsetBits);
+    DEBUGLOG(6, "ZSTD_encodeSequences: flushing LL state with %u bits", stateLitLength.stateLog);
+    FSE_flushCState(&blockStream, &stateLitLength);
+
+    {   size_t const streamSize = BIT_closeCStream(&blockStream);
+        if (streamSize==0) return ERROR(dstSize_tooSmall);   /* not enough space */
+        return streamSize;
+    }
+}
+
+static size_t
+ZSTD_encodeSequences_default(
+            void* dst, size_t dstCapacity,
+            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
+            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
+            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
+            seqDef const* sequences, size_t nbSeq, int longOffsets)
+{
+    return ZSTD_encodeSequences_body(dst, dstCapacity,
+                                    CTable_MatchLength, mlCodeTable,
+                                    CTable_OffsetBits, ofCodeTable,
+                                    CTable_LitLength, llCodeTable,
+                                    sequences, nbSeq, longOffsets);
+}
+
+
+#if DYNAMIC_BMI2
+
+static TARGET_ATTRIBUTE("bmi2") size_t
+ZSTD_encodeSequences_bmi2(
+            void* dst, size_t dstCapacity,
+            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
+            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
+            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
+            seqDef const* sequences, size_t nbSeq, int longOffsets)
+{
+    return ZSTD_encodeSequences_body(dst, dstCapacity,
+                                    CTable_MatchLength, mlCodeTable,
+                                    CTable_OffsetBits, ofCodeTable,
+                                    CTable_LitLength, llCodeTable,
+                                    sequences, nbSeq, longOffsets);
+}
+
+#endif
+
+size_t ZSTD_encodeSequences(
+            void* dst, size_t dstCapacity,
+            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
+            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
+            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
+            seqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return ZSTD_encodeSequences_bmi2(dst, dstCapacity,
+                                         CTable_MatchLength, mlCodeTable,
+                                         CTable_OffsetBits, ofCodeTable,
+                                         CTable_LitLength, llCodeTable,
+                                         sequences, nbSeq, longOffsets);
+    }
+#endif
+    (void)bmi2;
+    return ZSTD_encodeSequences_default(dst, dstCapacity,
+                                        CTable_MatchLength, mlCodeTable,
+                                        CTable_OffsetBits, ofCodeTable,
+                                        CTable_LitLength, llCodeTable,
+                                        sequences, nbSeq, longOffsets);
+}
+
+MEM_STATIC size_t ZSTD_compressSequences_internal(seqStore_t* seqStorePtr,
+                              ZSTD_entropyCTables_t const* prevEntropy,
+                              ZSTD_entropyCTables_t* nextEntropy,
+                              ZSTD_CCtx_params const* cctxParams,
+                              void* dst, size_t dstCapacity, U32* workspace,
+                              const int bmi2)
+{
+    const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
+    U32 count[MaxSeq+1];
+    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
+    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
+    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
+    U32 LLtype, Offtype, MLtype;   /* compressed, raw or rle */
+    const seqDef* const sequences = seqStorePtr->sequencesStart;
+    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
+    const BYTE* const llCodeTable = seqStorePtr->llCode;
+    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = ostart + dstCapacity;
+    BYTE* op = ostart;
+    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
+    BYTE* seqHead;
+
+    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
+
+    /* Compress literals */
+    {   const BYTE* const literals = seqStorePtr->litStart;
+        size_t const litSize = seqStorePtr->lit - literals;
+        size_t const cSize = ZSTD_compressLiterals(
+                                    prevEntropy, nextEntropy,
+                                    cctxParams->cParams.strategy, cctxParams->disableLiteralCompression,
+                                    op, dstCapacity,
+                                    literals, litSize,
+                                    workspace, bmi2);
+        if (ZSTD_isError(cSize))
+          return cSize;
+        assert(cSize <= dstCapacity);
+        op += cSize;
+    }
+
+    /* Sequences Header */
+    if ((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/) return ERROR(dstSize_tooSmall);
+    if (nbSeq < 0x7F)
+        *op++ = (BYTE)nbSeq;
+    else if (nbSeq < LONGNBSEQ)
+        op[0] = (BYTE)((nbSeq>>8) + 0x80), op[1] = (BYTE)nbSeq, op+=2;
+    else
+        op[0]=0xFF, MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ)), op+=3;
+    if (nbSeq==0) {
+      memcpy(nextEntropy->litlengthCTable, prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable));
+      nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
+      memcpy(nextEntropy->offcodeCTable, prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable));
+      nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
+      memcpy(nextEntropy->matchlengthCTable, prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable));
+      nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
+      return op - ostart;
+    }
+
+    /* seqHead : flags for FSE encoding type */
+    seqHead = op++;
+
+    /* convert length/distances into codes */
+    ZSTD_seqToCodes(seqStorePtr);
+    /* build CTable for Literal Lengths */
+    {   U32 max = MaxLL;
+        size_t const mostFrequent = FSE_countFast_wksp(count, &max, llCodeTable, nbSeq, workspace);
+        DEBUGLOG(5, "Building LL table");
+        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
+        LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode, mostFrequent, nbSeq, LL_defaultNormLog, ZSTD_defaultAllowed);
+        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
+                    count, max, llCodeTable, nbSeq, LL_defaultNorm, LL_defaultNormLog, MaxLL,
+                    prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable),
+                    workspace, HUF_WORKSPACE_SIZE);
+            if (ZSTD_isError(countSize)) return countSize;
+            op += countSize;
+    }   }
+    /* build CTable for Offsets */
+    {   U32 max = MaxOff;
+        size_t const mostFrequent = FSE_countFast_wksp(count, &max, ofCodeTable, nbSeq, workspace);
+        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
+        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
+        DEBUGLOG(5, "Building OF table");
+        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
+        Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode, mostFrequent, nbSeq, OF_defaultNormLog, defaultPolicy);
+        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
+                    count, max, ofCodeTable, nbSeq, OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
+                    prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable),
+                    workspace, HUF_WORKSPACE_SIZE);
+            if (ZSTD_isError(countSize)) return countSize;
+            op += countSize;
+    }   }
+    /* build CTable for MatchLengths */
+    {   U32 max = MaxML;
+        size_t const mostFrequent = FSE_countFast_wksp(count, &max, mlCodeTable, nbSeq, workspace);
+        DEBUGLOG(5, "Building ML table");
+        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
+        MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode, mostFrequent, nbSeq, ML_defaultNormLog, ZSTD_defaultAllowed);
+        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
+                    count, max, mlCodeTable, nbSeq, ML_defaultNorm, ML_defaultNormLog, MaxML,
+                    prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable),
+                    workspace, HUF_WORKSPACE_SIZE);
+            if (ZSTD_isError(countSize)) return countSize;
+            op += countSize;
+    }   }
+
+    *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
+
+    {   size_t const bitstreamSize = ZSTD_encodeSequences(
+                                        op, oend - op,
+                                        CTable_MatchLength, mlCodeTable,
+                                        CTable_OffsetBits, ofCodeTable,
+                                        CTable_LitLength, llCodeTable,
+                                        sequences, nbSeq,
+                                        longOffsets, bmi2);
+        if (ZSTD_isError(bitstreamSize)) return bitstreamSize;
+        op += bitstreamSize;
+    }
+
+    return op - ostart;
+}
+
+MEM_STATIC size_t ZSTD_compressSequences(seqStore_t* seqStorePtr,
+                              ZSTD_entropyCTables_t const* prevEntropy,
+                              ZSTD_entropyCTables_t* nextEntropy,
+                              ZSTD_CCtx_params const* cctxParams,
+                              void* dst, size_t dstCapacity,
+                              size_t srcSize, U32* workspace, int bmi2)
+{
+    size_t const cSize = ZSTD_compressSequences_internal(
+            seqStorePtr, prevEntropy, nextEntropy, cctxParams, dst, dstCapacity,
+            workspace, bmi2);
+    /* When srcSize <= dstCapacity, there is enough space to write a raw uncompressed block.
+     * Since we ran out of space, block must be not compressible, so fall back to raw uncompressed block.
+     */
+    if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity))
+        return 0;  /* block not compressed */
+    if (ZSTD_isError(cSize)) return cSize;
+
+    /* Check compressibility */
+    {   size_t const maxCSize = srcSize - ZSTD_minGain(srcSize);  /* note : fixed formula, maybe should depend on compression level, or strategy */
+        if (cSize >= maxCSize) return 0;  /* block not compressed */
+    }
+
+    /* We check that dictionaries have offset codes available for the first
+     * block. After the first block, the offcode table might not have large
+     * enough codes to represent the offsets in the data.
+     */
+    if (nextEntropy->offcode_repeatMode == FSE_repeat_valid)
+        nextEntropy->offcode_repeatMode = FSE_repeat_check;
+
+    return cSize;
+}
+
+/* ZSTD_selectBlockCompressor() :
+ * Not static, but internal use only (used by long distance matcher)
+ * assumption : strat is a valid strategy */
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, int extDict)
+{
+    static const ZSTD_blockCompressor blockCompressor[2][(unsigned)ZSTD_btultra+1] = {
+        { ZSTD_compressBlock_fast  /* default for 0 */,
+          ZSTD_compressBlock_fast, ZSTD_compressBlock_doubleFast, ZSTD_compressBlock_greedy,
+          ZSTD_compressBlock_lazy, ZSTD_compressBlock_lazy2, ZSTD_compressBlock_btlazy2,
+          ZSTD_compressBlock_btopt, ZSTD_compressBlock_btultra },
+        { ZSTD_compressBlock_fast_extDict  /* default for 0 */,
+          ZSTD_compressBlock_fast_extDict, ZSTD_compressBlock_doubleFast_extDict, ZSTD_compressBlock_greedy_extDict,
+          ZSTD_compressBlock_lazy_extDict,ZSTD_compressBlock_lazy2_extDict, ZSTD_compressBlock_btlazy2_extDict,
+          ZSTD_compressBlock_btopt_extDict, ZSTD_compressBlock_btultra_extDict }
+    };
+    ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1);
+
+    assert((U32)strat >= (U32)ZSTD_fast);
+    assert((U32)strat <= (U32)ZSTD_btultra);
+    return blockCompressor[extDict!=0][(U32)strat];
+}
+
+static void ZSTD_storeLastLiterals(seqStore_t* seqStorePtr,
+                                   const BYTE* anchor, size_t lastLLSize)
+{
+    memcpy(seqStorePtr->lit, anchor, lastLLSize);
+    seqStorePtr->lit += lastLLSize;
+}
+
+static void ZSTD_resetSeqStore(seqStore_t* ssPtr)
+{
+    ssPtr->lit = ssPtr->litStart;
+    ssPtr->sequences = ssPtr->sequencesStart;
+    ssPtr->longLengthID = 0;
+}
+
+static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
+                                        void* dst, size_t dstCapacity,
+                                        const void* src, size_t srcSize)
+{
+    ZSTD_matchState_t* const ms = &zc->blockState.matchState;
+    DEBUGLOG(5, "ZSTD_compressBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
+                (U32)dstCapacity, ms->window.dictLimit, ms->nextToUpdate);
+    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) {
+        ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.searchLength);
+        return 0;   /* don't even attempt compression below a certain srcSize */
+    }
+    ZSTD_resetSeqStore(&(zc->seqStore));
+
+    /* limited update after a very long match */
+    {   const BYTE* const base = ms->window.base;
+        const BYTE* const istart = (const BYTE*)src;
+        const U32 current = (U32)(istart-base);
+        if (current > ms->nextToUpdate + 384)
+            ms->nextToUpdate = current - MIN(192, (U32)(current - ms->nextToUpdate - 384));
+    }
+
+    /* select and store sequences */
+    {   U32 const extDict = ZSTD_window_hasExtDict(ms->window);
+        size_t lastLLSize;
+        {   int i;
+            for (i = 0; i < ZSTD_REP_NUM; ++i)
+                zc->blockState.nextCBlock->rep[i] = zc->blockState.prevCBlock->rep[i];
+        }
+        if (zc->externSeqStore.pos < zc->externSeqStore.size) {
+            assert(!zc->appliedParams.ldmParams.enableLdm);
+            /* Updates ldmSeqStore.pos */
+            lastLLSize =
+                ZSTD_ldm_blockCompress(&zc->externSeqStore,
+                                       ms, &zc->seqStore,
+                                       zc->blockState.nextCBlock->rep,
+                                       &zc->appliedParams.cParams,
+                                       src, srcSize, extDict);
+            assert(zc->externSeqStore.pos <= zc->externSeqStore.size);
+        } else if (zc->appliedParams.ldmParams.enableLdm) {
+            rawSeqStore_t ldmSeqStore = {NULL, 0, 0, 0};
+
+            ldmSeqStore.seq = zc->ldmSequences;
+            ldmSeqStore.capacity = zc->maxNbLdmSequences;
+            /* Updates ldmSeqStore.size */
+            CHECK_F(ZSTD_ldm_generateSequences(&zc->ldmState, &ldmSeqStore,
+                                               &zc->appliedParams.ldmParams,
+                                               src, srcSize));
+            /* Updates ldmSeqStore.pos */
+            lastLLSize =
+                ZSTD_ldm_blockCompress(&ldmSeqStore,
+                                       ms, &zc->seqStore,
+                                       zc->blockState.nextCBlock->rep,
+                                       &zc->appliedParams.cParams,
+                                       src, srcSize, extDict);
+            assert(ldmSeqStore.pos == ldmSeqStore.size);
+        } else {   /* not long range mode */
+            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, extDict);
+            lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, &zc->appliedParams.cParams, src, srcSize);
+        }
+        {   const BYTE* const lastLiterals = (const BYTE*)src + srcSize - lastLLSize;
+            ZSTD_storeLastLiterals(&zc->seqStore, lastLiterals, lastLLSize);
+    }   }
+
+    /* encode sequences and literals */
+    {   size_t const cSize = ZSTD_compressSequences(&zc->seqStore,
+                                &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
+                                &zc->appliedParams,
+                                dst, dstCapacity,
+                                srcSize, zc->entropyWorkspace, zc->bmi2);
+        if (ZSTD_isError(cSize) || cSize == 0) return cSize;
+        /* confirm repcodes and entropy tables */
+        {   ZSTD_compressedBlockState_t* const tmp = zc->blockState.prevCBlock;
+            zc->blockState.prevCBlock = zc->blockState.nextCBlock;
+            zc->blockState.nextCBlock = tmp;
+        }
+        return cSize;
+    }
+}
+
+
+/*! ZSTD_compress_frameChunk() :
+*   Compress a chunk of data into one or multiple blocks.
+*   All blocks will be terminated, all input will be consumed.
+*   Function will issue an error if there is not enough `dstCapacity` to hold the compressed content.
+*   Frame is supposed already started (header already produced)
+*   @return : compressed size, or an error code
+*/
+static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
+                                     void* dst, size_t dstCapacity,
+                               const void* src, size_t srcSize,
+                                     U32 lastFrameChunk)
+{
+    size_t blockSize = cctx->blockSize;
+    size_t remaining = srcSize;
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* op = ostart;
+    U32 const maxDist = (U32)1 << cctx->appliedParams.cParams.windowLog;
+    assert(cctx->appliedParams.cParams.windowLog <= 31);
+
+    DEBUGLOG(5, "ZSTD_compress_frameChunk (blockSize=%u)", (U32)blockSize);
+    if (cctx->appliedParams.fParams.checksumFlag && srcSize)
+        XXH64_update(&cctx->xxhState, src, srcSize);
+
+    while (remaining) {
+        ZSTD_matchState_t* const ms = &cctx->blockState.matchState;
+        U32 const lastBlock = lastFrameChunk & (blockSize >= remaining);
+
+        if (dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE)
+            return ERROR(dstSize_tooSmall);   /* not enough space to store compressed block */
+        if (remaining < blockSize) blockSize = remaining;
+
+        if (ZSTD_window_needOverflowCorrection(ms->window, ip + blockSize)) {
+            U32 const cycleLog = ZSTD_cycleLog(cctx->appliedParams.cParams.chainLog, cctx->appliedParams.cParams.strategy);
+            U32 const correction = ZSTD_window_correctOverflow(&ms->window, cycleLog, maxDist, ip);
+            ZSTD_STATIC_ASSERT(ZSTD_CHAINLOG_MAX <= 30);
+            ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX_32 <= 30);
+            ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
+
+            ZSTD_reduceIndex(cctx, correction);
+            if (ms->nextToUpdate < correction) ms->nextToUpdate = 0;
+            else ms->nextToUpdate -= correction;
+            ms->loadedDictEnd = 0;
+        }
+        ZSTD_window_enforceMaxDist(&ms->window, ip + blockSize, maxDist, &ms->loadedDictEnd);
+        if (ms->nextToUpdate < ms->window.lowLimit) ms->nextToUpdate = ms->window.lowLimit;
+
+        {   size_t cSize = ZSTD_compressBlock_internal(cctx,
+                                op+ZSTD_blockHeaderSize, dstCapacity-ZSTD_blockHeaderSize,
+                                ip, blockSize);
+            if (ZSTD_isError(cSize)) return cSize;
+
+            if (cSize == 0) {  /* block is not compressible */
+                U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(blockSize << 3);
+                if (blockSize + ZSTD_blockHeaderSize > dstCapacity) return ERROR(dstSize_tooSmall);
+                MEM_writeLE32(op, cBlockHeader24);   /* 4th byte will be overwritten */
+                memcpy(op + ZSTD_blockHeaderSize, ip, blockSize);
+                cSize = ZSTD_blockHeaderSize + blockSize;
+            } else {
+                U32 const cBlockHeader24 = lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
+                MEM_writeLE24(op, cBlockHeader24);
+                cSize += ZSTD_blockHeaderSize;
+            }
+
+            ip += blockSize;
+            assert(remaining >= blockSize);
+            remaining -= blockSize;
+            op += cSize;
+            assert(dstCapacity >= cSize);
+            dstCapacity -= cSize;
+            DEBUGLOG(5, "ZSTD_compress_frameChunk: adding a block of size %u",
+                        (U32)cSize);
+    }   }
+
+    if (lastFrameChunk && (op>ostart)) cctx->stage = ZSTDcs_ending;
+    return op-ostart;
+}
+
+
+static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
+                                    ZSTD_CCtx_params params, U64 pledgedSrcSize, U32 dictID)
+{   BYTE* const op = (BYTE*)dst;
+    U32   const dictIDSizeCodeLength = (dictID>0) + (dictID>=256) + (dictID>=65536);   /* 0-3 */
+    U32   const dictIDSizeCode = params.fParams.noDictIDFlag ? 0 : dictIDSizeCodeLength;   /* 0-3 */
+    U32   const checksumFlag = params.fParams.checksumFlag>0;
+    U32   const windowSize = (U32)1 << params.cParams.windowLog;
+    U32   const singleSegment = params.fParams.contentSizeFlag && (windowSize >= pledgedSrcSize);
+    BYTE  const windowLogByte = (BYTE)((params.cParams.windowLog - ZSTD_WINDOWLOG_ABSOLUTEMIN) << 3);
+    U32   const fcsCode = params.fParams.contentSizeFlag ?
+                     (pledgedSrcSize>=256) + (pledgedSrcSize>=65536+256) + (pledgedSrcSize>=0xFFFFFFFFU) : 0;  /* 0-3 */
+    BYTE  const frameHeaderDecriptionByte = (BYTE)(dictIDSizeCode + (checksumFlag<<2) + (singleSegment<<5) + (fcsCode<<6) );
+    size_t pos=0;
+
+    if (dstCapacity < ZSTD_frameHeaderSize_max) return ERROR(dstSize_tooSmall);
+    DEBUGLOG(4, "ZSTD_writeFrameHeader : dictIDFlag : %u ; dictID : %u ; dictIDSizeCode : %u",
+                !params.fParams.noDictIDFlag, dictID,  dictIDSizeCode);
+
+    if (params.format == ZSTD_f_zstd1) {
+        MEM_writeLE32(dst, ZSTD_MAGICNUMBER);
+        pos = 4;
+    }
+    op[pos++] = frameHeaderDecriptionByte;
+    if (!singleSegment) op[pos++] = windowLogByte;
+    switch(dictIDSizeCode)
+    {
+        default:  assert(0); /* impossible */
+        case 0 : break;
+        case 1 : op[pos] = (BYTE)(dictID); pos++; break;
+        case 2 : MEM_writeLE16(op+pos, (U16)dictID); pos+=2; break;
+        case 3 : MEM_writeLE32(op+pos, dictID); pos+=4; break;
+    }
+    switch(fcsCode)
+    {
+        default:  assert(0); /* impossible */
+        case 0 : if (singleSegment) op[pos++] = (BYTE)(pledgedSrcSize); break;
+        case 1 : MEM_writeLE16(op+pos, (U16)(pledgedSrcSize-256)); pos+=2; break;
+        case 2 : MEM_writeLE32(op+pos, (U32)(pledgedSrcSize)); pos+=4; break;
+        case 3 : MEM_writeLE64(op+pos, (U64)(pledgedSrcSize)); pos+=8; break;
+    }
+    return pos;
+}
+
+/* ZSTD_writeLastEmptyBlock() :
+ * output an empty Block with end-of-frame mark to complete a frame
+ * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
+ *           or an error code if `dstCapcity` is too small (<ZSTD_blockHeaderSize)
+ */
+size_t ZSTD_writeLastEmptyBlock(void* dst, size_t dstCapacity)
+{
+    if (dstCapacity < ZSTD_blockHeaderSize) return ERROR(dstSize_tooSmall);
+    {   U32 const cBlockHeader24 = 1 /*lastBlock*/ + (((U32)bt_raw)<<1);  /* 0 size */
+        MEM_writeLE24(dst, cBlockHeader24);
+        return ZSTD_blockHeaderSize;
+    }
+}
+
+size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq)
+{
+    if (cctx->stage != ZSTDcs_init)
+        return ERROR(stage_wrong);
+    if (cctx->appliedParams.ldmParams.enableLdm)
+        return ERROR(parameter_unsupported);
+    cctx->externSeqStore.seq = seq;
+    cctx->externSeqStore.size = nbSeq;
+    cctx->externSeqStore.capacity = nbSeq;
+    cctx->externSeqStore.pos = 0;
+    return 0;
+}
+
+
+static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx,
+                              void* dst, size_t dstCapacity,
+                        const void* src, size_t srcSize,
+                               U32 frame, U32 lastFrameChunk)
+{
+    ZSTD_matchState_t* ms = &cctx->blockState.matchState;
+    size_t fhSize = 0;
+
+    DEBUGLOG(5, "ZSTD_compressContinue_internal, stage: %u, srcSize: %u",
+                cctx->stage, (U32)srcSize);
+    if (cctx->stage==ZSTDcs_created) return ERROR(stage_wrong);   /* missing init (ZSTD_compressBegin) */
+
+    if (frame && (cctx->stage==ZSTDcs_init)) {
+        fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, cctx->appliedParams,
+                                       cctx->pledgedSrcSizePlusOne-1, cctx->dictID);
+        if (ZSTD_isError(fhSize)) return fhSize;
+        dstCapacity -= fhSize;
+        dst = (char*)dst + fhSize;
+        cctx->stage = ZSTDcs_ongoing;
+    }
+
+    if (!srcSize) return fhSize;  /* do not generate an empty block if no input */
+
+    if (!ZSTD_window_update(&ms->window, src, srcSize)) {
+        ms->nextToUpdate = ms->window.dictLimit;
+    }
+    if (cctx->appliedParams.ldmParams.enableLdm)
+        ZSTD_window_update(&cctx->ldmState.window, src, srcSize);
+
+    DEBUGLOG(5, "ZSTD_compressContinue_internal (blockSize=%u)", (U32)cctx->blockSize);
+    {   size_t const cSize = frame ?
+                             ZSTD_compress_frameChunk (cctx, dst, dstCapacity, src, srcSize, lastFrameChunk) :
+                             ZSTD_compressBlock_internal (cctx, dst, dstCapacity, src, srcSize);
+        if (ZSTD_isError(cSize)) return cSize;
+        cctx->consumedSrcSize += srcSize;
+        cctx->producedCSize += (cSize + fhSize);
+        if (cctx->appliedParams.fParams.contentSizeFlag) {  /* control src size */
+            if (cctx->consumedSrcSize+1 > cctx->pledgedSrcSizePlusOne) {
+                DEBUGLOG(4, "error : pledgedSrcSize = %u, while realSrcSize >= %u",
+                    (U32)cctx->pledgedSrcSizePlusOne-1, (U32)cctx->consumedSrcSize);
+                return ERROR(srcSize_wrong);
+            }
+        }
+        return cSize + fhSize;
+    }
+}
+
+size_t ZSTD_compressContinue (ZSTD_CCtx* cctx,
+                              void* dst, size_t dstCapacity,
+                        const void* src, size_t srcSize)
+{
+    DEBUGLOG(5, "ZSTD_compressContinue (srcSize=%u)", (U32)srcSize);
+    return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 1 /* frame mode */, 0 /* last chunk */);
+}
+
+
+size_t ZSTD_getBlockSize(const ZSTD_CCtx* cctx)
+{
+    ZSTD_compressionParameters const cParams = cctx->appliedParams.cParams;
+    assert(!ZSTD_checkCParams(cParams));
+    return MIN (ZSTD_BLOCKSIZE_MAX, (U32)1 << cParams.windowLog);
+}
+
+size_t ZSTD_compressBlock(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    size_t const blockSizeMax = ZSTD_getBlockSize(cctx);
+    if (srcSize > blockSizeMax) return ERROR(srcSize_wrong);
+    return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 0 /* frame mode */, 0 /* last chunk */);
+}
+
+/*! ZSTD_loadDictionaryContent() :
+ *  @return : 0, or an error code
+ */
+static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, ZSTD_CCtx_params const* params, const void* src, size_t srcSize)
+{
+    const BYTE* const ip = (const BYTE*) src;
+    const BYTE* const iend = ip + srcSize;
+    ZSTD_compressionParameters const* cParams = &params->cParams;
+
+    ZSTD_window_update(&ms->window, src, srcSize);
+    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
+
+    if (srcSize <= HASH_READ_SIZE) return 0;
+
+    switch(params->cParams.strategy)
+    {
+    case ZSTD_fast:
+        ZSTD_fillHashTable(ms, cParams, iend);
+        break;
+    case ZSTD_dfast:
+        ZSTD_fillDoubleHashTable(ms, cParams, iend);
+        break;
+
+    case ZSTD_greedy:
+    case ZSTD_lazy:
+    case ZSTD_lazy2:
+        if (srcSize >= HASH_READ_SIZE)
+            ZSTD_insertAndFindFirstIndex(ms, cParams, iend-HASH_READ_SIZE);
+        break;
+
+    case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
+    case ZSTD_btopt:
+    case ZSTD_btultra:
+        if (srcSize >= HASH_READ_SIZE)
+            ZSTD_updateTree(ms, cParams, iend-HASH_READ_SIZE, iend);
+        break;
+
+    default:
+        assert(0);  /* not possible : not a valid strategy id */
+    }
+
+    ms->nextToUpdate = (U32)(iend - ms->window.base);
+    return 0;
+}
+
+
+/* Dictionaries that assign zero probability to symbols that show up causes problems
+   when FSE encoding.  Refuse dictionaries that assign zero probability to symbols
+   that we may encounter during compression.
+   NOTE: This behavior is not standard and could be improved in the future. */
+static size_t ZSTD_checkDictNCount(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue) {
+    U32 s;
+    if (dictMaxSymbolValue < maxSymbolValue) return ERROR(dictionary_corrupted);
+    for (s = 0; s <= maxSymbolValue; ++s) {
+        if (normalizedCounter[s] == 0) return ERROR(dictionary_corrupted);
+    }
+    return 0;
+}
+
+
+/* Dictionary format :
+ * See :
+ * https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#dictionary-format
+ */
+/*! ZSTD_loadZstdDictionary() :
+ * @return : dictID, or an error code
+ *  assumptions : magic number supposed already checked
+ *                dictSize supposed > 8
+ */
+static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs, ZSTD_matchState_t* ms, ZSTD_CCtx_params const* params, const void* dict, size_t dictSize, void* workspace)
+{
+    const BYTE* dictPtr = (const BYTE*)dict;
+    const BYTE* const dictEnd = dictPtr + dictSize;
+    short offcodeNCount[MaxOff+1];
+    unsigned offcodeMaxValue = MaxOff;
+    size_t dictID;
+
+    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
+
+    dictPtr += 4;   /* skip magic number */
+    dictID = params->fParams.noDictIDFlag ? 0 :  MEM_readLE32(dictPtr);
+    dictPtr += 4;
+
+    {   unsigned maxSymbolValue = 255;
+        size_t const hufHeaderSize = HUF_readCTable((HUF_CElt*)bs->entropy.hufCTable, &maxSymbolValue, dictPtr, dictEnd-dictPtr);
+        if (HUF_isError(hufHeaderSize)) return ERROR(dictionary_corrupted);
+        if (maxSymbolValue < 255) return ERROR(dictionary_corrupted);
+        dictPtr += hufHeaderSize;
+    }
+
+    {   unsigned offcodeLog;
+        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(offcodeHeaderSize)) return ERROR(dictionary_corrupted);
+        if (offcodeLog > OffFSELog) return ERROR(dictionary_corrupted);
+        /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */
+        CHECK_E( FSE_buildCTable_wksp(bs->entropy.offcodeCTable, offcodeNCount, offcodeMaxValue, offcodeLog, workspace, HUF_WORKSPACE_SIZE),
+                 dictionary_corrupted);
+        dictPtr += offcodeHeaderSize;
+    }
+
+    {   short matchlengthNCount[MaxML+1];
+        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
+        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(matchlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (matchlengthLog > MLFSELog) return ERROR(dictionary_corrupted);
+        /* Every match length code must have non-zero probability */
+        CHECK_F( ZSTD_checkDictNCount(matchlengthNCount, matchlengthMaxValue, MaxML));
+        CHECK_E( FSE_buildCTable_wksp(bs->entropy.matchlengthCTable, matchlengthNCount, matchlengthMaxValue, matchlengthLog, workspace, HUF_WORKSPACE_SIZE),
+                 dictionary_corrupted);
+        dictPtr += matchlengthHeaderSize;
+    }
+
+    {   short litlengthNCount[MaxLL+1];
+        unsigned litlengthMaxValue = MaxLL, litlengthLog;
+        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(litlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (litlengthLog > LLFSELog) return ERROR(dictionary_corrupted);
+        /* Every literal length code must have non-zero probability */
+        CHECK_F( ZSTD_checkDictNCount(litlengthNCount, litlengthMaxValue, MaxLL));
+        CHECK_E( FSE_buildCTable_wksp(bs->entropy.litlengthCTable, litlengthNCount, litlengthMaxValue, litlengthLog, workspace, HUF_WORKSPACE_SIZE),
+                 dictionary_corrupted);
+        dictPtr += litlengthHeaderSize;
+    }
+
+    if (dictPtr+12 > dictEnd) return ERROR(dictionary_corrupted);
+    bs->rep[0] = MEM_readLE32(dictPtr+0);
+    bs->rep[1] = MEM_readLE32(dictPtr+4);
+    bs->rep[2] = MEM_readLE32(dictPtr+8);
+    dictPtr += 12;
+
+    {   size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
+        U32 offcodeMax = MaxOff;
+        if (dictContentSize <= ((U32)-1) - 128 KB) {
+            U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */
+            offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */
+        }
+        /* All offset values <= dictContentSize + 128 KB must be representable */
+        CHECK_F (ZSTD_checkDictNCount(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff)));
+        /* All repCodes must be <= dictContentSize and != 0*/
+        {   U32 u;
+            for (u=0; u<3; u++) {
+                if (bs->rep[u] == 0) return ERROR(dictionary_corrupted);
+                if (bs->rep[u] > dictContentSize) return ERROR(dictionary_corrupted);
+        }   }
+
+        bs->entropy.hufCTable_repeatMode = HUF_repeat_valid;
+        bs->entropy.offcode_repeatMode = FSE_repeat_valid;
+        bs->entropy.matchlength_repeatMode = FSE_repeat_valid;
+        bs->entropy.litlength_repeatMode = FSE_repeat_valid;
+        CHECK_F(ZSTD_loadDictionaryContent(ms, params, dictPtr, dictContentSize));
+        return dictID;
+    }
+}
+
+/** ZSTD_compress_insertDictionary() :
+*   @return : dictID, or an error code */
+static size_t ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs, ZSTD_matchState_t* ms,
+                                             ZSTD_CCtx_params const* params,
+                                       const void* dict, size_t dictSize,
+                                             ZSTD_dictContentType_e dictContentType,
+                                             void* workspace)
+{
+    DEBUGLOG(4, "ZSTD_compress_insertDictionary (dictSize=%u)", (U32)dictSize);
+    if ((dict==NULL) || (dictSize<=8)) return 0;
+
+    ZSTD_reset_compressedBlockState(bs);
+
+    /* dict restricted modes */
+    if (dictContentType == ZSTD_dct_rawContent)
+        return ZSTD_loadDictionaryContent(ms, params, dict, dictSize);
+
+    if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) {
+        if (dictContentType == ZSTD_dct_auto) {
+            DEBUGLOG(4, "raw content dictionary detected");
+            return ZSTD_loadDictionaryContent(ms, params, dict, dictSize);
+        }
+        if (dictContentType == ZSTD_dct_fullDict)
+            return ERROR(dictionary_wrong);
+        assert(0);   /* impossible */
+    }
+
+    /* dict as full zstd dictionary */
+    return ZSTD_loadZstdDictionary(bs, ms, params, dict, dictSize, workspace);
+}
+
+/*! ZSTD_compressBegin_internal() :
+ * @return : 0, or an error code */
+size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
+                             const void* dict, size_t dictSize,
+                             ZSTD_dictContentType_e dictContentType,
+                             const ZSTD_CDict* cdict,
+                             ZSTD_CCtx_params params, U64 pledgedSrcSize,
+                             ZSTD_buffered_policy_e zbuff)
+{
+    DEBUGLOG(4, "ZSTD_compressBegin_internal: wlog=%u", params.cParams.windowLog);
+    /* params are supposed to be fully validated at this point */
+    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
+
+    if (cdict && cdict->dictContentSize>0) {
+        cctx->requestedParams = params;
+        return ZSTD_resetCCtx_usingCDict(cctx, cdict, params.cParams.windowLog,
+                                         params.fParams, pledgedSrcSize, zbuff);
+    }
+
+    CHECK_F( ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+                                     ZSTDcrp_continue, zbuff) );
+    {
+        size_t const dictID = ZSTD_compress_insertDictionary(
+                cctx->blockState.prevCBlock, &cctx->blockState.matchState,
+                &params, dict, dictSize, dictContentType, cctx->entropyWorkspace);
+        if (ZSTD_isError(dictID)) return dictID;
+        assert(dictID <= (size_t)(U32)-1);
+        cctx->dictID = (U32)dictID;
+    }
+    return 0;
+}
+
+size_t ZSTD_compressBegin_advanced_internal(ZSTD_CCtx* cctx,
+                                    const void* dict, size_t dictSize,
+                                    ZSTD_dictContentType_e dictContentType,
+                                    const ZSTD_CDict* cdict,
+                                    ZSTD_CCtx_params params,
+                                    unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_compressBegin_advanced_internal: wlog=%u", params.cParams.windowLog);
+    /* compression parameters verification and optimization */
+    CHECK_F( ZSTD_checkCParams(params.cParams) );
+    return ZSTD_compressBegin_internal(cctx,
+                                       dict, dictSize, dictContentType,
+                                       cdict,
+                                       params, pledgedSrcSize,
+                                       ZSTDb_not_buffered);
+}
+
+/*! ZSTD_compressBegin_advanced() :
+*   @return : 0, or an error code */
+size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx,
+                             const void* dict, size_t dictSize,
+                                   ZSTD_parameters params, unsigned long long pledgedSrcSize)
+{
+    ZSTD_CCtx_params const cctxParams =
+            ZSTD_assignParamsToCCtxParams(cctx->requestedParams, params);
+    return ZSTD_compressBegin_advanced_internal(cctx,
+                                            dict, dictSize, ZSTD_dct_auto,
+                                            NULL /*cdict*/,
+                                            cctxParams, pledgedSrcSize);
+}
+
+size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
+{
+    ZSTD_parameters const params = ZSTD_getParams(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize);
+    ZSTD_CCtx_params const cctxParams =
+            ZSTD_assignParamsToCCtxParams(cctx->requestedParams, params);
+    DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (U32)dictSize);
+    return ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, NULL,
+                                       cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered);
+}
+
+size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel)
+{
+    return ZSTD_compressBegin_usingDict(cctx, NULL, 0, compressionLevel);
+}
+
+
+/*! ZSTD_writeEpilogue() :
+*   Ends a frame.
+*   @return : nb of bytes written into dst (or an error code) */
+static size_t ZSTD_writeEpilogue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity)
+{
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* op = ostart;
+    size_t fhSize = 0;
+
+    DEBUGLOG(4, "ZSTD_writeEpilogue");
+    if (cctx->stage == ZSTDcs_created) return ERROR(stage_wrong);  /* init missing */
+
+    /* special case : empty frame */
+    if (cctx->stage == ZSTDcs_init) {
+        fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, cctx->appliedParams, 0, 0);
+        if (ZSTD_isError(fhSize)) return fhSize;
+        dstCapacity -= fhSize;
+        op += fhSize;
+        cctx->stage = ZSTDcs_ongoing;
+    }
+
+    if (cctx->stage != ZSTDcs_ending) {
+        /* write one last empty block, make it the "last" block */
+        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1) + 0;
+        if (dstCapacity<4) return ERROR(dstSize_tooSmall);
+        MEM_writeLE32(op, cBlockHeader24);
+        op += ZSTD_blockHeaderSize;
+        dstCapacity -= ZSTD_blockHeaderSize;
+    }
+
+    if (cctx->appliedParams.fParams.checksumFlag) {
+        U32 const checksum = (U32) XXH64_digest(&cctx->xxhState);
+        if (dstCapacity<4) return ERROR(dstSize_tooSmall);
+        DEBUGLOG(4, "ZSTD_writeEpilogue: write checksum : %08X", checksum);
+        MEM_writeLE32(op, checksum);
+        op += 4;
+    }
+
+    cctx->stage = ZSTDcs_created;  /* return to "created but no init" status */
+    return op-ostart;
+}
+
+size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
+                         void* dst, size_t dstCapacity,
+                   const void* src, size_t srcSize)
+{
+    size_t endResult;
+    size_t const cSize = ZSTD_compressContinue_internal(cctx,
+                                dst, dstCapacity, src, srcSize,
+                                1 /* frame mode */, 1 /* last chunk */);
+    if (ZSTD_isError(cSize)) return cSize;
+    endResult = ZSTD_writeEpilogue(cctx, (char*)dst + cSize, dstCapacity-cSize);
+    if (ZSTD_isError(endResult)) return endResult;
+    if (cctx->appliedParams.fParams.contentSizeFlag) {  /* control src size */
+        DEBUGLOG(4, "end of frame : controlling src size");
+        if (cctx->pledgedSrcSizePlusOne != cctx->consumedSrcSize+1) {
+            DEBUGLOG(4, "error : pledgedSrcSize = %u, while realSrcSize = %u",
+                (U32)cctx->pledgedSrcSizePlusOne-1, (U32)cctx->consumedSrcSize);
+            return ERROR(srcSize_wrong);
+    }   }
+    return cSize + endResult;
+}
+
+
+static size_t ZSTD_compress_internal (ZSTD_CCtx* cctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize,
+                         const void* dict,size_t dictSize,
+                               ZSTD_parameters params)
+{
+    ZSTD_CCtx_params const cctxParams =
+            ZSTD_assignParamsToCCtxParams(cctx->requestedParams, params);
+    DEBUGLOG(4, "ZSTD_compress_internal");
+    return ZSTD_compress_advanced_internal(cctx,
+                                          dst, dstCapacity,
+                                          src, srcSize,
+                                          dict, dictSize,
+                                          cctxParams);
+}
+
+size_t ZSTD_compress_advanced (ZSTD_CCtx* ctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize,
+                         const void* dict,size_t dictSize,
+                               ZSTD_parameters params)
+{
+    DEBUGLOG(4, "ZSTD_compress_advanced");
+    CHECK_F(ZSTD_checkCParams(params.cParams));
+    return ZSTD_compress_internal(ctx, dst, dstCapacity, src, srcSize, dict, dictSize, params);
+}
+
+/* Internal */
+size_t ZSTD_compress_advanced_internal(
+        ZSTD_CCtx* cctx,
+        void* dst, size_t dstCapacity,
+        const void* src, size_t srcSize,
+        const void* dict,size_t dictSize,
+        ZSTD_CCtx_params params)
+{
+    DEBUGLOG(4, "ZSTD_compress_advanced_internal (srcSize:%u)",
+                (U32)srcSize);
+    CHECK_F( ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, NULL,
+                                         params, srcSize, ZSTDb_not_buffered) );
+    return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
+}
+
+size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize,
+                               const void* dict, size_t dictSize, int compressionLevel)
+{
+    ZSTD_parameters const params = ZSTD_getParams(compressionLevel, srcSize ? srcSize : 1, dict ? dictSize : 0);
+    ZSTD_CCtx_params cctxParams = ZSTD_assignParamsToCCtxParams(cctx->requestedParams, params);
+    assert(params.fParams.contentSizeFlag == 1);
+    ZSTD_CCtxParam_setParameter(&cctxParams, ZSTD_p_compressLiterals, compressionLevel>=0);
+    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, cctxParams);
+}
+
+size_t ZSTD_compressCCtx (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize, int compressionLevel)
+{
+    DEBUGLOG(4, "ZSTD_compressCCtx (srcSize=%u)", (U32)srcSize);
+    return ZSTD_compress_usingDict(cctx, dst, dstCapacity, src, srcSize, NULL, 0, compressionLevel);
+}
+
+size_t ZSTD_compress(void* dst, size_t dstCapacity, const void* src, size_t srcSize, int compressionLevel)
+{
+    size_t result;
+    ZSTD_CCtx ctxBody;
+    memset(&ctxBody, 0, sizeof(ctxBody));
+    ctxBody.customMem = ZSTD_defaultCMem;
+    result = ZSTD_compressCCtx(&ctxBody, dst, dstCapacity, src, srcSize, compressionLevel);
+    ZSTD_free(ctxBody.workSpace, ZSTD_defaultCMem);  /* can't free ctxBody itself, as it's on stack; free only heap content */
+    return result;
+}
+
+
+/* =====  Dictionary API  ===== */
+
+/*! ZSTD_estimateCDictSize_advanced() :
+ *  Estimate amount of memory that will be needed to create a dictionary with following arguments */
+size_t ZSTD_estimateCDictSize_advanced(
+        size_t dictSize, ZSTD_compressionParameters cParams,
+        ZSTD_dictLoadMethod_e dictLoadMethod)
+{
+    DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (U32)sizeof(ZSTD_CDict));
+    return sizeof(ZSTD_CDict) + HUF_WORKSPACE_SIZE + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0)
+           + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize);
+}
+
+size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel)
+{
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams(compressionLevel, 0, dictSize);
+    return ZSTD_estimateCDictSize_advanced(dictSize, cParams, ZSTD_dlm_byCopy);
+}
+
+size_t ZSTD_sizeof_CDict(const ZSTD_CDict* cdict)
+{
+    if (cdict==NULL) return 0;   /* support sizeof on NULL */
+    DEBUGLOG(5, "sizeof(*cdict) : %u", (U32)sizeof(*cdict));
+    return cdict->workspaceSize + (cdict->dictBuffer ? cdict->dictContentSize : 0) + sizeof(*cdict);
+}
+
+static size_t ZSTD_initCDict_internal(
+                    ZSTD_CDict* cdict,
+              const void* dictBuffer, size_t dictSize,
+                    ZSTD_dictLoadMethod_e dictLoadMethod,
+                    ZSTD_dictContentType_e dictContentType,
+                    ZSTD_compressionParameters cParams)
+{
+    DEBUGLOG(3, "ZSTD_initCDict_internal, dictContentType %u", (U32)dictContentType);
+    assert(!ZSTD_checkCParams(cParams));
+    cdict->cParams = cParams;
+    if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dictBuffer) || (!dictSize)) {
+        cdict->dictBuffer = NULL;
+        cdict->dictContent = dictBuffer;
+    } else {
+        void* const internalBuffer = ZSTD_malloc(dictSize, cdict->customMem);
+        cdict->dictBuffer = internalBuffer;
+        cdict->dictContent = internalBuffer;
+        if (!internalBuffer) return ERROR(memory_allocation);
+        memcpy(internalBuffer, dictBuffer, dictSize);
+    }
+    cdict->dictContentSize = dictSize;
+
+    /* Reset the state to no dictionary */
+    ZSTD_reset_compressedBlockState(&cdict->cBlockState);
+    {   void* const end = ZSTD_reset_matchState(
+                &cdict->matchState,
+                (U32*)cdict->workspace + HUF_WORKSPACE_SIZE_U32,
+                &cParams, ZSTDcrp_continue, /* forCCtx */ 0);
+        assert(end == (char*)cdict->workspace + cdict->workspaceSize);
+        (void)end;
+    }
+    /* (Maybe) load the dictionary
+     * Skips loading the dictionary if it is <= 8 bytes.
+     */
+    {   ZSTD_CCtx_params params;
+        memset(&params, 0, sizeof(params));
+        params.compressionLevel = ZSTD_CLEVEL_DEFAULT;
+        params.fParams.contentSizeFlag = 1;
+        params.cParams = cParams;
+        {   size_t const dictID = ZSTD_compress_insertDictionary(
+                    &cdict->cBlockState, &cdict->matchState, &params,
+                    cdict->dictContent, cdict->dictContentSize,
+                    dictContentType, cdict->workspace);
+            if (ZSTD_isError(dictID)) return dictID;
+            assert(dictID <= (size_t)(U32)-1);
+            cdict->dictID = (U32)dictID;
+        }
+    }
+
+    return 0;
+}
+
+ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize,
+                                      ZSTD_dictLoadMethod_e dictLoadMethod,
+                                      ZSTD_dictContentType_e dictContentType,
+                                      ZSTD_compressionParameters cParams, ZSTD_customMem customMem)
+{
+    DEBUGLOG(3, "ZSTD_createCDict_advanced, mode %u", (U32)dictContentType);
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+
+    {   ZSTD_CDict* const cdict = (ZSTD_CDict*)ZSTD_malloc(sizeof(ZSTD_CDict), customMem);
+        size_t const workspaceSize = HUF_WORKSPACE_SIZE + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0);
+        void* const workspace = ZSTD_malloc(workspaceSize, customMem);
+
+        if (!cdict || !workspace) {
+            ZSTD_free(cdict, customMem);
+            ZSTD_free(workspace, customMem);
+            return NULL;
+        }
+        cdict->customMem = customMem;
+        cdict->workspace = workspace;
+        cdict->workspaceSize = workspaceSize;
+        if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
+                                        dictBuffer, dictSize,
+                                        dictLoadMethod, dictContentType,
+                                        cParams) )) {
+            ZSTD_freeCDict(cdict);
+            return NULL;
+        }
+
+        return cdict;
+    }
+}
+
+ZSTD_CDict* ZSTD_createCDict(const void* dict, size_t dictSize, int compressionLevel)
+{
+    ZSTD_compressionParameters cParams = ZSTD_getCParams(compressionLevel, 0, dictSize);
+    return ZSTD_createCDict_advanced(dict, dictSize,
+                                     ZSTD_dlm_byCopy, ZSTD_dct_auto,
+                                     cParams, ZSTD_defaultCMem);
+}
+
+ZSTD_CDict* ZSTD_createCDict_byReference(const void* dict, size_t dictSize, int compressionLevel)
+{
+    ZSTD_compressionParameters cParams = ZSTD_getCParams(compressionLevel, 0, dictSize);
+    return ZSTD_createCDict_advanced(dict, dictSize,
+                                     ZSTD_dlm_byRef, ZSTD_dct_auto,
+                                     cParams, ZSTD_defaultCMem);
+}
+
+size_t ZSTD_freeCDict(ZSTD_CDict* cdict)
+{
+    if (cdict==NULL) return 0;   /* support free on NULL */
+    {   ZSTD_customMem const cMem = cdict->customMem;
+        ZSTD_free(cdict->workspace, cMem);
+        ZSTD_free(cdict->dictBuffer, cMem);
+        ZSTD_free(cdict, cMem);
+        return 0;
+    }
+}
+
+/*! ZSTD_initStaticCDict_advanced() :
+ *  Generate a digested dictionary in provided memory area.
+ *  workspace: The memory area to emplace the dictionary into.
+ *             Provided pointer must 8-bytes aligned.
+ *             It must outlive dictionary usage.
+ *  workspaceSize: Use ZSTD_estimateCDictSize()
+ *                 to determine how large workspace must be.
+ *  cParams : use ZSTD_getCParams() to transform a compression level
+ *            into its relevants cParams.
+ * @return : pointer to ZSTD_CDict*, or NULL if error (size too small)
+ *  Note : there is no corresponding "free" function.
+ *         Since workspace was allocated externally, it must be freed externally.
+ */
+const ZSTD_CDict* ZSTD_initStaticCDict(
+                                 void* workspace, size_t workspaceSize,
+                           const void* dict, size_t dictSize,
+                                 ZSTD_dictLoadMethod_e dictLoadMethod,
+                                 ZSTD_dictContentType_e dictContentType,
+                                 ZSTD_compressionParameters cParams)
+{
+    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0);
+    size_t const neededSize = sizeof(ZSTD_CDict) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize)
+                            + HUF_WORKSPACE_SIZE + matchStateSize;
+    ZSTD_CDict* const cdict = (ZSTD_CDict*) workspace;
+    void* ptr;
+    if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
+    DEBUGLOG(4, "(workspaceSize < neededSize) : (%u < %u) => %u",
+        (U32)workspaceSize, (U32)neededSize, (U32)(workspaceSize < neededSize));
+    if (workspaceSize < neededSize) return NULL;
+
+    if (dictLoadMethod == ZSTD_dlm_byCopy) {
+        memcpy(cdict+1, dict, dictSize);
+        dict = cdict+1;
+        ptr = (char*)workspace + sizeof(ZSTD_CDict) + dictSize;
+    } else {
+        ptr = cdict+1;
+    }
+    cdict->workspace = ptr;
+    cdict->workspaceSize = HUF_WORKSPACE_SIZE + matchStateSize;
+
+    if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
+                                              dict, dictSize,
+                                              ZSTD_dlm_byRef, dictContentType,
+                                              cParams) ))
+        return NULL;
+
+    return cdict;
+}
+
+ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict)
+{
+    assert(cdict != NULL);
+    return cdict->cParams;
+}
+
+/* ZSTD_compressBegin_usingCDict_advanced() :
+ * cdict must be != NULL */
+size_t ZSTD_compressBegin_usingCDict_advanced(
+    ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
+    ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_advanced");
+    if (cdict==NULL) return ERROR(dictionary_wrong);
+    {   ZSTD_CCtx_params params = cctx->requestedParams;
+        params.cParams = ZSTD_getCParamsFromCDict(cdict);
+        /* Increase window log to fit the entire dictionary and source if the
+         * source size is known. Limit the increase to 19, which is the
+         * window log for compression level 1 with the largest source size.
+         */
+        if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+            U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
+            U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
+            params.cParams.windowLog = MAX(params.cParams.windowLog, limitedSrcLog);
+        }
+        params.fParams = fParams;
+        return ZSTD_compressBegin_internal(cctx,
+                                           NULL, 0, ZSTD_dct_auto,
+                                           cdict,
+                                           params, pledgedSrcSize,
+                                           ZSTDb_not_buffered);
+    }
+}
+
+/* ZSTD_compressBegin_usingCDict() :
+ * pledgedSrcSize=0 means "unknown"
+ * if pledgedSrcSize>0, it will enable contentSizeFlag */
+size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
+{
+    ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
+    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict : dictIDFlag == %u", !fParams.noDictIDFlag);
+    return ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, 0);
+}
+
+size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize,
+                                const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
+{
+    CHECK_F (ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, srcSize));   /* will check if cdict != NULL */
+    return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
+}
+
+/*! ZSTD_compress_usingCDict() :
+ *  Compression using a digested Dictionary.
+ *  Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
+ *  Note that compression parameters are decided at CDict creation time
+ *  while frame parameters are hardcoded */
+size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize,
+                                const ZSTD_CDict* cdict)
+{
+    ZSTD_frameParameters const fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
+    return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
+}
+
+
+
+/* ******************************************************************
+*  Streaming
+********************************************************************/
+
+ZSTD_CStream* ZSTD_createCStream(void)
+{
+    DEBUGLOG(3, "ZSTD_createCStream");
+    return ZSTD_createCStream_advanced(ZSTD_defaultCMem);
+}
+
+ZSTD_CStream* ZSTD_initStaticCStream(void *workspace, size_t workspaceSize)
+{
+    return ZSTD_initStaticCCtx(workspace, workspaceSize);
+}
+
+ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem)
+{   /* CStream and CCtx are now same object */
+    return ZSTD_createCCtx_advanced(customMem);
+}
+
+size_t ZSTD_freeCStream(ZSTD_CStream* zcs)
+{
+    return ZSTD_freeCCtx(zcs);   /* same object */
+}
+
+
+
+/*======   Initialization   ======*/
+
+size_t ZSTD_CStreamInSize(void)  { return ZSTD_BLOCKSIZE_MAX; }
+
+size_t ZSTD_CStreamOutSize(void)
+{
+    return ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + 4 /* 32-bits hash */ ;
+}
+
+static size_t ZSTD_resetCStream_internal(ZSTD_CStream* cctx,
+                    const void* const dict, size_t const dictSize, ZSTD_dictContentType_e const dictContentType,
+                    const ZSTD_CDict* const cdict,
+                    ZSTD_CCtx_params const params, unsigned long long const pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_resetCStream_internal (disableLiteralCompression=%i)",
+                params.disableLiteralCompression);
+    /* params are supposed to be fully validated at this point */
+    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
+
+    CHECK_F( ZSTD_compressBegin_internal(cctx,
+                                         dict, dictSize, dictContentType,
+                                         cdict,
+                                         params, pledgedSrcSize,
+                                         ZSTDb_buffered) );
+
+    cctx->inToCompress = 0;
+    cctx->inBuffPos = 0;
+    cctx->inBuffTarget = cctx->blockSize
+                      + (cctx->blockSize == pledgedSrcSize);   /* for small input: avoid automatic flush on reaching end of block, since it would require to add a 3-bytes null block to end frame */
+    cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0;
+    cctx->streamStage = zcss_load;
+    cctx->frameEnded = 0;
+    return 0;   /* ready to go */
+}
+
+/* ZSTD_resetCStream():
+ * pledgedSrcSize == 0 means "unknown" */
+size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize)
+{
+    ZSTD_CCtx_params params = zcs->requestedParams;
+    DEBUGLOG(4, "ZSTD_resetCStream: pledgedSrcSize = %u", (U32)pledgedSrcSize);
+    if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    params.fParams.contentSizeFlag = 1;
+    params.cParams = ZSTD_getCParamsFromCCtxParams(&params, pledgedSrcSize, 0);
+    return ZSTD_resetCStream_internal(zcs, NULL, 0, ZSTD_dct_auto, zcs->cdict, params, pledgedSrcSize);
+}
+
+/*! ZSTD_initCStream_internal() :
+ *  Note : for lib/compress only. Used by zstdmt_compress.c.
+ *  Assumption 1 : params are valid
+ *  Assumption 2 : either dict, or cdict, is defined, not both */
+size_t ZSTD_initCStream_internal(ZSTD_CStream* zcs,
+                    const void* dict, size_t dictSize, const ZSTD_CDict* cdict,
+                    ZSTD_CCtx_params params, unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_initCStream_internal");
+    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
+
+    if (dict && dictSize >= 8) {
+        DEBUGLOG(4, "loading dictionary of size %u", (U32)dictSize);
+        if (zcs->staticSize) {   /* static CCtx : never uses malloc */
+            /* incompatible with internal cdict creation */
+            return ERROR(memory_allocation);
+        }
+        ZSTD_freeCDict(zcs->cdictLocal);
+        zcs->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
+                                            ZSTD_dlm_byCopy, ZSTD_dct_auto,
+                                            params.cParams, zcs->customMem);
+        zcs->cdict = zcs->cdictLocal;
+        if (zcs->cdictLocal == NULL) return ERROR(memory_allocation);
+    } else {
+        if (cdict) {
+            params.cParams = ZSTD_getCParamsFromCDict(cdict);  /* cParams are enforced from cdict; it includes windowLog */
+        }
+        ZSTD_freeCDict(zcs->cdictLocal);
+        zcs->cdictLocal = NULL;
+        zcs->cdict = cdict;
+    }
+
+    return ZSTD_resetCStream_internal(zcs, NULL, 0, ZSTD_dct_auto, zcs->cdict, params, pledgedSrcSize);
+}
+
+/* ZSTD_initCStream_usingCDict_advanced() :
+ * same as ZSTD_initCStream_usingCDict(), with control over frame parameters */
+size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
+                                            const ZSTD_CDict* cdict,
+                                            ZSTD_frameParameters fParams,
+                                            unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_initCStream_usingCDict_advanced");
+    if (!cdict) return ERROR(dictionary_wrong); /* cannot handle NULL cdict (does not know what to do) */
+    {   ZSTD_CCtx_params params = zcs->requestedParams;
+        params.cParams = ZSTD_getCParamsFromCDict(cdict);
+        params.fParams = fParams;
+        return ZSTD_initCStream_internal(zcs,
+                                NULL, 0, cdict,
+                                params, pledgedSrcSize);
+    }
+}
+
+/* note : cdict must outlive compression session */
+size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict)
+{
+    ZSTD_frameParameters const fParams = { 0 /* contentSizeFlag */, 0 /* checksum */, 0 /* hideDictID */ };
+    DEBUGLOG(4, "ZSTD_initCStream_usingCDict");
+    return ZSTD_initCStream_usingCDict_advanced(zcs, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);  /* note : will check that cdict != NULL */
+}
+
+
+/* ZSTD_initCStream_advanced() :
+ * pledgedSrcSize must be exact.
+ * if srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN.
+ * dict is loaded with default parameters ZSTD_dm_auto and ZSTD_dlm_byCopy. */
+size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
+                                 const void* dict, size_t dictSize,
+                                 ZSTD_parameters params, unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_initCStream_advanced: pledgedSrcSize=%u, flag=%u",
+                (U32)pledgedSrcSize, params.fParams.contentSizeFlag);
+    CHECK_F( ZSTD_checkCParams(params.cParams) );
+    if ((pledgedSrcSize==0) && (params.fParams.contentSizeFlag==0)) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;  /* for compatibility with older programs relying on this behavior. Users should now specify ZSTD_CONTENTSIZE_UNKNOWN. This line will be removed in the future. */
+    {   ZSTD_CCtx_params const cctxParams = ZSTD_assignParamsToCCtxParams(zcs->requestedParams, params);
+        return ZSTD_initCStream_internal(zcs, dict, dictSize, NULL /*cdict*/, cctxParams, pledgedSrcSize);
+    }
+}
+
+size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs, const void* dict, size_t dictSize, int compressionLevel)
+{
+    ZSTD_parameters const params = ZSTD_getParams(compressionLevel, 0, dictSize);
+    ZSTD_CCtx_params const cctxParams =
+            ZSTD_assignParamsToCCtxParams(zcs->requestedParams, params);
+    return ZSTD_initCStream_internal(zcs, dict, dictSize, NULL, cctxParams, ZSTD_CONTENTSIZE_UNKNOWN);
+}
+
+size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs, int compressionLevel, unsigned long long pss)
+{
+    U64 const pledgedSrcSize = (pss==0) ? ZSTD_CONTENTSIZE_UNKNOWN : pss;  /* temporary : 0 interpreted as "unknown" during transition period. Users willing to specify "unknown" **must** use ZSTD_CONTENTSIZE_UNKNOWN. `0` will be interpreted as "empty" in the future */
+    ZSTD_parameters const params = ZSTD_getParams(compressionLevel, pledgedSrcSize, 0);
+    ZSTD_CCtx_params const cctxParams = ZSTD_assignParamsToCCtxParams(zcs->requestedParams, params);
+    return ZSTD_initCStream_internal(zcs, NULL, 0, NULL, cctxParams, pledgedSrcSize);
+}
+
+size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel)
+{
+    DEBUGLOG(4, "ZSTD_initCStream");
+    return ZSTD_initCStream_srcSize(zcs, compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN);
+}
+
+/*======   Compression   ======*/
+
+MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity,
+                           const void* src, size_t srcSize)
+{
+    size_t const length = MIN(dstCapacity, srcSize);
+    if (length) memcpy(dst, src, length);
+    return length;
+}
+
+/** ZSTD_compressStream_generic():
+ *  internal function for all *compressStream*() variants and *compress_generic()
+ *  non-static, because can be called from zstdmt_compress.c
+ * @return : hint size for next input */
+size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
+                                   ZSTD_outBuffer* output,
+                                   ZSTD_inBuffer* input,
+                                   ZSTD_EndDirective const flushMode)
+{
+    const char* const istart = (const char*)input->src;
+    const char* const iend = istart + input->size;
+    const char* ip = istart + input->pos;
+    char* const ostart = (char*)output->dst;
+    char* const oend = ostart + output->size;
+    char* op = ostart + output->pos;
+    U32 someMoreWork = 1;
+
+    /* check expectations */
+    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%u", (U32)flushMode);
+    assert(zcs->inBuff != NULL);
+    assert(zcs->inBuffSize > 0);
+    assert(zcs->outBuff !=  NULL);
+    assert(zcs->outBuffSize > 0);
+    assert(output->pos <= output->size);
+    assert(input->pos <= input->size);
+
+    while (someMoreWork) {
+        switch(zcs->streamStage)
+        {
+        case zcss_init:
+            /* call ZSTD_initCStream() first ! */
+            return ERROR(init_missing);
+
+        case zcss_load:
+            if ( (flushMode == ZSTD_e_end)
+              && ((size_t)(oend-op) >= ZSTD_compressBound(iend-ip))  /* enough dstCapacity */
+              && (zcs->inBuffPos == 0) ) {
+                /* shortcut to compression pass directly into output buffer */
+                size_t const cSize = ZSTD_compressEnd(zcs,
+                                                op, oend-op, ip, iend-ip);
+                DEBUGLOG(4, "ZSTD_compressEnd : %u", (U32)cSize);
+                if (ZSTD_isError(cSize)) return cSize;
+                ip = iend;
+                op += cSize;
+                zcs->frameEnded = 1;
+                ZSTD_startNewCompression(zcs);
+                someMoreWork = 0; break;
+            }
+            /* complete loading into inBuffer */
+            {   size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
+                size_t const loaded = ZSTD_limitCopy(
+                                        zcs->inBuff + zcs->inBuffPos, toLoad,
+                                        ip, iend-ip);
+                zcs->inBuffPos += loaded;
+                ip += loaded;
+                if ( (flushMode == ZSTD_e_continue)
+                  && (zcs->inBuffPos < zcs->inBuffTarget) ) {
+                    /* not enough input to fill full block : stop here */
+                    someMoreWork = 0; break;
+                }
+                if ( (flushMode == ZSTD_e_flush)
+                  && (zcs->inBuffPos == zcs->inToCompress) ) {
+                    /* empty */
+                    someMoreWork = 0; break;
+                }
+            }
+            /* compress current block (note : this stage cannot be stopped in the middle) */
+            DEBUGLOG(5, "stream compression stage (flushMode==%u)", flushMode);
+            {   void* cDst;
+                size_t cSize;
+                size_t const iSize = zcs->inBuffPos - zcs->inToCompress;
+                size_t oSize = oend-op;
+                unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend);
+                if (oSize >= ZSTD_compressBound(iSize))
+                    cDst = op;   /* compress into output buffer, to skip flush stage */
+                else
+                    cDst = zcs->outBuff, oSize = zcs->outBuffSize;
+                cSize = lastBlock ?
+                        ZSTD_compressEnd(zcs, cDst, oSize,
+                                    zcs->inBuff + zcs->inToCompress, iSize) :
+                        ZSTD_compressContinue(zcs, cDst, oSize,
+                                    zcs->inBuff + zcs->inToCompress, iSize);
+                if (ZSTD_isError(cSize)) return cSize;
+                zcs->frameEnded = lastBlock;
+                /* prepare next block */
+                zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize;
+                if (zcs->inBuffTarget > zcs->inBuffSize)
+                    zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize;
+                DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u",
+                         (U32)zcs->inBuffTarget, (U32)zcs->inBuffSize);
+                if (!lastBlock)
+                    assert(zcs->inBuffTarget <= zcs->inBuffSize);
+                zcs->inToCompress = zcs->inBuffPos;
+                if (cDst == op) {  /* no need to flush */
+                    op += cSize;
+                    if (zcs->frameEnded) {
+                        DEBUGLOG(5, "Frame completed directly in outBuffer");
+                        someMoreWork = 0;
+                        ZSTD_startNewCompression(zcs);
+                    }
+                    break;
+                }
+                zcs->outBuffContentSize = cSize;
+                zcs->outBuffFlushedSize = 0;
+                zcs->streamStage = zcss_flush; /* pass-through to flush stage */
+            }
+	    /* fall-through */
+        case zcss_flush:
+            DEBUGLOG(5, "flush stage");
+            {   size_t const toFlush = zcs->outBuffContentSize - zcs->outBuffFlushedSize;
+                size_t const flushed = ZSTD_limitCopy(op, oend-op,
+                            zcs->outBuff + zcs->outBuffFlushedSize, toFlush);
+                DEBUGLOG(5, "toFlush: %u into %u ==> flushed: %u",
+                            (U32)toFlush, (U32)(oend-op), (U32)flushed);
+                op += flushed;
+                zcs->outBuffFlushedSize += flushed;
+                if (toFlush!=flushed) {
+                    /* flush not fully completed, presumably because dst is too small */
+                    assert(op==oend);
+                    someMoreWork = 0;
+                    break;
+                }
+                zcs->outBuffContentSize = zcs->outBuffFlushedSize = 0;
+                if (zcs->frameEnded) {
+                    DEBUGLOG(5, "Frame completed on flush");
+                    someMoreWork = 0;
+                    ZSTD_startNewCompression(zcs);
+                    break;
+                }
+                zcs->streamStage = zcss_load;
+                break;
+            }
+
+        default: /* impossible */
+            assert(0);
+        }
+    }
+
+    input->pos = ip - istart;
+    output->pos = op - ostart;
+    if (zcs->frameEnded) return 0;
+    {   size_t hintInSize = zcs->inBuffTarget - zcs->inBuffPos;
+        if (hintInSize==0) hintInSize = zcs->blockSize;
+        return hintInSize;
+    }
+}
+
+size_t ZSTD_compressStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
+{
+    /* check conditions */
+    if (output->pos > output->size) return ERROR(GENERIC);
+    if (input->pos  > input->size)  return ERROR(GENERIC);
+
+    return ZSTD_compressStream_generic(zcs, output, input, ZSTD_e_continue);
+}
+
+
+size_t ZSTD_compress_generic (ZSTD_CCtx* cctx,
+                              ZSTD_outBuffer* output,
+                              ZSTD_inBuffer* input,
+                              ZSTD_EndDirective endOp)
+{
+    DEBUGLOG(5, "ZSTD_compress_generic, endOp=%u ", (U32)endOp);
+    /* check conditions */
+    if (output->pos > output->size) return ERROR(GENERIC);
+    if (input->pos  > input->size)  return ERROR(GENERIC);
+    assert(cctx!=NULL);
+
+    /* transparent initialization stage */
+    if (cctx->streamStage == zcss_init) {
+        ZSTD_CCtx_params params = cctx->requestedParams;
+        ZSTD_prefixDict const prefixDict = cctx->prefixDict;
+        memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));  /* single usage */
+        assert(prefixDict.dict==NULL || cctx->cdict==NULL);   /* only one can be set */
+        DEBUGLOG(4, "ZSTD_compress_generic : transparent init stage");
+        if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = input->size + 1;  /* auto-fix pledgedSrcSize */
+        params.cParams = ZSTD_getCParamsFromCCtxParams(
+                &cctx->requestedParams, cctx->pledgedSrcSizePlusOne-1, 0 /*dictSize*/);
+
+#ifdef ZSTD_MULTITHREAD
+        if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
+            params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
+        }
+        if (params.nbWorkers > 0) {
+            /* mt context creation */
+            if (cctx->mtctx == NULL || (params.nbWorkers != ZSTDMT_getNbWorkers(cctx->mtctx))) {
+                DEBUGLOG(4, "ZSTD_compress_generic: creating new mtctx for nbWorkers=%u",
+                            params.nbWorkers);
+                if (cctx->mtctx != NULL)
+                    DEBUGLOG(4, "ZSTD_compress_generic: previous nbWorkers was %u",
+                                ZSTDMT_getNbWorkers(cctx->mtctx));
+                ZSTDMT_freeCCtx(cctx->mtctx);
+                cctx->mtctx = ZSTDMT_createCCtx_advanced(params.nbWorkers, cctx->customMem);
+                if (cctx->mtctx == NULL) return ERROR(memory_allocation);
+            }
+            /* mt compression */
+            DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers);
+            CHECK_F( ZSTDMT_initCStream_internal(
+                        cctx->mtctx,
+                        prefixDict.dict, prefixDict.dictSize, ZSTD_dct_rawContent,
+                        cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) );
+            cctx->streamStage = zcss_load;
+            cctx->appliedParams.nbWorkers = params.nbWorkers;
+        } else
+#endif
+        {   CHECK_F( ZSTD_resetCStream_internal(cctx,
+                            prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
+                            cctx->cdict,
+                            params, cctx->pledgedSrcSizePlusOne-1) );
+            assert(cctx->streamStage == zcss_load);
+            assert(cctx->appliedParams.nbWorkers == 0);
+    }   }
+
+    /* compression stage */
+#ifdef ZSTD_MULTITHREAD
+    if (cctx->appliedParams.nbWorkers > 0) {
+        if (cctx->cParamsChanged) {
+            ZSTDMT_updateCParams_whileCompressing(cctx->mtctx, &cctx->requestedParams);
+            cctx->cParamsChanged = 0;
+        }
+        {   size_t const flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
+            if ( ZSTD_isError(flushMin)
+              || (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */
+                ZSTD_startNewCompression(cctx);
+            }
+            return flushMin;
+    }   }
+#endif
+    CHECK_F( ZSTD_compressStream_generic(cctx, output, input, endOp) );
+    DEBUGLOG(5, "completed ZSTD_compress_generic");
+    return cctx->outBuffContentSize - cctx->outBuffFlushedSize; /* remaining to flush */
+}
+
+size_t ZSTD_compress_generic_simpleArgs (
+                            ZSTD_CCtx* cctx,
+                            void* dst, size_t dstCapacity, size_t* dstPos,
+                      const void* src, size_t srcSize, size_t* srcPos,
+                            ZSTD_EndDirective endOp)
+{
+    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
+    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
+    /* ZSTD_compress_generic() will check validity of dstPos and srcPos */
+    size_t const cErr = ZSTD_compress_generic(cctx, &output, &input, endOp);
+    *dstPos = output.pos;
+    *srcPos = input.pos;
+    return cErr;
+}
+
+
+/*======   Finalize   ======*/
+
+/*! ZSTD_flushStream() :
+ * @return : amount of data remaining to flush */
+size_t ZSTD_flushStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
+{
+    ZSTD_inBuffer input = { NULL, 0, 0 };
+    if (output->pos > output->size) return ERROR(GENERIC);
+    CHECK_F( ZSTD_compressStream_generic(zcs, output, &input, ZSTD_e_flush) );
+    return zcs->outBuffContentSize - zcs->outBuffFlushedSize;  /* remaining to flush */
+}
+
+
+size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
+{
+    ZSTD_inBuffer input = { NULL, 0, 0 };
+    if (output->pos > output->size) return ERROR(GENERIC);
+    CHECK_F( ZSTD_compressStream_generic(zcs, output, &input, ZSTD_e_end) );
+    {   size_t const lastBlockSize = zcs->frameEnded ? 0 : ZSTD_BLOCKHEADERSIZE;
+        size_t const checksumSize = zcs->frameEnded ? 0 : zcs->appliedParams.fParams.checksumFlag * 4;
+        size_t const toFlush = zcs->outBuffContentSize - zcs->outBuffFlushedSize + lastBlockSize + checksumSize;
+        DEBUGLOG(4, "ZSTD_endStream : remaining to flush : %u", (U32)toFlush);
+        return toFlush;
+    }
+}
+
+
+/*-=====  Pre-defined compression levels  =====-*/
+
+#define ZSTD_MAX_CLEVEL     22
+int ZSTD_maxCLevel(void) { return ZSTD_MAX_CLEVEL; }
+
+static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
+{   /* "default" - guarantees a monotonically increasing memory budget */
+    /* W,  C,  H,  S,  L, TL, strat */
+    { 19, 12, 13,  1,  6,  1, ZSTD_fast    },  /* base for negative levels */
+    { 19, 13, 14,  1,  7,  1, ZSTD_fast    },  /* level  1 */
+    { 19, 15, 16,  1,  6,  1, ZSTD_fast    },  /* level  2 */
+    { 20, 16, 17,  1,  5,  8, ZSTD_dfast   },  /* level  3 */
+    { 20, 17, 18,  1,  5,  8, ZSTD_dfast   },  /* level  4 */
+    { 20, 17, 18,  2,  5, 16, ZSTD_greedy  },  /* level  5 */
+    { 21, 17, 19,  2,  5, 16, ZSTD_lazy    },  /* level  6 */
+    { 21, 18, 19,  3,  5, 16, ZSTD_lazy    },  /* level  7 */
+    { 21, 18, 20,  3,  5, 16, ZSTD_lazy2   },  /* level  8 */
+    { 21, 19, 20,  3,  5, 16, ZSTD_lazy2   },  /* level  9 */
+    { 21, 19, 21,  4,  5, 16, ZSTD_lazy2   },  /* level 10 */
+    { 22, 20, 22,  4,  5, 16, ZSTD_lazy2   },  /* level 11 */
+    { 22, 20, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 12 */
+    { 22, 21, 22,  4,  5, 32, ZSTD_btlazy2 },  /* level 13 */
+    { 22, 21, 22,  5,  5, 32, ZSTD_btlazy2 },  /* level 14 */
+    { 22, 22, 22,  6,  5, 32, ZSTD_btlazy2 },  /* level 15 */
+    { 22, 21, 22,  4,  5, 48, ZSTD_btopt   },  /* level 16 */
+    { 23, 22, 22,  4,  4, 48, ZSTD_btopt   },  /* level 17 */
+    { 23, 22, 22,  5,  3, 64, ZSTD_btopt   },  /* level 18 */
+    { 23, 23, 22,  7,  3,128, ZSTD_btopt   },  /* level 19 */
+    { 25, 25, 23,  7,  3,128, ZSTD_btultra },  /* level 20 */
+    { 26, 26, 24,  7,  3,256, ZSTD_btultra },  /* level 21 */
+    { 27, 27, 25,  9,  3,512, ZSTD_btultra },  /* level 22 */
+},
+{   /* for srcSize <= 256 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 18, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 18, 13, 14,  1,  6,  1, ZSTD_fast    },  /* level  1 */
+    { 18, 14, 13,  1,  5,  8, ZSTD_dfast   },  /* level  2 */
+    { 18, 16, 15,  1,  5,  8, ZSTD_dfast   },  /* level  3 */
+    { 18, 15, 17,  1,  5,  8, ZSTD_greedy  },  /* level  4.*/
+    { 18, 16, 17,  4,  5,  8, ZSTD_greedy  },  /* level  5.*/
+    { 18, 16, 17,  3,  5,  8, ZSTD_lazy    },  /* level  6.*/
+    { 18, 17, 17,  4,  4,  8, ZSTD_lazy    },  /* level  7 */
+    { 18, 17, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 18, 17, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 18, 17, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 18, 18, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 11.*/
+    { 18, 18, 17,  5,  4,  8, ZSTD_btlazy2 },  /* level 12.*/
+    { 18, 19, 17,  7,  4,  8, ZSTD_btlazy2 },  /* level 13 */
+    { 18, 18, 18,  4,  4, 16, ZSTD_btopt   },  /* level 14.*/
+    { 18, 18, 18,  4,  3, 16, ZSTD_btopt   },  /* level 15.*/
+    { 18, 19, 18,  6,  3, 32, ZSTD_btopt   },  /* level 16.*/
+    { 18, 19, 18,  8,  3, 64, ZSTD_btopt   },  /* level 17.*/
+    { 18, 19, 18,  9,  3,128, ZSTD_btopt   },  /* level 18.*/
+    { 18, 19, 18, 10,  3,256, ZSTD_btopt   },  /* level 19.*/
+    { 18, 19, 18, 11,  3,512, ZSTD_btultra },  /* level 20.*/
+    { 18, 19, 18, 12,  3,512, ZSTD_btultra },  /* level 21.*/
+    { 18, 19, 18, 13,  3,512, ZSTD_btultra },  /* level 22.*/
+},
+{   /* for srcSize <= 128 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 17, 12, 12,  1,  5,  1, ZSTD_fast    },  /* level  0 - not used */
+    { 17, 12, 13,  1,  6,  1, ZSTD_fast    },  /* level  1 */
+    { 17, 13, 16,  1,  5,  1, ZSTD_fast    },  /* level  2 */
+    { 17, 16, 16,  2,  5,  8, ZSTD_dfast   },  /* level  3 */
+    { 17, 13, 15,  3,  4,  8, ZSTD_greedy  },  /* level  4 */
+    { 17, 15, 17,  4,  4,  8, ZSTD_greedy  },  /* level  5 */
+    { 17, 16, 17,  3,  4,  8, ZSTD_lazy    },  /* level  6 */
+    { 17, 15, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  7 */
+    { 17, 17, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 17, 17, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 17, 17, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 17, 17, 17,  7,  4,  8, ZSTD_lazy2   },  /* level 11 */
+    { 17, 17, 17,  8,  4,  8, ZSTD_lazy2   },  /* level 12 */
+    { 17, 18, 17,  6,  4,  8, ZSTD_btlazy2 },  /* level 13.*/
+    { 17, 17, 17,  7,  3,  8, ZSTD_btopt   },  /* level 14.*/
+    { 17, 17, 17,  7,  3, 16, ZSTD_btopt   },  /* level 15.*/
+    { 17, 18, 17,  7,  3, 32, ZSTD_btopt   },  /* level 16.*/
+    { 17, 18, 17,  7,  3, 64, ZSTD_btopt   },  /* level 17.*/
+    { 17, 18, 17,  7,  3,256, ZSTD_btopt   },  /* level 18.*/
+    { 17, 18, 17,  8,  3,256, ZSTD_btopt   },  /* level 19.*/
+    { 17, 18, 17,  9,  3,256, ZSTD_btultra },  /* level 20.*/
+    { 17, 18, 17, 10,  3,256, ZSTD_btultra },  /* level 21.*/
+    { 17, 18, 17, 11,  3,512, ZSTD_btultra },  /* level 22.*/
+},
+{   /* for srcSize <= 16 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 14, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 14, 14, 14,  1,  6,  1, ZSTD_fast    },  /* level  1 */
+    { 14, 14, 14,  1,  4,  1, ZSTD_fast    },  /* level  2 */
+    { 14, 14, 14,  1,  4,  6, ZSTD_dfast   },  /* level  3.*/
+    { 14, 14, 14,  4,  4,  6, ZSTD_greedy  },  /* level  4.*/
+    { 14, 14, 14,  3,  4,  6, ZSTD_lazy    },  /* level  5.*/
+    { 14, 14, 14,  4,  4,  6, ZSTD_lazy2   },  /* level  6 */
+    { 14, 14, 14,  5,  4,  6, ZSTD_lazy2   },  /* level  7 */
+    { 14, 14, 14,  6,  4,  6, ZSTD_lazy2   },  /* level  8.*/
+    { 14, 15, 14,  6,  4,  6, ZSTD_btlazy2 },  /* level  9.*/
+    { 14, 15, 14,  3,  3,  6, ZSTD_btopt   },  /* level 10.*/
+    { 14, 15, 14,  6,  3,  8, ZSTD_btopt   },  /* level 11.*/
+    { 14, 15, 14,  6,  3, 16, ZSTD_btopt   },  /* level 12.*/
+    { 14, 15, 14,  6,  3, 24, ZSTD_btopt   },  /* level 13.*/
+    { 14, 15, 15,  6,  3, 48, ZSTD_btopt   },  /* level 14.*/
+    { 14, 15, 15,  6,  3, 64, ZSTD_btopt   },  /* level 15.*/
+    { 14, 15, 15,  6,  3, 96, ZSTD_btopt   },  /* level 16.*/
+    { 14, 15, 15,  6,  3,128, ZSTD_btopt   },  /* level 17.*/
+    { 14, 15, 15,  6,  3,256, ZSTD_btopt   },  /* level 18.*/
+    { 14, 15, 15,  7,  3,256, ZSTD_btopt   },  /* level 19.*/
+    { 14, 15, 15,  8,  3,256, ZSTD_btultra },  /* level 20.*/
+    { 14, 15, 15,  9,  3,256, ZSTD_btultra },  /* level 21.*/
+    { 14, 15, 15, 10,  3,256, ZSTD_btultra },  /* level 22.*/
+},
+};
+
+/*! ZSTD_getCParams() :
+*  @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize.
+*   Size values are optional, provide 0 if not known or unused */
+ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
+{
+    size_t const addedSize = srcSizeHint ? 0 : 500;
+    U64 const rSize = srcSizeHint+dictSize ? srcSizeHint+dictSize+addedSize : (U64)-1;
+    U32 const tableID = (rSize <= 256 KB) + (rSize <= 128 KB) + (rSize <= 16 KB);   /* intentional underflow for srcSizeHint == 0 */
+    int row = compressionLevel;
+    DEBUGLOG(5, "ZSTD_getCParams (cLevel=%i)", compressionLevel);
+    if (compressionLevel == 0) row = ZSTD_CLEVEL_DEFAULT;   /* 0 == default */
+    if (compressionLevel < 0) row = 0;   /* entry 0 is baseline for fast mode */
+    if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL;
+    {   ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row];
+        if (compressionLevel < 0) cp.targetLength = (unsigned)(-compressionLevel);   /* acceleration factor */
+        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize); }
+
+}
+
+/*! ZSTD_getParams() :
+*   same as ZSTD_getCParams(), but @return a `ZSTD_parameters` object (instead of `ZSTD_compressionParameters`).
+*   All fields of `ZSTD_frameParameters` are set to default (0) */
+ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) {
+    ZSTD_parameters params;
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams(compressionLevel, srcSizeHint, dictSize);
+    DEBUGLOG(5, "ZSTD_getParams (cLevel=%i)", compressionLevel);
+    memset(&params, 0, sizeof(params));
+    params.cParams = cParams;
+    params.fParams.contentSizeFlag = 1;
+    return params;
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_compress_internal.h b/vendor/github.com/DataDog/zstd/zstd_compress_internal.h
new file mode 100644
index 0000000..81f12ca
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_compress_internal.h
@@ -0,0 +1,709 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* This header contains definitions
+ * that shall **only** be used by modules within lib/compress.
+ */
+
+#ifndef ZSTD_COMPRESS_H
+#define ZSTD_COMPRESS_H
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include "zstd_internal.h"
+#ifdef ZSTD_MULTITHREAD
+#  include "zstdmt_compress.h"
+#endif
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-*************************************
+*  Constants
+***************************************/
+#define kSearchStrength      8
+#define HASH_READ_SIZE       8
+#define ZSTD_DUBT_UNSORTED_MARK 1   /* For btlazy2 strategy, index 1 now means "unsorted".
+                                       It could be confused for a real successor at index "1", if sorted as larger than its predecessor.
+                                       It's not a big deal though : candidate will just be sorted again.
+                                       Additionnally, candidate position 1 will be lost.
+                                       But candidate 1 cannot hide a large tree of candidates, so it's a minimal loss.
+                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be misdhandled after table re-use with a different strategy */
+
+
+/*-*************************************
+*  Context memory management
+***************************************/
+typedef enum { ZSTDcs_created=0, ZSTDcs_init, ZSTDcs_ongoing, ZSTDcs_ending } ZSTD_compressionStage_e;
+typedef enum { zcss_init=0, zcss_load, zcss_flush } ZSTD_cStreamStage;
+
+typedef struct ZSTD_prefixDict_s {
+    const void* dict;
+    size_t dictSize;
+    ZSTD_dictContentType_e dictContentType;
+} ZSTD_prefixDict;
+
+typedef struct {
+    U32 hufCTable[HUF_CTABLE_SIZE_U32(255)];
+    FSE_CTable offcodeCTable[FSE_CTABLE_SIZE_U32(OffFSELog, MaxOff)];
+    FSE_CTable matchlengthCTable[FSE_CTABLE_SIZE_U32(MLFSELog, MaxML)];
+    FSE_CTable litlengthCTable[FSE_CTABLE_SIZE_U32(LLFSELog, MaxLL)];
+    HUF_repeat hufCTable_repeatMode;
+    FSE_repeat offcode_repeatMode;
+    FSE_repeat matchlength_repeatMode;
+    FSE_repeat litlength_repeatMode;
+} ZSTD_entropyCTables_t;
+
+typedef struct {
+    U32 off;
+    U32 len;
+} ZSTD_match_t;
+
+typedef struct {
+    int price;
+    U32 off;
+    U32 mlen;
+    U32 litlen;
+    U32 rep[ZSTD_REP_NUM];
+} ZSTD_optimal_t;
+
+typedef struct {
+    /* All tables are allocated inside cctx->workspace by ZSTD_resetCCtx_internal() */
+    U32* litFreq;               /* table of literals statistics, of size 256 */
+    U32* litLengthFreq;         /* table of litLength statistics, of size (MaxLL+1) */
+    U32* matchLengthFreq;       /* table of matchLength statistics, of size (MaxML+1) */
+    U32* offCodeFreq;           /* table of offCode statistics, of size (MaxOff+1) */
+    ZSTD_match_t* matchTable;   /* list of found matches, of size ZSTD_OPT_NUM+1 */
+    ZSTD_optimal_t* priceTable; /* All positions tracked by optimal parser, of size ZSTD_OPT_NUM+1 */
+
+    U32  litSum;                 /* nb of literals */
+    U32  litLengthSum;           /* nb of litLength codes */
+    U32  matchLengthSum;         /* nb of matchLength codes */
+    U32  offCodeSum;             /* nb of offset codes */
+    /* begin updated by ZSTD_setLog2Prices */
+    U32  log2litSum;             /* pow2 to compare log2(litfreq) to */
+    U32  log2litLengthSum;       /* pow2 to compare log2(llfreq) to */
+    U32  log2matchLengthSum;     /* pow2 to compare log2(mlfreq) to */
+    U32  log2offCodeSum;         /* pow2 to compare log2(offreq) to */
+    /* end : updated by ZSTD_setLog2Prices */
+    U32  staticPrices;           /* prices follow a pre-defined cost structure, statistics are irrelevant */
+} optState_t;
+
+typedef struct {
+  ZSTD_entropyCTables_t entropy;
+  U32 rep[ZSTD_REP_NUM];
+} ZSTD_compressedBlockState_t;
+
+typedef struct {
+    BYTE const* nextSrc;    /* next block here to continue on current prefix */
+    BYTE const* base;       /* All regular indexes relative to this position */
+    BYTE const* dictBase;   /* extDict indexes relative to this position */
+    U32 dictLimit;          /* below that point, need extDict */
+    U32 lowLimit;           /* below that point, no more data */
+} ZSTD_window_t;
+
+typedef struct {
+    ZSTD_window_t window;      /* State for window round buffer management */
+    U32 loadedDictEnd;         /* index of end of dictionary */
+    U32 nextToUpdate;          /* index from which to continue table update */
+    U32 nextToUpdate3;         /* index from which to continue table update */
+    U32 hashLog3;              /* dispatch table : larger == faster, more memory */
+    U32* hashTable;
+    U32* hashTable3;
+    U32* chainTable;
+    optState_t opt;         /* optimal parser state */
+} ZSTD_matchState_t;
+
+typedef struct {
+    ZSTD_compressedBlockState_t* prevCBlock;
+    ZSTD_compressedBlockState_t* nextCBlock;
+    ZSTD_matchState_t matchState;
+} ZSTD_blockState_t;
+
+typedef struct {
+    U32 offset;
+    U32 checksum;
+} ldmEntry_t;
+
+typedef struct {
+    ZSTD_window_t window;   /* State for the window round buffer management */
+    ldmEntry_t* hashTable;
+    BYTE* bucketOffsets;    /* Next position in bucket to insert entry */
+    U64 hashPower;          /* Used to compute the rolling hash.
+                             * Depends on ldmParams.minMatchLength */
+} ldmState_t;
+
+typedef struct {
+    U32 enableLdm;          /* 1 if enable long distance matching */
+    U32 hashLog;            /* Log size of hashTable */
+    U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
+    U32 minMatchLength;     /* Minimum match length */
+    U32 hashEveryLog;       /* Log number of entries to skip */
+    U32 windowLog;          /* Window log for the LDM */
+} ldmParams_t;
+
+typedef struct {
+    U32 offset;
+    U32 litLength;
+    U32 matchLength;
+} rawSeq;
+
+typedef struct {
+  rawSeq* seq;     /* The start of the sequences */
+  size_t pos;      /* The position where reading stopped. <= size. */
+  size_t size;     /* The number of sequences. <= capacity. */
+  size_t capacity; /* The capacity of the `seq` pointer */
+} rawSeqStore_t;
+
+struct ZSTD_CCtx_params_s {
+    ZSTD_format_e format;
+    ZSTD_compressionParameters cParams;
+    ZSTD_frameParameters fParams;
+
+    int compressionLevel;
+    int disableLiteralCompression;
+    int forceWindow;           /* force back-references to respect limit of
+                                * 1<<wLog, even for dictionary */
+
+    /* Multithreading: used to pass parameters to mtctx */
+    unsigned nbWorkers;
+    unsigned jobSize;
+    unsigned overlapSizeLog;
+
+    /* Long distance matching parameters */
+    ldmParams_t ldmParams;
+
+    /* Internal use, for createCCtxParams() and freeCCtxParams() only */
+    ZSTD_customMem customMem;
+};  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
+
+struct ZSTD_CCtx_s {
+    ZSTD_compressionStage_e stage;
+    int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
+    int bmi2;                            /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
+    ZSTD_CCtx_params requestedParams;
+    ZSTD_CCtx_params appliedParams;
+    U32   dictID;
+    void* workSpace;
+    size_t workSpaceSize;
+    size_t blockSize;
+    unsigned long long pledgedSrcSizePlusOne;  /* this way, 0 (default) == unknown */
+    unsigned long long consumedSrcSize;
+    unsigned long long producedCSize;
+    XXH64_state_t xxhState;
+    ZSTD_customMem customMem;
+    size_t staticSize;
+
+    seqStore_t seqStore;      /* sequences storage ptrs */
+    ldmState_t ldmState;      /* long distance matching state */
+    rawSeq* ldmSequences;     /* Storage for the ldm output sequences */
+    size_t maxNbLdmSequences;
+    rawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
+    ZSTD_blockState_t blockState;
+    U32* entropyWorkspace;  /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
+
+    /* streaming */
+    char*  inBuff;
+    size_t inBuffSize;
+    size_t inToCompress;
+    size_t inBuffPos;
+    size_t inBuffTarget;
+    char*  outBuff;
+    size_t outBuffSize;
+    size_t outBuffContentSize;
+    size_t outBuffFlushedSize;
+    ZSTD_cStreamStage streamStage;
+    U32    frameEnded;
+
+    /* Dictionary */
+    ZSTD_CDict* cdictLocal;
+    const ZSTD_CDict* cdict;
+    ZSTD_prefixDict prefixDict;   /* single-usage dictionary */
+
+    /* Multi-threading */
+#ifdef ZSTD_MULTITHREAD
+    ZSTDMT_CCtx* mtctx;
+#endif
+};
+
+
+typedef size_t (*ZSTD_blockCompressor) (
+        ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, int extDict);
+
+
+MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
+{
+    static const BYTE LL_Code[64] = {  0,  1,  2,  3,  4,  5,  6,  7,
+                                       8,  9, 10, 11, 12, 13, 14, 15,
+                                      16, 16, 17, 17, 18, 18, 19, 19,
+                                      20, 20, 20, 20, 21, 21, 21, 21,
+                                      22, 22, 22, 22, 22, 22, 22, 22,
+                                      23, 23, 23, 23, 23, 23, 23, 23,
+                                      24, 24, 24, 24, 24, 24, 24, 24,
+                                      24, 24, 24, 24, 24, 24, 24, 24 };
+    static const U32 LL_deltaCode = 19;
+    return (litLength > 63) ? ZSTD_highbit32(litLength) + LL_deltaCode : LL_Code[litLength];
+}
+
+/* ZSTD_MLcode() :
+ * note : mlBase = matchLength - MINMATCH;
+ *        because it's the format it's stored in seqStore->sequences */
+MEM_STATIC U32 ZSTD_MLcode(U32 mlBase)
+{
+    static const BYTE ML_Code[128] = { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+                                      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+                                      32, 32, 33, 33, 34, 34, 35, 35, 36, 36, 36, 36, 37, 37, 37, 37,
+                                      38, 38, 38, 38, 38, 38, 38, 38, 39, 39, 39, 39, 39, 39, 39, 39,
+                                      40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40,
+                                      41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41,
+                                      42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42,
+                                      42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42 };
+    static const U32 ML_deltaCode = 36;
+    return (mlBase > 127) ? ZSTD_highbit32(mlBase) + ML_deltaCode : ML_Code[mlBase];
+}
+
+/*! ZSTD_storeSeq() :
+ *  Store a sequence (literal length, literals, offset code and match length code) into seqStore_t.
+ *  `offsetCode` : distance to match + 3 (values 1-3 are repCodes).
+ *  `mlBase` : matchLength - MINMATCH
+*/
+MEM_STATIC void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const void* literals, U32 offsetCode, size_t mlBase)
+{
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG >= 6)
+    static const BYTE* g_start = NULL;
+    if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
+    {   U32 const pos = (U32)((const BYTE*)literals - g_start);
+        DEBUGLOG(6, "Cpos%7u :%3u literals, match%3u bytes at dist.code%7u",
+               pos, (U32)litLength, (U32)mlBase+MINMATCH, (U32)offsetCode);
+    }
+#endif
+    /* copy Literals */
+    assert(seqStorePtr->lit + litLength <= seqStorePtr->litStart + 128 KB);
+    ZSTD_wildcopy(seqStorePtr->lit, literals, litLength);
+    seqStorePtr->lit += litLength;
+
+    /* literal Length */
+    if (litLength>0xFFFF) {
+        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
+        seqStorePtr->longLengthID = 1;
+        seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    }
+    seqStorePtr->sequences[0].litLength = (U16)litLength;
+
+    /* match offset */
+    seqStorePtr->sequences[0].offset = offsetCode + 1;
+
+    /* match Length */
+    if (mlBase>0xFFFF) {
+        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
+        seqStorePtr->longLengthID = 2;
+        seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    }
+    seqStorePtr->sequences[0].matchLength = (U16)mlBase;
+
+    seqStorePtr->sequences++;
+}
+
+
+/*-*************************************
+*  Match length counter
+***************************************/
+static unsigned ZSTD_NbCommonBytes (size_t val)
+{
+    if (MEM_isLittleEndian()) {
+        if (MEM_64bits()) {
+#       if defined(_MSC_VER) && defined(_WIN64)
+            unsigned long r = 0;
+            _BitScanForward64( &r, (U64)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 4)
+            return (__builtin_ctzll((U64)val) >> 3);
+#       else
+            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2,
+                                                     0, 3, 1, 3, 1, 4, 2, 7,
+                                                     0, 2, 3, 6, 1, 5, 3, 5,
+                                                     1, 3, 4, 4, 2, 5, 6, 7,
+                                                     7, 0, 1, 2, 3, 3, 4, 6,
+                                                     2, 6, 5, 5, 3, 4, 5, 6,
+                                                     7, 1, 2, 4, 6, 4, 4, 5,
+                                                     7, 2, 6, 5, 7, 6, 7, 7 };
+            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
+#       endif
+        } else { /* 32 bits */
+#       if defined(_MSC_VER)
+            unsigned long r=0;
+            _BitScanForward( &r, (U32)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_ctz((U32)val) >> 3);
+#       else
+            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0,
+                                                     3, 2, 2, 1, 3, 2, 0, 1,
+                                                     3, 3, 1, 2, 2, 2, 2, 0,
+                                                     3, 1, 2, 0, 1, 0, 1, 1 };
+            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
+#       endif
+        }
+    } else {  /* Big Endian CPU */
+        if (MEM_64bits()) {
+#       if defined(_MSC_VER) && defined(_WIN64)
+            unsigned long r = 0;
+            _BitScanReverse64( &r, val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 4)
+            return (__builtin_clzll(val) >> 3);
+#       else
+            unsigned r;
+            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
+            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
+            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
+            r += (!val);
+            return r;
+#       endif
+        } else { /* 32 bits */
+#       if defined(_MSC_VER)
+            unsigned long r = 0;
+            _BitScanReverse( &r, (unsigned long)val );
+            return (unsigned)(r>>3);
+#       elif defined(__GNUC__) && (__GNUC__ >= 3)
+            return (__builtin_clz((U32)val) >> 3);
+#       else
+            unsigned r;
+            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
+            r += (!val);
+            return r;
+#       endif
+    }   }
+}
+
+
+MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
+{
+    const BYTE* const pStart = pIn;
+    const BYTE* const pInLoopLimit = pInLimit - (sizeof(size_t)-1);
+
+    if (pIn < pInLoopLimit) {
+        { size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
+          if (diff) return ZSTD_NbCommonBytes(diff); }
+        pIn+=sizeof(size_t); pMatch+=sizeof(size_t);
+        while (pIn < pInLoopLimit) {
+            size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
+            if (!diff) { pIn+=sizeof(size_t); pMatch+=sizeof(size_t); continue; }
+            pIn += ZSTD_NbCommonBytes(diff);
+            return (size_t)(pIn - pStart);
+    }   }
+    if (MEM_64bits() && (pIn<(pInLimit-3)) && (MEM_read32(pMatch) == MEM_read32(pIn))) { pIn+=4; pMatch+=4; }
+    if ((pIn<(pInLimit-1)) && (MEM_read16(pMatch) == MEM_read16(pIn))) { pIn+=2; pMatch+=2; }
+    if ((pIn<pInLimit) && (*pMatch == *pIn)) pIn++;
+    return (size_t)(pIn - pStart);
+}
+
+/** ZSTD_count_2segments() :
+ *  can count match length with `ip` & `match` in 2 different segments.
+ *  convention : on reaching mEnd, match count continue starting from iStart
+ */
+MEM_STATIC size_t
+ZSTD_count_2segments(const BYTE* ip, const BYTE* match,
+                     const BYTE* iEnd, const BYTE* mEnd, const BYTE* iStart)
+{
+    const BYTE* const vEnd = MIN( ip + (mEnd - match), iEnd);
+    size_t const matchLength = ZSTD_count(ip, match, vEnd);
+    if (match + matchLength != mEnd) return matchLength;
+    return matchLength + ZSTD_count(ip+matchLength, iStart, iEnd);
+}
+
+
+/*-*************************************
+ *  Hashes
+ ***************************************/
+static const U32 prime3bytes = 506832829U;
+static U32    ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
+MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */
+
+static const U32 prime4bytes = 2654435761U;
+static U32    ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; }
+static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); }
+
+static const U64 prime5bytes = 889523592379ULL;
+static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
+static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); }
+
+static const U64 prime6bytes = 227718039650203ULL;
+static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
+static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
+
+static const U64 prime7bytes = 58295818150454627ULL;
+static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
+static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); }
+
+static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
+static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
+static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
+
+MEM_STATIC size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
+{
+    switch(mls)
+    {
+    default:
+    case 4: return ZSTD_hash4Ptr(p, hBits);
+    case 5: return ZSTD_hash5Ptr(p, hBits);
+    case 6: return ZSTD_hash6Ptr(p, hBits);
+    case 7: return ZSTD_hash7Ptr(p, hBits);
+    case 8: return ZSTD_hash8Ptr(p, hBits);
+    }
+}
+
+/*-*************************************
+*  Round buffer management
+***************************************/
+/* Max current allowed */
+#define ZSTD_CURRENT_MAX ((3U << 29) + (1U << ZSTD_WINDOWLOG_MAX))
+/* Maximum chunk size before overflow correction needs to be called again */
+#define ZSTD_CHUNKSIZE_MAX                                                     \
+    ( ((U32)-1)                  /* Maximum ending current index */            \
+    - ZSTD_CURRENT_MAX)          /* Maximum beginning lowLimit */
+
+/**
+ * ZSTD_window_clear():
+ * Clears the window containing the history by simply setting it to empty.
+ */
+MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window)
+{
+    size_t const endT = (size_t)(window->nextSrc - window->base);
+    U32 const end = (U32)endT;
+
+    window->lowLimit = end;
+    window->dictLimit = end;
+}
+
+/**
+ * ZSTD_window_hasExtDict():
+ * Returns non-zero if the window has a non-empty extDict.
+ */
+MEM_STATIC U32 ZSTD_window_hasExtDict(ZSTD_window_t const window)
+{
+    return window.lowLimit < window.dictLimit;
+}
+
+/**
+ * ZSTD_window_needOverflowCorrection():
+ * Returns non-zero if the indices are getting too large and need overflow
+ * protection.
+ */
+MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
+                                                  void const* srcEnd)
+{
+    U32 const current = (U32)((BYTE const*)srcEnd - window.base);
+    return current > ZSTD_CURRENT_MAX;
+}
+
+/**
+ * ZSTD_window_correctOverflow():
+ * Reduces the indices to protect from index overflow.
+ * Returns the correction made to the indices, which must be applied to every
+ * stored index.
+ *
+ * The least significant cycleLog bits of the indices must remain the same,
+ * which may be 0. Every index up to maxDist in the past must be valid.
+ * NOTE: (maxDist & cycleMask) must be zero.
+ */
+MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
+                                           U32 maxDist, void const* src)
+{
+    /* preemptive overflow correction:
+     * 1. correction is large enough:
+     *    lowLimit > (3<<29) ==> current > 3<<29 + 1<<windowLog
+     *    1<<windowLog <= newCurrent < 1<<chainLog + 1<<windowLog
+     *
+     *    current - newCurrent
+     *    > (3<<29 + 1<<windowLog) - (1<<windowLog + 1<<chainLog)
+     *    > (3<<29) - (1<<chainLog)
+     *    > (3<<29) - (1<<30)             (NOTE: chainLog <= 30)
+     *    > 1<<29
+     *
+     * 2. (ip+ZSTD_CHUNKSIZE_MAX - cctx->base) doesn't overflow:
+     *    After correction, current is less than (1<<chainLog + 1<<windowLog).
+     *    In 64-bit mode we are safe, because we have 64-bit ptrdiff_t.
+     *    In 32-bit mode we are safe, because (chainLog <= 29), so
+     *    ip+ZSTD_CHUNKSIZE_MAX - cctx->base < 1<<32.
+     * 3. (cctx->lowLimit + 1<<windowLog) < 1<<32:
+     *    windowLog <= 31 ==> 3<<29 + 1<<windowLog < 7<<29 < 1<<32.
+     */
+    U32 const cycleMask = (1U << cycleLog) - 1;
+    U32 const current = (U32)((BYTE const*)src - window->base);
+    U32 const newCurrent = (current & cycleMask) + maxDist;
+    U32 const correction = current - newCurrent;
+    assert((maxDist & cycleMask) == 0);
+    assert(current > newCurrent);
+    /* Loose bound, should be around 1<<29 (see above) */
+    assert(correction > 1<<28);
+
+    window->base += correction;
+    window->dictBase += correction;
+    window->lowLimit -= correction;
+    window->dictLimit -= correction;
+
+    DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction,
+             window->lowLimit);
+    return correction;
+}
+
+/**
+ * ZSTD_window_enforceMaxDist():
+ * Updates lowLimit so that:
+ *    (srcEnd - base) - lowLimit == maxDist + loadedDictEnd
+ * This allows a simple check that index >= lowLimit to see if index is valid.
+ * This must be called before a block compression call, with srcEnd as the block
+ * source end.
+ * If loadedDictEndPtr is not NULL, we set it to zero once we update lowLimit.
+ * This is because dictionaries are allowed to be referenced as long as the last
+ * byte of the dictionary is in the window, but once they are out of range,
+ * they cannot be referenced. If loadedDictEndPtr is NULL, we use
+ * loadedDictEnd == 0.
+ */
+MEM_STATIC void ZSTD_window_enforceMaxDist(ZSTD_window_t* window,
+                                           void const* srcEnd, U32 maxDist,
+                                           U32* loadedDictEndPtr)
+{
+    U32 const current = (U32)((BYTE const*)srcEnd - window->base);
+    U32 loadedDictEnd = loadedDictEndPtr != NULL ? *loadedDictEndPtr : 0;
+    if (current > maxDist + loadedDictEnd) {
+        U32 const newLowLimit = current - maxDist;
+        if (window->lowLimit < newLowLimit) window->lowLimit = newLowLimit;
+        if (window->dictLimit < window->lowLimit) {
+            DEBUGLOG(5, "Update dictLimit from %u to %u", window->dictLimit,
+                     window->lowLimit);
+            window->dictLimit = window->lowLimit;
+        }
+        if (loadedDictEndPtr)
+            *loadedDictEndPtr = 0;
+    }
+}
+
+/**
+ * ZSTD_window_update():
+ * Updates the window by appending [src, src + srcSize) to the window.
+ * If it is not contiguous, the current prefix becomes the extDict, and we
+ * forget about the extDict. Handles overlap of the prefix and extDict.
+ * Returns non-zero if the segment is contiguous.
+ */
+MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
+                                  void const* src, size_t srcSize)
+{
+    BYTE const* const ip = (BYTE const*)src;
+    U32 contiguous = 1;
+    /* Check if blocks follow each other */
+    if (src != window->nextSrc) {
+        /* not contiguous */
+        size_t const distanceFromBase = (size_t)(window->nextSrc - window->base);
+        DEBUGLOG(5, "Non contiguous blocks, new segment starts at %u",
+                 window->dictLimit);
+        window->lowLimit = window->dictLimit;
+        assert(distanceFromBase == (size_t)(U32)distanceFromBase);  /* should never overflow */
+        window->dictLimit = (U32)distanceFromBase;
+        window->dictBase = window->base;
+        window->base = ip - distanceFromBase;
+        // ms->nextToUpdate = window->dictLimit;
+        if (window->dictLimit - window->lowLimit < HASH_READ_SIZE) window->lowLimit = window->dictLimit;   /* too small extDict */
+        contiguous = 0;
+    }
+    window->nextSrc = ip + srcSize;
+    /* if input and dictionary overlap : reduce dictionary (area presumed modified by input) */
+    if ( (ip+srcSize > window->dictBase + window->lowLimit)
+       & (ip < window->dictBase + window->dictLimit)) {
+        ptrdiff_t const highInputIdx = (ip + srcSize) - window->dictBase;
+        U32 const lowLimitMax = (highInputIdx > (ptrdiff_t)window->dictLimit) ? window->dictLimit : (U32)highInputIdx;
+        window->lowLimit = lowLimitMax;
+    }
+    return contiguous;
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+
+/* ==============================================================
+ * Private declarations
+ * These prototypes shall only be called from within lib/compress
+ * ============================================================== */
+
+/* ZSTD_getCParamsFromCCtxParams() :
+ * cParams are built depending on compressionLevel, src size hints, 
+ * LDM and manually set compression parameters.
+ */
+ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize);
+
+/*! ZSTD_initCStream_internal() :
+ *  Private use only. Init streaming operation.
+ *  expects params to be valid.
+ *  must receive dict, or cdict, or none, but not both.
+ *  @return : 0, or an error code */
+size_t ZSTD_initCStream_internal(ZSTD_CStream* zcs,
+                     const void* dict, size_t dictSize,
+                     const ZSTD_CDict* cdict,
+                     ZSTD_CCtx_params  params, unsigned long long pledgedSrcSize);
+
+/*! ZSTD_compressStream_generic() :
+ *  Private use only. To be called from zstdmt_compress.c in single-thread mode. */
+size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
+                                   ZSTD_outBuffer* output,
+                                   ZSTD_inBuffer* input,
+                                   ZSTD_EndDirective const flushMode);
+
+/*! ZSTD_getCParamsFromCDict() :
+ *  as the name implies */
+ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict);
+
+/* ZSTD_compressBegin_advanced_internal() :
+ * Private use only. To be called from zstdmt_compress.c. */
+size_t ZSTD_compressBegin_advanced_internal(ZSTD_CCtx* cctx,
+                                    const void* dict, size_t dictSize,
+                                    ZSTD_dictContentType_e dictContentType,
+                                    const ZSTD_CDict* cdict,
+                                    ZSTD_CCtx_params params,
+                                    unsigned long long pledgedSrcSize);
+
+/* ZSTD_compress_advanced_internal() :
+ * Private use only. To be called from zstdmt_compress.c. */
+size_t ZSTD_compress_advanced_internal(ZSTD_CCtx* cctx,
+                                       void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize,
+                                 const void* dict,size_t dictSize,
+                                 ZSTD_CCtx_params params);
+
+
+/* ZSTD_writeLastEmptyBlock() :
+ * output an empty Block with end-of-frame mark to complete a frame
+ * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
+ *           or an error code if `dstCapcity` is too small (<ZSTD_blockHeaderSize)
+ */
+size_t ZSTD_writeLastEmptyBlock(void* dst, size_t dstCapacity);
+
+
+/* ZSTD_referenceExternalSequences() :
+ * Must be called before starting a compression operation.
+ * seqs must parse a prefix of the source.
+ * This cannot be used when long range matching is enabled.
+ * Zstd will use these sequences, and pass the literals to a secondary block
+ * compressor.
+ * @return : An error code on failure.
+ * NOTE: seqs are not verified! Invalid sequences can cause out-of-bounds memory
+ * access and data corruption.
+ */
+size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq);
+
+
+#endif /* ZSTD_COMPRESS_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_decompress.c b/vendor/github.com/DataDog/zstd/zstd_decompress.c
new file mode 100644
index 0000000..3ec6a1c
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_decompress.c
@@ -0,0 +1,3003 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * HEAPMODE :
+ * Select how default decompression function ZSTD_decompress() allocates its context,
+ * on stack (0), or into heap (1, default; requires malloc()).
+ * Note that functions with explicit context such as ZSTD_decompressDCtx() are unaffected.
+ */
+#ifndef ZSTD_HEAPMODE
+#  define ZSTD_HEAPMODE 1
+#endif
+
+/*!
+*  LEGACY_SUPPORT :
+*  if set to 1+, ZSTD_decompress() can decode older formats (v0.1+)
+*/
+#ifndef ZSTD_LEGACY_SUPPORT
+#  define ZSTD_LEGACY_SUPPORT 0
+#endif
+
+/*!
+ *  MAXWINDOWSIZE_DEFAULT :
+ *  maximum window size accepted by DStream __by default__.
+ *  Frames requiring more memory will be rejected.
+ *  It's possible to set a different limit using ZSTD_DCtx_setMaxWindowSize().
+ */
+#ifndef ZSTD_MAXWINDOWSIZE_DEFAULT
+#  define ZSTD_MAXWINDOWSIZE_DEFAULT (((U32)1 << ZSTD_WINDOWLOG_DEFAULTMAX) + 1)
+#endif
+
+
+/*-*******************************************************
+*  Dependencies
+*********************************************************/
+#include <string.h>      /* memcpy, memmove, memset */
+#include "cpu.h"
+#include "mem.h"         /* low level memory routines */
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"
+#include "zstd_internal.h"
+
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
+#  include "zstd_legacy.h"
+#endif
+
+
+/*-*************************************
+*  Errors
+***************************************/
+#define ZSTD_isError ERR_isError   /* for inlining */
+#define FSE_isError  ERR_isError
+#define HUF_isError  ERR_isError
+
+
+/*_*******************************************************
+*  Memory operations
+**********************************************************/
+static void ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+
+/*-*************************************************************
+*   Context management
+***************************************************************/
+typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
+               ZSTDds_decodeBlockHeader, ZSTDds_decompressBlock,
+               ZSTDds_decompressLastBlock, ZSTDds_checkChecksum,
+               ZSTDds_decodeSkippableHeader, ZSTDds_skipFrame } ZSTD_dStage;
+
+typedef enum { zdss_init=0, zdss_loadHeader,
+               zdss_read, zdss_load, zdss_flush } ZSTD_dStreamStage;
+
+
+typedef struct {
+    U32 fastMode;
+    U32 tableLog;
+} ZSTD_seqSymbol_header;
+
+typedef struct {
+    U16  nextState;
+    BYTE nbAdditionalBits;
+    BYTE nbBits;
+    U32  baseValue;
+} ZSTD_seqSymbol;
+
+#define SEQSYMBOL_TABLE_SIZE(log)   (1 + (1 << (log)))
+
+typedef struct {
+    ZSTD_seqSymbol LLTable[SEQSYMBOL_TABLE_SIZE(LLFSELog)];
+    ZSTD_seqSymbol OFTable[SEQSYMBOL_TABLE_SIZE(OffFSELog)];
+    ZSTD_seqSymbol MLTable[SEQSYMBOL_TABLE_SIZE(MLFSELog)];
+    HUF_DTable hufTable[HUF_DTABLE_SIZE(HufLog)];  /* can accommodate HUF_decompress4X */
+    U32 workspace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
+    U32 rep[ZSTD_REP_NUM];
+} ZSTD_entropyDTables_t;
+
+struct ZSTD_DCtx_s
+{
+    const ZSTD_seqSymbol* LLTptr;
+    const ZSTD_seqSymbol* MLTptr;
+    const ZSTD_seqSymbol* OFTptr;
+    const HUF_DTable* HUFptr;
+    ZSTD_entropyDTables_t entropy;
+    const void* previousDstEnd;   /* detect continuity */
+    const void* base;             /* start of current segment */
+    const void* vBase;            /* virtual start of previous segment if it was just before current one */
+    const void* dictEnd;          /* end of previous segment */
+    size_t expected;
+    ZSTD_frameHeader fParams;
+    U64 decodedSize;
+    blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
+    ZSTD_dStage stage;
+    U32 litEntropy;
+    U32 fseEntropy;
+    XXH64_state_t xxhState;
+    size_t headerSize;
+    U32 dictID;
+    ZSTD_format_e format;
+    const BYTE* litPtr;
+    ZSTD_customMem customMem;
+    size_t litSize;
+    size_t rleSize;
+    size_t staticSize;
+    int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
+
+    /* streaming */
+    ZSTD_DDict* ddictLocal;
+    const ZSTD_DDict* ddict;
+    ZSTD_dStreamStage streamStage;
+    char*  inBuff;
+    size_t inBuffSize;
+    size_t inPos;
+    size_t maxWindowSize;
+    char*  outBuff;
+    size_t outBuffSize;
+    size_t outStart;
+    size_t outEnd;
+    size_t lhSize;
+    void* legacyContext;
+    U32 previousLegacyVersion;
+    U32 legacyVersion;
+    U32 hostageByte;
+
+    /* workspace */
+    BYTE litBuffer[ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH];
+    BYTE headerBuffer[ZSTD_FRAMEHEADERSIZE_MAX];
+};  /* typedef'd to ZSTD_DCtx within "zstd.h" */
+
+size_t ZSTD_sizeof_DCtx (const ZSTD_DCtx* dctx)
+{
+    if (dctx==NULL) return 0;   /* support sizeof NULL */
+    return sizeof(*dctx)
+           + ZSTD_sizeof_DDict(dctx->ddictLocal)
+           + dctx->inBuffSize + dctx->outBuffSize;
+}
+
+size_t ZSTD_estimateDCtxSize(void) { return sizeof(ZSTD_DCtx); }
+
+
+static size_t ZSTD_startingInputLength(ZSTD_format_e format)
+{
+    size_t const startingInputLength = (format==ZSTD_f_zstd1_magicless) ?
+                    ZSTD_frameHeaderSize_prefix - ZSTD_frameIdSize :
+                    ZSTD_frameHeaderSize_prefix;
+    ZSTD_STATIC_ASSERT(ZSTD_FRAMEHEADERSIZE_PREFIX >= ZSTD_FRAMEIDSIZE);
+    /* only supports formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless */
+    assert( (format == ZSTD_f_zstd1) || (format == ZSTD_f_zstd1_magicless) );
+    return startingInputLength;
+}
+
+static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
+{
+    dctx->format = ZSTD_f_zstd1;  /* ZSTD_decompressBegin() invokes ZSTD_startingInputLength() with argument dctx->format */
+    dctx->staticSize  = 0;
+    dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
+    dctx->ddict       = NULL;
+    dctx->ddictLocal  = NULL;
+    dctx->inBuff      = NULL;
+    dctx->inBuffSize  = 0;
+    dctx->outBuffSize = 0;
+    dctx->streamStage = zdss_init;
+    dctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+}
+
+ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize)
+{
+    ZSTD_DCtx* const dctx = (ZSTD_DCtx*) workspace;
+
+    if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
+    if (workspaceSize < sizeof(ZSTD_DCtx)) return NULL;  /* minimum size */
+
+    ZSTD_initDCtx_internal(dctx);
+    dctx->staticSize = workspaceSize;
+    dctx->inBuff = (char*)(dctx+1);
+    return dctx;
+}
+
+ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
+{
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+
+    {   ZSTD_DCtx* const dctx = (ZSTD_DCtx*)ZSTD_malloc(sizeof(*dctx), customMem);
+        if (!dctx) return NULL;
+        dctx->customMem = customMem;
+        dctx->legacyContext = NULL;
+        dctx->previousLegacyVersion = 0;
+        ZSTD_initDCtx_internal(dctx);
+        return dctx;
+    }
+}
+
+ZSTD_DCtx* ZSTD_createDCtx(void)
+{
+    DEBUGLOG(3, "ZSTD_createDCtx");
+    return ZSTD_createDCtx_advanced(ZSTD_defaultCMem);
+}
+
+size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
+{
+    if (dctx==NULL) return 0;   /* support free on NULL */
+    if (dctx->staticSize) return ERROR(memory_allocation);   /* not compatible with static DCtx */
+    {   ZSTD_customMem const cMem = dctx->customMem;
+        ZSTD_freeDDict(dctx->ddictLocal);
+        dctx->ddictLocal = NULL;
+        ZSTD_free(dctx->inBuff, cMem);
+        dctx->inBuff = NULL;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
+        if (dctx->legacyContext)
+            ZSTD_freeLegacyStreamContext(dctx->legacyContext, dctx->previousLegacyVersion);
+#endif
+        ZSTD_free(dctx, cMem);
+        return 0;
+    }
+}
+
+/* no longer useful */
+void ZSTD_copyDCtx(ZSTD_DCtx* dstDCtx, const ZSTD_DCtx* srcDCtx)
+{
+    size_t const toCopy = (size_t)((char*)(&dstDCtx->inBuff) - (char*)dstDCtx);
+    memcpy(dstDCtx, srcDCtx, toCopy);  /* no need to copy workspace */
+}
+
+
+/*-*************************************************************
+ *   Frame header decoding
+ ***************************************************************/
+
+/*! ZSTD_isFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier.
+ *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
+ *  Note 2 : Legacy Frame Identifiers are considered valid only if Legacy Support is enabled.
+ *  Note 3 : Skippable Frame Identifiers are considered valid. */
+unsigned ZSTD_isFrame(const void* buffer, size_t size)
+{
+    if (size < ZSTD_frameIdSize) return 0;
+    {   U32 const magic = MEM_readLE32(buffer);
+        if (magic == ZSTD_MAGICNUMBER) return 1;
+        if ((magic & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) return 1;
+    }
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
+    if (ZSTD_isLegacy(buffer, size)) return 1;
+#endif
+    return 0;
+}
+
+/** ZSTD_frameHeaderSize_internal() :
+ *  srcSize must be large enough to reach header size fields.
+ *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless.
+ * @return : size of the Frame Header
+ *           or an error code, which can be tested with ZSTD_isError() */
+static size_t ZSTD_frameHeaderSize_internal(const void* src, size_t srcSize, ZSTD_format_e format)
+{
+    size_t const minInputSize = ZSTD_startingInputLength(format);
+    if (srcSize < minInputSize) return ERROR(srcSize_wrong);
+
+    {   BYTE const fhd = ((const BYTE*)src)[minInputSize-1];
+        U32 const dictID= fhd & 3;
+        U32 const singleSegment = (fhd >> 5) & 1;
+        U32 const fcsId = fhd >> 6;
+        return minInputSize + !singleSegment
+             + ZSTD_did_fieldSize[dictID] + ZSTD_fcs_fieldSize[fcsId]
+             + (singleSegment && !fcsId);
+    }
+}
+
+/** ZSTD_frameHeaderSize() :
+ *  srcSize must be >= ZSTD_frameHeaderSize_prefix.
+ * @return : size of the Frame Header */
+size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize)
+{
+    return ZSTD_frameHeaderSize_internal(src, srcSize, ZSTD_f_zstd1);
+}
+
+
+/** ZSTD_getFrameHeader_internal() :
+ *  decode Frame Header, or require larger `srcSize`.
+ *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless
+ * @return : 0, `zfhPtr` is correctly filled,
+ *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
+ *           or an error code, which can be tested using ZSTD_isError() */
+static size_t ZSTD_getFrameHeader_internal(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t const minInputSize = ZSTD_startingInputLength(format);
+
+    if (srcSize < minInputSize) return minInputSize;
+
+    if ( (format != ZSTD_f_zstd1_magicless)
+      && (MEM_readLE32(src) != ZSTD_MAGICNUMBER) ) {
+        if ((MEM_readLE32(src) & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) {
+            /* skippable frame */
+            if (srcSize < ZSTD_skippableHeaderSize)
+                return ZSTD_skippableHeaderSize; /* magic number + frame length */
+            memset(zfhPtr, 0, sizeof(*zfhPtr));
+            zfhPtr->frameContentSize = MEM_readLE32((const char *)src + ZSTD_frameIdSize);
+            zfhPtr->frameType = ZSTD_skippableFrame;
+            return 0;
+        }
+        return ERROR(prefix_unknown);
+    }
+
+    /* ensure there is enough `srcSize` to fully read/decode frame header */
+    {   size_t const fhsize = ZSTD_frameHeaderSize_internal(src, srcSize, format);
+        if (srcSize < fhsize) return fhsize;
+        zfhPtr->headerSize = (U32)fhsize;
+    }
+
+    {   BYTE const fhdByte = ip[minInputSize-1];
+        size_t pos = minInputSize;
+        U32 const dictIDSizeCode = fhdByte&3;
+        U32 const checksumFlag = (fhdByte>>2)&1;
+        U32 const singleSegment = (fhdByte>>5)&1;
+        U32 const fcsID = fhdByte>>6;
+        U64 windowSize = 0;
+        U32 dictID = 0;
+        U64 frameContentSize = ZSTD_CONTENTSIZE_UNKNOWN;
+        if ((fhdByte & 0x08) != 0)
+            return ERROR(frameParameter_unsupported); /* reserved bits, must be zero */
+
+        if (!singleSegment) {
+            BYTE const wlByte = ip[pos++];
+            U32 const windowLog = (wlByte >> 3) + ZSTD_WINDOWLOG_ABSOLUTEMIN;
+            if (windowLog > ZSTD_WINDOWLOG_MAX)
+                return ERROR(frameParameter_windowTooLarge);
+            windowSize = (1ULL << windowLog);
+            windowSize += (windowSize >> 3) * (wlByte&7);
+        }
+        switch(dictIDSizeCode)
+        {
+            default: assert(0);  /* impossible */
+            case 0 : break;
+            case 1 : dictID = ip[pos]; pos++; break;
+            case 2 : dictID = MEM_readLE16(ip+pos); pos+=2; break;
+            case 3 : dictID = MEM_readLE32(ip+pos); pos+=4; break;
+        }
+        switch(fcsID)
+        {
+            default: assert(0);  /* impossible */
+            case 0 : if (singleSegment) frameContentSize = ip[pos]; break;
+            case 1 : frameContentSize = MEM_readLE16(ip+pos)+256; break;
+            case 2 : frameContentSize = MEM_readLE32(ip+pos); break;
+            case 3 : frameContentSize = MEM_readLE64(ip+pos); break;
+        }
+        if (singleSegment) windowSize = frameContentSize;
+
+        zfhPtr->frameType = ZSTD_frame;
+        zfhPtr->frameContentSize = frameContentSize;
+        zfhPtr->windowSize = windowSize;
+        zfhPtr->blockSizeMax = (unsigned) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
+        zfhPtr->dictID = dictID;
+        zfhPtr->checksumFlag = checksumFlag;
+    }
+    return 0;
+}
+
+/** ZSTD_getFrameHeader() :
+ *  decode Frame Header, or require larger `srcSize`.
+ *  note : this function does not consume input, it only reads it.
+ * @return : 0, `zfhPtr` is correctly filled,
+ *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
+ *           or an error code, which can be tested using ZSTD_isError() */
+size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize)
+{
+    return ZSTD_getFrameHeader_internal(zfhPtr, src, srcSize, ZSTD_f_zstd1);
+}
+
+
+/** ZSTD_getFrameContentSize() :
+ *  compatible with legacy mode
+ * @return : decompressed size of the single frame pointed to be `src` if known, otherwise
+ *         - ZSTD_CONTENTSIZE_UNKNOWN if the size cannot be determined
+ *         - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small) */
+unsigned long long ZSTD_getFrameContentSize(const void *src, size_t srcSize)
+{
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
+    if (ZSTD_isLegacy(src, srcSize)) {
+        unsigned long long const ret = ZSTD_getDecompressedSize_legacy(src, srcSize);
+        return ret == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : ret;
+    }
+#endif
+    {   ZSTD_frameHeader zfh;
+        if (ZSTD_getFrameHeader(&zfh, src, srcSize) != 0)
+            return ZSTD_CONTENTSIZE_ERROR;
+        if (zfh.frameType == ZSTD_skippableFrame) {
+            return 0;
+        } else {
+            return zfh.frameContentSize;
+    }   }
+}
+
+/** ZSTD_findDecompressedSize() :
+ *  compatible with legacy mode
+ *  `srcSize` must be the exact length of some number of ZSTD compressed and/or
+ *      skippable frames
+ *  @return : decompressed size of the frames contained */
+unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize)
+{
+    unsigned long long totalDstSize = 0;
+
+    while (srcSize >= ZSTD_frameHeaderSize_prefix) {
+        U32 const magicNumber = MEM_readLE32(src);
+
+        if ((magicNumber & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) {
+            size_t skippableSize;
+            if (srcSize < ZSTD_skippableHeaderSize)
+                return ERROR(srcSize_wrong);
+            skippableSize = MEM_readLE32((const BYTE *)src + ZSTD_frameIdSize)
+                          + ZSTD_skippableHeaderSize;
+            if (srcSize < skippableSize) {
+                return ZSTD_CONTENTSIZE_ERROR;
+            }
+
+            src = (const BYTE *)src + skippableSize;
+            srcSize -= skippableSize;
+            continue;
+        }
+
+        {   unsigned long long const ret = ZSTD_getFrameContentSize(src, srcSize);
+            if (ret >= ZSTD_CONTENTSIZE_ERROR) return ret;
+
+            /* check for overflow */
+            if (totalDstSize + ret < totalDstSize) return ZSTD_CONTENTSIZE_ERROR;
+            totalDstSize += ret;
+        }
+        {   size_t const frameSrcSize = ZSTD_findFrameCompressedSize(src, srcSize);
+            if (ZSTD_isError(frameSrcSize)) {
+                return ZSTD_CONTENTSIZE_ERROR;
+            }
+
+            src = (const BYTE *)src + frameSrcSize;
+            srcSize -= frameSrcSize;
+        }
+    }  /* while (srcSize >= ZSTD_frameHeaderSize_prefix) */
+
+    if (srcSize) return ZSTD_CONTENTSIZE_ERROR;
+
+    return totalDstSize;
+}
+
+/** ZSTD_getDecompressedSize() :
+*   compatible with legacy mode
+*   @return : decompressed size if known, 0 otherwise
+              note : 0 can mean any of the following :
+                   - frame content is empty
+                   - decompressed size field is not present in frame header
+                   - frame header unknown / not supported
+                   - frame header not complete (`srcSize` too small) */
+unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize)
+{
+    unsigned long long const ret = ZSTD_getFrameContentSize(src, srcSize);
+    ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_ERROR < ZSTD_CONTENTSIZE_UNKNOWN);
+    return (ret >= ZSTD_CONTENTSIZE_ERROR) ? 0 : ret;
+}
+
+
+/** ZSTD_decodeFrameHeader() :
+*   `headerSize` must be the size provided by ZSTD_frameHeaderSize().
+*   @return : 0 if success, or an error code, which can be tested using ZSTD_isError() */
+static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t headerSize)
+{
+    size_t const result = ZSTD_getFrameHeader_internal(&(dctx->fParams), src, headerSize, dctx->format);
+    if (ZSTD_isError(result)) return result;    /* invalid header */
+    if (result>0) return ERROR(srcSize_wrong);  /* headerSize too small */
+    if (dctx->fParams.dictID && (dctx->dictID != dctx->fParams.dictID))
+        return ERROR(dictionary_wrong);
+    if (dctx->fParams.checksumFlag) XXH64_reset(&dctx->xxhState, 0);
+    return 0;
+}
+
+
+/*-*************************************************************
+ *   Block decoding
+ ***************************************************************/
+
+/*! ZSTD_getcBlockSize() :
+*   Provides the size of compressed block from block header `src` */
+size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
+                          blockProperties_t* bpPtr)
+{
+    if (srcSize < ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    {   U32 const cBlockHeader = MEM_readLE24(src);
+        U32 const cSize = cBlockHeader >> 3;
+        bpPtr->lastBlock = cBlockHeader & 1;
+        bpPtr->blockType = (blockType_e)((cBlockHeader >> 1) & 3);
+        bpPtr->origSize = cSize;   /* only useful for RLE */
+        if (bpPtr->blockType == bt_rle) return 1;
+        if (bpPtr->blockType == bt_reserved) return ERROR(corruption_detected);
+        return cSize;
+    }
+}
+
+
+static size_t ZSTD_copyRawBlock(void* dst, size_t dstCapacity,
+                          const void* src, size_t srcSize)
+{
+    if (srcSize > dstCapacity) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+static size_t ZSTD_setRleBlock(void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize,
+                               size_t regenSize)
+{
+    if (srcSize != 1) return ERROR(srcSize_wrong);
+    if (regenSize > dstCapacity) return ERROR(dstSize_tooSmall);
+    memset(dst, *(const BYTE*)src, regenSize);
+    return regenSize;
+}
+
+/*! ZSTD_decodeLiteralsBlock() :
+ * @return : nb of bytes read from src (< srcSize )
+ *  note : symbol not declared but exposed for fullbench */
+size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
+                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+{
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    {   const BYTE* const istart = (const BYTE*) src;
+        symbolEncodingType_e const litEncType = (symbolEncodingType_e)(istart[0] & 3);
+
+        switch(litEncType)
+        {
+        case set_repeat:
+            if (dctx->litEntropy==0) return ERROR(dictionary_corrupted);
+            /* fall-through */
+        case set_compressed:
+            if (srcSize < 5) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for case 3 */
+            {   size_t lhSize, litSize, litCSize;
+                U32 singleStream=0;
+                U32 const lhlCode = (istart[0] >> 2) & 3;
+                U32 const lhc = MEM_readLE32(istart);
+                switch(lhlCode)
+                {
+                case 0: case 1: default:   /* note : default is impossible, since lhlCode into [0..3] */
+                    /* 2 - 2 - 10 - 10 */
+                    singleStream = !lhlCode;
+                    lhSize = 3;
+                    litSize  = (lhc >> 4) & 0x3FF;
+                    litCSize = (lhc >> 14) & 0x3FF;
+                    break;
+                case 2:
+                    /* 2 - 2 - 14 - 14 */
+                    lhSize = 4;
+                    litSize  = (lhc >> 4) & 0x3FFF;
+                    litCSize = lhc >> 18;
+                    break;
+                case 3:
+                    /* 2 - 2 - 18 - 18 */
+                    lhSize = 5;
+                    litSize  = (lhc >> 4) & 0x3FFFF;
+                    litCSize = (lhc >> 22) + (istart[4] << 10);
+                    break;
+                }
+                if (litSize > ZSTD_BLOCKSIZE_MAX) return ERROR(corruption_detected);
+                if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+                if (HUF_isError((litEncType==set_repeat) ?
+                                    ( singleStream ?
+                                        HUF_decompress1X_usingDTable_bmi2(dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->HUFptr, dctx->bmi2) :
+                                        HUF_decompress4X_usingDTable_bmi2(dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->HUFptr, dctx->bmi2) ) :
+                                    ( singleStream ?
+                                        HUF_decompress1X2_DCtx_wksp_bmi2(dctx->entropy.hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize,
+                                                                         dctx->entropy.workspace, sizeof(dctx->entropy.workspace), dctx->bmi2) :
+                                        HUF_decompress4X_hufOnly_wksp_bmi2(dctx->entropy.hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize,
+                                                                           dctx->entropy.workspace, sizeof(dctx->entropy.workspace), dctx->bmi2))))
+                    return ERROR(corruption_detected);
+
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                dctx->litEntropy = 1;
+                if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
+                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                return litCSize + lhSize;
+            }
+
+        case set_basic:
+            {   size_t litSize, lhSize;
+                U32 const lhlCode = ((istart[0]) >> 2) & 3;
+                switch(lhlCode)
+                {
+                case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
+                    lhSize = 1;
+                    litSize = istart[0] >> 3;
+                    break;
+                case 1:
+                    lhSize = 2;
+                    litSize = MEM_readLE16(istart) >> 4;
+                    break;
+                case 3:
+                    lhSize = 3;
+                    litSize = MEM_readLE24(istart) >> 4;
+                    break;
+                }
+
+                if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
+                    if (litSize+lhSize > srcSize) return ERROR(corruption_detected);
+                    memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                    dctx->litPtr = dctx->litBuffer;
+                    dctx->litSize = litSize;
+                    memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                    return lhSize+litSize;
+                }
+                /* direct reference into compressed stream */
+                dctx->litPtr = istart+lhSize;
+                dctx->litSize = litSize;
+                return lhSize+litSize;
+            }
+
+        case set_rle:
+            {   U32 const lhlCode = ((istart[0]) >> 2) & 3;
+                size_t litSize, lhSize;
+                switch(lhlCode)
+                {
+                case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
+                    lhSize = 1;
+                    litSize = istart[0] >> 3;
+                    break;
+                case 1:
+                    lhSize = 2;
+                    litSize = MEM_readLE16(istart) >> 4;
+                    break;
+                case 3:
+                    lhSize = 3;
+                    litSize = MEM_readLE24(istart) >> 4;
+                    if (srcSize<4) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4 */
+                    break;
+                }
+                if (litSize > ZSTD_BLOCKSIZE_MAX) return ERROR(corruption_detected);
+                memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                return lhSize+1;
+            }
+        default:
+            return ERROR(corruption_detected);   /* impossible */
+        }
+    }
+}
+
+/* Default FSE distribution tables.
+ * These are pre-calculated FSE decoding tables using default distributions as defined in specification :
+ * https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#default-distributions
+ * They were generated programmatically with following method :
+ * - start from default distributions, present in /lib/common/zstd_internal.h
+ * - generate tables normally, using ZSTD_buildFSETable()
+ * - printout the content of tables
+ * - pretify output, report below, test with fuzzer to ensure it's correct */
+
+/* Default FSE distribution table for Literal Lengths */
+static const ZSTD_seqSymbol LL_defaultDTable[(1<<LL_DEFAULTNORMLOG)+1] = {
+     {  1,  1,  1, LL_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
+     /* nextState, nbAddBits, nbBits, baseVal */
+     {  0,  0,  4,    0},  { 16,  0,  4,    0},
+     { 32,  0,  5,    1},  {  0,  0,  5,    3},
+     {  0,  0,  5,    4},  {  0,  0,  5,    6},
+     {  0,  0,  5,    7},  {  0,  0,  5,    9},
+     {  0,  0,  5,   10},  {  0,  0,  5,   12},
+     {  0,  0,  6,   14},  {  0,  1,  5,   16},
+     {  0,  1,  5,   20},  {  0,  1,  5,   22},
+     {  0,  2,  5,   28},  {  0,  3,  5,   32},
+     {  0,  4,  5,   48},  { 32,  6,  5,   64},
+     {  0,  7,  5,  128},  {  0,  8,  6,  256},
+     {  0, 10,  6, 1024},  {  0, 12,  6, 4096},
+     { 32,  0,  4,    0},  {  0,  0,  4,    1},
+     {  0,  0,  5,    2},  { 32,  0,  5,    4},
+     {  0,  0,  5,    5},  { 32,  0,  5,    7},
+     {  0,  0,  5,    8},  { 32,  0,  5,   10},
+     {  0,  0,  5,   11},  {  0,  0,  6,   13},
+     { 32,  1,  5,   16},  {  0,  1,  5,   18},
+     { 32,  1,  5,   22},  {  0,  2,  5,   24},
+     { 32,  3,  5,   32},  {  0,  3,  5,   40},
+     {  0,  6,  4,   64},  { 16,  6,  4,   64},
+     { 32,  7,  5,  128},  {  0,  9,  6,  512},
+     {  0, 11,  6, 2048},  { 48,  0,  4,    0},
+     { 16,  0,  4,    1},  { 32,  0,  5,    2},
+     { 32,  0,  5,    3},  { 32,  0,  5,    5},
+     { 32,  0,  5,    6},  { 32,  0,  5,    8},
+     { 32,  0,  5,    9},  { 32,  0,  5,   11},
+     { 32,  0,  5,   12},  {  0,  0,  6,   15},
+     { 32,  1,  5,   18},  { 32,  1,  5,   20},
+     { 32,  2,  5,   24},  { 32,  2,  5,   28},
+     { 32,  3,  5,   40},  { 32,  4,  5,   48},
+     {  0, 16,  6,65536},  {  0, 15,  6,32768},
+     {  0, 14,  6,16384},  {  0, 13,  6, 8192},
+};   /* LL_defaultDTable */
+
+/* Default FSE distribution table for Offset Codes */
+static const ZSTD_seqSymbol OF_defaultDTable[(1<<OF_DEFAULTNORMLOG)+1] = {
+    {  1,  1,  1, OF_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
+    /* nextState, nbAddBits, nbBits, baseVal */
+    {  0,  0,  5,    0},     {  0,  6,  4,   61},
+    {  0,  9,  5,  509},     {  0, 15,  5,32765},
+    {  0, 21,  5,2097149},   {  0,  3,  5,    5},
+    {  0,  7,  4,  125},     {  0, 12,  5, 4093},
+    {  0, 18,  5,262141},    {  0, 23,  5,8388605},
+    {  0,  5,  5,   29},     {  0,  8,  4,  253},
+    {  0, 14,  5,16381},     {  0, 20,  5,1048573},
+    {  0,  2,  5,    1},     { 16,  7,  4,  125},
+    {  0, 11,  5, 2045},     {  0, 17,  5,131069},
+    {  0, 22,  5,4194301},   {  0,  4,  5,   13},
+    { 16,  8,  4,  253},     {  0, 13,  5, 8189},
+    {  0, 19,  5,524285},    {  0,  1,  5,    1},
+    { 16,  6,  4,   61},     {  0, 10,  5, 1021},
+    {  0, 16,  5,65533},     {  0, 28,  5,268435453},
+    {  0, 27,  5,134217725}, {  0, 26,  5,67108861},
+    {  0, 25,  5,33554429},  {  0, 24,  5,16777213},
+};   /* OF_defaultDTable */
+
+
+/* Default FSE distribution table for Match Lengths */
+static const ZSTD_seqSymbol ML_defaultDTable[(1<<ML_DEFAULTNORMLOG)+1] = {
+    {  1,  1,  1, ML_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
+    /* nextState, nbAddBits, nbBits, baseVal */
+    {  0,  0,  6,    3},  {  0,  0,  4,    4},
+    { 32,  0,  5,    5},  {  0,  0,  5,    6},
+    {  0,  0,  5,    8},  {  0,  0,  5,    9},
+    {  0,  0,  5,   11},  {  0,  0,  6,   13},
+    {  0,  0,  6,   16},  {  0,  0,  6,   19},
+    {  0,  0,  6,   22},  {  0,  0,  6,   25},
+    {  0,  0,  6,   28},  {  0,  0,  6,   31},
+    {  0,  0,  6,   34},  {  0,  1,  6,   37},
+    {  0,  1,  6,   41},  {  0,  2,  6,   47},
+    {  0,  3,  6,   59},  {  0,  4,  6,   83},
+    {  0,  7,  6,  131},  {  0,  9,  6,  515},
+    { 16,  0,  4,    4},  {  0,  0,  4,    5},
+    { 32,  0,  5,    6},  {  0,  0,  5,    7},
+    { 32,  0,  5,    9},  {  0,  0,  5,   10},
+    {  0,  0,  6,   12},  {  0,  0,  6,   15},
+    {  0,  0,  6,   18},  {  0,  0,  6,   21},
+    {  0,  0,  6,   24},  {  0,  0,  6,   27},
+    {  0,  0,  6,   30},  {  0,  0,  6,   33},
+    {  0,  1,  6,   35},  {  0,  1,  6,   39},
+    {  0,  2,  6,   43},  {  0,  3,  6,   51},
+    {  0,  4,  6,   67},  {  0,  5,  6,   99},
+    {  0,  8,  6,  259},  { 32,  0,  4,    4},
+    { 48,  0,  4,    4},  { 16,  0,  4,    5},
+    { 32,  0,  5,    7},  { 32,  0,  5,    8},
+    { 32,  0,  5,   10},  { 32,  0,  5,   11},
+    {  0,  0,  6,   14},  {  0,  0,  6,   17},
+    {  0,  0,  6,   20},  {  0,  0,  6,   23},
+    {  0,  0,  6,   26},  {  0,  0,  6,   29},
+    {  0,  0,  6,   32},  {  0, 16,  6,65539},
+    {  0, 15,  6,32771},  {  0, 14,  6,16387},
+    {  0, 13,  6, 8195},  {  0, 12,  6, 4099},
+    {  0, 11,  6, 2051},  {  0, 10,  6, 1027},
+};   /* ML_defaultDTable */
+
+
+static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U32 nbAddBits)
+{
+    void* ptr = dt;
+    ZSTD_seqSymbol_header* const DTableH = (ZSTD_seqSymbol_header*)ptr;
+    ZSTD_seqSymbol* const cell = dt + 1;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->nbBits = 0;
+    cell->nextState = 0;
+    assert(nbAddBits < 255);
+    cell->nbAdditionalBits = (BYTE)nbAddBits;
+    cell->baseValue = baseValue;
+}
+
+
+/* ZSTD_buildFSETable() :
+ * generate FSE decoding table for one symbol (ll, ml or off) */
+static void
+ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
+    const short* normalizedCounter, unsigned maxSymbolValue,
+    const U32* baseValue, const U32* nbAdditionalBits,
+    unsigned tableLog)
+{
+    ZSTD_seqSymbol* const tableDecode = dt+1;
+    U16 symbolNext[MaxSeq+1];
+
+    U32 const maxSV1 = maxSymbolValue + 1;
+    U32 const tableSize = 1 << tableLog;
+    U32 highThreshold = tableSize-1;
+
+    /* Sanity Checks */
+    assert(maxSymbolValue <= MaxSeq);
+    assert(tableLog <= MaxFSELog);
+
+    /* Init, lay down lowprob symbols */
+    {   ZSTD_seqSymbol_header DTableH;
+        DTableH.tableLog = tableLog;
+        DTableH.fastMode = 1;
+        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
+            U32 s;
+            for (s=0; s<maxSV1; s++) {
+                if (normalizedCounter[s]==-1) {
+                    tableDecode[highThreshold--].baseValue = s;
+                    symbolNext[s] = 1;
+                } else {
+                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+                    symbolNext[s] = normalizedCounter[s];
+        }   }   }
+        memcpy(dt, &DTableH, sizeof(DTableH));
+    }
+
+    /* Spread symbols */
+    {   U32 const tableMask = tableSize-1;
+        U32 const step = FSE_TABLESTEP(tableSize);
+        U32 s, position = 0;
+        for (s=0; s<maxSV1; s++) {
+            int i;
+            for (i=0; i<normalizedCounter[s]; i++) {
+                tableDecode[position].baseValue = s;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }   }
+        assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+    }
+
+    /* Build Decoding table */
+    {   U32 u;
+        for (u=0; u<tableSize; u++) {
+            U32 const symbol = tableDecode[u].baseValue;
+            U32 const nextState = symbolNext[symbol]++;
+            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+            assert(nbAdditionalBits[symbol] < 255);
+            tableDecode[u].nbAdditionalBits = (BYTE)nbAdditionalBits[symbol];
+            tableDecode[u].baseValue = baseValue[symbol];
+    }   }
+}
+
+
+/*! ZSTD_buildSeqTable() :
+ * @return : nb bytes read from src,
+ *           or an error code if it fails */
+static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymbol** DTablePtr,
+                                 symbolEncodingType_e type, U32 max, U32 maxLog,
+                                 const void* src, size_t srcSize,
+                                 const U32* baseValue, const U32* nbAdditionalBits,
+                                 const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable)
+{
+    switch(type)
+    {
+    case set_rle :
+        if (!srcSize) return ERROR(srcSize_wrong);
+        if ( (*(const BYTE*)src) > max) return ERROR(corruption_detected);
+        {   U32 const symbol = *(const BYTE*)src;
+            U32 const baseline = baseValue[symbol];
+            U32 const nbBits = nbAdditionalBits[symbol];
+            ZSTD_buildSeqTable_rle(DTableSpace, baseline, nbBits);
+        }
+        *DTablePtr = DTableSpace;
+        return 1;
+    case set_basic :
+        *DTablePtr = defaultTable;
+        return 0;
+    case set_repeat:
+        if (!flagRepeatTable) return ERROR(corruption_detected);
+        return 0;
+    case set_compressed :
+        {   U32 tableLog;
+            S16 norm[MaxSeq+1];
+            size_t const headerSize = FSE_readNCount(norm, &max, &tableLog, src, srcSize);
+            if (FSE_isError(headerSize)) return ERROR(corruption_detected);
+            if (tableLog > maxLog) return ERROR(corruption_detected);
+            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog);
+            *DTablePtr = DTableSpace;
+            return headerSize;
+        }
+    default :   /* impossible */
+        assert(0);
+        return ERROR(GENERIC);
+    }
+}
+
+static const U32 LL_base[MaxLL+1] = {
+                 0,    1,    2,     3,     4,     5,     6,      7,
+                 8,    9,   10,    11,    12,    13,    14,     15,
+                16,   18,   20,    22,    24,    28,    32,     40,
+                48,   64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
+                0x2000, 0x4000, 0x8000, 0x10000 };
+
+static const U32 OF_base[MaxOff+1] = {
+                 0,        1,       1,       5,     0xD,     0x1D,     0x3D,     0x7D,
+                 0xFD,   0x1FD,   0x3FD,   0x7FD,   0xFFD,   0x1FFD,   0x3FFD,   0x7FFD,
+                 0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
+                 0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD, 0x1FFFFFFD, 0x3FFFFFFD, 0x7FFFFFFD };
+
+static const U32 OF_bits[MaxOff+1] = {
+                     0,  1,  2,  3,  4,  5,  6,  7,
+                     8,  9, 10, 11, 12, 13, 14, 15,
+                    16, 17, 18, 19, 20, 21, 22, 23,
+                    24, 25, 26, 27, 28, 29, 30, 31 };
+
+static const U32 ML_base[MaxML+1] = {
+                     3,  4,  5,    6,     7,     8,     9,    10,
+                    11, 12, 13,   14,    15,    16,    17,    18,
+                    19, 20, 21,   22,    23,    24,    25,    26,
+                    27, 28, 29,   30,    31,    32,    33,    34,
+                    35, 37, 39,   41,    43,    47,    51,    59,
+                    67, 83, 99, 0x83, 0x103, 0x203, 0x403, 0x803,
+                    0x1003, 0x2003, 0x4003, 0x8003, 0x10003 };
+
+
+size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
+                             const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip = istart;
+    DEBUGLOG(5, "ZSTD_decodeSeqHeaders");
+
+    /* check */
+    if (srcSize < MIN_SEQUENCES_SIZE) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    {   int nbSeq = *ip++;
+        if (!nbSeq) { *nbSeqPtr=0; return 1; }
+        if (nbSeq > 0x7F) {
+            if (nbSeq == 0xFF) {
+                if (ip+2 > iend) return ERROR(srcSize_wrong);
+                nbSeq = MEM_readLE16(ip) + LONGNBSEQ, ip+=2;
+            } else {
+                if (ip >= iend) return ERROR(srcSize_wrong);
+                nbSeq = ((nbSeq-0x80)<<8) + *ip++;
+            }
+        }
+        *nbSeqPtr = nbSeq;
+    }
+
+    /* FSE table descriptors */
+    if (ip+4 > iend) return ERROR(srcSize_wrong); /* minimum possible size */
+    {   symbolEncodingType_e const LLtype = (symbolEncodingType_e)(*ip >> 6);
+        symbolEncodingType_e const OFtype = (symbolEncodingType_e)((*ip >> 4) & 3);
+        symbolEncodingType_e const MLtype = (symbolEncodingType_e)((*ip >> 2) & 3);
+        ip++;
+
+        /* Build DTables */
+        {   size_t const llhSize = ZSTD_buildSeqTable(dctx->entropy.LLTable, &dctx->LLTptr,
+                                                      LLtype, MaxLL, LLFSELog,
+                                                      ip, iend-ip,
+                                                      LL_base, LL_bits,
+                                                      LL_defaultDTable, dctx->fseEntropy);
+            if (ZSTD_isError(llhSize)) return ERROR(corruption_detected);
+            ip += llhSize;
+        }
+
+        {   size_t const ofhSize = ZSTD_buildSeqTable(dctx->entropy.OFTable, &dctx->OFTptr,
+                                                      OFtype, MaxOff, OffFSELog,
+                                                      ip, iend-ip,
+                                                      OF_base, OF_bits,
+                                                      OF_defaultDTable, dctx->fseEntropy);
+            if (ZSTD_isError(ofhSize)) return ERROR(corruption_detected);
+            ip += ofhSize;
+        }
+
+        {   size_t const mlhSize = ZSTD_buildSeqTable(dctx->entropy.MLTable, &dctx->MLTptr,
+                                                      MLtype, MaxML, MLFSELog,
+                                                      ip, iend-ip,
+                                                      ML_base, ML_bits,
+                                                      ML_defaultDTable, dctx->fseEntropy);
+            if (ZSTD_isError(mlhSize)) return ERROR(corruption_detected);
+            ip += mlhSize;
+        }
+    }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t matchLength;
+    size_t offset;
+    const BYTE* match;
+} seq_t;
+
+typedef struct {
+    size_t state;
+    const ZSTD_seqSymbol* table;
+} ZSTD_fseState;
+
+typedef struct {
+    BIT_DStream_t DStream;
+    ZSTD_fseState stateLL;
+    ZSTD_fseState stateOffb;
+    ZSTD_fseState stateML;
+    size_t prevOffset[ZSTD_REP_NUM];
+    const BYTE* prefixStart;
+    const BYTE* dictEnd;
+    size_t pos;
+} seqState_t;
+
+
+FORCE_NOINLINE
+size_t ZSTD_execSequenceLast7(BYTE* op,
+                              BYTE* const oend, seq_t sequence,
+                              const BYTE** litPtr, const BYTE* const litLimit,
+                              const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if (oMatchEnd>oend) return ERROR(dstSize_tooSmall); /* last match must start at a minimum distance of WILDCOPY_OVERLENGTH from oend */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+    if (oLitEnd <= oend_w) return ERROR(GENERIC);   /* Precondition */
+
+    /* copy literals */
+    if (op < oend_w) {
+        ZSTD_wildcopy(op, *litPtr, oend_w - op);
+        *litPtr += oend_w - op;
+        op = oend_w;
+    }
+    while (op < oLitEnd) *op++ = *(*litPtr)++;
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base)) {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - vBase)) return ERROR(corruption_detected);
+        match = dictEnd - (base-match);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+    }   }
+    while (op < oMatchEnd) *op++ = *match++;
+    return sequenceLength;
+}
+
+
+HINT_INLINE
+size_t ZSTD_execSequence(BYTE* op,
+                         BYTE* const oend, seq_t sequence,
+                         const BYTE** litPtr, const BYTE* const litLimit,
+                         const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if (oMatchEnd>oend) return ERROR(dstSize_tooSmall); /* last match must start at a minimum distance of WILDCOPY_OVERLENGTH from oend */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+    if (oLitEnd>oend_w) return ZSTD_execSequenceLast7(op, oend, sequence, litPtr, litLimit, base, vBase, dictEnd);
+
+    /* copy Literals */
+    ZSTD_copy8(op, *litPtr);
+    if (sequence.litLength > 8)
+        ZSTD_wildcopy(op+8, (*litPtr)+8, sequence.litLength - 8);   /* note : since oLitEnd <= oend-WILDCOPY_OVERLENGTH, no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = iLitEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base)) {
+        /* offset beyond prefix -> go into extDict */
+        if (sequence.offset > (size_t)(oLitEnd - vBase))
+            return ERROR(corruption_detected);
+        match = dictEnd + (match - base);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+            if (op > oend_w || sequence.matchLength < MINMATCH) {
+              U32 i;
+              for (i = 0; i < sequence.matchLength; ++i) op[i] = match[i];
+              return sequenceLength;
+            }
+    }   }
+    /* Requirement: op <= oend_w && sequence.matchLength >= MINMATCH */
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+        static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* subtracted */
+        int const sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTD_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTD_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH)) {
+        if (op < oend_w) {
+            ZSTD_wildcopy(op, match, oend_w - op);
+            match += oend_w - op;
+            op = oend_w;
+        }
+        while (op < oMatchEnd) *op++ = *match++;
+    } else {
+        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+
+HINT_INLINE
+size_t ZSTD_execSequenceLong(BYTE* op,
+                             BYTE* const oend, seq_t sequence,
+                             const BYTE** litPtr, const BYTE* const litLimit,
+                             const BYTE* const prefixStart, const BYTE* const dictStart, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = sequence.match;
+
+    /* check */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall); /* last match must start at a minimum distance of WILDCOPY_OVERLENGTH from oend */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+    if (oLitEnd > oend_w) return ZSTD_execSequenceLast7(op, oend, sequence, litPtr, litLimit, prefixStart, dictStart, dictEnd);
+
+    /* copy Literals */
+    ZSTD_copy8(op, *litPtr);  /* note : op <= oLitEnd <= oend_w == oend - 8 */
+    if (sequence.litLength > 8)
+        ZSTD_wildcopy(op+8, (*litPtr)+8, sequence.litLength - 8);   /* note : since oLitEnd <= oend-WILDCOPY_OVERLENGTH, no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = iLitEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - dictStart)) return ERROR(corruption_detected);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = prefixStart;
+            if (op > oend_w || sequence.matchLength < MINMATCH) {
+              U32 i;
+              for (i = 0; i < sequence.matchLength; ++i) op[i] = match[i];
+              return sequenceLength;
+            }
+    }   }
+    assert(op <= oend_w);
+    assert(sequence.matchLength >= MINMATCH);
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+        static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* subtracted */
+        int const sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTD_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTD_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH)) {
+        if (op < oend_w) {
+            ZSTD_wildcopy(op, match, oend_w - op);
+            match += oend_w - op;
+            op = oend_w;
+        }
+        while (op < oMatchEnd) *op++ = *match++;
+    } else {
+        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+static void
+ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqSymbol* dt)
+{
+    const void* ptr = dt;
+    const ZSTD_seqSymbol_header* const DTableH = (const ZSTD_seqSymbol_header*)ptr;
+    DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
+    DEBUGLOG(6, "ZSTD_initFseState : val=%u using %u bits",
+                (U32)DStatePtr->state, DTableH->tableLog);
+    BIT_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+FORCE_INLINE_TEMPLATE void
+ZSTD_updateFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD)
+{
+    ZSTD_seqSymbol const DInfo = DStatePtr->table[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    size_t const lowBits = BIT_readBits(bitD, nbBits);
+    DStatePtr->state = DInfo.nextState + lowBits;
+}
+
+/* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
+ * offset bits. But we can only read at most (STREAM_ACCUMULATOR_MIN_32 - 1)
+ * bits before reloading. This value is the maximum number of bytes we read
+ * after reloading when we are decoding long offets.
+ */
+#define LONG_OFFSETS_MAX_EXTRA_BITS_32                       \
+    (ZSTD_WINDOWLOG_MAX_32 > STREAM_ACCUMULATOR_MIN_32       \
+        ? ZSTD_WINDOWLOG_MAX_32 - STREAM_ACCUMULATOR_MIN_32  \
+        : 0)
+
+typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
+
+FORCE_INLINE_TEMPLATE seq_t
+ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
+{
+    seq_t seq;
+    U32 const llBits = seqState->stateLL.table[seqState->stateLL.state].nbAdditionalBits;
+    U32 const mlBits = seqState->stateML.table[seqState->stateML.state].nbAdditionalBits;
+    U32 const ofBits = seqState->stateOffb.table[seqState->stateOffb.state].nbAdditionalBits;
+    U32 const totalBits = llBits+mlBits+ofBits;
+    U32 const llBase = seqState->stateLL.table[seqState->stateLL.state].baseValue;
+    U32 const mlBase = seqState->stateML.table[seqState->stateML.state].baseValue;
+    U32 const ofBase = seqState->stateOffb.table[seqState->stateOffb.state].baseValue;
+
+    /* sequence */
+    {   size_t offset;
+        if (!ofBits)
+            offset = 0;
+        else {
+            ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
+            ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
+            assert(ofBits <= MaxOff);
+            if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
+                U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
+                offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
+                BIT_reloadDStream(&seqState->DStream);
+                if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
+                assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
+            } else {
+                offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
+                if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
+            }
+        }
+
+        if (ofBits <= 1) {
+            offset += (llBase==0);
+            if (offset) {
+                size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
+                temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
+                if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset = temp;
+            } else {  /* offset == 0 */
+                offset = seqState->prevOffset[0];
+            }
+        } else {
+            seqState->prevOffset[2] = seqState->prevOffset[1];
+            seqState->prevOffset[1] = seqState->prevOffset[0];
+            seqState->prevOffset[0] = offset;
+        }
+        seq.offset = offset;
+    }
+
+    seq.matchLength = mlBase
+                    + ((mlBits>0) ? BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/) : 0);  /* <=  16 bits */
+    if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
+        BIT_reloadDStream(&seqState->DStream);
+    if (MEM_64bits() && (totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
+        BIT_reloadDStream(&seqState->DStream);
+    /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
+    ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
+
+    seq.litLength = llBase
+                  + ((llBits>0) ? BIT_readBitsFast(&seqState->DStream, llBits/*>0*/) : 0);    /* <=  16 bits */
+    if (MEM_32bits())
+        BIT_reloadDStream(&seqState->DStream);
+
+    DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
+                (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
+
+    /* ANS state update */
+    ZSTD_updateFseState(&seqState->stateLL, &seqState->DStream);    /* <=  9 bits */
+    ZSTD_updateFseState(&seqState->stateML, &seqState->DStream);    /* <=  9 bits */
+    if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
+    ZSTD_updateFseState(&seqState->stateOffb, &seqState->DStream);  /* <=  8 bits */
+
+    return seq;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize, int nbSeq,
+                         const ZSTD_longOffset_e isLongOffset)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + maxDstSize;
+    BYTE* op = ostart;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    const BYTE* const base = (const BYTE*) (dctx->base);
+    const BYTE* const vBase = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+    DEBUGLOG(5, "ZSTD_decompressSequences");
+
+    /* Regen sequences */
+    if (nbSeq) {
+        seqState_t seqState;
+        dctx->fseEntropy = 1;
+        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
+        CHECK_E(BIT_initDStream(&seqState.DStream, ip, iend-ip), corruption_detected);
+        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
+        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
+        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
+
+        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && nbSeq ; ) {
+            nbSeq--;
+            {   seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+                size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, base, vBase, dictEnd);
+                DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+                op += oneSeqSize;
+        }   }
+
+        /* check if reached exact end */
+        DEBUGLOG(5, "ZSTD_decompressSequences: after decode loop, remaining nbSeq : %i", nbSeq);
+        if (nbSeq) return ERROR(corruption_detected);
+        /* save reps for next block */
+        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
+    }
+
+    /* last literal segment */
+    {   size_t const lastLLSize = litEnd - litPtr;
+        if (lastLLSize > (size_t)(oend-op)) return ERROR(dstSize_tooSmall);
+        memcpy(op, litPtr, lastLLSize);
+        op += lastLLSize;
+    }
+
+    return op-ostart;
+}
+
+static size_t
+ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                           const void* seqStart, size_t seqSize, int nbSeq,
+                           const ZSTD_longOffset_e isLongOffset)
+{
+    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+
+
+FORCE_INLINE_TEMPLATE seq_t
+ZSTD_decodeSequenceLong(seqState_t* seqState, ZSTD_longOffset_e const longOffsets)
+{
+    seq_t seq;
+    U32 const llBits = seqState->stateLL.table[seqState->stateLL.state].nbAdditionalBits;
+    U32 const mlBits = seqState->stateML.table[seqState->stateML.state].nbAdditionalBits;
+    U32 const ofBits = seqState->stateOffb.table[seqState->stateOffb.state].nbAdditionalBits;
+    U32 const totalBits = llBits+mlBits+ofBits;
+    U32 const llBase = seqState->stateLL.table[seqState->stateLL.state].baseValue;
+    U32 const mlBase = seqState->stateML.table[seqState->stateML.state].baseValue;
+    U32 const ofBase = seqState->stateOffb.table[seqState->stateOffb.state].baseValue;
+
+    /* sequence */
+    {   size_t offset;
+        if (!ofBits)
+            offset = 0;
+        else {
+            ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
+            ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
+            assert(ofBits <= MaxOff);
+            if (MEM_32bits() && longOffsets) {
+                U32 const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN_32-1);
+                offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
+                if (MEM_32bits() || extraBits) BIT_reloadDStream(&seqState->DStream);
+                if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
+            } else {
+                offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
+                if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
+            }
+        }
+
+        if (ofBits <= 1) {
+            offset += (llBase==0);
+            if (offset) {
+                size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
+                temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
+                if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset = temp;
+            } else {
+                offset = seqState->prevOffset[0];
+            }
+        } else {
+            seqState->prevOffset[2] = seqState->prevOffset[1];
+            seqState->prevOffset[1] = seqState->prevOffset[0];
+            seqState->prevOffset[0] = offset;
+        }
+        seq.offset = offset;
+    }
+
+    seq.matchLength = mlBase + ((mlBits>0) ? BIT_readBitsFast(&seqState->DStream, mlBits) : 0);  /* <=  16 bits */
+    if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
+        BIT_reloadDStream(&seqState->DStream);
+    if (MEM_64bits() && (totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
+        BIT_reloadDStream(&seqState->DStream);
+    /* Verify that there is enough bits to read the rest of the data in 64-bit mode. */
+    ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
+
+    seq.litLength = llBase + ((llBits>0) ? BIT_readBitsFast(&seqState->DStream, llBits) : 0);    /* <=  16 bits */
+    if (MEM_32bits())
+        BIT_reloadDStream(&seqState->DStream);
+
+    {   size_t const pos = seqState->pos + seq.litLength;
+        const BYTE* const matchBase = (seq.offset > pos) ? seqState->dictEnd : seqState->prefixStart;
+        seq.match = matchBase + pos - seq.offset;  /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
+                                                    * No consequence though : no memory access will occur, overly large offset will be detected in ZSTD_execSequenceLong() */
+        seqState->pos = pos + seq.matchLength;
+    }
+
+    /* ANS state update */
+    ZSTD_updateFseState(&seqState->stateLL, &seqState->DStream);    /* <=  9 bits */
+    ZSTD_updateFseState(&seqState->stateML, &seqState->DStream);    /* <=  9 bits */
+    if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
+    ZSTD_updateFseState(&seqState->stateOffb, &seqState->DStream);  /* <=  8 bits */
+
+    return seq;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_decompressSequencesLong_body(
+                               ZSTD_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize, int nbSeq,
+                         const ZSTD_longOffset_e isLongOffset)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + maxDstSize;
+    BYTE* op = ostart;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    const BYTE* const prefixStart = (const BYTE*) (dctx->base);
+    const BYTE* const dictStart = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+
+    /* Regen sequences */
+    if (nbSeq) {
+#define STORED_SEQS 4
+#define STOSEQ_MASK (STORED_SEQS-1)
+#define ADVANCED_SEQS 4
+        seq_t sequences[STORED_SEQS];
+        int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
+        seqState_t seqState;
+        int seqNb;
+        dctx->fseEntropy = 1;
+        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
+        seqState.prefixStart = prefixStart;
+        seqState.pos = (size_t)(op-prefixStart);
+        seqState.dictEnd = dictEnd;
+        CHECK_E(BIT_initDStream(&seqState.DStream, ip, iend-ip), corruption_detected);
+        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
+        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
+        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
+
+        /* prepare in advance */
+        for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
+            sequences[seqNb] = ZSTD_decodeSequenceLong(&seqState, isLongOffset);
+        }
+        if (seqNb<seqAdvance) return ERROR(corruption_detected);
+
+        /* decode and decompress */
+        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb<nbSeq) ; seqNb++) {
+            seq_t const sequence = ZSTD_decodeSequenceLong(&seqState, isLongOffset);
+            size_t const oneSeqSize = ZSTD_execSequenceLong(op, oend, sequences[(seqNb-ADVANCED_SEQS) & STOSEQ_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+            PREFETCH(sequence.match);  /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+            sequences[seqNb&STOSEQ_MASK] = sequence;
+            op += oneSeqSize;
+        }
+        if (seqNb<nbSeq) return ERROR(corruption_detected);
+
+        /* finish queue */
+        seqNb -= seqAdvance;
+        for ( ; seqNb<nbSeq ; seqNb++) {
+            size_t const oneSeqSize = ZSTD_execSequenceLong(op, oend, sequences[seqNb&STOSEQ_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* save reps for next block */
+        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
+#undef STORED_SEQS
+#undef STOSEQ_MASK
+#undef ADVANCED_SEQS
+    }
+
+    /* last literal segment */
+    {   size_t const lastLLSize = litEnd - litPtr;
+        if (lastLLSize > (size_t)(oend-op)) return ERROR(dstSize_tooSmall);
+        memcpy(op, litPtr, lastLLSize);
+        op += lastLLSize;
+    }
+
+    return op-ostart;
+}
+
+static size_t
+ZSTD_decompressSequencesLong_default(ZSTD_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                           const void* seqStart, size_t seqSize, int nbSeq,
+                           const ZSTD_longOffset_e isLongOffset)
+{
+    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+
+
+#if DYNAMIC_BMI2
+
+static TARGET_ATTRIBUTE("bmi2") size_t
+ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                           const void* seqStart, size_t seqSize, int nbSeq,
+                           const ZSTD_longOffset_e isLongOffset)
+{
+    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+static TARGET_ATTRIBUTE("bmi2") size_t
+ZSTD_decompressSequencesLong_bmi2(ZSTD_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                           const void* seqStart, size_t seqSize, int nbSeq,
+                           const ZSTD_longOffset_e isLongOffset)
+{
+    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+#endif
+
+typedef size_t (*ZSTD_decompressSequences_t)(
+    ZSTD_DCtx *dctx, void *dst, size_t maxDstSize,
+    const void *seqStart, size_t seqSize, int nbSeq,
+    const ZSTD_longOffset_e isLongOffset);
+
+static size_t ZSTD_decompressSequences(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
+                                const void* seqStart, size_t seqSize, int nbSeq,
+                                const ZSTD_longOffset_e isLongOffset)
+{
+    DEBUGLOG(5, "ZSTD_decompressSequences");
+#if DYNAMIC_BMI2
+    if (dctx->bmi2) {
+        return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+    }
+#endif
+  return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+static size_t ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
+                                void* dst, size_t maxDstSize,
+                                const void* seqStart, size_t seqSize, int nbSeq,
+                                const ZSTD_longOffset_e isLongOffset)
+{
+    DEBUGLOG(5, "ZSTD_decompressSequencesLong");
+#if DYNAMIC_BMI2
+    if (dctx->bmi2) {
+        return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+    }
+#endif
+  return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
+}
+
+/* ZSTD_getLongOffsetsShare() :
+ * condition : offTable must be valid
+ * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
+ *           compared to maximum possible of (1<<OffFSELog) */
+static unsigned
+ZSTD_getLongOffsetsShare(const ZSTD_seqSymbol* offTable)
+{
+    const void* ptr = offTable;
+    U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
+    const ZSTD_seqSymbol* table = offTable + 1;
+    U32 const max = 1 << tableLog;
+    U32 u, total = 0;
+    DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
+
+    assert(max <= (1 << OffFSELog));  /* max not too large */
+    for (u=0; u<max; u++) {
+        if (table[u].nbAdditionalBits > 22) total += 1;
+    }
+
+    assert(tableLog <= OffFSELog);
+    total <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
+
+    return total;
+}
+
+
+static size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize, const int frame)
+{   /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+    /* isLongOffset must be true if there are long offsets.
+     * Offsets are long if they are larger than 2^STREAM_ACCUMULATOR_MIN.
+     * We don't expect that to be the case in 64-bit mode.
+     * In block mode, window size is not known, so we have to be conservative. (note: but it could be evaluated from current-lowLimit)
+     */
+    ZSTD_longOffset_e const isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (!frame || dctx->fParams.windowSize > (1ULL << STREAM_ACCUMULATOR_MIN)));
+    DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
+
+    if (srcSize >= ZSTD_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);
+
+    /* Decode literals section */
+    {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize);
+        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize);
+        if (ZSTD_isError(litCSize)) return litCSize;
+        ip += litCSize;
+        srcSize -= litCSize;
+    }
+
+    /* Build Decoding Tables */
+    {   int nbSeq;
+        size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
+        if (ZSTD_isError(seqHSize)) return seqHSize;
+        ip += seqHSize;
+        srcSize -= seqHSize;
+
+        if ( (!frame || dctx->fParams.windowSize > (1<<24))
+          && (nbSeq>0) ) {  /* could probably use a larger nbSeq limit */
+            U32 const shareLongOffsets = ZSTD_getLongOffsetsShare(dctx->OFTptr);
+            U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
+            if (shareLongOffsets >= minShare)
+                return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
+        }
+
+        return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
+    }
+}
+
+
+static void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst)
+{
+    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+        dctx->dictEnd = dctx->previousDstEnd;
+        dctx->vBase = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+        dctx->base = dst;
+        dctx->previousDstEnd = dst;
+    }
+}
+
+size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{
+    size_t dSize;
+    ZSTD_checkContinuity(dctx, dst);
+    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0);
+    dctx->previousDstEnd = (char*)dst + dSize;
+    return dSize;
+}
+
+
+/** ZSTD_insertBlock() :
+    insert `src` block into `dctx` history. Useful to track uncompressed blocks. */
+ZSTDLIB_API size_t ZSTD_insertBlock(ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize)
+{
+    ZSTD_checkContinuity(dctx, blockStart);
+    dctx->previousDstEnd = (const char*)blockStart + blockSize;
+    return blockSize;
+}
+
+
+static size_t ZSTD_generateNxBytes(void* dst, size_t dstCapacity, BYTE byte, size_t length)
+{
+    if (length > dstCapacity) return ERROR(dstSize_tooSmall);
+    memset(dst, byte, length);
+    return length;
+}
+
+/** ZSTD_findFrameCompressedSize() :
+ *  compatible with legacy mode
+ *  `src` must point to the start of a ZSTD frame, ZSTD legacy frame, or skippable frame
+ *  `srcSize` must be at least as large as the frame contained
+ *  @return : the compressed size of the frame starting at `src` */
+size_t ZSTD_findFrameCompressedSize(const void *src, size_t srcSize)
+{
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
+    if (ZSTD_isLegacy(src, srcSize))
+        return ZSTD_findFrameCompressedSizeLegacy(src, srcSize);
+#endif
+    if ( (srcSize >= ZSTD_skippableHeaderSize)
+      && (MEM_readLE32(src) & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START ) {
+        return ZSTD_skippableHeaderSize + MEM_readLE32((const BYTE*)src + ZSTD_frameIdSize);
+    } else {
+        const BYTE* ip = (const BYTE*)src;
+        const BYTE* const ipstart = ip;
+        size_t remainingSize = srcSize;
+        ZSTD_frameHeader zfh;
+
+        /* Extract Frame Header */
+        {   size_t const ret = ZSTD_getFrameHeader(&zfh, src, srcSize);
+            if (ZSTD_isError(ret)) return ret;
+            if (ret > 0) return ERROR(srcSize_wrong);
+        }
+
+        ip += zfh.headerSize;
+        remainingSize -= zfh.headerSize;
+
+        /* Loop on each block */
+        while (1) {
+            blockProperties_t blockProperties;
+            size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
+            if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+            if (ZSTD_blockHeaderSize + cBlockSize > remainingSize)
+                return ERROR(srcSize_wrong);
+
+            ip += ZSTD_blockHeaderSize + cBlockSize;
+            remainingSize -= ZSTD_blockHeaderSize + cBlockSize;
+
+            if (blockProperties.lastBlock) break;
+        }
+
+        if (zfh.checksumFlag) {   /* Final frame content checksum */
+            if (remainingSize < 4) return ERROR(srcSize_wrong);
+            ip += 4;
+            remainingSize -= 4;
+        }
+
+        return ip - ipstart;
+    }
+}
+
+/*! ZSTD_decompressFrame() :
+*   @dctx must be properly initialized */
+static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
+                                   void* dst, size_t dstCapacity,
+                             const void** srcPtr, size_t *srcSizePtr)
+{
+    const BYTE* ip = (const BYTE*)(*srcPtr);
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + dstCapacity;
+    BYTE* op = ostart;
+    size_t remainingSize = *srcSizePtr;
+
+    /* check */
+    if (remainingSize < ZSTD_frameHeaderSize_min+ZSTD_blockHeaderSize)
+        return ERROR(srcSize_wrong);
+
+    /* Frame Header */
+    {   size_t const frameHeaderSize = ZSTD_frameHeaderSize(ip, ZSTD_frameHeaderSize_prefix);
+        if (ZSTD_isError(frameHeaderSize)) return frameHeaderSize;
+        if (remainingSize < frameHeaderSize+ZSTD_blockHeaderSize)
+            return ERROR(srcSize_wrong);
+        CHECK_F( ZSTD_decodeFrameHeader(dctx, ip, frameHeaderSize) );
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1) {
+        size_t decodedSize;
+        blockProperties_t blockProperties;
+        size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTD_decompressBlock_internal(dctx, op, oend-op, ip, cBlockSize, /* frame */ 1);
+            break;
+        case bt_raw :
+            decodedSize = ZSTD_copyRawBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            decodedSize = ZSTD_generateNxBytes(op, oend-op, *ip, blockProperties.origSize);
+            break;
+        case bt_reserved :
+        default:
+            return ERROR(corruption_detected);
+        }
+
+        if (ZSTD_isError(decodedSize)) return decodedSize;
+        if (dctx->fParams.checksumFlag)
+            XXH64_update(&dctx->xxhState, op, decodedSize);
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+        if (blockProperties.lastBlock) break;
+    }
+
+    if (dctx->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+        if ((U64)(op-ostart) != dctx->fParams.frameContentSize) {
+            return ERROR(corruption_detected);
+    }   }
+    if (dctx->fParams.checksumFlag) { /* Frame content checksum verification */
+        U32 const checkCalc = (U32)XXH64_digest(&dctx->xxhState);
+        U32 checkRead;
+        if (remainingSize<4) return ERROR(checksum_wrong);
+        checkRead = MEM_readLE32(ip);
+        if (checkRead != checkCalc) return ERROR(checksum_wrong);
+        ip += 4;
+        remainingSize -= 4;
+    }
+
+    /* Allow caller to get size read */
+    *srcPtr = ip;
+    *srcSizePtr = remainingSize;
+    return op-ostart;
+}
+
+static const void* ZSTD_DDictDictContent(const ZSTD_DDict* ddict);
+static size_t ZSTD_DDictDictSize(const ZSTD_DDict* ddict);
+
+static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
+                                        void* dst, size_t dstCapacity,
+                                  const void* src, size_t srcSize,
+                                  const void* dict, size_t dictSize,
+                                  const ZSTD_DDict* ddict)
+{
+    void* const dststart = dst;
+    assert(dict==NULL || ddict==NULL);  /* either dict or ddict set, not both */
+
+    if (ddict) {
+        dict = ZSTD_DDictDictContent(ddict);
+        dictSize = ZSTD_DDictDictSize(ddict);
+    }
+
+    while (srcSize >= ZSTD_frameHeaderSize_prefix) {
+        U32 magicNumber;
+
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
+        if (ZSTD_isLegacy(src, srcSize)) {
+            size_t decodedSize;
+            size_t const frameSize = ZSTD_findFrameCompressedSizeLegacy(src, srcSize);
+            if (ZSTD_isError(frameSize)) return frameSize;
+            /* legacy support is not compatible with static dctx */
+            if (dctx->staticSize) return ERROR(memory_allocation);
+
+            decodedSize = ZSTD_decompressLegacy(dst, dstCapacity, src, frameSize, dict, dictSize);
+
+            dst = (BYTE*)dst + decodedSize;
+            dstCapacity -= decodedSize;
+
+            src = (const BYTE*)src + frameSize;
+            srcSize -= frameSize;
+
+            continue;
+        }
+#endif
+
+        magicNumber = MEM_readLE32(src);
+        DEBUGLOG(4, "reading magic number %08X (expecting %08X)",
+                    (U32)magicNumber, (U32)ZSTD_MAGICNUMBER);
+        if (magicNumber != ZSTD_MAGICNUMBER) {
+            if ((magicNumber & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) {
+                size_t skippableSize;
+                if (srcSize < ZSTD_skippableHeaderSize)
+                    return ERROR(srcSize_wrong);
+                skippableSize = MEM_readLE32((const BYTE*)src + ZSTD_frameIdSize)
+                              + ZSTD_skippableHeaderSize;
+                if (srcSize < skippableSize) return ERROR(srcSize_wrong);
+
+                src = (const BYTE *)src + skippableSize;
+                srcSize -= skippableSize;
+                continue;
+            }
+            return ERROR(prefix_unknown);
+        }
+
+        if (ddict) {
+            /* we were called from ZSTD_decompress_usingDDict */
+            CHECK_F(ZSTD_decompressBegin_usingDDict(dctx, ddict));
+        } else {
+            /* this will initialize correctly with no dict if dict == NULL, so
+             * use this in all cases but ddict */
+            CHECK_F(ZSTD_decompressBegin_usingDict(dctx, dict, dictSize));
+        }
+        ZSTD_checkContinuity(dctx, dst);
+
+        {   const size_t res = ZSTD_decompressFrame(dctx, dst, dstCapacity,
+                                                    &src, &srcSize);
+            if (ZSTD_isError(res)) return res;
+            /* no need to bound check, ZSTD_decompressFrame already has */
+            dst = (BYTE*)dst + res;
+            dstCapacity -= res;
+        }
+    }  /* while (srcSize >= ZSTD_frameHeaderSize_prefix) */
+
+    if (srcSize) return ERROR(srcSize_wrong); /* input not entirely consumed */
+
+    return (BYTE*)dst - (BYTE*)dststart;
+}
+
+size_t ZSTD_decompress_usingDict(ZSTD_DCtx* dctx,
+                                 void* dst, size_t dstCapacity,
+                           const void* src, size_t srcSize,
+                           const void* dict, size_t dictSize)
+{
+    return ZSTD_decompressMultiFrame(dctx, dst, dstCapacity, src, srcSize, dict, dictSize, NULL);
+}
+
+
+size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    return ZSTD_decompress_usingDict(dctx, dst, dstCapacity, src, srcSize, NULL, 0);
+}
+
+
+size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+#if defined(ZSTD_HEAPMODE) && (ZSTD_HEAPMODE>=1)
+    size_t regenSize;
+    ZSTD_DCtx* const dctx = ZSTD_createDCtx();
+    if (dctx==NULL) return ERROR(memory_allocation);
+    regenSize = ZSTD_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
+    ZSTD_freeDCtx(dctx);
+    return regenSize;
+#else   /* stack mode */
+    ZSTD_DCtx dctx;
+    return ZSTD_decompressDCtx(&dctx, dst, dstCapacity, src, srcSize);
+#endif
+}
+
+
+/*-**************************************
+*   Advanced Streaming Decompression API
+*   Bufferless and synchronous
+****************************************/
+size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx) { return dctx->expected; }
+
+ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
+    switch(dctx->stage)
+    {
+    default:   /* should not happen */
+        assert(0);
+    case ZSTDds_getFrameHeaderSize:
+    case ZSTDds_decodeFrameHeader:
+        return ZSTDnit_frameHeader;
+    case ZSTDds_decodeBlockHeader:
+        return ZSTDnit_blockHeader;
+    case ZSTDds_decompressBlock:
+        return ZSTDnit_block;
+    case ZSTDds_decompressLastBlock:
+        return ZSTDnit_lastBlock;
+    case ZSTDds_checkChecksum:
+        return ZSTDnit_checksum;
+    case ZSTDds_decodeSkippableHeader:
+    case ZSTDds_skipFrame:
+        return ZSTDnit_skippableFrame;
+    }
+}
+
+static int ZSTD_isSkipFrame(ZSTD_DCtx* dctx) { return dctx->stage == ZSTDds_skipFrame; }
+
+/** ZSTD_decompressContinue() :
+ *  srcSize : must be the exact nb of bytes expected (see ZSTD_nextSrcSizeToDecompress())
+ *  @return : nb of bytes generated into `dst` (necessarily <= `dstCapacity)
+ *            or an error code, which can be tested using ZSTD_isError() */
+size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    DEBUGLOG(5, "ZSTD_decompressContinue (srcSize:%u)", (U32)srcSize);
+    /* Sanity check */
+    if (srcSize != dctx->expected) return ERROR(srcSize_wrong);  /* not allowed */
+    if (dstCapacity) ZSTD_checkContinuity(dctx, dst);
+
+    switch (dctx->stage)
+    {
+    case ZSTDds_getFrameHeaderSize :
+        assert(src != NULL);
+        if (dctx->format == ZSTD_f_zstd1) {  /* allows header */
+            assert(srcSize >= ZSTD_frameIdSize);  /* to read skippable magic number */
+            if ((MEM_readLE32(src) & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) {        /* skippable frame */
+                memcpy(dctx->headerBuffer, src, srcSize);
+                dctx->expected = ZSTD_skippableHeaderSize - srcSize;  /* remaining to load to get full skippable frame header */
+                dctx->stage = ZSTDds_decodeSkippableHeader;
+                return 0;
+        }   }
+        dctx->headerSize = ZSTD_frameHeaderSize_internal(src, srcSize, dctx->format);
+        if (ZSTD_isError(dctx->headerSize)) return dctx->headerSize;
+        memcpy(dctx->headerBuffer, src, srcSize);
+        dctx->expected = dctx->headerSize - srcSize;
+        dctx->stage = ZSTDds_decodeFrameHeader;
+        return 0;
+
+    case ZSTDds_decodeFrameHeader:
+        assert(src != NULL);
+        memcpy(dctx->headerBuffer + (dctx->headerSize - srcSize), src, srcSize);
+        CHECK_F(ZSTD_decodeFrameHeader(dctx, dctx->headerBuffer, dctx->headerSize));
+        dctx->expected = ZSTD_blockHeaderSize;
+        dctx->stage = ZSTDds_decodeBlockHeader;
+        return 0;
+
+    case ZSTDds_decodeBlockHeader:
+        {   blockProperties_t bp;
+            size_t const cBlockSize = ZSTD_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
+            if (ZSTD_isError(cBlockSize)) return cBlockSize;
+            dctx->expected = cBlockSize;
+            dctx->bType = bp.blockType;
+            dctx->rleSize = bp.origSize;
+            if (cBlockSize) {
+                dctx->stage = bp.lastBlock ? ZSTDds_decompressLastBlock : ZSTDds_decompressBlock;
+                return 0;
+            }
+            /* empty block */
+            if (bp.lastBlock) {
+                if (dctx->fParams.checksumFlag) {
+                    dctx->expected = 4;
+                    dctx->stage = ZSTDds_checkChecksum;
+                } else {
+                    dctx->expected = 0; /* end of frame */
+                    dctx->stage = ZSTDds_getFrameHeaderSize;
+                }
+            } else {
+                dctx->expected = ZSTD_blockHeaderSize;  /* jump to next header */
+                dctx->stage = ZSTDds_decodeBlockHeader;
+            }
+            return 0;
+        }
+
+    case ZSTDds_decompressLastBlock:
+    case ZSTDds_decompressBlock:
+        DEBUGLOG(5, "ZSTD_decompressContinue: case ZSTDds_decompressBlock");
+        {   size_t rSize;
+            switch(dctx->bType)
+            {
+            case bt_compressed:
+                DEBUGLOG(5, "ZSTD_decompressContinue: case bt_compressed");
+                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1);
+                break;
+            case bt_raw :
+                rSize = ZSTD_copyRawBlock(dst, dstCapacity, src, srcSize);
+                break;
+            case bt_rle :
+                rSize = ZSTD_setRleBlock(dst, dstCapacity, src, srcSize, dctx->rleSize);
+                break;
+            case bt_reserved :   /* should never happen */
+            default:
+                return ERROR(corruption_detected);
+            }
+            if (ZSTD_isError(rSize)) return rSize;
+            DEBUGLOG(5, "ZSTD_decompressContinue: decoded size from block : %u", (U32)rSize);
+            dctx->decodedSize += rSize;
+            if (dctx->fParams.checksumFlag) XXH64_update(&dctx->xxhState, dst, rSize);
+
+            if (dctx->stage == ZSTDds_decompressLastBlock) {   /* end of frame */
+                DEBUGLOG(4, "ZSTD_decompressContinue: decoded size from frame : %u", (U32)dctx->decodedSize);
+                if (dctx->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+                    if (dctx->decodedSize != dctx->fParams.frameContentSize) {
+                        return ERROR(corruption_detected);
+                }   }
+                if (dctx->fParams.checksumFlag) {  /* another round for frame checksum */
+                    dctx->expected = 4;
+                    dctx->stage = ZSTDds_checkChecksum;
+                } else {
+                    dctx->expected = 0;   /* ends here */
+                    dctx->stage = ZSTDds_getFrameHeaderSize;
+                }
+            } else {
+                dctx->stage = ZSTDds_decodeBlockHeader;
+                dctx->expected = ZSTD_blockHeaderSize;
+                dctx->previousDstEnd = (char*)dst + rSize;
+            }
+            return rSize;
+        }
+
+    case ZSTDds_checkChecksum:
+        assert(srcSize == 4);  /* guaranteed by dctx->expected */
+        {   U32 const h32 = (U32)XXH64_digest(&dctx->xxhState);
+            U32 const check32 = MEM_readLE32(src);
+            DEBUGLOG(4, "ZSTD_decompressContinue: checksum : calculated %08X :: %08X read", h32, check32);
+            if (check32 != h32) return ERROR(checksum_wrong);
+            dctx->expected = 0;
+            dctx->stage = ZSTDds_getFrameHeaderSize;
+            return 0;
+        }
+
+    case ZSTDds_decodeSkippableHeader:
+        assert(src != NULL);
+        assert(srcSize <= ZSTD_skippableHeaderSize);
+        memcpy(dctx->headerBuffer + (ZSTD_skippableHeaderSize - srcSize), src, srcSize);   /* complete skippable header */
+        dctx->expected = MEM_readLE32(dctx->headerBuffer + ZSTD_frameIdSize);   /* note : dctx->expected can grow seriously large, beyond local buffer size */
+        dctx->stage = ZSTDds_skipFrame;
+        return 0;
+
+    case ZSTDds_skipFrame:
+        dctx->expected = 0;
+        dctx->stage = ZSTDds_getFrameHeaderSize;
+        return 0;
+
+    default:
+        return ERROR(GENERIC);   /* impossible */
+    }
+}
+
+
+static size_t ZSTD_refDictContent(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    dctx->dictEnd = dctx->previousDstEnd;
+    dctx->vBase = (const char*)dict - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+    dctx->base = dict;
+    dctx->previousDstEnd = (const char*)dict + dictSize;
+    return 0;
+}
+
+/* ZSTD_loadEntropy() :
+ * dict : must point at beginning of a valid zstd dictionary
+ * @return : size of entropy tables read */
+static size_t ZSTD_loadEntropy(ZSTD_entropyDTables_t* entropy, const void* const dict, size_t const dictSize)
+{
+    const BYTE* dictPtr = (const BYTE*)dict;
+    const BYTE* const dictEnd = dictPtr + dictSize;
+
+    if (dictSize <= 8) return ERROR(dictionary_corrupted);
+    dictPtr += 8;   /* skip header = magic + dictID */
+
+
+    {   size_t const hSize = HUF_readDTableX4_wksp(
+            entropy->hufTable, dictPtr, dictEnd - dictPtr,
+            entropy->workspace, sizeof(entropy->workspace));
+        if (HUF_isError(hSize)) return ERROR(dictionary_corrupted);
+        dictPtr += hSize;
+    }
+
+    {   short offcodeNCount[MaxOff+1];
+        U32 offcodeMaxValue = MaxOff, offcodeLog;
+        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(offcodeHeaderSize)) return ERROR(dictionary_corrupted);
+        if (offcodeMaxValue > MaxOff) return ERROR(dictionary_corrupted);
+        if (offcodeLog > OffFSELog) return ERROR(dictionary_corrupted);
+        ZSTD_buildFSETable(entropy->OFTable,
+                            offcodeNCount, offcodeMaxValue,
+                            OF_base, OF_bits,
+                            offcodeLog);
+        dictPtr += offcodeHeaderSize;
+    }
+
+    {   short matchlengthNCount[MaxML+1];
+        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
+        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(matchlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (matchlengthMaxValue > MaxML) return ERROR(dictionary_corrupted);
+        if (matchlengthLog > MLFSELog) return ERROR(dictionary_corrupted);
+        ZSTD_buildFSETable(entropy->MLTable,
+                            matchlengthNCount, matchlengthMaxValue,
+                            ML_base, ML_bits,
+                            matchlengthLog);
+        dictPtr += matchlengthHeaderSize;
+    }
+
+    {   short litlengthNCount[MaxLL+1];
+        unsigned litlengthMaxValue = MaxLL, litlengthLog;
+        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSE_isError(litlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (litlengthMaxValue > MaxLL) return ERROR(dictionary_corrupted);
+        if (litlengthLog > LLFSELog) return ERROR(dictionary_corrupted);
+        ZSTD_buildFSETable(entropy->LLTable,
+                            litlengthNCount, litlengthMaxValue,
+                            LL_base, LL_bits,
+                            litlengthLog);
+        dictPtr += litlengthHeaderSize;
+    }
+
+    if (dictPtr+12 > dictEnd) return ERROR(dictionary_corrupted);
+    {   int i;
+        size_t const dictContentSize = (size_t)(dictEnd - (dictPtr+12));
+        for (i=0; i<3; i++) {
+            U32 const rep = MEM_readLE32(dictPtr); dictPtr += 4;
+            if (rep==0 || rep >= dictContentSize) return ERROR(dictionary_corrupted);
+            entropy->rep[i] = rep;
+    }   }
+
+    return dictPtr - (const BYTE*)dict;
+}
+
+static size_t ZSTD_decompress_insertDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    if (dictSize < 8) return ZSTD_refDictContent(dctx, dict, dictSize);
+    {   U32 const magic = MEM_readLE32(dict);
+        if (magic != ZSTD_MAGIC_DICTIONARY) {
+            return ZSTD_refDictContent(dctx, dict, dictSize);   /* pure content mode */
+    }   }
+    dctx->dictID = MEM_readLE32((const char*)dict + ZSTD_frameIdSize);
+
+    /* load entropy tables */
+    {   size_t const eSize = ZSTD_loadEntropy(&dctx->entropy, dict, dictSize);
+        if (ZSTD_isError(eSize)) return ERROR(dictionary_corrupted);
+        dict = (const char*)dict + eSize;
+        dictSize -= eSize;
+    }
+    dctx->litEntropy = dctx->fseEntropy = 1;
+
+    /* reference dictionary content */
+    return ZSTD_refDictContent(dctx, dict, dictSize);
+}
+
+/* Note : this function cannot fail */
+size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx)
+{
+    assert(dctx != NULL);
+    dctx->expected = ZSTD_startingInputLength(dctx->format);  /* dctx->format must be properly set */
+    dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->decodedSize = 0;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    dctx->vBase = NULL;
+    dctx->dictEnd = NULL;
+    dctx->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    dctx->litEntropy = dctx->fseEntropy = 0;
+    dctx->dictID = 0;
+    ZSTD_STATIC_ASSERT(sizeof(dctx->entropy.rep) == sizeof(repStartValue));
+    memcpy(dctx->entropy.rep, repStartValue, sizeof(repStartValue));  /* initial repcodes */
+    dctx->LLTptr = dctx->entropy.LLTable;
+    dctx->MLTptr = dctx->entropy.MLTable;
+    dctx->OFTptr = dctx->entropy.OFTable;
+    dctx->HUFptr = dctx->entropy.hufTable;
+    return 0;
+}
+
+size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    CHECK_F( ZSTD_decompressBegin(dctx) );
+    if (dict && dictSize)
+        CHECK_E(ZSTD_decompress_insertDictionary(dctx, dict, dictSize), dictionary_corrupted);
+    return 0;
+}
+
+
+/* ======   ZSTD_DDict   ====== */
+
+struct ZSTD_DDict_s {
+    void* dictBuffer;
+    const void* dictContent;
+    size_t dictSize;
+    ZSTD_entropyDTables_t entropy;
+    U32 dictID;
+    U32 entropyPresent;
+    ZSTD_customMem cMem;
+};  /* typedef'd to ZSTD_DDict within "zstd.h" */
+
+static const void* ZSTD_DDictDictContent(const ZSTD_DDict* ddict)
+{
+    return ddict->dictContent;
+}
+
+static size_t ZSTD_DDictDictSize(const ZSTD_DDict* ddict)
+{
+    return ddict->dictSize;
+}
+
+size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dstDCtx, const ZSTD_DDict* ddict)
+{
+    CHECK_F( ZSTD_decompressBegin(dstDCtx) );
+    if (ddict) {   /* support begin on NULL */
+        dstDCtx->dictID = ddict->dictID;
+        dstDCtx->base = ddict->dictContent;
+        dstDCtx->vBase = ddict->dictContent;
+        dstDCtx->dictEnd = (const BYTE*)ddict->dictContent + ddict->dictSize;
+        dstDCtx->previousDstEnd = dstDCtx->dictEnd;
+        if (ddict->entropyPresent) {
+            dstDCtx->litEntropy = 1;
+            dstDCtx->fseEntropy = 1;
+            dstDCtx->LLTptr = ddict->entropy.LLTable;
+            dstDCtx->MLTptr = ddict->entropy.MLTable;
+            dstDCtx->OFTptr = ddict->entropy.OFTable;
+            dstDCtx->HUFptr = ddict->entropy.hufTable;
+            dstDCtx->entropy.rep[0] = ddict->entropy.rep[0];
+            dstDCtx->entropy.rep[1] = ddict->entropy.rep[1];
+            dstDCtx->entropy.rep[2] = ddict->entropy.rep[2];
+        } else {
+            dstDCtx->litEntropy = 0;
+            dstDCtx->fseEntropy = 0;
+        }
+    }
+    return 0;
+}
+
+static size_t ZSTD_loadEntropy_inDDict(ZSTD_DDict* ddict, ZSTD_dictContentType_e dictContentType)
+{
+    ddict->dictID = 0;
+    ddict->entropyPresent = 0;
+    if (dictContentType == ZSTD_dct_rawContent) return 0;
+
+    if (ddict->dictSize < 8) {
+        if (dictContentType == ZSTD_dct_fullDict)
+            return ERROR(dictionary_corrupted);   /* only accept specified dictionaries */
+        return 0;   /* pure content mode */
+    }
+    {   U32 const magic = MEM_readLE32(ddict->dictContent);
+        if (magic != ZSTD_MAGIC_DICTIONARY) {
+            if (dictContentType == ZSTD_dct_fullDict)
+                return ERROR(dictionary_corrupted);   /* only accept specified dictionaries */
+            return 0;   /* pure content mode */
+        }
+    }
+    ddict->dictID = MEM_readLE32((const char*)ddict->dictContent + ZSTD_frameIdSize);
+
+    /* load entropy tables */
+    CHECK_E( ZSTD_loadEntropy(&ddict->entropy, ddict->dictContent, ddict->dictSize), dictionary_corrupted );
+    ddict->entropyPresent = 1;
+    return 0;
+}
+
+
+static size_t ZSTD_initDDict_internal(ZSTD_DDict* ddict,
+                                      const void* dict, size_t dictSize,
+                                      ZSTD_dictLoadMethod_e dictLoadMethod,
+                                      ZSTD_dictContentType_e dictContentType)
+{
+    if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dict) || (!dictSize)) {
+        ddict->dictBuffer = NULL;
+        ddict->dictContent = dict;
+    } else {
+        void* const internalBuffer = ZSTD_malloc(dictSize, ddict->cMem);
+        ddict->dictBuffer = internalBuffer;
+        ddict->dictContent = internalBuffer;
+        if (!internalBuffer) return ERROR(memory_allocation);
+        memcpy(internalBuffer, dict, dictSize);
+    }
+    ddict->dictSize = dictSize;
+    ddict->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+
+    /* parse dictionary content */
+    CHECK_F( ZSTD_loadEntropy_inDDict(ddict, dictContentType) );
+
+    return 0;
+}
+
+ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
+                                      ZSTD_dictLoadMethod_e dictLoadMethod,
+                                      ZSTD_dictContentType_e dictContentType,
+                                      ZSTD_customMem customMem)
+{
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+
+    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_malloc(sizeof(ZSTD_DDict), customMem);
+        if (!ddict) return NULL;
+        ddict->cMem = customMem;
+
+        if (ZSTD_isError( ZSTD_initDDict_internal(ddict, dict, dictSize, dictLoadMethod, dictContentType) )) {
+            ZSTD_freeDDict(ddict);
+            return NULL;
+        }
+
+        return ddict;
+    }
+}
+
+/*! ZSTD_createDDict() :
+*   Create a digested dictionary, to start decompression without startup delay.
+*   `dict` content is copied inside DDict.
+*   Consequently, `dict` can be released after `ZSTD_DDict` creation */
+ZSTD_DDict* ZSTD_createDDict(const void* dict, size_t dictSize)
+{
+    ZSTD_customMem const allocator = { NULL, NULL, NULL };
+    return ZSTD_createDDict_advanced(dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto, allocator);
+}
+
+/*! ZSTD_createDDict_byReference() :
+ *  Create a digested dictionary, to start decompression without startup delay.
+ *  Dictionary content is simply referenced, it will be accessed during decompression.
+ *  Warning : dictBuffer must outlive DDict (DDict must be freed before dictBuffer) */
+ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize)
+{
+    ZSTD_customMem const allocator = { NULL, NULL, NULL };
+    return ZSTD_createDDict_advanced(dictBuffer, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto, allocator);
+}
+
+
+const ZSTD_DDict* ZSTD_initStaticDDict(
+                                void* workspace, size_t workspaceSize,
+                                const void* dict, size_t dictSize,
+                                ZSTD_dictLoadMethod_e dictLoadMethod,
+                                ZSTD_dictContentType_e dictContentType)
+{
+    size_t const neededSpace =
+            sizeof(ZSTD_DDict) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize);
+    ZSTD_DDict* const ddict = (ZSTD_DDict*)workspace;
+    assert(workspace != NULL);
+    assert(dict != NULL);
+    if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
+    if (workspaceSize < neededSpace) return NULL;
+    if (dictLoadMethod == ZSTD_dlm_byCopy) {
+        memcpy(ddict+1, dict, dictSize);  /* local copy */
+        dict = ddict+1;
+    }
+    if (ZSTD_isError( ZSTD_initDDict_internal(ddict, dict, dictSize, ZSTD_dlm_byRef, dictContentType) ))
+        return NULL;
+    return ddict;
+}
+
+
+size_t ZSTD_freeDDict(ZSTD_DDict* ddict)
+{
+    if (ddict==NULL) return 0;   /* support free on NULL */
+    {   ZSTD_customMem const cMem = ddict->cMem;
+        ZSTD_free(ddict->dictBuffer, cMem);
+        ZSTD_free(ddict, cMem);
+        return 0;
+    }
+}
+
+/*! ZSTD_estimateDDictSize() :
+ *  Estimate amount of memory that will be needed to create a dictionary for decompression.
+ *  Note : dictionary created by reference using ZSTD_dlm_byRef are smaller */
+size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod)
+{
+    return sizeof(ZSTD_DDict) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize);
+}
+
+size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict)
+{
+    if (ddict==NULL) return 0;   /* support sizeof on NULL */
+    return sizeof(*ddict) + (ddict->dictBuffer ? ddict->dictSize : 0) ;
+}
+
+/*! ZSTD_getDictID_fromDict() :
+ *  Provides the dictID stored within dictionary.
+ *  if @return == 0, the dictionary is not conformant with Zstandard specification.
+ *  It can still be loaded, but as a content-only dictionary. */
+unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
+{
+    if (dictSize < 8) return 0;
+    if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) return 0;
+    return MEM_readLE32((const char*)dict + ZSTD_frameIdSize);
+}
+
+/*! ZSTD_getDictID_fromDDict() :
+ *  Provides the dictID of the dictionary loaded into `ddict`.
+ *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
+ *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
+unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
+{
+    if (ddict==NULL) return 0;
+    return ZSTD_getDictID_fromDict(ddict->dictContent, ddict->dictSize);
+}
+
+/*! ZSTD_getDictID_fromFrame() :
+ *  Provides the dictID required to decompresse frame stored within `src`.
+ *  If @return == 0, the dictID could not be decoded.
+ *  This could for one of the following reasons :
+ *  - The frame does not require a dictionary (most common case).
+ *  - The frame was built with dictID intentionally removed.
+ *    Needed dictionary is a hidden information.
+ *    Note : this use case also happens when using a non-conformant dictionary.
+ *  - `srcSize` is too small, and as a result, frame header could not be decoded.
+ *    Note : possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`.
+ *  - This is not a Zstandard frame.
+ *  When identifying the exact failure cause, it's possible to use
+ *  ZSTD_getFrameHeader(), which will provide a more precise error code. */
+unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize)
+{
+    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0 };
+    size_t const hError = ZSTD_getFrameHeader(&zfp, src, srcSize);
+    if (ZSTD_isError(hError)) return 0;
+    return zfp.dictID;
+}
+
+
+/*! ZSTD_decompress_usingDDict() :
+*   Decompression using a pre-digested Dictionary
+*   Use dictionary without significant overhead. */
+size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
+                                  void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize,
+                            const ZSTD_DDict* ddict)
+{
+    /* pass content and size in case legacy frames are encountered */
+    return ZSTD_decompressMultiFrame(dctx, dst, dstCapacity, src, srcSize,
+                                     NULL, 0,
+                                     ddict);
+}
+
+
+/*=====================================
+*   Streaming decompression
+*====================================*/
+
+ZSTD_DStream* ZSTD_createDStream(void)
+{
+    DEBUGLOG(3, "ZSTD_createDStream");
+    return ZSTD_createDStream_advanced(ZSTD_defaultCMem);
+}
+
+ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
+{
+    return ZSTD_initStaticDCtx(workspace, workspaceSize);
+}
+
+ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem)
+{
+    return ZSTD_createDCtx_advanced(customMem);
+}
+
+size_t ZSTD_freeDStream(ZSTD_DStream* zds)
+{
+    return ZSTD_freeDCtx(zds);
+}
+
+
+/* *** Initialization *** */
+
+size_t ZSTD_DStreamInSize(void)  { return ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize; }
+size_t ZSTD_DStreamOutSize(void) { return ZSTD_BLOCKSIZE_MAX; }
+
+size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType)
+{
+    if (dctx->streamStage != zdss_init) return ERROR(stage_wrong);
+    ZSTD_freeDDict(dctx->ddictLocal);
+    if (dict && dictSize >= 8) {
+        dctx->ddictLocal = ZSTD_createDDict_advanced(dict, dictSize, dictLoadMethod, dictContentType, dctx->customMem);
+        if (dctx->ddictLocal == NULL) return ERROR(memory_allocation);
+    } else {
+        dctx->ddictLocal = NULL;
+    }
+    dctx->ddict = dctx->ddictLocal;
+    return 0;
+}
+
+size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    return ZSTD_DCtx_loadDictionary_advanced(dctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto);
+}
+
+size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    return ZSTD_DCtx_loadDictionary_advanced(dctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto);
+}
+
+size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType)
+{
+    return ZSTD_DCtx_loadDictionary_advanced(dctx, prefix, prefixSize, ZSTD_dlm_byRef, dictContentType);
+}
+
+size_t ZSTD_DCtx_refPrefix(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize)
+{
+    return ZSTD_DCtx_refPrefix_advanced(dctx, prefix, prefixSize, ZSTD_dct_rawContent);
+}
+
+
+/* ZSTD_initDStream_usingDict() :
+ * return : expected size, aka ZSTD_frameHeaderSize_prefix.
+ * this function cannot fail */
+size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize)
+{
+    DEBUGLOG(4, "ZSTD_initDStream_usingDict");
+    zds->streamStage = zdss_init;
+    CHECK_F( ZSTD_DCtx_loadDictionary(zds, dict, dictSize) );
+    return ZSTD_frameHeaderSize_prefix;
+}
+
+/* note : this variant can't fail */
+size_t ZSTD_initDStream(ZSTD_DStream* zds)
+{
+    DEBUGLOG(4, "ZSTD_initDStream");
+    return ZSTD_initDStream_usingDict(zds, NULL, 0);
+}
+
+size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
+{
+    if (dctx->streamStage != zdss_init) return ERROR(stage_wrong);
+    dctx->ddict = ddict;
+    return 0;
+}
+
+/* ZSTD_initDStream_usingDDict() :
+ * ddict will just be referenced, and must outlive decompression session
+ * this function cannot fail */
+size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
+{
+    size_t const initResult = ZSTD_initDStream(dctx);
+    dctx->ddict = ddict;
+    return initResult;
+}
+
+/* ZSTD_resetDStream() :
+ * return : expected size, aka ZSTD_frameHeaderSize_prefix.
+ * this function cannot fail */
+size_t ZSTD_resetDStream(ZSTD_DStream* dctx)
+{
+    DEBUGLOG(4, "ZSTD_resetDStream");
+    dctx->streamStage = zdss_loadHeader;
+    dctx->lhSize = dctx->inPos = dctx->outStart = dctx->outEnd = 0;
+    dctx->legacyVersion = 0;
+    dctx->hostageByte = 0;
+    return ZSTD_frameHeaderSize_prefix;
+}
+
+size_t ZSTD_setDStreamParameter(ZSTD_DStream* dctx,
+                                ZSTD_DStreamParameter_e paramType, unsigned paramValue)
+{
+    if (dctx->streamStage != zdss_init) return ERROR(stage_wrong);
+    switch(paramType)
+    {
+        default : return ERROR(parameter_unsupported);
+        case DStream_p_maxWindowSize :
+            DEBUGLOG(4, "setting maxWindowSize = %u KB", paramValue >> 10);
+            dctx->maxWindowSize = paramValue ? paramValue : (U32)(-1);
+            break;
+    }
+    return 0;
+}
+
+size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize)
+{
+    if (dctx->streamStage != zdss_init) return ERROR(stage_wrong);
+    dctx->maxWindowSize = maxWindowSize;
+    return 0;
+}
+
+size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format)
+{
+    DEBUGLOG(4, "ZSTD_DCtx_setFormat : %u", (unsigned)format);
+    if (dctx->streamStage != zdss_init) return ERROR(stage_wrong);
+    dctx->format = format;
+    return 0;
+}
+
+
+size_t ZSTD_sizeof_DStream(const ZSTD_DStream* dctx)
+{
+    return ZSTD_sizeof_DCtx(dctx);
+}
+
+size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
+{
+    size_t const blockSize = (size_t) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
+    unsigned long long const neededRBSize = windowSize + blockSize + (WILDCOPY_OVERLENGTH * 2);
+    unsigned long long const neededSize = MIN(frameContentSize, neededRBSize);
+    size_t const minRBSize = (size_t) neededSize;
+    if ((unsigned long long)minRBSize != neededSize) return ERROR(frameParameter_windowTooLarge);
+    return minRBSize;
+}
+
+size_t ZSTD_estimateDStreamSize(size_t windowSize)
+{
+    size_t const blockSize = MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
+    size_t const inBuffSize = blockSize;  /* no block can be larger */
+    size_t const outBuffSize = ZSTD_decodingBufferSize_min(windowSize, ZSTD_CONTENTSIZE_UNKNOWN);
+    return ZSTD_estimateDCtxSize() + inBuffSize + outBuffSize;
+}
+
+size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize)
+{
+    U32 const windowSizeMax = 1U << ZSTD_WINDOWLOG_MAX;   /* note : should be user-selectable */
+    ZSTD_frameHeader zfh;
+    size_t const err = ZSTD_getFrameHeader(&zfh, src, srcSize);
+    if (ZSTD_isError(err)) return err;
+    if (err>0) return ERROR(srcSize_wrong);
+    if (zfh.windowSize > windowSizeMax)
+        return ERROR(frameParameter_windowTooLarge);
+    return ZSTD_estimateDStreamSize((size_t)zfh.windowSize);
+}
+
+
+/* *****   Decompression   ***** */
+
+MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    size_t const length = MIN(dstCapacity, srcSize);
+    memcpy(dst, src, length);
+    return length;
+}
+
+
+size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
+{
+    const char* const istart = (const char*)(input->src) + input->pos;
+    const char* const iend = (const char*)(input->src) + input->size;
+    const char* ip = istart;
+    char* const ostart = (char*)(output->dst) + output->pos;
+    char* const oend = (char*)(output->dst) + output->size;
+    char* op = ostart;
+    U32 someMoreWork = 1;
+
+    DEBUGLOG(5, "ZSTD_decompressStream");
+    if (input->pos > input->size) {  /* forbidden */
+        DEBUGLOG(5, "in: pos: %u   vs size: %u",
+                    (U32)input->pos, (U32)input->size);
+        return ERROR(srcSize_wrong);
+    }
+    if (output->pos > output->size) {  /* forbidden */
+        DEBUGLOG(5, "out: pos: %u   vs size: %u",
+                    (U32)output->pos, (U32)output->size);
+        return ERROR(dstSize_tooSmall);
+    }
+    DEBUGLOG(5, "input size : %u", (U32)(input->size - input->pos));
+
+    while (someMoreWork) {
+        switch(zds->streamStage)
+        {
+        case zdss_init :
+            DEBUGLOG(5, "stage zdss_init => transparent reset ");
+            ZSTD_resetDStream(zds);   /* transparent reset on starting decoding a new frame */
+            /* fall-through */
+
+        case zdss_loadHeader :
+            DEBUGLOG(5, "stage zdss_loadHeader (srcSize : %u)", (U32)(iend - ip));
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
+            if (zds->legacyVersion) {
+                /* legacy support is incompatible with static dctx */
+                if (zds->staticSize) return ERROR(memory_allocation);
+                {   size_t const hint = ZSTD_decompressLegacyStream(zds->legacyContext, zds->legacyVersion, output, input);
+                    if (hint==0) zds->streamStage = zdss_init;
+                    return hint;
+            }   }
+#endif
+            {   size_t const hSize = ZSTD_getFrameHeader_internal(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format);
+                DEBUGLOG(5, "header size : %u", (U32)hSize);
+                if (ZSTD_isError(hSize)) {
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
+                    U32 const legacyVersion = ZSTD_isLegacy(istart, iend-istart);
+                    if (legacyVersion) {
+                        const void* const dict = zds->ddict ? zds->ddict->dictContent : NULL;
+                        size_t const dictSize = zds->ddict ? zds->ddict->dictSize : 0;
+                        DEBUGLOG(5, "ZSTD_decompressStream: detected legacy version v0.%u", legacyVersion);
+                        /* legacy support is incompatible with static dctx */
+                        if (zds->staticSize) return ERROR(memory_allocation);
+                        CHECK_F(ZSTD_initLegacyStream(&zds->legacyContext,
+                                    zds->previousLegacyVersion, legacyVersion,
+                                    dict, dictSize));
+                        zds->legacyVersion = zds->previousLegacyVersion = legacyVersion;
+                        {   size_t const hint = ZSTD_decompressLegacyStream(zds->legacyContext, legacyVersion, output, input);
+                            if (hint==0) zds->streamStage = zdss_init;   /* or stay in stage zdss_loadHeader */
+                            return hint;
+                    }   }
+#endif
+                    return hSize;   /* error */
+                }
+                if (hSize != 0) {   /* need more input */
+                    size_t const toLoad = hSize - zds->lhSize;   /* if hSize!=0, hSize > zds->lhSize */
+                    size_t const remainingInput = (size_t)(iend-ip);
+                    assert(iend >= ip);
+                    if (toLoad > remainingInput) {   /* not enough input to load full header */
+                        if (remainingInput > 0) {
+                            memcpy(zds->headerBuffer + zds->lhSize, ip, remainingInput);
+                            zds->lhSize += remainingInput;
+                        }
+                        input->pos = input->size;
+                        return (MAX(ZSTD_frameHeaderSize_min, hSize) - zds->lhSize) + ZSTD_blockHeaderSize;   /* remaining header bytes + next block header */
+                    }
+                    assert(ip != NULL);
+                    memcpy(zds->headerBuffer + zds->lhSize, ip, toLoad); zds->lhSize = hSize; ip += toLoad;
+                    break;
+            }   }
+
+            /* check for single-pass mode opportunity */
+            if (zds->fParams.frameContentSize && zds->fParams.windowSize /* skippable frame if == 0 */
+                && (U64)(size_t)(oend-op) >= zds->fParams.frameContentSize) {
+                size_t const cSize = ZSTD_findFrameCompressedSize(istart, iend-istart);
+                if (cSize <= (size_t)(iend-istart)) {
+                    /* shortcut : using single-pass mode */
+                    size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, oend-op, istart, cSize, zds->ddict);
+                    if (ZSTD_isError(decompressedSize)) return decompressedSize;
+                    DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()")
+                    ip = istart + cSize;
+                    op += decompressedSize;
+                    zds->expected = 0;
+                    zds->streamStage = zdss_init;
+                    someMoreWork = 0;
+                    break;
+            }   }
+
+            /* Consume header (see ZSTDds_decodeFrameHeader) */
+            DEBUGLOG(4, "Consume header");
+            CHECK_F(ZSTD_decompressBegin_usingDDict(zds, zds->ddict));
+
+            if ((MEM_readLE32(zds->headerBuffer) & 0xFFFFFFF0U) == ZSTD_MAGIC_SKIPPABLE_START) {  /* skippable frame */
+                zds->expected = MEM_readLE32(zds->headerBuffer + ZSTD_frameIdSize);
+                zds->stage = ZSTDds_skipFrame;
+            } else {
+                CHECK_F(ZSTD_decodeFrameHeader(zds, zds->headerBuffer, zds->lhSize));
+                zds->expected = ZSTD_blockHeaderSize;
+                zds->stage = ZSTDds_decodeBlockHeader;
+            }
+
+            /* control buffer memory usage */
+            DEBUGLOG(4, "Control max memory usage (%u KB <= max %u KB)",
+                        (U32)(zds->fParams.windowSize >>10),
+                        (U32)(zds->maxWindowSize >> 10) );
+            zds->fParams.windowSize = MAX(zds->fParams.windowSize, 1U << ZSTD_WINDOWLOG_ABSOLUTEMIN);
+            if (zds->fParams.windowSize > zds->maxWindowSize) return ERROR(frameParameter_windowTooLarge);
+
+            /* Adapt buffer sizes to frame header instructions */
+            {   size_t const neededInBuffSize = MAX(zds->fParams.blockSizeMax, 4 /* frame checksum */);
+                size_t const neededOutBuffSize = ZSTD_decodingBufferSize_min(zds->fParams.windowSize, zds->fParams.frameContentSize);
+                if ((zds->inBuffSize < neededInBuffSize) || (zds->outBuffSize < neededOutBuffSize)) {
+                    size_t const bufferSize = neededInBuffSize + neededOutBuffSize;
+                    DEBUGLOG(4, "inBuff  : from %u to %u",
+                                (U32)zds->inBuffSize, (U32)neededInBuffSize);
+                    DEBUGLOG(4, "outBuff : from %u to %u",
+                                (U32)zds->outBuffSize, (U32)neededOutBuffSize);
+                    if (zds->staticSize) {  /* static DCtx */
+                        DEBUGLOG(4, "staticSize : %u", (U32)zds->staticSize);
+                        assert(zds->staticSize >= sizeof(ZSTD_DCtx));  /* controlled at init */
+                        if (bufferSize > zds->staticSize - sizeof(ZSTD_DCtx))
+                            return ERROR(memory_allocation);
+                    } else {
+                        ZSTD_free(zds->inBuff, zds->customMem);
+                        zds->inBuffSize = 0;
+                        zds->outBuffSize = 0;
+                        zds->inBuff = (char*)ZSTD_malloc(bufferSize, zds->customMem);
+                        if (zds->inBuff == NULL) return ERROR(memory_allocation);
+                    }
+                    zds->inBuffSize = neededInBuffSize;
+                    zds->outBuff = zds->inBuff + zds->inBuffSize;
+                    zds->outBuffSize = neededOutBuffSize;
+            }   }
+            zds->streamStage = zdss_read;
+            /* fall-through */
+
+        case zdss_read:
+            DEBUGLOG(5, "stage zdss_read");
+            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompress(zds);
+                DEBUGLOG(5, "neededInSize = %u", (U32)neededInSize);
+                if (neededInSize==0) {  /* end of frame */
+                    zds->streamStage = zdss_init;
+                    someMoreWork = 0;
+                    break;
+                }
+                if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
+                    int const isSkipFrame = ZSTD_isSkipFrame(zds);
+                    size_t const decodedSize = ZSTD_decompressContinue(zds,
+                        zds->outBuff + zds->outStart, (isSkipFrame ? 0 : zds->outBuffSize - zds->outStart),
+                        ip, neededInSize);
+                    if (ZSTD_isError(decodedSize)) return decodedSize;
+                    ip += neededInSize;
+                    if (!decodedSize && !isSkipFrame) break;   /* this was just a header */
+                    zds->outEnd = zds->outStart + decodedSize;
+                    zds->streamStage = zdss_flush;
+                    break;
+            }   }
+            if (ip==iend) { someMoreWork = 0; break; }   /* no more input */
+            zds->streamStage = zdss_load;
+            /* fall-through */
+
+        case zdss_load:
+            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompress(zds);
+                size_t const toLoad = neededInSize - zds->inPos;
+                int const isSkipFrame = ZSTD_isSkipFrame(zds);
+                size_t loadedSize;
+                if (isSkipFrame) {
+                    loadedSize = MIN(toLoad, (size_t)(iend-ip));
+                } else {
+                    if (toLoad > zds->inBuffSize - zds->inPos) return ERROR(corruption_detected);   /* should never happen */
+                    loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, iend-ip);
+                }
+                ip += loadedSize;
+                zds->inPos += loadedSize;
+                if (loadedSize < toLoad) { someMoreWork = 0; break; }   /* not enough input, wait for more */
+
+                /* decode loaded input */
+                {   size_t const decodedSize = ZSTD_decompressContinue(zds,
+                        zds->outBuff + zds->outStart, zds->outBuffSize - zds->outStart,
+                        zds->inBuff, neededInSize);
+                    if (ZSTD_isError(decodedSize)) return decodedSize;
+                    zds->inPos = 0;   /* input is consumed */
+                    if (!decodedSize && !isSkipFrame) { zds->streamStage = zdss_read; break; }   /* this was just a header */
+                    zds->outEnd = zds->outStart +  decodedSize;
+            }   }
+            zds->streamStage = zdss_flush;
+            /* fall-through */
+
+        case zdss_flush:
+            {   size_t const toFlushSize = zds->outEnd - zds->outStart;
+                size_t const flushedSize = ZSTD_limitCopy(op, oend-op, zds->outBuff + zds->outStart, toFlushSize);
+                op += flushedSize;
+                zds->outStart += flushedSize;
+                if (flushedSize == toFlushSize) {  /* flush completed */
+                    zds->streamStage = zdss_read;
+                    if ( (zds->outBuffSize < zds->fParams.frameContentSize)
+                      && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
+                        DEBUGLOG(5, "restart filling outBuff from beginning (left:%i, needed:%u)",
+                                (int)(zds->outBuffSize - zds->outStart),
+                                (U32)zds->fParams.blockSizeMax);
+                        zds->outStart = zds->outEnd = 0;
+                    }
+                    break;
+            }   }
+            /* cannot complete flush */
+            someMoreWork = 0;
+            break;
+
+        default: return ERROR(GENERIC);   /* impossible */
+    }   }
+
+    /* result */
+    input->pos += (size_t)(ip-istart);
+    output->pos += (size_t)(op-ostart);
+    {   size_t nextSrcSizeHint = ZSTD_nextSrcSizeToDecompress(zds);
+        if (!nextSrcSizeHint) {   /* frame fully decoded */
+            if (zds->outEnd == zds->outStart) {  /* output fully flushed */
+                if (zds->hostageByte) {
+                    if (input->pos >= input->size) {
+                        /* can't release hostage (not present) */
+                        zds->streamStage = zdss_read;
+                        return 1;
+                    }
+                    input->pos++;  /* release hostage */
+                }   /* zds->hostageByte */
+                return 0;
+            }  /* zds->outEnd == zds->outStart */
+            if (!zds->hostageByte) { /* output not fully flushed; keep last byte as hostage; will be released when all output is flushed */
+                input->pos--;   /* note : pos > 0, otherwise, impossible to finish reading last block */
+                zds->hostageByte=1;
+            }
+            return 1;
+        }  /* nextSrcSizeHint==0 */
+        nextSrcSizeHint += ZSTD_blockHeaderSize * (ZSTD_nextInputType(zds) == ZSTDnit_block);   /* preload header of next block */
+        assert(zds->inPos <= nextSrcSizeHint);
+        nextSrcSizeHint -= zds->inPos;   /* part already loaded*/
+        return nextSrcSizeHint;
+    }
+}
+
+
+size_t ZSTD_decompress_generic(ZSTD_DCtx* dctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
+{
+    return ZSTD_decompressStream(dctx, output, input);
+}
+
+size_t ZSTD_decompress_generic_simpleArgs (
+                            ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity, size_t* dstPos,
+                      const void* src, size_t srcSize, size_t* srcPos)
+{
+    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
+    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
+    /* ZSTD_compress_generic() will check validity of dstPos and srcPos */
+    size_t const cErr = ZSTD_decompress_generic(dctx, &output, &input);
+    *dstPos = output.pos;
+    *srcPos = input.pos;
+    return cErr;
+}
+
+void ZSTD_DCtx_reset(ZSTD_DCtx* dctx)
+{
+    (void)ZSTD_initDStream(dctx);
+    dctx->format = ZSTD_f_zstd1;
+    dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_double_fast.c b/vendor/github.com/DataDog/zstd/zstd_double_fast.c
new file mode 100644
index 0000000..86e6b39
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_double_fast.c
@@ -0,0 +1,327 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#include "zstd_compress_internal.h"
+#include "zstd_double_fast.h"
+
+
+void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+                              ZSTD_compressionParameters const* cParams,
+                              void const* end)
+{
+    U32* const hashLarge = ms->hashTable;
+    U32  const hBitsL = cParams->hashLog;
+    U32  const mls = cParams->searchLength;
+    U32* const hashSmall = ms->chainTable;
+    U32  const hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
+
+    /* Always insert every fastHashFillStep position into the hash tables.
+     * Insert the other positions into the large hash table if their entry
+     * is empty.
+     */
+    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
+        U32 const current = (U32)(ip - base);
+        U32 i;
+        for (i = 0; i < fastHashFillStep; ++i) {
+            size_t const smHash = ZSTD_hashPtr(ip + i, hBitsS, mls);
+            size_t const lgHash = ZSTD_hashPtr(ip + i, hBitsL, 8);
+            if (i == 0)
+                hashSmall[smHash] = current + i;
+            if (i == 0 || hashLarge[lgHash] == 0)
+                hashLarge[lgHash] = current + i;
+        }
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_doubleFast_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize,
+        U32 const mls /* template */)
+{
+    U32* const hashLong = ms->hashTable;
+    const U32 hBitsL = cParams->hashLog;
+    U32* const hashSmall = ms->chainTable;
+    const U32 hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const U32 lowestIndex = ms->window.dictLimit;
+    const BYTE* const lowest = base + lowestIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved = 0;
+
+    /* init */
+    ip += (ip==lowest);
+    {   U32 const maxRep = (U32)(ip-lowest);
+        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+    }
+
+    /* Main Search Loop */
+    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+        size_t mLength;
+        size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8);
+        size_t const h = ZSTD_hashPtr(ip, hBitsS, mls);
+        U32 const current = (U32)(ip-base);
+        U32 const matchIndexL = hashLong[h2];
+        U32 const matchIndexS = hashSmall[h];
+        const BYTE* matchLong = base + matchIndexL;
+        const BYTE* match = base + matchIndexS;
+        hashLong[h2] = hashSmall[h] = current;   /* update hash tables */
+
+        assert(offset_1 <= current);   /* supposed guaranteed by construction */
+        if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
+            /* favor repcode */
+            mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+            ip++;
+            ZSTD_storeSeq(seqStore, ip-anchor, anchor, 0, mLength-MINMATCH);
+        } else {
+            U32 offset;
+            if ( (matchIndexL > lowestIndex) && (MEM_read64(matchLong) == MEM_read64(ip)) ) {
+                mLength = ZSTD_count(ip+8, matchLong+8, iend) + 8;
+                offset = (U32)(ip-matchLong);
+                while (((ip>anchor) & (matchLong>lowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
+            } else if ( (matchIndexS > lowestIndex) && (MEM_read32(match) == MEM_read32(ip)) ) {
+                size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
+                U32 const matchIndexL3 = hashLong[hl3];
+                const BYTE* matchL3 = base + matchIndexL3;
+                hashLong[hl3] = current + 1;
+                if ( (matchIndexL3 > lowestIndex) && (MEM_read64(matchL3) == MEM_read64(ip+1)) ) {
+                    mLength = ZSTD_count(ip+9, matchL3+8, iend) + 8;
+                    ip++;
+                    offset = (U32)(ip-matchL3);
+                    while (((ip>anchor) & (matchL3>lowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
+                } else {
+                    mLength = ZSTD_count(ip+4, match+4, iend) + 4;
+                    offset = (U32)(ip-match);
+                    while (((ip>anchor) & (match>lowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
+                }
+            } else {
+                ip += ((ip-anchor) >> kSearchStrength) + 1;
+                continue;
+            }
+
+            offset_2 = offset_1;
+            offset_1 = offset;
+
+            ZSTD_storeSeq(seqStore, ip-anchor, anchor, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+        }
+
+        /* match found */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Fill Table */
+            hashLong[ZSTD_hashPtr(base+current+2, hBitsL, 8)] =
+                hashSmall[ZSTD_hashPtr(base+current+2, hBitsS, mls)] = current+2;  /* here because current+2 could be > iend-8 */
+            hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] =
+                hashSmall[ZSTD_hashPtr(ip-2, hBitsS, mls)] = (U32)(ip-2-base);
+
+            /* check immediate repcode */
+            while ( (ip <= ilimit)
+                 && ( (offset_2>0)
+                 & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+                { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */
+                hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
+                hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
+                ZSTD_storeSeq(seqStore, 0, anchor, 0, rLength-MINMATCH);
+                ip += rLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+    }   }   }
+
+    /* save reps for next block */
+    rep[0] = offset_1 ? offset_1 : offsetSaved;
+    rep[1] = offset_2 ? offset_2 : offsetSaved;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_doubleFast(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    const U32 mls = cParams->searchLength;
+    switch(mls)
+    {
+    default: /* includes case 3 */
+    case 4 :
+        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, cParams, src, srcSize, 4);
+    case 5 :
+        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, cParams, src, srcSize, 5);
+    case 6 :
+        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, cParams, src, srcSize, 6);
+    case 7 :
+        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, cParams, src, srcSize, 7);
+    }
+}
+
+
+static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize,
+        U32 const mls /* template */)
+{
+    U32* const hashLong = ms->hashTable;
+    U32  const hBitsL = cParams->hashLog;
+    U32* const hashSmall = ms->chainTable;
+    U32  const hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const U32   lowestIndex = ms->window.lowLimit;
+    const BYTE* const dictStart = dictBase + lowestIndex;
+    const U32   dictLimit = ms->window.dictLimit;
+    const BYTE* const lowPrefixPtr = base + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    U32 offset_1=rep[0], offset_2=rep[1];
+
+    /* Search Loop */
+    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
+        const size_t hSmall = ZSTD_hashPtr(ip, hBitsS, mls);
+        const U32 matchIndex = hashSmall[hSmall];
+        const BYTE* matchBase = matchIndex < dictLimit ? dictBase : base;
+        const BYTE* match = matchBase + matchIndex;
+
+        const size_t hLong = ZSTD_hashPtr(ip, hBitsL, 8);
+        const U32 matchLongIndex = hashLong[hLong];
+        const BYTE* matchLongBase = matchLongIndex < dictLimit ? dictBase : base;
+        const BYTE* matchLong = matchLongBase + matchLongIndex;
+
+        const U32 current = (U32)(ip-base);
+        const U32 repIndex = current + 1 - offset_1;   /* offset_1 expected <= current +1 */
+        const BYTE* repBase = repIndex < dictLimit ? dictBase : base;
+        const BYTE* repMatch = repBase + repIndex;
+        size_t mLength;
+        hashSmall[hSmall] = hashLong[hLong] = current;   /* update hash table */
+
+        if ( (((U32)((dictLimit-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > lowestIndex))
+           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
+            const BYTE* repMatchEnd = repIndex < dictLimit ? dictEnd : iend;
+            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, lowPrefixPtr) + 4;
+            ip++;
+            ZSTD_storeSeq(seqStore, ip-anchor, anchor, 0, mLength-MINMATCH);
+        } else {
+            if ((matchLongIndex > lowestIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
+                const BYTE* matchEnd = matchLongIndex < dictLimit ? dictEnd : iend;
+                const BYTE* lowMatchPtr = matchLongIndex < dictLimit ? dictStart : lowPrefixPtr;
+                U32 offset;
+                mLength = ZSTD_count_2segments(ip+8, matchLong+8, iend, matchEnd, lowPrefixPtr) + 8;
+                offset = current - matchLongIndex;
+                while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; }   /* catch up */
+                offset_2 = offset_1;
+                offset_1 = offset;
+                ZSTD_storeSeq(seqStore, ip-anchor, anchor, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+
+            } else if ((matchIndex > lowestIndex) && (MEM_read32(match) == MEM_read32(ip))) {
+                size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
+                U32 const matchIndex3 = hashLong[h3];
+                const BYTE* const match3Base = matchIndex3 < dictLimit ? dictBase : base;
+                const BYTE* match3 = match3Base + matchIndex3;
+                U32 offset;
+                hashLong[h3] = current + 1;
+                if ( (matchIndex3 > lowestIndex) && (MEM_read64(match3) == MEM_read64(ip+1)) ) {
+                    const BYTE* matchEnd = matchIndex3 < dictLimit ? dictEnd : iend;
+                    const BYTE* lowMatchPtr = matchIndex3 < dictLimit ? dictStart : lowPrefixPtr;
+                    mLength = ZSTD_count_2segments(ip+9, match3+8, iend, matchEnd, lowPrefixPtr) + 8;
+                    ip++;
+                    offset = current+1 - matchIndex3;
+                    while (((ip>anchor) & (match3>lowMatchPtr)) && (ip[-1] == match3[-1])) { ip--; match3--; mLength++; } /* catch up */
+                } else {
+                    const BYTE* matchEnd = matchIndex < dictLimit ? dictEnd : iend;
+                    const BYTE* lowMatchPtr = matchIndex < dictLimit ? dictStart : lowPrefixPtr;
+                    mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, lowPrefixPtr) + 4;
+                    offset = current - matchIndex;
+                    while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
+                }
+                offset_2 = offset_1;
+                offset_1 = offset;
+                ZSTD_storeSeq(seqStore, ip-anchor, anchor, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+
+            } else {
+                ip += ((ip-anchor) >> kSearchStrength) + 1;
+                continue;
+        }   }
+
+        /* found a match : store it */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Fill Table */
+            hashSmall[ZSTD_hashPtr(base+current+2, hBitsS, mls)] = current+2;
+            hashLong[ZSTD_hashPtr(base+current+2, hBitsL, 8)] = current+2;
+            hashSmall[ZSTD_hashPtr(ip-2, hBitsS, mls)] = (U32)(ip-2-base);
+            hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
+            /* check immediate repcode */
+            while (ip <= ilimit) {
+                U32 const current2 = (U32)(ip-base);
+                U32 const repIndex2 = current2 - offset_2;
+                const BYTE* repMatch2 = repIndex2 < dictLimit ? dictBase + repIndex2 : base + repIndex2;
+                if ( (((U32)((dictLimit-1) - repIndex2) >= 3) & (repIndex2 > lowestIndex))  /* intentional overflow */
+                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                    const BYTE* const repEnd2 = repIndex2 < dictLimit ? dictEnd : iend;
+                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, lowPrefixPtr) + 4;
+                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, 0, repLength2-MINMATCH);
+                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
+                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
+                    ip += repLength2;
+                    anchor = ip;
+                    continue;
+                }
+                break;
+    }   }   }
+
+    /* save reps for next block */
+    rep[0] = offset_1;
+    rep[1] = offset_2;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_doubleFast_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    U32 const mls = cParams->searchLength;
+    switch(mls)
+    {
+    default: /* includes case 3 */
+    case 4 :
+        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 4);
+    case 5 :
+        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 5);
+    case 6 :
+        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 6);
+    case 7 :
+        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 7);
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_double_fast.h b/vendor/github.com/DataDog/zstd/zstd_double_fast.h
new file mode 100644
index 0000000..6d80b27
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_double_fast.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_DOUBLE_FAST_H
+#define ZSTD_DOUBLE_FAST_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include "mem.h"      /* U32 */
+#include "zstd_compress_internal.h"     /* ZSTD_CCtx, size_t */
+
+void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+                              ZSTD_compressionParameters const* cParams,
+                              void const* end);
+size_t ZSTD_compressBlock_doubleFast(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_doubleFast_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_DOUBLE_FAST_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_errors.h b/vendor/github.com/DataDog/zstd/zstd_errors.h
new file mode 100644
index 0000000..57533f2
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_errors.h
@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_ERRORS_H_398273423
+#define ZSTD_ERRORS_H_398273423
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*===== dependency =====*/
+#include <stddef.h>   /* size_t */
+
+
+/* =====   ZSTDERRORLIB_API : control library symbols visibility   ===== */
+#ifndef ZSTDERRORLIB_VISIBILITY
+#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#    define ZSTDERRORLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  else
+#    define ZSTDERRORLIB_VISIBILITY
+#  endif
+#endif
+#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBILITY
+#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBILITY
+#endif
+
+/*-*********************************************
+ *  Error codes list
+ *-*********************************************
+ *  Error codes _values_ are pinned down since v1.3.1 only.
+ *  Therefore, don't rely on values if you may link to any version < v1.3.1.
+ *
+ *  Only values < 100 are considered stable.
+ *
+ *  note 1 : this API shall be used with static linking only.
+ *           dynamic linking is not yet officially supported.
+ *  note 2 : Prefer relying on the enum than on its value whenever possible
+ *           This is the only supported way to use the error list < v1.3.1
+ *  note 3 : ZSTD_isError() is always correct, whatever the library version.
+ **********************************************/
+typedef enum {
+  ZSTD_error_no_error = 0,
+  ZSTD_error_GENERIC  = 1,
+  ZSTD_error_prefix_unknown                = 10,
+  ZSTD_error_version_unsupported           = 12,
+  ZSTD_error_frameParameter_unsupported    = 14,
+  ZSTD_error_frameParameter_windowTooLarge = 16,
+  ZSTD_error_corruption_detected = 20,
+  ZSTD_error_checksum_wrong      = 22,
+  ZSTD_error_dictionary_corrupted      = 30,
+  ZSTD_error_dictionary_wrong          = 32,
+  ZSTD_error_dictionaryCreation_failed = 34,
+  ZSTD_error_parameter_unsupported   = 40,
+  ZSTD_error_parameter_outOfBound    = 42,
+  ZSTD_error_tableLog_tooLarge       = 44,
+  ZSTD_error_maxSymbolValue_tooLarge = 46,
+  ZSTD_error_maxSymbolValue_tooSmall = 48,
+  ZSTD_error_stage_wrong       = 60,
+  ZSTD_error_init_missing      = 62,
+  ZSTD_error_memory_allocation = 64,
+  ZSTD_error_workSpace_tooSmall= 66,
+  ZSTD_error_dstSize_tooSmall = 70,
+  ZSTD_error_srcSize_wrong    = 72,
+  /* following error codes are __NOT STABLE__, they can be removed or changed in future versions */
+  ZSTD_error_frameIndex_tooLarge = 100,
+  ZSTD_error_seekableIO          = 102,
+  ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
+} ZSTD_ErrorCode;
+
+/*! ZSTD_getErrorCode() :
+    convert a `size_t` function result into a `ZSTD_ErrorCode` enum type,
+    which can be used to compare with enum list published above */
+ZSTDERRORLIB_API ZSTD_ErrorCode ZSTD_getErrorCode(size_t functionResult);
+ZSTDERRORLIB_API const char* ZSTD_getErrorString(ZSTD_ErrorCode code);   /**< Same as ZSTD_getErrorName, but using a `ZSTD_ErrorCode` enum argument */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_ERRORS_H_398273423 */
diff --git a/vendor/github.com/DataDog/zstd/zstd_fast.c b/vendor/github.com/DataDog/zstd/zstd_fast.c
new file mode 100644
index 0000000..df4d28b
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_fast.c
@@ -0,0 +1,259 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#include "zstd_compress_internal.h"
+#include "zstd_fast.h"
+
+
+void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+                        ZSTD_compressionParameters const* cParams,
+                        void const* end)
+{
+    U32* const hashTable = ms->hashTable;
+    U32  const hBits = cParams->hashLog;
+    U32  const mls = cParams->searchLength;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
+
+    /* Always insert every fastHashFillStep position into the hash table.
+     * Insert the other positions if their hash entry is empty.
+     */
+    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
+        U32 const current = (U32)(ip - base);
+        U32 i;
+        for (i = 0; i < fastHashFillStep; ++i) {
+            size_t const hash = ZSTD_hashPtr(ip + i, hBits, mls);
+            if (i == 0 || hashTable[hash] == 0)
+                hashTable[hash] = current + i;
+        }
+    }
+}
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_fast_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize,
+        U32 const hlog, U32 const stepSize, U32 const mls)
+{
+    U32* const hashTable = ms->hashTable;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const U32   lowestIndex = ms->window.dictLimit;
+    const BYTE* const lowest = base + lowestIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved = 0;
+
+    /* init */
+    ip += (ip==lowest);
+    {   U32 const maxRep = (U32)(ip-lowest);
+        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+    }
+
+    /* Main Search Loop */
+    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+        size_t mLength;
+        size_t const h = ZSTD_hashPtr(ip, hlog, mls);
+        U32 const current = (U32)(ip-base);
+        U32 const matchIndex = hashTable[h];
+        const BYTE* match = base + matchIndex;
+        hashTable[h] = current;   /* update hash table */
+
+        if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
+            mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+            ip++;
+            ZSTD_storeSeq(seqStore, ip-anchor, anchor, 0, mLength-MINMATCH);
+        } else {
+            if ( (matchIndex <= lowestIndex)
+              || (MEM_read32(match) != MEM_read32(ip)) ) {
+                assert(stepSize >= 1);
+                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
+                continue;
+            }
+            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
+            {   U32 const offset = (U32)(ip-match);
+                while (((ip>anchor) & (match>lowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
+                offset_2 = offset_1;
+                offset_1 = offset;
+                ZSTD_storeSeq(seqStore, ip-anchor, anchor, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+        }   }
+
+        /* match found */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Fill Table */
+            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;  /* here because current+2 could be > iend-8 */
+            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+            /* check immediate repcode */
+            while ( (ip <= ilimit)
+                 && ( (offset_2>0)
+                 & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+                { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; }  /* swap offset_2 <=> offset_1 */
+                hashTable[ZSTD_hashPtr(ip, hlog, mls)] = (U32)(ip-base);
+                ZSTD_storeSeq(seqStore, 0, anchor, 0, rLength-MINMATCH);
+                ip += rLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+    }   }   }
+
+    /* save reps for next block */
+    rep[0] = offset_1 ? offset_1 : offsetSaved;
+    rep[1] = offset_2 ? offset_2 : offsetSaved;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_fast(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    U32 const hlog = cParams->hashLog;
+    U32 const mls = cParams->searchLength;
+    U32 const stepSize = cParams->targetLength;
+    switch(mls)
+    {
+    default: /* includes case 3 */
+    case 4 :
+        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 4);
+    case 5 :
+        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 5);
+    case 6 :
+        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 6);
+    case 7 :
+        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 7);
+    }
+}
+
+
+static size_t ZSTD_compressBlock_fast_extDict_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize,
+        U32 const hlog, U32 const stepSize, U32 const mls)
+{
+    U32* hashTable = ms->hashTable;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const U32   lowestIndex = ms->window.lowLimit;
+    const BYTE* const dictStart = dictBase + lowestIndex;
+    const U32   dictLimit = ms->window.dictLimit;
+    const BYTE* const lowPrefixPtr = base + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    U32 offset_1=rep[0], offset_2=rep[1];
+
+    /* Search Loop */
+    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
+        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
+        const U32 matchIndex = hashTable[h];
+        const BYTE* matchBase = matchIndex < dictLimit ? dictBase : base;
+        const BYTE* match = matchBase + matchIndex;
+        const U32 current = (U32)(ip-base);
+        const U32 repIndex = current + 1 - offset_1;   /* offset_1 expected <= current +1 */
+        const BYTE* repBase = repIndex < dictLimit ? dictBase : base;
+        const BYTE* repMatch = repBase + repIndex;
+        size_t mLength;
+        hashTable[h] = current;   /* update hash table */
+
+        if ( (((U32)((dictLimit-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > lowestIndex))
+           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
+            const BYTE* repMatchEnd = repIndex < dictLimit ? dictEnd : iend;
+            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, lowPrefixPtr) + 4;
+            ip++;
+            ZSTD_storeSeq(seqStore, ip-anchor, anchor, 0, mLength-MINMATCH);
+        } else {
+            if ( (matchIndex < lowestIndex) ||
+                 (MEM_read32(match) != MEM_read32(ip)) ) {
+                assert(stepSize >= 1);
+                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
+                continue;
+            }
+            {   const BYTE* matchEnd = matchIndex < dictLimit ? dictEnd : iend;
+                const BYTE* lowMatchPtr = matchIndex < dictLimit ? dictStart : lowPrefixPtr;
+                U32 offset;
+                mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, lowPrefixPtr) + 4;
+                while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
+                offset = current - matchIndex;
+                offset_2 = offset_1;
+                offset_1 = offset;
+                ZSTD_storeSeq(seqStore, ip-anchor, anchor, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+        }   }
+
+        /* found a match : store it */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Fill Table */
+            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;
+            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+            /* check immediate repcode */
+            while (ip <= ilimit) {
+                U32 const current2 = (U32)(ip-base);
+                U32 const repIndex2 = current2 - offset_2;
+                const BYTE* repMatch2 = repIndex2 < dictLimit ? dictBase + repIndex2 : base + repIndex2;
+                if ( (((U32)((dictLimit-1) - repIndex2) >= 3) & (repIndex2 > lowestIndex))  /* intentional overflow */
+                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                    const BYTE* const repEnd2 = repIndex2 < dictLimit ? dictEnd : iend;
+                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, lowPrefixPtr) + 4;
+                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, 0, repLength2-MINMATCH);
+                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
+                    ip += repLength2;
+                    anchor = ip;
+                    continue;
+                }
+                break;
+    }   }   }
+
+    /* save reps for next block */
+    rep[0] = offset_1;
+    rep[1] = offset_2;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_fast_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    U32 const hlog = cParams->hashLog;
+    U32 const mls = cParams->searchLength;
+    U32 const stepSize = cParams->targetLength;
+    switch(mls)
+    {
+    default: /* includes case 3 */
+    case 4 :
+        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 4);
+    case 5 :
+        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 5);
+    case 6 :
+        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 6);
+    case 7 :
+        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, hlog, stepSize, 7);
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_fast.h b/vendor/github.com/DataDog/zstd/zstd_fast.h
new file mode 100644
index 0000000..f0438ad
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_fast.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_FAST_H
+#define ZSTD_FAST_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include "mem.h"      /* U32 */
+#include "zstd_compress_internal.h"
+
+void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+                        ZSTD_compressionParameters const* cParams,
+                        void const* end);
+size_t ZSTD_compressBlock_fast(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_fast_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_FAST_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_internal.h b/vendor/github.com/DataDog/zstd/zstd_internal.h
new file mode 100644
index 0000000..65c08a8
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_internal.h
@@ -0,0 +1,290 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_CCOMMON_H_MODULE
+#define ZSTD_CCOMMON_H_MODULE
+
+/* this module contains definitions which must be identical
+ * across compression, decompression and dictBuilder.
+ * It also contains a few functions useful to at least 2 of them
+ * and which benefit from being inlined */
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include "compiler.h"
+#include "mem.h"
+#include "error_private.h"
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#define HUF_STATIC_LINKING_ONLY
+#include "huf.h"
+#ifndef XXH_STATIC_LINKING_ONLY
+#  define XXH_STATIC_LINKING_ONLY  /* XXH64_state_t */
+#endif
+#include "xxhash.h"                /* XXH_reset, update, digest */
+
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*-*************************************
+*  Debug
+***************************************/
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
+#define ZSTD_STATIC_ASSERT(c) { enum { ZSTD_static_assert = 1/(int)(!!(c)) }; }
+
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
+#  include <stdio.h>
+extern int g_debuglog_enable;
+/* recommended values for ZSTD_DEBUG display levels :
+ * 1 : no display, enables assert() only
+ * 2 : reserved for currently active debug path
+ * 3 : events once per object lifetime (CCtx, CDict, etc.)
+ * 4 : events once per frame
+ * 5 : events once per block
+ * 6 : events once per sequence (*very* verbose) */
+#  define RAWLOG(l, ...) {                                      \
+                if ((g_debuglog_enable) & (l<=ZSTD_DEBUG)) {    \
+                    fprintf(stderr, __VA_ARGS__);               \
+            }   }
+#  define DEBUGLOG(l, ...) {                                    \
+                if ((g_debuglog_enable) & (l<=ZSTD_DEBUG)) {    \
+                    fprintf(stderr, __FILE__ ": " __VA_ARGS__); \
+                    fprintf(stderr, " \n");                     \
+            }   }
+#else
+#  define RAWLOG(l, ...)      {}    /* disabled */
+#  define DEBUGLOG(l, ...)    {}    /* disabled */
+#endif
+
+
+/*-*************************************
+*  shared macros
+***************************************/
+#undef MIN
+#undef MAX
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+#define CHECK_F(f) { size_t const errcod = f; if (ERR_isError(errcod)) return errcod; }  /* check and Forward error code */
+#define CHECK_E(f, e) { size_t const errcod = f; if (ERR_isError(errcod)) return ERROR(e); }  /* check and send Error code */
+
+
+/*-*************************************
+*  Common constants
+***************************************/
+#define ZSTD_OPT_NUM    (1<<12)
+
+#define ZSTD_REP_NUM      3                 /* number of repcodes */
+#define ZSTD_REP_MOVE     (ZSTD_REP_NUM-1)
+static const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define ZSTD_WINDOWLOG_ABSOLUTEMIN 10
+#define ZSTD_WINDOWLOG_DEFAULTMAX 27 /* Default maximum allowed window log */
+static const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
+static const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
+
+#define ZSTD_FRAMEIDSIZE 4
+static const size_t ZSTD_frameIdSize = ZSTD_FRAMEIDSIZE;  /* magic number size */
+
+#define ZSTD_BLOCKHEADERSIZE 3   /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
+static const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
+typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;
+
+#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+
+#define HufLog 12
+typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingType_e;
+
+#define LONGNBSEQ 0x7F00
+
+#define MINMATCH 3
+
+#define Litbits  8
+#define MaxLit ((1<<Litbits) - 1)
+#define MaxML   52
+#define MaxLL   35
+#define DefaultMaxOff 28
+#define MaxOff  31
+#define MaxSeq MAX(MaxLL, MaxML)   /* Assumption : MaxOff < MaxLL,MaxML */
+#define MLFSELog    9
+#define LLFSELog    9
+#define OffFSELog   8
+#define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
+
+static const U32 LL_bits[MaxLL+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3,
+                                      4, 6, 7, 8, 9,10,11,12,
+                                     13,14,15,16 };
+static const S16 LL_defaultNorm[MaxLL+1] = { 4, 3, 2, 2, 2, 2, 2, 2,
+                                             2, 2, 2, 2, 2, 1, 1, 1,
+                                             2, 2, 2, 2, 2, 2, 2, 2,
+                                             2, 3, 2, 1, 1, 1, 1, 1,
+                                            -1,-1,-1,-1 };
+#define LL_DEFAULTNORMLOG 6  /* for static allocation */
+static const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
+
+static const U32 ML_bits[MaxML+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3,
+                                      4, 4, 5, 7, 8, 9,10,11,
+                                     12,13,14,15,16 };
+static const S16 ML_defaultNorm[MaxML+1] = { 1, 4, 3, 2, 2, 2, 2, 2,
+                                             2, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1,-1,-1,
+                                            -1,-1,-1,-1,-1 };
+#define ML_DEFAULTNORMLOG 6  /* for static allocation */
+static const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
+
+static const S16 OF_defaultNorm[DefaultMaxOff+1] = { 1, 1, 1, 1, 1, 1, 2, 2,
+                                                     2, 1, 1, 1, 1, 1, 1, 1,
+                                                     1, 1, 1, 1, 1, 1, 1, 1,
+                                                    -1,-1,-1,-1,-1 };
+#define OF_DEFAULTNORMLOG 5  /* for static allocation */
+static const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
+
+
+/*-*******************************************
+*  Shared functions to include for inlining
+*********************************************/
+static void ZSTD_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTD_wildcopy() :
+ *  custom version of memcpy(), can overwrite up to WILDCOPY_OVERLENGTH bytes (if length==0) */
+#define WILDCOPY_OVERLENGTH 8
+MEM_STATIC void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+MEM_STATIC void ZSTD_wildcopy_e(void* dst, const void* src, void* dstEnd)   /* should be faster for decoding, but strangely, not verified on all platform */
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = (BYTE*)dstEnd;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+
+/*-*******************************************
+*  Private declarations
+*********************************************/
+typedef struct seqDef_s {
+    U32 offset;
+    U16 litLength;
+    U16 matchLength;
+} seqDef;
+
+typedef struct {
+    seqDef* sequencesStart;
+    seqDef* sequences;
+    BYTE* litStart;
+    BYTE* lit;
+    BYTE* llCode;
+    BYTE* mlCode;
+    BYTE* ofCode;
+    U32   longLengthID;   /* 0 == no longLength; 1 == Lit.longLength; 2 == Match.longLength; */
+    U32   longLengthPos;
+} seqStore_t;
+
+const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
+void ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
+
+/* custom memory allocation functions */
+void* ZSTD_malloc(size_t size, ZSTD_customMem customMem);
+void* ZSTD_calloc(size_t size, ZSTD_customMem customMem);
+void ZSTD_free(void* ptr, ZSTD_customMem customMem);
+
+
+MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
+{
+    assert(val != 0);
+    {
+#   if defined(_MSC_VER)   /* Visual */
+        unsigned long r=0;
+        _BitScanReverse(&r, val);
+        return (unsigned)r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* GCC Intrinsic */
+        return 31 - __builtin_clz(val);
+#   else   /* Software version */
+        static const U32 DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+        U32 v = val;
+        v |= v >> 1;
+        v |= v >> 2;
+        v |= v >> 4;
+        v |= v >> 8;
+        v |= v >> 16;
+        return DeBruijnClz[(v * 0x07C4ACDDU) >> 27];
+#   endif
+    }
+}
+
+
+/* ZSTD_invalidateRepCodes() :
+ * ensures next compression will not use repcodes from previous block.
+ * Note : only works with regular variant;
+ *        do not use with extDict variant ! */
+void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx);   /* zstdmt, adaptive_compression (shouldn't get this definition from here) */
+
+
+typedef struct {
+    blockType_e blockType;
+    U32 lastBlock;
+    U32 origSize;
+} blockProperties_t;
+
+/*! ZSTD_getcBlockSize() :
+ *  Provides the size of compressed block from block header `src` */
+/* Used by: decompress, fullbench (does not get its definition from here) */
+size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
+                          blockProperties_t* bpPtr);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* ZSTD_CCOMMON_H_MODULE */
diff --git a/vendor/github.com/DataDog/zstd/zstd_lazy.c b/vendor/github.com/DataDog/zstd/zstd_lazy.c
new file mode 100644
index 0000000..9f15812
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_lazy.c
@@ -0,0 +1,824 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#include "zstd_compress_internal.h"
+#include "zstd_lazy.h"
+
+
+/*-*************************************
+*  Binary Tree search
+***************************************/
+
+void ZSTD_updateDUBT(
+                ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                const BYTE* ip, const BYTE* iend,
+                U32 mls)
+{
+    U32* const hashTable = ms->hashTable;
+    U32  const hashLog = cParams->hashLog;
+
+    U32* const bt = ms->chainTable;
+    U32  const btLog  = cParams->chainLog - 1;
+    U32  const btMask = (1 << btLog) - 1;
+
+    const BYTE* const base = ms->window.base;
+    U32 const target = (U32)(ip - base);
+    U32 idx = ms->nextToUpdate;
+
+    if (idx != target)
+        DEBUGLOG(7, "ZSTD_updateDUBT, from %u to %u (dictLimit:%u)",
+                    idx, target, ms->window.dictLimit);
+    assert(ip + 8 <= iend);   /* condition for ZSTD_hashPtr */
+    (void)iend;
+
+    assert(idx >= ms->window.dictLimit);   /* condition for valid base+idx */
+    for ( ; idx < target ; idx++) {
+        size_t const h  = ZSTD_hashPtr(base + idx, hashLog, mls);   /* assumption : ip + 8 <= iend */
+        U32    const matchIndex = hashTable[h];
+
+        U32*   const nextCandidatePtr = bt + 2*(idx&btMask);
+        U32*   const sortMarkPtr  = nextCandidatePtr + 1;
+
+        DEBUGLOG(8, "ZSTD_updateDUBT: insert %u", idx);
+        hashTable[h] = idx;   /* Update Hash Table */
+        *nextCandidatePtr = matchIndex;   /* update BT like a chain */
+        *sortMarkPtr = ZSTD_DUBT_UNSORTED_MARK;
+    }
+    ms->nextToUpdate = target;
+}
+
+
+/** ZSTD_insertDUBT1() :
+ *  sort one already inserted but unsorted position
+ *  assumption : current >= btlow == (current - btmask)
+ *  doesn't fail */
+static void ZSTD_insertDUBT1(
+                 ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                 U32 current, const BYTE* inputEnd,
+                 U32 nbCompares, U32 btLow, int extDict)
+{
+    U32*   const bt = ms->chainTable;
+    U32    const btLog  = cParams->chainLog - 1;
+    U32    const btMask = (1 << btLog) - 1;
+    size_t commonLengthSmaller=0, commonLengthLarger=0;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const U32 dictLimit = ms->window.dictLimit;
+    const BYTE* const ip = (current>=dictLimit) ? base + current : dictBase + current;
+    const BYTE* const iend = (current>=dictLimit) ? inputEnd : dictBase + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* match;
+    U32* smallerPtr = bt + 2*(current&btMask);
+    U32* largerPtr  = smallerPtr + 1;
+    U32 matchIndex = *smallerPtr;
+    U32 dummy32;   /* to be nullified at the end */
+    U32 const windowLow = ms->window.lowLimit;
+
+    DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
+                current, dictLimit, windowLow);
+    assert(current >= btLow);
+    assert(ip < iend);   /* condition for ZSTD_count */
+
+    while (nbCompares-- && (matchIndex > windowLow)) {
+        U32* const nextPtr = bt + 2*(matchIndex & btMask);
+        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
+        assert(matchIndex < current);
+
+        if ( (!extDict)
+          || (matchIndex+matchLength >= dictLimit)  /* both in current segment*/
+          || (current < dictLimit) /* both in extDict */) {
+            const BYTE* const mBase = !extDict || ((matchIndex+matchLength) >= dictLimit) ? base : dictBase;
+            assert( (matchIndex+matchLength >= dictLimit)   /* might be wrong if extDict is incorrectly set to 0 */
+                 || (current < dictLimit) );
+            match = mBase + matchIndex;
+            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
+        } else {
+            match = dictBase + matchIndex;
+            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
+            if (matchIndex+matchLength >= dictLimit)
+                match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
+        }
+
+        DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
+                    current, matchIndex, (U32)matchLength);
+
+        if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
+            break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
+        }
+
+        if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
+            /* match is smaller than current */
+            *smallerPtr = matchIndex;             /* update smaller idx */
+            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
+            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
+            DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is smaller : next => %u",
+                        matchIndex, btLow, nextPtr[1]);
+            smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
+            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
+        } else {
+            /* match is larger than current */
+            *largerPtr = matchIndex;
+            commonLengthLarger = matchLength;
+            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
+            DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is larger => %u",
+                        matchIndex, btLow, nextPtr[0]);
+            largerPtr = nextPtr;
+            matchIndex = nextPtr[0];
+    }   }
+
+    *smallerPtr = *largerPtr = 0;
+}
+
+
+static size_t ZSTD_DUBT_findBestMatch (
+                            ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                            const BYTE* const ip, const BYTE* const iend,
+                            size_t* offsetPtr,
+                            U32 const mls,
+                            U32 const extDict)
+{
+    U32*   const hashTable = ms->hashTable;
+    U32    const hashLog = cParams->hashLog;
+    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
+    U32          matchIndex  = hashTable[h];
+
+    const BYTE* const base = ms->window.base;
+    U32    const current = (U32)(ip-base);
+    U32    const windowLow = ms->window.lowLimit;
+
+    U32*   const bt = ms->chainTable;
+    U32    const btLog  = cParams->chainLog - 1;
+    U32    const btMask = (1 << btLog) - 1;
+    U32    const btLow = (btMask >= current) ? 0 : current - btMask;
+    U32    const unsortLimit = MAX(btLow, windowLow);
+
+    U32*         nextCandidate = bt + 2*(matchIndex&btMask);
+    U32*         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
+    U32          nbCompares = 1U << cParams->searchLog;
+    U32          nbCandidates = nbCompares;
+    U32          previousCandidate = 0;
+
+    DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", current);
+    assert(ip <= iend-8);   /* required for h calculation */
+
+    /* reach end of unsorted candidates list */
+    while ( (matchIndex > unsortLimit)
+         && (*unsortedMark == ZSTD_DUBT_UNSORTED_MARK)
+         && (nbCandidates > 1) ) {
+        DEBUGLOG(8, "ZSTD_DUBT_findBestMatch: candidate %u is unsorted",
+                    matchIndex);
+        *unsortedMark = previousCandidate;
+        previousCandidate = matchIndex;
+        matchIndex = *nextCandidate;
+        nextCandidate = bt + 2*(matchIndex&btMask);
+        unsortedMark = bt + 2*(matchIndex&btMask) + 1;
+        nbCandidates --;
+    }
+
+    if ( (matchIndex > unsortLimit)
+      && (*unsortedMark==ZSTD_DUBT_UNSORTED_MARK) ) {
+        DEBUGLOG(7, "ZSTD_DUBT_findBestMatch: nullify last unsorted candidate %u",
+                    matchIndex);
+        *nextCandidate = *unsortedMark = 0;   /* nullify next candidate if it's still unsorted (note : simplification, detrimental to compression ratio, beneficial for speed) */
+    }
+
+    /* batch sort stacked candidates */
+    matchIndex = previousCandidate;
+    while (matchIndex) {  /* will end on matchIndex == 0 */
+        U32* const nextCandidateIdxPtr = bt + 2*(matchIndex&btMask) + 1;
+        U32 const nextCandidateIdx = *nextCandidateIdxPtr;
+        ZSTD_insertDUBT1(ms, cParams, matchIndex, iend,
+                         nbCandidates, unsortLimit, extDict);
+        matchIndex = nextCandidateIdx;
+        nbCandidates++;
+    }
+
+    /* find longest match */
+    {   size_t commonLengthSmaller=0, commonLengthLarger=0;
+        const BYTE* const dictBase = ms->window.dictBase;
+        const U32 dictLimit = ms->window.dictLimit;
+        const BYTE* const dictEnd = dictBase + dictLimit;
+        const BYTE* const prefixStart = base + dictLimit;
+        U32* smallerPtr = bt + 2*(current&btMask);
+        U32* largerPtr  = bt + 2*(current&btMask) + 1;
+        U32 matchEndIdx = current+8+1;
+        U32 dummy32;   /* to be nullified at the end */
+        size_t bestLength = 0;
+
+        matchIndex  = hashTable[h];
+        hashTable[h] = current;   /* Update Hash Table */
+
+        while (nbCompares-- && (matchIndex > windowLow)) {
+            U32* const nextPtr = bt + 2*(matchIndex & btMask);
+            size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
+            const BYTE* match;
+
+            if ((!extDict) || (matchIndex+matchLength >= dictLimit)) {
+                match = base + matchIndex;
+                matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
+            } else {
+                match = dictBase + matchIndex;
+                matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
+                if (matchIndex+matchLength >= dictLimit)
+                    match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
+            }
+
+            if (matchLength > bestLength) {
+                if (matchLength > matchEndIdx - matchIndex)
+                    matchEndIdx = matchIndex + (U32)matchLength;
+                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
+                    bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex;
+                if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
+                    break;   /* drop, to guarantee consistency (miss a little bit of compression) */
+                }
+            }
+
+            if (match[matchLength] < ip[matchLength]) {
+                /* match is smaller than current */
+                *smallerPtr = matchIndex;             /* update smaller idx */
+                commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
+                if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+                smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
+                matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
+            } else {
+                /* match is larger than current */
+                *largerPtr = matchIndex;
+                commonLengthLarger = matchLength;
+                if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+                largerPtr = nextPtr;
+                matchIndex = nextPtr[0];
+        }   }
+
+        *smallerPtr = *largerPtr = 0;
+
+        assert(matchEndIdx > current+8); /* ensure nextToUpdate is increased */
+        ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
+        if (bestLength >= MINMATCH) {
+            U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
+            DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
+                        current, (U32)bestLength, (U32)*offsetPtr, mIndex);
+        }
+        return bestLength;
+    }
+}
+
+
+/** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
+static size_t ZSTD_BtFindBestMatch (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls /* template */)
+{
+    DEBUGLOG(7, "ZSTD_BtFindBestMatch");
+    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
+    ZSTD_updateDUBT(ms, cParams, ip, iLimit, mls);
+    return ZSTD_DUBT_findBestMatch(ms, cParams, ip, iLimit, offsetPtr, mls, 0);
+}
+
+
+static size_t ZSTD_BtFindBestMatch_selectMLS (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    switch(cParams->searchLength)
+    {
+    default : /* includes case 3 */
+    case 4 : return ZSTD_BtFindBestMatch(ms, cParams, ip, iLimit, offsetPtr, 4);
+    case 5 : return ZSTD_BtFindBestMatch(ms, cParams, ip, iLimit, offsetPtr, 5);
+    case 7 :
+    case 6 : return ZSTD_BtFindBestMatch(ms, cParams, ip, iLimit, offsetPtr, 6);
+    }
+}
+
+
+/** Tree updater, providing best match */
+static size_t ZSTD_BtFindBestMatch_extDict (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls)
+{
+    DEBUGLOG(7, "ZSTD_BtFindBestMatch_extDict");
+    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
+    ZSTD_updateDUBT(ms, cParams, ip, iLimit, mls);
+    return ZSTD_DUBT_findBestMatch(ms, cParams, ip, iLimit, offsetPtr, mls, 1);
+}
+
+
+static size_t ZSTD_BtFindBestMatch_selectMLS_extDict (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    switch(cParams->searchLength)
+    {
+    default : /* includes case 3 */
+    case 4 : return ZSTD_BtFindBestMatch_extDict(ms, cParams, ip, iLimit, offsetPtr, 4);
+    case 5 : return ZSTD_BtFindBestMatch_extDict(ms, cParams, ip, iLimit, offsetPtr, 5);
+    case 7 :
+    case 6 : return ZSTD_BtFindBestMatch_extDict(ms, cParams, ip, iLimit, offsetPtr, 6);
+    }
+}
+
+
+
+/* *********************************
+*  Hash Chain
+***********************************/
+#define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & mask]
+
+/* Update chains up to ip (excluded)
+   Assumption : always within prefix (i.e. not within extDict) */
+static U32 ZSTD_insertAndFindFirstIndex_internal(
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, U32 const mls)
+{
+    U32* const hashTable  = ms->hashTable;
+    const U32 hashLog = cParams->hashLog;
+    U32* const chainTable = ms->chainTable;
+    const U32 chainMask = (1 << cParams->chainLog) - 1;
+    const BYTE* const base = ms->window.base;
+    const U32 target = (U32)(ip - base);
+    U32 idx = ms->nextToUpdate;
+
+    while(idx < target) { /* catch up */
+        size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
+        NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
+        hashTable[h] = idx;
+        idx++;
+    }
+
+    ms->nextToUpdate = target;
+    return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
+}
+
+U32 ZSTD_insertAndFindFirstIndex(
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip)
+{
+    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, cParams->searchLength);
+}
+
+
+/* inlining is important to hardwire a hot branch (template emulation) */
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_HcFindBestMatch_generic (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls, const U32 extDict)
+{
+    U32* const chainTable = ms->chainTable;
+    const U32 chainSize = (1 << cParams->chainLog);
+    const U32 chainMask = chainSize-1;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const U32 dictLimit = ms->window.dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const U32 lowLimit = ms->window.lowLimit;
+    const U32 current = (U32)(ip-base);
+    const U32 minChain = current > chainSize ? current - chainSize : 0;
+    U32 nbAttempts = 1U << cParams->searchLog;
+    size_t ml=4-1;
+
+    /* HC4 match finder */
+    U32 matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
+
+    for ( ; (matchIndex>lowLimit) & (nbAttempts>0) ; nbAttempts--) {
+        size_t currentMl=0;
+        if ((!extDict) || matchIndex >= dictLimit) {
+            const BYTE* const match = base + matchIndex;
+            if (match[ml] == ip[ml])   /* potentially better */
+                currentMl = ZSTD_count(ip, match, iLimit);
+        } else {
+            const BYTE* const match = dictBase + matchIndex;
+            assert(match+4 <= dictEnd);
+            if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
+        }
+
+        /* save best solution */
+        if (currentMl > ml) {
+            ml = currentMl;
+            *offsetPtr = current - matchIndex + ZSTD_REP_MOVE;
+            if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+        }
+
+        if (matchIndex <= minChain) break;
+        matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
+    }
+
+    return ml;
+}
+
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    switch(cParams->searchLength)
+    {
+    default : /* includes case 3 */
+    case 4 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 4, 0);
+    case 5 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 5, 0);
+    case 7 :
+    case 6 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 6, 0);
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* const offsetPtr)
+{
+    switch(cParams->searchLength)
+    {
+    default : /* includes case 3 */
+    case 4 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 4, 1);
+    case 5 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 5, 1);
+    case 7 :
+    case 6 : return ZSTD_HcFindBestMatch_generic(ms, cParams, ip, iLimit, offsetPtr, 6, 1);
+    }
+}
+
+
+/* *******************************
+*  Common parser - lazy strategy
+*********************************/
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_lazy_generic(
+                        ZSTD_matchState_t* ms, seqStore_t* seqStore,
+                        U32 rep[ZSTD_REP_NUM],
+                        ZSTD_compressionParameters const* cParams,
+                        const void* src, size_t srcSize,
+                        const U32 searchMethod, const U32 depth)
+{
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    const BYTE* const base = ms->window.base + ms->window.dictLimit;
+
+    typedef size_t (*searchMax_f)(
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
+    searchMax_f const searchMax = searchMethod ? ZSTD_BtFindBestMatch_selectMLS : ZSTD_HcFindBestMatch_selectMLS;
+    U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
+
+    /* init */
+    ip += (ip==base);
+    ms->nextToUpdate3 = ms->nextToUpdate;
+    {   U32 const maxRep = (U32)(ip-base);
+        if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
+    }
+
+    /* Match Loop */
+    while (ip < ilimit) {
+        size_t matchLength=0;
+        size_t offset=0;
+        const BYTE* start=ip+1;
+
+        /* check repCode */
+        if ((offset_1>0) & (MEM_read32(ip+1) == MEM_read32(ip+1 - offset_1))) {
+            /* repcode : we take it */
+            matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+            if (depth==0) goto _storeSequence;
+        }
+
+        /* first search (depth 0) */
+        {   size_t offsetFound = 99999999;
+            size_t const ml2 = searchMax(ms, cParams, ip, iend, &offsetFound);
+            if (ml2 > matchLength)
+                matchLength = ml2, start = ip, offset=offsetFound;
+        }
+
+        if (matchLength < 4) {
+            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+            continue;
+        }
+
+        /* let's try to find a better solution */
+        if (depth>=1)
+        while (ip<ilimit) {
+            ip ++;
+            if ((offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
+                int const gain2 = (int)(mlRep * 3);
+                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                if ((mlRep >= 4) && (gain2 > gain1))
+                    matchLength = mlRep, offset = 0, start = ip;
+            }
+            {   size_t offset2=99999999;
+                size_t const ml2 = searchMax(ms, cParams, ip, iend, &offset2);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+                if ((ml2 >= 4) && (gain2 > gain1)) {
+                    matchLength = ml2, offset = offset2, start = ip;
+                    continue;   /* search a better one */
+            }   }
+
+            /* let's find an even better one */
+            if ((depth==2) && (ip<ilimit)) {
+                ip ++;
+                if ((offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                    size_t const ml2 = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
+                    int const gain2 = (int)(ml2 * 4);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                    if ((ml2 >= 4) && (gain2 > gain1))
+                        matchLength = ml2, offset = 0, start = ip;
+                }
+                {   size_t offset2=99999999;
+                    size_t const ml2 = searchMax(ms, cParams, ip, iend, &offset2);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                    if ((ml2 >= 4) && (gain2 > gain1)) {
+                        matchLength = ml2, offset = offset2, start = ip;
+                        continue;
+            }   }   }
+            break;  /* nothing found : store previous solution */
+        }
+
+        /* NOTE:
+         * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
+         * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
+         * overflows the pointer, which is undefined behavior.
+         */
+        /* catch up */
+        if (offset) {
+            while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > base))
+                 && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) )  /* only search for offset within prefix */
+                { start--; matchLength++; }
+            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+        }
+        /* store sequence */
+_storeSequence:
+        {   size_t const litLength = start - anchor;
+            ZSTD_storeSeq(seqStore, litLength, anchor, (U32)offset, matchLength-MINMATCH);
+            anchor = ip = start + matchLength;
+        }
+
+        /* check immediate repcode */
+        while ( ((ip <= ilimit) & (offset_2>0))
+             && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
+            /* store sequence */
+            matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+            offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
+            ZSTD_storeSeq(seqStore, 0, anchor, 0, matchLength-MINMATCH);
+            ip += matchLength;
+            anchor = ip;
+            continue;   /* faster when present ... (?) */
+    }   }
+
+    /* Save reps for next block */
+    rep[0] = offset_1 ? offset_1 : savedOffset;
+    rep[1] = offset_2 ? offset_2 : savedOffset;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_btlazy2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, cParams, src, srcSize, 1, 2);
+}
+
+size_t ZSTD_compressBlock_lazy2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 2);
+}
+
+size_t ZSTD_compressBlock_lazy(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 1);
+}
+
+size_t ZSTD_compressBlock_greedy(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 0);
+}
+
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_lazy_extDict_generic(
+                        ZSTD_matchState_t* ms, seqStore_t* seqStore,
+                        U32 rep[ZSTD_REP_NUM],
+                        ZSTD_compressionParameters const* cParams,
+                        const void* src, size_t srcSize,
+                        const U32 searchMethod, const U32 depth)
+{
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    const BYTE* const base = ms->window.base;
+    const U32 dictLimit = ms->window.dictLimit;
+    const U32 lowestIndex = ms->window.lowLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const BYTE* const dictEnd  = dictBase + dictLimit;
+    const BYTE* const dictStart  = dictBase + lowestIndex;
+
+    typedef size_t (*searchMax_f)(
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
+    searchMax_f searchMax = searchMethod ? ZSTD_BtFindBestMatch_selectMLS_extDict : ZSTD_HcFindBestMatch_extDict_selectMLS;
+
+    U32 offset_1 = rep[0], offset_2 = rep[1];
+
+    /* init */
+    ms->nextToUpdate3 = ms->nextToUpdate;
+    ip += (ip == prefixStart);
+
+    /* Match Loop */
+    while (ip < ilimit) {
+        size_t matchLength=0;
+        size_t offset=0;
+        const BYTE* start=ip+1;
+        U32 current = (U32)(ip-base);
+
+        /* check repCode */
+        {   const U32 repIndex = (U32)(current+1 - offset_1);
+            const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
+            const BYTE* const repMatch = repBase + repIndex;
+            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex))   /* intentional overflow */
+            if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
+                /* repcode detected we should take it */
+                const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
+                matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
+                if (depth==0) goto _storeSequence;
+        }   }
+
+        /* first search (depth 0) */
+        {   size_t offsetFound = 99999999;
+            size_t const ml2 = searchMax(ms, cParams, ip, iend, &offsetFound);
+            if (ml2 > matchLength)
+                matchLength = ml2, start = ip, offset=offsetFound;
+        }
+
+         if (matchLength < 4) {
+            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+            continue;
+        }
+
+        /* let's try to find a better solution */
+        if (depth>=1)
+        while (ip<ilimit) {
+            ip ++;
+            current++;
+            /* check repCode */
+            if (offset) {
+                const U32 repIndex = (U32)(current - offset_1);
+                const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
+                const BYTE* const repMatch = repBase + repIndex;
+                if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex))  /* intentional overflow */
+                if (MEM_read32(ip) == MEM_read32(repMatch)) {
+                    /* repcode detected */
+                    const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
+                    size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
+                    int const gain2 = (int)(repLength * 3);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                    if ((repLength >= 4) && (gain2 > gain1))
+                        matchLength = repLength, offset = 0, start = ip;
+            }   }
+
+            /* search match, depth 1 */
+            {   size_t offset2=99999999;
+                size_t const ml2 = searchMax(ms, cParams, ip, iend, &offset2);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+                if ((ml2 >= 4) && (gain2 > gain1)) {
+                    matchLength = ml2, offset = offset2, start = ip;
+                    continue;   /* search a better one */
+            }   }
+
+            /* let's find an even better one */
+            if ((depth==2) && (ip<ilimit)) {
+                ip ++;
+                current++;
+                /* check repCode */
+                if (offset) {
+                    const U32 repIndex = (U32)(current - offset_1);
+                    const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
+                    const BYTE* const repMatch = repBase + repIndex;
+                    if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex))  /* intentional overflow */
+                    if (MEM_read32(ip) == MEM_read32(repMatch)) {
+                        /* repcode detected */
+                        const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
+                        size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
+                        int const gain2 = (int)(repLength * 4);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                        if ((repLength >= 4) && (gain2 > gain1))
+                            matchLength = repLength, offset = 0, start = ip;
+                }   }
+
+                /* search match, depth 2 */
+                {   size_t offset2=99999999;
+                    size_t const ml2 = searchMax(ms, cParams, ip, iend, &offset2);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                    if ((ml2 >= 4) && (gain2 > gain1)) {
+                        matchLength = ml2, offset = offset2, start = ip;
+                        continue;
+            }   }   }
+            break;  /* nothing found : store previous solution */
+        }
+
+        /* catch up */
+        if (offset) {
+            U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
+            const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
+            const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
+            while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
+            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+        }
+
+        /* store sequence */
+_storeSequence:
+        {   size_t const litLength = start - anchor;
+            ZSTD_storeSeq(seqStore, litLength, anchor, (U32)offset, matchLength-MINMATCH);
+            anchor = ip = start + matchLength;
+        }
+
+        /* check immediate repcode */
+        while (ip <= ilimit) {
+            const U32 repIndex = (U32)((ip-base) - offset_2);
+            const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
+            const BYTE* const repMatch = repBase + repIndex;
+            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > lowestIndex))  /* intentional overflow */
+            if (MEM_read32(ip) == MEM_read32(repMatch)) {
+                /* repcode detected we should take it */
+                const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
+                matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
+                offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset history */
+                ZSTD_storeSeq(seqStore, 0, anchor, 0, matchLength-MINMATCH);
+                ip += matchLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+            }
+            break;
+    }   }
+
+    /* Save reps for next block */
+    rep[0] = offset_1;
+    rep[1] = offset_2;
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_greedy_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 0);
+}
+
+size_t ZSTD_compressBlock_lazy_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 1);
+}
+
+size_t ZSTD_compressBlock_lazy2_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 0, 2);
+}
+
+size_t ZSTD_compressBlock_btlazy2_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, cParams, src, srcSize, 1, 2);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_lazy.h b/vendor/github.com/DataDog/zstd/zstd_lazy.h
new file mode 100644
index 0000000..bda064f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_lazy.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_LAZY_H
+#define ZSTD_LAZY_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include "zstd_compress_internal.h"
+
+U32 ZSTD_insertAndFindFirstIndex(
+        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+        const BYTE* ip);
+
+void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue);  /*! used in ZSTD_reduceIndex(). pre-emptively increase value of ZSTD_DUBT_UNSORTED_MARK */
+
+size_t ZSTD_compressBlock_btlazy2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_greedy_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_btlazy2_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_LAZY_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_ldm.c b/vendor/github.com/DataDog/zstd/zstd_ldm.c
new file mode 100644
index 0000000..bffd8a3
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_ldm.c
@@ -0,0 +1,653 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ */
+
+#include "zstd_ldm.h"
+
+#include "zstd_fast.h"          /* ZSTD_fillHashTable() */
+#include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
+
+#define LDM_BUCKET_SIZE_LOG 3
+#define LDM_MIN_MATCH_LENGTH 64
+#define LDM_HASH_RLOG 7
+#define LDM_HASH_CHAR_OFFSET 10
+
+void ZSTD_ldm_adjustParameters(ldmParams_t* params,
+                               ZSTD_compressionParameters const* cParams)
+{
+    U32 const windowLog = cParams->windowLog;
+    ZSTD_STATIC_ASSERT(LDM_BUCKET_SIZE_LOG <= ZSTD_LDM_BUCKETSIZELOG_MAX);
+    DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
+    if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG;
+    if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH;
+    if (cParams->strategy >= ZSTD_btopt) {
+      /* Get out of the way of the optimal parser */
+      U32 const minMatch = MAX(cParams->targetLength, params->minMatchLength);
+      assert(minMatch >= ZSTD_LDM_MINMATCH_MIN);
+      assert(minMatch <= ZSTD_LDM_MINMATCH_MAX);
+      params->minMatchLength = minMatch;
+    }
+    if (params->hashLog == 0) {
+        params->hashLog = MAX(ZSTD_HASHLOG_MIN, windowLog - LDM_HASH_RLOG);
+        assert(params->hashLog <= ZSTD_HASHLOG_MAX);
+    }
+    if (params->hashEveryLog == 0) {
+        params->hashEveryLog =
+                windowLog < params->hashLog ? 0 : windowLog - params->hashLog;
+    }
+    params->bucketSizeLog = MIN(params->bucketSizeLog, params->hashLog);
+}
+
+size_t ZSTD_ldm_getTableSize(ldmParams_t params)
+{
+    size_t const ldmHSize = ((size_t)1) << params.hashLog;
+    size_t const ldmBucketSizeLog = MIN(params.bucketSizeLog, params.hashLog);
+    size_t const ldmBucketSize =
+        ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
+    size_t const totalSize = ldmBucketSize + ldmHSize * sizeof(ldmEntry_t);
+    return params.enableLdm ? totalSize : 0;
+}
+
+size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
+{
+    return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0;
+}
+
+/** ZSTD_ldm_getSmallHash() :
+ *  numBits should be <= 32
+ *  If numBits==0, returns 0.
+ *  @return : the most significant numBits of value. */
+static U32 ZSTD_ldm_getSmallHash(U64 value, U32 numBits)
+{
+    assert(numBits <= 32);
+    return numBits == 0 ? 0 : (U32)(value >> (64 - numBits));
+}
+
+/** ZSTD_ldm_getChecksum() :
+ *  numBitsToDiscard should be <= 32
+ *  @return : the next most significant 32 bits after numBitsToDiscard */
+static U32 ZSTD_ldm_getChecksum(U64 hash, U32 numBitsToDiscard)
+{
+    assert(numBitsToDiscard <= 32);
+    return (hash >> (64 - 32 - numBitsToDiscard)) & 0xFFFFFFFF;
+}
+
+/** ZSTD_ldm_getTag() ;
+ *  Given the hash, returns the most significant numTagBits bits
+ *  after (32 + hbits) bits.
+ *
+ *  If there are not enough bits remaining, return the last
+ *  numTagBits bits. */
+static U32 ZSTD_ldm_getTag(U64 hash, U32 hbits, U32 numTagBits)
+{
+    assert(numTagBits < 32 && hbits <= 32);
+    if (32 - hbits < numTagBits) {
+        return hash & (((U32)1 << numTagBits) - 1);
+    } else {
+        return (hash >> (32 - hbits - numTagBits)) & (((U32)1 << numTagBits) - 1);
+    }
+}
+
+/** ZSTD_ldm_getBucket() :
+ *  Returns a pointer to the start of the bucket associated with hash. */
+static ldmEntry_t* ZSTD_ldm_getBucket(
+        ldmState_t* ldmState, size_t hash, ldmParams_t const ldmParams)
+{
+    return ldmState->hashTable + (hash << ldmParams.bucketSizeLog);
+}
+
+/** ZSTD_ldm_insertEntry() :
+ *  Insert the entry with corresponding hash into the hash table */
+static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
+                                 size_t const hash, const ldmEntry_t entry,
+                                 ldmParams_t const ldmParams)
+{
+    BYTE* const bucketOffsets = ldmState->bucketOffsets;
+    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + bucketOffsets[hash]) = entry;
+    bucketOffsets[hash]++;
+    bucketOffsets[hash] &= ((U32)1 << ldmParams.bucketSizeLog) - 1;
+}
+
+/** ZSTD_ldm_makeEntryAndInsertByTag() :
+ *
+ *  Gets the small hash, checksum, and tag from the rollingHash.
+ *
+ *  If the tag matches (1 << ldmParams.hashEveryLog)-1, then
+ *  creates an ldmEntry from the offset, and inserts it into the hash table.
+ *
+ *  hBits is the length of the small hash, which is the most significant hBits
+ *  of rollingHash. The checksum is the next 32 most significant bits, followed
+ *  by ldmParams.hashEveryLog bits that make up the tag. */
+static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
+                                             U64 const rollingHash,
+                                             U32 const hBits,
+                                             U32 const offset,
+                                             ldmParams_t const ldmParams)
+{
+    U32 const tag = ZSTD_ldm_getTag(rollingHash, hBits, ldmParams.hashEveryLog);
+    U32 const tagMask = ((U32)1 << ldmParams.hashEveryLog) - 1;
+    if (tag == tagMask) {
+        U32 const hash = ZSTD_ldm_getSmallHash(rollingHash, hBits);
+        U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
+        ldmEntry_t entry;
+        entry.offset = offset;
+        entry.checksum = checksum;
+        ZSTD_ldm_insertEntry(ldmState, hash, entry, ldmParams);
+    }
+}
+
+/** ZSTD_ldm_getRollingHash() :
+ *  Get a 64-bit hash using the first len bytes from buf.
+ *
+ *  Giving bytes s = s_1, s_2, ... s_k, the hash is defined to be
+ *  H(s) = s_1*(a^(k-1)) + s_2*(a^(k-2)) + ... + s_k*(a^0)
+ *
+ *  where the constant a is defined to be prime8bytes.
+ *
+ *  The implementation adds an offset to each byte, so
+ *  H(s) = (s_1 + HASH_CHAR_OFFSET)*(a^(k-1)) + ... */
+static U64 ZSTD_ldm_getRollingHash(const BYTE* buf, U32 len)
+{
+    U64 ret = 0;
+    U32 i;
+    for (i = 0; i < len; i++) {
+        ret *= prime8bytes;
+        ret += buf[i] + LDM_HASH_CHAR_OFFSET;
+    }
+    return ret;
+}
+
+/** ZSTD_ldm_ipow() :
+ *  Return base^exp. */
+static U64 ZSTD_ldm_ipow(U64 base, U64 exp)
+{
+    U64 ret = 1;
+    while (exp) {
+        if (exp & 1) { ret *= base; }
+        exp >>= 1;
+        base *= base;
+    }
+    return ret;
+}
+
+U64 ZSTD_ldm_getHashPower(U32 minMatchLength) {
+    DEBUGLOG(4, "ZSTD_ldm_getHashPower: mml=%u", minMatchLength);
+    assert(minMatchLength >= ZSTD_LDM_MINMATCH_MIN);
+    return ZSTD_ldm_ipow(prime8bytes, minMatchLength - 1);
+}
+
+/** ZSTD_ldm_updateHash() :
+ *  Updates hash by removing toRemove and adding toAdd. */
+static U64 ZSTD_ldm_updateHash(U64 hash, BYTE toRemove, BYTE toAdd, U64 hashPower)
+{
+    hash -= ((toRemove + LDM_HASH_CHAR_OFFSET) * hashPower);
+    hash *= prime8bytes;
+    hash += toAdd + LDM_HASH_CHAR_OFFSET;
+    return hash;
+}
+
+/** ZSTD_ldm_countBackwardsMatch() :
+ *  Returns the number of bytes that match backwards before pIn and pMatch.
+ *
+ *  We count only bytes where pMatch >= pBase and pIn >= pAnchor. */
+static size_t ZSTD_ldm_countBackwardsMatch(
+            const BYTE* pIn, const BYTE* pAnchor,
+            const BYTE* pMatch, const BYTE* pBase)
+{
+    size_t matchLength = 0;
+    while (pIn > pAnchor && pMatch > pBase && pIn[-1] == pMatch[-1]) {
+        pIn--;
+        pMatch--;
+        matchLength++;
+    }
+    return matchLength;
+}
+
+/** ZSTD_ldm_fillFastTables() :
+ *
+ *  Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies.
+ *  This is similar to ZSTD_loadDictionaryContent.
+ *
+ *  The tables for the other strategies are filled within their
+ *  block compressors. */
+static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
+                                      ZSTD_compressionParameters const* cParams,
+                                      void const* end)
+{
+    const BYTE* const iend = (const BYTE*)end;
+
+    switch(cParams->strategy)
+    {
+    case ZSTD_fast:
+        ZSTD_fillHashTable(ms, cParams, iend);
+        ms->nextToUpdate = (U32)(iend - ms->window.base);
+        break;
+
+    case ZSTD_dfast:
+        ZSTD_fillDoubleHashTable(ms, cParams, iend);
+        ms->nextToUpdate = (U32)(iend - ms->window.base);
+        break;
+
+    case ZSTD_greedy:
+    case ZSTD_lazy:
+    case ZSTD_lazy2:
+    case ZSTD_btlazy2:
+    case ZSTD_btopt:
+    case ZSTD_btultra:
+        break;
+    default:
+        assert(0);  /* not possible : not a valid strategy id */
+    }
+
+    return 0;
+}
+
+/** ZSTD_ldm_fillLdmHashTable() :
+ *
+ *  Fills hashTable from (lastHashed + 1) to iend (non-inclusive).
+ *  lastHash is the rolling hash that corresponds to lastHashed.
+ *
+ *  Returns the rolling hash corresponding to position iend-1. */
+static U64 ZSTD_ldm_fillLdmHashTable(ldmState_t* state,
+                                     U64 lastHash, const BYTE* lastHashed,
+                                     const BYTE* iend, const BYTE* base,
+                                     U32 hBits, ldmParams_t const ldmParams)
+{
+    U64 rollingHash = lastHash;
+    const BYTE* cur = lastHashed + 1;
+
+    while (cur < iend) {
+        rollingHash = ZSTD_ldm_updateHash(rollingHash, cur[-1],
+                                          cur[ldmParams.minMatchLength-1],
+                                          state->hashPower);
+        ZSTD_ldm_makeEntryAndInsertByTag(state,
+                                         rollingHash, hBits,
+                                         (U32)(cur - base), ldmParams);
+        ++cur;
+    }
+    return rollingHash;
+}
+
+
+/** ZSTD_ldm_limitTableUpdate() :
+ *
+ *  Sets cctx->nextToUpdate to a position corresponding closer to anchor
+ *  if it is far way
+ *  (after a long match, only update tables a limited amount). */
+static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor)
+{
+    U32 const current = (U32)(anchor - ms->window.base);
+    if (current > ms->nextToUpdate + 1024) {
+        ms->nextToUpdate =
+            current - MIN(512, current - ms->nextToUpdate - 1024);
+    }
+}
+
+static size_t ZSTD_ldm_generateSequences_internal(
+        ldmState_t* ldmState, rawSeqStore_t* rawSeqStore,
+        ldmParams_t const* params, void const* src, size_t srcSize)
+{
+    /* LDM parameters */
+    int const extDict = ZSTD_window_hasExtDict(ldmState->window);
+    U32 const minMatchLength = params->minMatchLength;
+    U64 const hashPower = ldmState->hashPower;
+    U32 const hBits = params->hashLog - params->bucketSizeLog;
+    U32 const ldmBucketSize = 1U << params->bucketSizeLog;
+    U32 const hashEveryLog = params->hashEveryLog;
+    U32 const ldmTagMask = (1U << params->hashEveryLog) - 1;
+    /* Prefix and extDict parameters */
+    U32 const dictLimit = ldmState->window.dictLimit;
+    U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
+    BYTE const* const base = ldmState->window.base;
+    BYTE const* const dictBase = extDict ? ldmState->window.dictBase : NULL;
+    BYTE const* const dictStart = extDict ? dictBase + lowestIndex : NULL;
+    BYTE const* const dictEnd = extDict ? dictBase + dictLimit : NULL;
+    BYTE const* const lowPrefixPtr = base + dictLimit;
+    /* Input bounds */
+    BYTE const* const istart = (BYTE const*)src;
+    BYTE const* const iend = istart + srcSize;
+    BYTE const* const ilimit = iend - MAX(minMatchLength, HASH_READ_SIZE);
+    /* Input positions */
+    BYTE const* anchor = istart;
+    BYTE const* ip = istart;
+    /* Rolling hash */
+    BYTE const* lastHashed = NULL;
+    U64 rollingHash = 0;
+
+    while (ip <= ilimit) {
+        size_t mLength;
+        U32 const current = (U32)(ip - base);
+        size_t forwardMatchLength = 0, backwardMatchLength = 0;
+        ldmEntry_t* bestEntry = NULL;
+        if (ip != istart) {
+            rollingHash = ZSTD_ldm_updateHash(rollingHash, lastHashed[0],
+                                              lastHashed[minMatchLength],
+                                              hashPower);
+        } else {
+            rollingHash = ZSTD_ldm_getRollingHash(ip, minMatchLength);
+        }
+        lastHashed = ip;
+
+        /* Do not insert and do not look for a match */
+        if (ZSTD_ldm_getTag(rollingHash, hBits, hashEveryLog) != ldmTagMask) {
+           ip++;
+           continue;
+        }
+
+        /* Get the best entry and compute the match lengths */
+        {
+            ldmEntry_t* const bucket =
+                ZSTD_ldm_getBucket(ldmState,
+                                   ZSTD_ldm_getSmallHash(rollingHash, hBits),
+                                   *params);
+            ldmEntry_t* cur;
+            size_t bestMatchLength = 0;
+            U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
+
+            for (cur = bucket; cur < bucket + ldmBucketSize; ++cur) {
+                size_t curForwardMatchLength, curBackwardMatchLength,
+                       curTotalMatchLength;
+                if (cur->checksum != checksum || cur->offset <= lowestIndex) {
+                    continue;
+                }
+                if (extDict) {
+                    BYTE const* const curMatchBase =
+                        cur->offset < dictLimit ? dictBase : base;
+                    BYTE const* const pMatch = curMatchBase + cur->offset;
+                    BYTE const* const matchEnd =
+                        cur->offset < dictLimit ? dictEnd : iend;
+                    BYTE const* const lowMatchPtr =
+                        cur->offset < dictLimit ? dictStart : lowPrefixPtr;
+
+                    curForwardMatchLength = ZSTD_count_2segments(
+                                                ip, pMatch, iend,
+                                                matchEnd, lowPrefixPtr);
+                    if (curForwardMatchLength < minMatchLength) {
+                        continue;
+                    }
+                    curBackwardMatchLength =
+                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
+                                                     lowMatchPtr);
+                    curTotalMatchLength = curForwardMatchLength +
+                                          curBackwardMatchLength;
+                } else { /* !extDict */
+                    BYTE const* const pMatch = base + cur->offset;
+                    curForwardMatchLength = ZSTD_count(ip, pMatch, iend);
+                    if (curForwardMatchLength < minMatchLength) {
+                        continue;
+                    }
+                    curBackwardMatchLength =
+                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
+                                                     lowPrefixPtr);
+                    curTotalMatchLength = curForwardMatchLength +
+                                          curBackwardMatchLength;
+                }
+
+                if (curTotalMatchLength > bestMatchLength) {
+                    bestMatchLength = curTotalMatchLength;
+                    forwardMatchLength = curForwardMatchLength;
+                    backwardMatchLength = curBackwardMatchLength;
+                    bestEntry = cur;
+                }
+            }
+        }
+
+        /* No match found -- continue searching */
+        if (bestEntry == NULL) {
+            ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash,
+                                             hBits, current,
+                                             *params);
+            ip++;
+            continue;
+        }
+
+        /* Match found */
+        mLength = forwardMatchLength + backwardMatchLength;
+        ip -= backwardMatchLength;
+
+        {
+            /* Store the sequence:
+             * ip = current - backwardMatchLength
+             * The match is at (bestEntry->offset - backwardMatchLength)
+             */
+            U32 const matchIndex = bestEntry->offset;
+            U32 const offset = current - matchIndex;
+            rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
+
+            /* Out of sequence storage */
+            if (rawSeqStore->size == rawSeqStore->capacity)
+                return ERROR(dstSize_tooSmall);
+            seq->litLength = (U32)(ip - anchor);
+            seq->matchLength = (U32)mLength;
+            seq->offset = offset;
+            rawSeqStore->size++;
+        }
+
+        /* Insert the current entry into the hash table */
+        ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, hBits,
+                                         (U32)(lastHashed - base),
+                                         *params);
+
+        assert(ip + backwardMatchLength == lastHashed);
+
+        /* Fill the hash table from lastHashed+1 to ip+mLength*/
+        /* Heuristic: don't need to fill the entire table at end of block */
+        if (ip + mLength <= ilimit) {
+            rollingHash = ZSTD_ldm_fillLdmHashTable(
+                              ldmState, rollingHash, lastHashed,
+                              ip + mLength, base, hBits, *params);
+            lastHashed = ip + mLength - 1;
+        }
+        ip += mLength;
+        anchor = ip;
+    }
+    return iend - anchor;
+}
+
+/*! ZSTD_ldm_reduceTable() :
+ *  reduce table indexes by `reducerValue` */
+static void ZSTD_ldm_reduceTable(ldmEntry_t* const table, U32 const size,
+                                 U32 const reducerValue)
+{
+    U32 u;
+    for (u = 0; u < size; u++) {
+        if (table[u].offset < reducerValue) table[u].offset = 0;
+        else table[u].offset -= reducerValue;
+    }
+}
+
+size_t ZSTD_ldm_generateSequences(
+        ldmState_t* ldmState, rawSeqStore_t* sequences,
+        ldmParams_t const* params, void const* src, size_t srcSize)
+{
+    U32 const maxDist = 1U << params->windowLog;
+    BYTE const* const istart = (BYTE const*)src;
+    BYTE const* const iend = istart + srcSize;
+    size_t const kMaxChunkSize = 1 << 20;
+    size_t const nbChunks = (srcSize / kMaxChunkSize) + ((srcSize % kMaxChunkSize) != 0);
+    size_t chunk;
+    size_t leftoverSize = 0;
+
+    assert(ZSTD_CHUNKSIZE_MAX >= kMaxChunkSize);
+    /* Check that ZSTD_window_update() has been called for this chunk prior
+     * to passing it to this function.
+     */
+    assert(ldmState->window.nextSrc >= (BYTE const*)src + srcSize);
+    /* The input could be very large (in zstdmt), so it must be broken up into
+     * chunks to enforce the maximmum distance and handle overflow correction.
+     */
+    assert(sequences->pos <= sequences->size);
+    assert(sequences->size <= sequences->capacity);
+    for (chunk = 0; chunk < nbChunks && sequences->size < sequences->capacity; ++chunk) {
+        BYTE const* const chunkStart = istart + chunk * kMaxChunkSize;
+        size_t const remaining = (size_t)(iend - chunkStart);
+        BYTE const *const chunkEnd =
+            (remaining < kMaxChunkSize) ? iend : chunkStart + kMaxChunkSize;
+        size_t const chunkSize = chunkEnd - chunkStart;
+        size_t newLeftoverSize;
+        size_t const prevSize = sequences->size;
+
+        assert(chunkStart < iend);
+        /* 1. Perform overflow correction if necessary. */
+        if (ZSTD_window_needOverflowCorrection(ldmState->window, chunkEnd)) {
+            U32 const ldmHSize = 1U << params->hashLog;
+            U32 const correction = ZSTD_window_correctOverflow(
+                &ldmState->window, /* cycleLog */ 0, maxDist, src);
+            ZSTD_ldm_reduceTable(ldmState->hashTable, ldmHSize, correction);
+        }
+        /* 2. We enforce the maximum offset allowed.
+         *
+         * kMaxChunkSize should be small enough that we don't lose too much of
+         * the window through early invalidation.
+         * TODO: * Test the chunk size.
+         *       * Try invalidation after the sequence generation and test the
+         *         the offset against maxDist directly.
+         */
+        ZSTD_window_enforceMaxDist(&ldmState->window, chunkEnd, maxDist, NULL);
+        /* 3. Generate the sequences for the chunk, and get newLeftoverSize. */
+        newLeftoverSize = ZSTD_ldm_generateSequences_internal(
+            ldmState, sequences, params, chunkStart, chunkSize);
+        if (ZSTD_isError(newLeftoverSize))
+            return newLeftoverSize;
+        /* 4. We add the leftover literals from previous iterations to the first
+         *    newly generated sequence, or add the `newLeftoverSize` if none are
+         *    generated.
+         */
+        /* Prepend the leftover literals from the last call */
+        if (prevSize < sequences->size) {
+            sequences->seq[prevSize].litLength += (U32)leftoverSize;
+            leftoverSize = newLeftoverSize;
+        } else {
+            assert(newLeftoverSize == chunkSize);
+            leftoverSize += chunkSize;
+        }
+    }
+    return 0;
+}
+
+void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch) {
+    while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
+        rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
+        if (srcSize <= seq->litLength) {
+            /* Skip past srcSize literals */
+            seq->litLength -= (U32)srcSize;
+            return;
+        }
+        srcSize -= seq->litLength;
+        seq->litLength = 0;
+        if (srcSize < seq->matchLength) {
+            /* Skip past the first srcSize of the match */
+            seq->matchLength -= (U32)srcSize;
+            if (seq->matchLength < minMatch) {
+                /* The match is too short, omit it */
+                if (rawSeqStore->pos + 1 < rawSeqStore->size) {
+                    seq[1].litLength += seq[0].matchLength;
+                }
+                rawSeqStore->pos++;
+            }
+            return;
+        }
+        srcSize -= seq->matchLength;
+        seq->matchLength = 0;
+        rawSeqStore->pos++;
+    }
+}
+
+/**
+ * If the sequence length is longer than remaining then the sequence is split
+ * between this block and the next.
+ *
+ * Returns the current sequence to handle, or if the rest of the block should
+ * be literals, it returns a sequence with offset == 0.
+ */
+static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore,
+                                 U32 const remaining, U32 const minMatch)
+{
+    rawSeq sequence = rawSeqStore->seq[rawSeqStore->pos];
+    assert(sequence.offset > 0);
+    /* Likely: No partial sequence */
+    if (remaining >= sequence.litLength + sequence.matchLength) {
+        rawSeqStore->pos++;
+        return sequence;
+    }
+    /* Cut the sequence short (offset == 0 ==> rest is literals). */
+    if (remaining <= sequence.litLength) {
+        sequence.offset = 0;
+    } else if (remaining < sequence.litLength + sequence.matchLength) {
+        sequence.matchLength = remaining - sequence.litLength;
+        if (sequence.matchLength < minMatch) {
+            sequence.offset = 0;
+        }
+    }
+    /* Skip past `remaining` bytes for the future sequences. */
+    ZSTD_ldm_skipSequences(rawSeqStore, remaining, minMatch);
+    return sequence;
+}
+
+size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
+    ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+    ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize,
+    int const extDict)
+{
+    unsigned const minMatch = cParams->searchLength;
+    ZSTD_blockCompressor const blockCompressor =
+        ZSTD_selectBlockCompressor(cParams->strategy, extDict);
+    BYTE const* const base = ms->window.base;
+    /* Input bounds */
+    BYTE const* const istart = (BYTE const*)src;
+    BYTE const* const iend = istart + srcSize;
+    /* Input positions */
+    BYTE const* ip = istart;
+
+    assert(rawSeqStore->pos <= rawSeqStore->size);
+    assert(rawSeqStore->size <= rawSeqStore->capacity);
+    /* Loop through each sequence and apply the block compressor to the lits */
+    while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
+        /* maybeSplitSequence updates rawSeqStore->pos */
+        rawSeq const sequence = maybeSplitSequence(rawSeqStore,
+                                                   (U32)(iend - ip), minMatch);
+        int i;
+        /* End signal */
+        if (sequence.offset == 0)
+            break;
+
+        assert(sequence.offset <= (1U << cParams->windowLog));
+        assert(ip + sequence.litLength + sequence.matchLength <= iend);
+
+        /* Fill tables for block compressor */
+        ZSTD_ldm_limitTableUpdate(ms, ip);
+        ZSTD_ldm_fillFastTables(ms, cParams, ip);
+        /* Run the block compressor */
+        {
+            size_t const newLitLength =
+                blockCompressor(ms, seqStore, rep, cParams, ip,
+                                sequence.litLength);
+            ip += sequence.litLength;
+            ms->nextToUpdate = (U32)(ip - base);
+            /* Update the repcodes */
+            for (i = ZSTD_REP_NUM - 1; i > 0; i--)
+                rep[i] = rep[i-1];
+            rep[0] = sequence.offset;
+            /* Store the sequence */
+            ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength,
+                          sequence.offset + ZSTD_REP_MOVE,
+                          sequence.matchLength - MINMATCH);
+            ip += sequence.matchLength;
+        }
+    }
+    /* Fill the tables for the block compressor */
+    ZSTD_ldm_limitTableUpdate(ms, ip);
+    ZSTD_ldm_fillFastTables(ms, cParams, ip);
+    /* Compress the last literals */
+    {
+        size_t const lastLiterals = blockCompressor(ms, seqStore, rep, cParams,
+                                                    ip, iend - ip);
+        ms->nextToUpdate = (U32)(iend - base);
+        return lastLiterals;
+    }
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_ldm.h b/vendor/github.com/DataDog/zstd/zstd_ldm.h
new file mode 100644
index 0000000..0c3789f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_ldm.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ */
+
+#ifndef ZSTD_LDM_H
+#define ZSTD_LDM_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include "zstd_compress_internal.h"   /* ldmParams_t, U32 */
+#include "zstd.h"   /* ZSTD_CCtx, size_t */
+
+/*-*************************************
+*  Long distance matching
+***************************************/
+
+#define ZSTD_LDM_DEFAULT_WINDOW_LOG ZSTD_WINDOWLOG_DEFAULTMAX
+
+/**
+ * ZSTD_ldm_generateSequences():
+ *
+ * Generates the sequences using the long distance match finder.
+ * Generates long range matching sequences in `sequences`, which parse a prefix
+ * of the source. `sequences` must be large enough to store every sequence,
+ * which can be checked with `ZSTD_ldm_getMaxNbSeq()`.
+ * @returns 0 or an error code.
+ *
+ * NOTE: The user must have called ZSTD_window_update() for all of the input
+ * they have, even if they pass it to ZSTD_ldm_generateSequences() in chunks.
+ * NOTE: This function returns an error if it runs out of space to store
+ *       sequences.
+ */
+size_t ZSTD_ldm_generateSequences(
+            ldmState_t* ldms, rawSeqStore_t* sequences,
+            ldmParams_t const* params, void const* src, size_t srcSize);
+
+/**
+ * ZSTD_ldm_blockCompress():
+ *
+ * Compresses a block using the predefined sequences, along with a secondary
+ * block compressor. The literals section of every sequence is passed to the
+ * secondary block compressor, and those sequences are interspersed with the
+ * predefined sequences. Returns the length of the last literals.
+ * Updates `rawSeqStore.pos` to indicate how many sequences have been consumed.
+ * `rawSeqStore.seq` may also be updated to split the last sequence between two
+ * blocks.
+ * @return The length of the last literals.
+ *
+ * NOTE: The source must be at most the maximum block size, but the predefined
+ * sequences can be any size, and may be longer than the block. In the case that
+ * they are longer than the block, the last sequences may need to be split into
+ * two. We handle that case correctly, and update `rawSeqStore` appropriately.
+ * NOTE: This function does not return any errors.
+ */
+size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+            ZSTD_compressionParameters const* cParams,
+            void const* src, size_t srcSize,
+            int const extDict);
+
+/**
+ * ZSTD_ldm_skipSequences():
+ *
+ * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`.
+ * Avoids emitting matches less than `minMatch` bytes.
+ * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ */
+void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize,
+    U32 const minMatch);
+
+
+/** ZSTD_ldm_getTableSize() :
+ *  Estimate the space needed for long distance matching tables or 0 if LDM is
+ *  disabled.
+ */
+size_t ZSTD_ldm_getTableSize(ldmParams_t params);
+
+/** ZSTD_ldm_getSeqSpace() :
+ *  Return an upper bound on the number of sequences that can be produced by
+ *  the long distance matcher, or 0 if LDM is disabled.
+ */
+size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize);
+
+/** ZSTD_ldm_getTableSize() :
+ *  Return prime8bytes^(minMatchLength-1) */
+U64 ZSTD_ldm_getHashPower(U32 minMatchLength);
+
+/** ZSTD_ldm_adjustParameters() :
+ *  If the params->hashEveryLog is not set, set it to its default value based on
+ *  windowLog and params->hashLog.
+ *
+ *  Ensures that params->bucketSizeLog is <= params->hashLog (setting it to
+ *  params->hashLog if it is not).
+ *
+ *  Ensures that the minMatchLength >= targetLength during optimal parsing.
+ */
+void ZSTD_ldm_adjustParameters(ldmParams_t* params,
+                               ZSTD_compressionParameters const* cParams);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_FAST_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_legacy.h b/vendor/github.com/DataDog/zstd/zstd_legacy.h
new file mode 100644
index 0000000..5893cb9
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_legacy.h
@@ -0,0 +1,381 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_LEGACY_H
+#define ZSTD_LEGACY_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include "mem.h"            /* MEM_STATIC */
+#include "error_private.h"  /* ERROR */
+#include "zstd.h"           /* ZSTD_inBuffer, ZSTD_outBuffer */
+
+#if !defined (ZSTD_LEGACY_SUPPORT) || (ZSTD_LEGACY_SUPPORT == 0)
+#  undef ZSTD_LEGACY_SUPPORT
+#  define ZSTD_LEGACY_SUPPORT 8
+#endif
+
+#if (ZSTD_LEGACY_SUPPORT <= 1)
+#  include "zstd_v01.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 2)
+#  include "zstd_v02.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 3)
+#  include "zstd_v03.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+#  include "zstd_v04.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+#  include "zstd_v05.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+#  include "zstd_v06.h"
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+#  include "zstd_v07.h"
+#endif
+
+/** ZSTD_isLegacy() :
+    @return : > 0 if supported by legacy decoder. 0 otherwise.
+              return value is the version.
+*/
+MEM_STATIC unsigned ZSTD_isLegacy(const void* src, size_t srcSize)
+{
+    U32 magicNumberLE;
+    if (srcSize<4) return 0;
+    magicNumberLE = MEM_readLE32(src);
+    switch(magicNumberLE)
+    {
+#if (ZSTD_LEGACY_SUPPORT <= 1)
+        case ZSTDv01_magicNumberLE:return 1;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 2)
+        case ZSTDv02_magicNumber : return 2;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 3)
+        case ZSTDv03_magicNumber : return 3;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case ZSTDv04_magicNumber : return 4;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case ZSTDv05_MAGICNUMBER : return 5;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case ZSTDv06_MAGICNUMBER : return 6;
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case ZSTDv07_MAGICNUMBER : return 7;
+#endif
+        default : return 0;
+    }
+}
+
+
+MEM_STATIC unsigned long long ZSTD_getDecompressedSize_legacy(const void* src, size_t srcSize)
+{
+    U32 const version = ZSTD_isLegacy(src, srcSize);
+    if (version < 5) return 0;  /* no decompressed size in frame header, or not a legacy format */
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+    if (version==5) {
+        ZSTDv05_parameters fParams;
+        size_t const frResult = ZSTDv05_getFrameParams(&fParams, src, srcSize);
+        if (frResult != 0) return 0;
+        return fParams.srcSize;
+    }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+    if (version==6) {
+        ZSTDv06_frameParams fParams;
+        size_t const frResult = ZSTDv06_getFrameParams(&fParams, src, srcSize);
+        if (frResult != 0) return 0;
+        return fParams.frameContentSize;
+    }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+    if (version==7) {
+        ZSTDv07_frameParams fParams;
+        size_t const frResult = ZSTDv07_getFrameParams(&fParams, src, srcSize);
+        if (frResult != 0) return 0;
+        return fParams.frameContentSize;
+    }
+#endif
+    return 0;   /* should not be possible */
+}
+
+
+MEM_STATIC size_t ZSTD_decompressLegacy(
+                     void* dst, size_t dstCapacity,
+               const void* src, size_t compressedSize,
+               const void* dict,size_t dictSize)
+{
+    U32 const version = ZSTD_isLegacy(src, compressedSize);
+    (void)dst; (void)dstCapacity; (void)dict; (void)dictSize;  /* unused when ZSTD_LEGACY_SUPPORT >= 8 */
+    switch(version)
+    {
+#if (ZSTD_LEGACY_SUPPORT <= 1)
+        case 1 :
+            return ZSTDv01_decompress(dst, dstCapacity, src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 2)
+        case 2 :
+            return ZSTDv02_decompress(dst, dstCapacity, src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 3)
+        case 3 :
+            return ZSTDv03_decompress(dst, dstCapacity, src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case 4 :
+            return ZSTDv04_decompress(dst, dstCapacity, src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case 5 :
+            {   size_t result;
+                ZSTDv05_DCtx* const zd = ZSTDv05_createDCtx();
+                if (zd==NULL) return ERROR(memory_allocation);
+                result = ZSTDv05_decompress_usingDict(zd, dst, dstCapacity, src, compressedSize, dict, dictSize);
+                ZSTDv05_freeDCtx(zd);
+                return result;
+            }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case 6 :
+            {   size_t result;
+                ZSTDv06_DCtx* const zd = ZSTDv06_createDCtx();
+                if (zd==NULL) return ERROR(memory_allocation);
+                result = ZSTDv06_decompress_usingDict(zd, dst, dstCapacity, src, compressedSize, dict, dictSize);
+                ZSTDv06_freeDCtx(zd);
+                return result;
+            }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case 7 :
+            {   size_t result;
+                ZSTDv07_DCtx* const zd = ZSTDv07_createDCtx();
+                if (zd==NULL) return ERROR(memory_allocation);
+                result = ZSTDv07_decompress_usingDict(zd, dst, dstCapacity, src, compressedSize, dict, dictSize);
+                ZSTDv07_freeDCtx(zd);
+                return result;
+            }
+#endif
+        default :
+            return ERROR(prefix_unknown);
+    }
+}
+
+MEM_STATIC size_t ZSTD_findFrameCompressedSizeLegacy(const void *src,
+                                             size_t compressedSize)
+{
+    U32 const version = ZSTD_isLegacy(src, compressedSize);
+    switch(version)
+    {
+#if (ZSTD_LEGACY_SUPPORT <= 1)
+        case 1 :
+            return ZSTDv01_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 2)
+        case 2 :
+            return ZSTDv02_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 3)
+        case 3 :
+            return ZSTDv03_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case 4 :
+            return ZSTDv04_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case 5 :
+            return ZSTDv05_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case 6 :
+            return ZSTDv06_findFrameCompressedSize(src, compressedSize);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case 7 :
+            return ZSTDv07_findFrameCompressedSize(src, compressedSize);
+#endif
+        default :
+            return ERROR(prefix_unknown);
+    }
+}
+
+MEM_STATIC size_t ZSTD_freeLegacyStreamContext(void* legacyContext, U32 version)
+{
+    switch(version)
+    {
+        default :
+        case 1 :
+        case 2 :
+        case 3 :
+            (void)legacyContext;
+            return ERROR(version_unsupported);
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case 4 : return ZBUFFv04_freeDCtx((ZBUFFv04_DCtx*)legacyContext);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case 5 : return ZBUFFv05_freeDCtx((ZBUFFv05_DCtx*)legacyContext);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case 6 : return ZBUFFv06_freeDCtx((ZBUFFv06_DCtx*)legacyContext);
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case 7 : return ZBUFFv07_freeDCtx((ZBUFFv07_DCtx*)legacyContext);
+#endif
+    }
+}
+
+
+MEM_STATIC size_t ZSTD_initLegacyStream(void** legacyContext, U32 prevVersion, U32 newVersion,
+                                        const void* dict, size_t dictSize)
+{
+    DEBUGLOG(5, "ZSTD_initLegacyStream for v0.%u", newVersion);
+    if (prevVersion != newVersion) ZSTD_freeLegacyStreamContext(*legacyContext, prevVersion);
+    switch(newVersion)
+    {
+        default :
+        case 1 :
+        case 2 :
+        case 3 :
+            (void)dict; (void)dictSize;
+            return 0;
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case 4 :
+        {
+            ZBUFFv04_DCtx* dctx = (prevVersion != newVersion) ? ZBUFFv04_createDCtx() : (ZBUFFv04_DCtx*)*legacyContext;
+            if (dctx==NULL) return ERROR(memory_allocation);
+            ZBUFFv04_decompressInit(dctx);
+            ZBUFFv04_decompressWithDictionary(dctx, dict, dictSize);
+            *legacyContext = dctx;
+            return 0;
+        }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case 5 :
+        {
+            ZBUFFv05_DCtx* dctx = (prevVersion != newVersion) ? ZBUFFv05_createDCtx() : (ZBUFFv05_DCtx*)*legacyContext;
+            if (dctx==NULL) return ERROR(memory_allocation);
+            ZBUFFv05_decompressInitDictionary(dctx, dict, dictSize);
+            *legacyContext = dctx;
+            return 0;
+        }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case 6 :
+        {
+            ZBUFFv06_DCtx* dctx = (prevVersion != newVersion) ? ZBUFFv06_createDCtx() : (ZBUFFv06_DCtx*)*legacyContext;
+            if (dctx==NULL) return ERROR(memory_allocation);
+            ZBUFFv06_decompressInitDictionary(dctx, dict, dictSize);
+            *legacyContext = dctx;
+            return 0;
+        }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case 7 :
+        {
+            ZBUFFv07_DCtx* dctx = (prevVersion != newVersion) ? ZBUFFv07_createDCtx() : (ZBUFFv07_DCtx*)*legacyContext;
+            if (dctx==NULL) return ERROR(memory_allocation);
+            ZBUFFv07_decompressInitDictionary(dctx, dict, dictSize);
+            *legacyContext = dctx;
+            return 0;
+        }
+#endif
+    }
+}
+
+
+
+MEM_STATIC size_t ZSTD_decompressLegacyStream(void* legacyContext, U32 version,
+                                              ZSTD_outBuffer* output, ZSTD_inBuffer* input)
+{
+    DEBUGLOG(5, "ZSTD_decompressLegacyStream for v0.%u", version);
+    switch(version)
+    {
+        default :
+        case 1 :
+        case 2 :
+        case 3 :
+            (void)legacyContext; (void)output; (void)input;
+            return ERROR(version_unsupported);
+#if (ZSTD_LEGACY_SUPPORT <= 4)
+        case 4 :
+            {
+                ZBUFFv04_DCtx* dctx = (ZBUFFv04_DCtx*) legacyContext;
+                const void* src = (const char*)input->src + input->pos;
+                size_t readSize = input->size - input->pos;
+                void* dst = (char*)output->dst + output->pos;
+                size_t decodedSize = output->size - output->pos;
+                size_t const hintSize = ZBUFFv04_decompressContinue(dctx, dst, &decodedSize, src, &readSize);
+                output->pos += decodedSize;
+                input->pos += readSize;
+                return hintSize;
+            }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 5)
+        case 5 :
+            {
+                ZBUFFv05_DCtx* dctx = (ZBUFFv05_DCtx*) legacyContext;
+                const void* src = (const char*)input->src + input->pos;
+                size_t readSize = input->size - input->pos;
+                void* dst = (char*)output->dst + output->pos;
+                size_t decodedSize = output->size - output->pos;
+                size_t const hintSize = ZBUFFv05_decompressContinue(dctx, dst, &decodedSize, src, &readSize);
+                output->pos += decodedSize;
+                input->pos += readSize;
+                return hintSize;
+            }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 6)
+        case 6 :
+            {
+                ZBUFFv06_DCtx* dctx = (ZBUFFv06_DCtx*) legacyContext;
+                const void* src = (const char*)input->src + input->pos;
+                size_t readSize = input->size - input->pos;
+                void* dst = (char*)output->dst + output->pos;
+                size_t decodedSize = output->size - output->pos;
+                size_t const hintSize = ZBUFFv06_decompressContinue(dctx, dst, &decodedSize, src, &readSize);
+                output->pos += decodedSize;
+                input->pos += readSize;
+                return hintSize;
+            }
+#endif
+#if (ZSTD_LEGACY_SUPPORT <= 7)
+        case 7 :
+            {
+                ZBUFFv07_DCtx* dctx = (ZBUFFv07_DCtx*) legacyContext;
+                const void* src = (const char*)input->src + input->pos;
+                size_t readSize = input->size - input->pos;
+                void* dst = (char*)output->dst + output->pos;
+                size_t decodedSize = output->size - output->pos;
+                size_t const hintSize = ZBUFFv07_decompressContinue(dctx, dst, &decodedSize, src, &readSize);
+                output->pos += decodedSize;
+                input->pos += readSize;
+                return hintSize;
+            }
+#endif
+    }
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* ZSTD_LEGACY_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_opt.c b/vendor/github.com/DataDog/zstd/zstd_opt.c
new file mode 100644
index 0000000..f63f0c5
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_opt.c
@@ -0,0 +1,923 @@
+/*
+ * Copyright (c) 2016-present, Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#include "zstd_compress_internal.h"
+#include "zstd_opt.h"
+
+
+#define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats. Also used for matchSum (?) */
+#define ZSTD_FREQ_DIV       4   /* log factor when using previous stats to init next stats */
+#define ZSTD_MAX_PRICE     (1<<30)
+
+
+/*-*************************************
+*  Price functions for optimal parser
+***************************************/
+static void ZSTD_setLog2Prices(optState_t* optPtr)
+{
+    optPtr->log2litSum = ZSTD_highbit32(optPtr->litSum+1);
+    optPtr->log2litLengthSum = ZSTD_highbit32(optPtr->litLengthSum+1);
+    optPtr->log2matchLengthSum = ZSTD_highbit32(optPtr->matchLengthSum+1);
+    optPtr->log2offCodeSum = ZSTD_highbit32(optPtr->offCodeSum+1);
+}
+
+
+static void ZSTD_rescaleFreqs(optState_t* const optPtr,
+                              const BYTE* const src, size_t const srcSize)
+{
+    optPtr->staticPrices = 0;
+
+    if (optPtr->litLengthSum == 0) {  /* first init */
+        unsigned u;
+        if (srcSize <= 1024) optPtr->staticPrices = 1;
+
+        assert(optPtr->litFreq!=NULL);
+        for (u=0; u<=MaxLit; u++)
+            optPtr->litFreq[u] = 0;
+        for (u=0; u<srcSize; u++)
+            optPtr->litFreq[src[u]]++;
+        optPtr->litSum = 0;
+        for (u=0; u<=MaxLit; u++) {
+            optPtr->litFreq[u] = 1 + (optPtr->litFreq[u] >> ZSTD_FREQ_DIV);
+            optPtr->litSum += optPtr->litFreq[u];
+        }
+
+        for (u=0; u<=MaxLL; u++)
+            optPtr->litLengthFreq[u] = 1;
+        optPtr->litLengthSum = MaxLL+1;
+        for (u=0; u<=MaxML; u++)
+            optPtr->matchLengthFreq[u] = 1;
+        optPtr->matchLengthSum = MaxML+1;
+        for (u=0; u<=MaxOff; u++)
+            optPtr->offCodeFreq[u] = 1;
+        optPtr->offCodeSum = (MaxOff+1);
+
+    } else {
+        unsigned u;
+
+        optPtr->litSum = 0;
+        for (u=0; u<=MaxLit; u++) {
+            optPtr->litFreq[u] = 1 + (optPtr->litFreq[u] >> (ZSTD_FREQ_DIV+1));
+            optPtr->litSum += optPtr->litFreq[u];
+        }
+        optPtr->litLengthSum = 0;
+        for (u=0; u<=MaxLL; u++) {
+            optPtr->litLengthFreq[u] = 1 + (optPtr->litLengthFreq[u]>>(ZSTD_FREQ_DIV+1));
+            optPtr->litLengthSum += optPtr->litLengthFreq[u];
+        }
+        optPtr->matchLengthSum = 0;
+        for (u=0; u<=MaxML; u++) {
+            optPtr->matchLengthFreq[u] = 1 + (optPtr->matchLengthFreq[u]>>ZSTD_FREQ_DIV);
+            optPtr->matchLengthSum += optPtr->matchLengthFreq[u];
+        }
+        optPtr->offCodeSum = 0;
+        for (u=0; u<=MaxOff; u++) {
+            optPtr->offCodeFreq[u] = 1 + (optPtr->offCodeFreq[u]>>ZSTD_FREQ_DIV);
+            optPtr->offCodeSum += optPtr->offCodeFreq[u];
+        }
+    }
+
+    ZSTD_setLog2Prices(optPtr);
+}
+
+
+/* ZSTD_rawLiteralsCost() :
+ * cost of literals (only) in given segment (which length can be null)
+ * does not include cost of literalLength symbol */
+static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
+                                const optState_t* const optPtr)
+{
+    if (optPtr->staticPrices) return (litLength*6);  /* 6 bit per literal - no statistic used */
+    if (litLength == 0) return 0;
+
+    /* literals */
+    {   U32 u;
+        U32 cost = litLength * optPtr->log2litSum;
+        for (u=0; u < litLength; u++)
+            cost -= ZSTD_highbit32(optPtr->litFreq[literals[u]]+1);
+        return cost;
+    }
+}
+
+/* ZSTD_litLengthPrice() :
+ * cost of literalLength symbol */
+static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr)
+{
+    if (optPtr->staticPrices) return ZSTD_highbit32((U32)litLength+1);
+
+    /* literal Length */
+    {   U32 const llCode = ZSTD_LLcode(litLength);
+        U32 const price = LL_bits[llCode] + optPtr->log2litLengthSum - ZSTD_highbit32(optPtr->litLengthFreq[llCode]+1);
+        return price;
+    }
+}
+
+/* ZSTD_litLengthPrice() :
+ * cost of the literal part of a sequence,
+ * including literals themselves, and literalLength symbol */
+static U32 ZSTD_fullLiteralsCost(const BYTE* const literals, U32 const litLength,
+                                 const optState_t* const optPtr)
+{
+    return ZSTD_rawLiteralsCost(literals, litLength, optPtr)
+         + ZSTD_litLengthPrice(litLength, optPtr);
+}
+
+/* ZSTD_litLengthContribution() :
+ * @return ( cost(litlength) - cost(0) )
+ * this value can then be added to rawLiteralsCost()
+ * to provide a cost which is directly comparable to a match ending at same position */
+static int ZSTD_litLengthContribution(U32 const litLength, const optState_t* const optPtr)
+{
+    if (optPtr->staticPrices) return ZSTD_highbit32(litLength+1);
+
+    /* literal Length */
+    {   U32 const llCode = ZSTD_LLcode(litLength);
+        int const contribution = LL_bits[llCode]
+                        + ZSTD_highbit32(optPtr->litLengthFreq[0]+1)
+                        - ZSTD_highbit32(optPtr->litLengthFreq[llCode]+1);
+#if 1
+        return contribution;
+#else
+        return MAX(0, contribution); /* sometimes better, sometimes not ... */
+#endif
+    }
+}
+
+/* ZSTD_literalsContribution() :
+ * creates a fake cost for the literals part of a sequence
+ * which can be compared to the ending cost of a match
+ * should a new match start at this position */
+static int ZSTD_literalsContribution(const BYTE* const literals, U32 const litLength,
+                                     const optState_t* const optPtr)
+{
+    int const contribution = ZSTD_rawLiteralsCost(literals, litLength, optPtr)
+                           + ZSTD_litLengthContribution(litLength, optPtr);
+    return contribution;
+}
+
+/* ZSTD_getMatchPrice() :
+ * Provides the cost of the match part (offset + matchLength) of a sequence
+ * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
+ * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */
+FORCE_INLINE_TEMPLATE U32 ZSTD_getMatchPrice(
+                                    U32 const offset, U32 const matchLength,
+                                    const optState_t* const optPtr,
+                                    int const optLevel)
+{
+    U32 price;
+    U32 const offCode = ZSTD_highbit32(offset+1);
+    U32 const mlBase = matchLength - MINMATCH;
+    assert(matchLength >= MINMATCH);
+
+    if (optPtr->staticPrices)  /* fixed scheme, do not use statistics */
+        return ZSTD_highbit32((U32)mlBase+1) + 16 + offCode;
+
+    price = offCode + optPtr->log2offCodeSum - ZSTD_highbit32(optPtr->offCodeFreq[offCode]+1);
+    if ((optLevel<2) /*static*/ && offCode >= 20) price += (offCode-19)*2; /* handicap for long distance offsets, favor decompression speed */
+
+    /* match Length */
+    {   U32 const mlCode = ZSTD_MLcode(mlBase);
+        price += ML_bits[mlCode] + optPtr->log2matchLengthSum - ZSTD_highbit32(optPtr->matchLengthFreq[mlCode]+1);
+    }
+
+    DEBUGLOG(8, "ZSTD_getMatchPrice(ml:%u) = %u", matchLength, price);
+    return price;
+}
+
+static void ZSTD_updateStats(optState_t* const optPtr,
+                             U32 litLength, const BYTE* literals,
+                             U32 offsetCode, U32 matchLength)
+{
+    /* literals */
+    {   U32 u;
+        for (u=0; u < litLength; u++)
+            optPtr->litFreq[literals[u]] += ZSTD_LITFREQ_ADD;
+        optPtr->litSum += litLength*ZSTD_LITFREQ_ADD;
+    }
+
+    /* literal Length */
+    {   U32 const llCode = ZSTD_LLcode(litLength);
+        optPtr->litLengthFreq[llCode]++;
+        optPtr->litLengthSum++;
+    }
+
+    /* match offset code (0-2=>repCode; 3+=>offset+2) */
+    {   U32 const offCode = ZSTD_highbit32(offsetCode+1);
+        assert(offCode <= MaxOff);
+        optPtr->offCodeFreq[offCode]++;
+        optPtr->offCodeSum++;
+    }
+
+    /* match Length */
+    {   U32 const mlBase = matchLength - MINMATCH;
+        U32 const mlCode = ZSTD_MLcode(mlBase);
+        optPtr->matchLengthFreq[mlCode]++;
+        optPtr->matchLengthSum++;
+    }
+}
+
+
+/* ZSTD_readMINMATCH() :
+ * function safe only for comparisons
+ * assumption : memPtr must be at least 4 bytes before end of buffer */
+MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
+{
+    switch (length)
+    {
+    default :
+    case 4 : return MEM_read32(memPtr);
+    case 3 : if (MEM_isLittleEndian())
+                return MEM_read32(memPtr)<<8;
+             else
+                return MEM_read32(memPtr)>>8;
+    }
+}
+
+
+/* Update hashTable3 up to ip (excluded)
+   Assumption : always within prefix (i.e. not within extDict) */
+static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms, const BYTE* const ip)
+{
+    U32* const hashTable3 = ms->hashTable3;
+    U32 const hashLog3 = ms->hashLog3;
+    const BYTE* const base = ms->window.base;
+    U32 idx = ms->nextToUpdate3;
+    U32 const target = ms->nextToUpdate3 = (U32)(ip - base);
+    size_t const hash3 = ZSTD_hash3Ptr(ip, hashLog3);
+    assert(hashLog3 > 0);
+
+    while(idx < target) {
+        hashTable3[ZSTD_hash3Ptr(base+idx, hashLog3)] = idx;
+        idx++;
+    }
+
+    return hashTable3[hash3];
+}
+
+
+/*-*************************************
+*  Binary Tree search
+***************************************/
+/** ZSTD_insertBt1() : add one or multiple positions to tree.
+ *  ip : assumed <= iend-8 .
+ * @return : nb of positions added */
+static U32 ZSTD_insertBt1(
+                ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                const BYTE* const ip, const BYTE* const iend,
+                U32 const mls, U32 const extDict)
+{
+    U32*   const hashTable = ms->hashTable;
+    U32    const hashLog = cParams->hashLog;
+    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
+    U32*   const bt = ms->chainTable;
+    U32    const btLog  = cParams->chainLog - 1;
+    U32    const btMask = (1 << btLog) - 1;
+    U32 matchIndex = hashTable[h];
+    size_t commonLengthSmaller=0, commonLengthLarger=0;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const U32 dictLimit = ms->window.dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* match;
+    const U32 current = (U32)(ip-base);
+    const U32 btLow = btMask >= current ? 0 : current - btMask;
+    U32* smallerPtr = bt + 2*(current&btMask);
+    U32* largerPtr  = smallerPtr + 1;
+    U32 dummy32;   /* to be nullified at the end */
+    U32 const windowLow = ms->window.lowLimit;
+    U32 matchEndIdx = current+8+1;
+    size_t bestLength = 8;
+    U32 nbCompares = 1U << cParams->searchLog;
+#ifdef ZSTD_C_PREDICT
+    U32 predictedSmall = *(bt + 2*((current-1)&btMask) + 0);
+    U32 predictedLarge = *(bt + 2*((current-1)&btMask) + 1);
+    predictedSmall += (predictedSmall>0);
+    predictedLarge += (predictedLarge>0);
+#endif /* ZSTD_C_PREDICT */
+
+    DEBUGLOG(8, "ZSTD_insertBt1 (%u)", current);
+
+    assert(ip <= iend-8);   /* required for h calculation */
+    hashTable[h] = current;   /* Update Hash Table */
+
+    while (nbCompares-- && (matchIndex > windowLow)) {
+        U32* const nextPtr = bt + 2*(matchIndex & btMask);
+        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
+        assert(matchIndex < current);
+
+#ifdef ZSTD_C_PREDICT   /* note : can create issues when hlog small <= 11 */
+        const U32* predictPtr = bt + 2*((matchIndex-1) & btMask);   /* written this way, as bt is a roll buffer */
+        if (matchIndex == predictedSmall) {
+            /* no need to check length, result known */
+            *smallerPtr = matchIndex;
+            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+            smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
+            matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
+            predictedSmall = predictPtr[1] + (predictPtr[1]>0);
+            continue;
+        }
+        if (matchIndex == predictedLarge) {
+            *largerPtr = matchIndex;
+            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+            largerPtr = nextPtr;
+            matchIndex = nextPtr[0];
+            predictedLarge = predictPtr[0] + (predictPtr[0]>0);
+            continue;
+        }
+#endif
+
+        if ((!extDict) || (matchIndex+matchLength >= dictLimit)) {
+            assert(matchIndex+matchLength >= dictLimit);   /* might be wrong if extDict is incorrectly set to 0 */
+            match = base + matchIndex;
+            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
+        } else {
+            match = dictBase + matchIndex;
+            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
+            if (matchIndex+matchLength >= dictLimit)
+                match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
+        }
+
+        if (matchLength > bestLength) {
+            bestLength = matchLength;
+            if (matchLength > matchEndIdx - matchIndex)
+                matchEndIdx = matchIndex + (U32)matchLength;
+        }
+
+        if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
+            break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
+        }
+
+        if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
+            /* match is smaller than current */
+            *smallerPtr = matchIndex;             /* update smaller idx */
+            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
+            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
+            smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
+            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
+        } else {
+            /* match is larger than current */
+            *largerPtr = matchIndex;
+            commonLengthLarger = matchLength;
+            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
+            largerPtr = nextPtr;
+            matchIndex = nextPtr[0];
+    }   }
+
+    *smallerPtr = *largerPtr = 0;
+    if (bestLength > 384) return MIN(192, (U32)(bestLength - 384));   /* speed optimization */
+    assert(matchEndIdx > current + 8);
+    return matchEndIdx - (current + 8);
+}
+
+FORCE_INLINE_TEMPLATE
+void ZSTD_updateTree_internal(
+                ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                const BYTE* const ip, const BYTE* const iend,
+                const U32 mls, const U32 extDict)
+{
+    const BYTE* const base = ms->window.base;
+    U32 const target = (U32)(ip - base);
+    U32 idx = ms->nextToUpdate;
+    DEBUGLOG(7, "ZSTD_updateTree_internal, from %u to %u  (extDict:%u)",
+                idx, target, extDict);
+
+    while(idx < target)
+        idx += ZSTD_insertBt1(ms, cParams, base+idx, iend, mls, extDict);
+    ms->nextToUpdate = target;
+}
+
+void ZSTD_updateTree(
+                ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                const BYTE* ip, const BYTE* iend)
+{
+    ZSTD_updateTree_internal(ms, cParams, ip, iend, cParams->searchLength, 0 /*extDict*/);
+}
+
+FORCE_INLINE_TEMPLATE
+U32 ZSTD_insertBtAndGetAllMatches (
+                    ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                    const BYTE* const ip, const BYTE* const iLimit, int const extDict,
+                    U32 rep[ZSTD_REP_NUM], U32 const ll0,
+                    ZSTD_match_t* matches, const U32 lengthToBeat, U32 const mls /* template */)
+{
+    U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
+    const BYTE* const base = ms->window.base;
+    U32 const current = (U32)(ip-base);
+    U32 const hashLog = cParams->hashLog;
+    U32 const minMatch = (mls==3) ? 3 : 4;
+    U32* const hashTable = ms->hashTable;
+    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
+    U32 matchIndex  = hashTable[h];
+    U32* const bt   = ms->chainTable;
+    U32 const btLog = cParams->chainLog - 1;
+    U32 const btMask= (1U << btLog) - 1;
+    size_t commonLengthSmaller=0, commonLengthLarger=0;
+    const BYTE* const dictBase = ms->window.dictBase;
+    U32 const dictLimit = ms->window.dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    U32 const btLow = btMask >= current ? 0 : current - btMask;
+    U32 const windowLow = ms->window.lowLimit;
+    U32* smallerPtr = bt + 2*(current&btMask);
+    U32* largerPtr  = bt + 2*(current&btMask) + 1;
+    U32 matchEndIdx = current+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
+    U32 dummy32;   /* to be nullified at the end */
+    U32 mnum = 0;
+    U32 nbCompares = 1U << cParams->searchLog;
+
+    size_t bestLength = lengthToBeat-1;
+    DEBUGLOG(7, "ZSTD_insertBtAndGetAllMatches");
+
+    /* check repCode */
+    {   U32 const lastR = ZSTD_REP_NUM + ll0;
+        U32 repCode;
+        for (repCode = ll0; repCode < lastR; repCode++) {
+            U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
+            U32 const repIndex = current - repOffset;
+            U32 repLen = 0;
+            assert(current >= dictLimit);
+            if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < current-dictLimit) {  /* equivalent to `current > repIndex >= dictLimit` */
+                if (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch)) {
+                    repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch;
+                }
+            } else {  /* repIndex < dictLimit || repIndex >= current */
+                const BYTE* const repMatch = dictBase + repIndex;
+                assert(current >= windowLow);
+                if ( extDict /* this case only valid in extDict mode */
+                  && ( ((repOffset-1) /*intentional overflow*/ < current - windowLow)  /* equivalent to `current > repIndex >= windowLow` */
+                     & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */)
+                  && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
+                    repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
+            }   }
+            /* save longer solution */
+            if (repLen > bestLength) {
+                DEBUGLOG(8, "found rep-match %u of length %u",
+                            repCode - ll0, (U32)repLen);
+                bestLength = repLen;
+                matches[mnum].off = repCode - ll0;
+                matches[mnum].len = (U32)repLen;
+                mnum++;
+                if ( (repLen > sufficient_len)
+                   | (ip+repLen == iLimit) ) {  /* best possible */
+                    return mnum;
+    }   }   }   }
+
+    /* HC3 match finder */
+    if ((mls == 3) /*static*/ && (bestLength < mls)) {
+        U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, ip);
+        if ((matchIndex3 > windowLow)
+          & (current - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
+            size_t mlen;
+            if ((!extDict) /*static*/ || (matchIndex3 >= dictLimit)) {
+                const BYTE* const match = base + matchIndex3;
+                mlen = ZSTD_count(ip, match, iLimit);
+            } else {
+                const BYTE* const match = dictBase + matchIndex3;
+                mlen = ZSTD_count_2segments(ip, match, iLimit, dictEnd, prefixStart);
+            }
+
+            /* save best solution */
+            if (mlen >= mls /* == 3 > bestLength */) {
+                DEBUGLOG(8, "found small match with hlog3, of length %u",
+                            (U32)mlen);
+                bestLength = mlen;
+                assert(current > matchIndex3);
+                assert(mnum==0);  /* no prior solution */
+                matches[0].off = (current - matchIndex3) + ZSTD_REP_MOVE;
+                matches[0].len = (U32)mlen;
+                mnum = 1;
+                if ( (mlen > sufficient_len) |
+                     (ip+mlen == iLimit) ) {  /* best possible length */
+                    ms->nextToUpdate = current+1;  /* skip insertion */
+                    return 1;
+    }   }   }   }
+
+    hashTable[h] = current;   /* Update Hash Table */
+
+    while (nbCompares-- && (matchIndex > windowLow)) {
+        U32* const nextPtr = bt + 2*(matchIndex & btMask);
+        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
+        const BYTE* match;
+        assert(current > matchIndex);
+
+        if ((!extDict) || (matchIndex+matchLength >= dictLimit)) {
+            assert(matchIndex+matchLength >= dictLimit);  /* ensure the condition is correct when !extDict */
+            match = base + matchIndex;
+            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iLimit);
+        } else {
+            match = dictBase + matchIndex;
+            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dictEnd, prefixStart);
+            if (matchIndex+matchLength >= dictLimit)
+                match = base + matchIndex;   /* prepare for match[matchLength] */
+        }
+
+        if (matchLength > bestLength) {
+            DEBUGLOG(8, "found match of length %u at distance %u",
+                    (U32)matchLength, current - matchIndex);
+            assert(matchEndIdx > matchIndex);
+            if (matchLength > matchEndIdx - matchIndex)
+                matchEndIdx = matchIndex + (U32)matchLength;
+            bestLength = matchLength;
+            matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE;
+            matches[mnum].len = (U32)matchLength;
+            mnum++;
+            if (matchLength > ZSTD_OPT_NUM) break;
+            if (ip+matchLength == iLimit) {  /* equal : no way to know if inf or sup */
+                break;   /* drop, to preserve bt consistency (miss a little bit of compression) */
+            }
+        }
+
+        if (match[matchLength] < ip[matchLength]) {
+            /* match smaller than current */
+            *smallerPtr = matchIndex;             /* update smaller idx */
+            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
+            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+            smallerPtr = nextPtr+1;               /* new candidate => larger than match, which was smaller than current */
+            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous, closer to current */
+        } else {
+            *largerPtr = matchIndex;
+            commonLengthLarger = matchLength;
+            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
+            largerPtr = nextPtr;
+            matchIndex = nextPtr[0];
+    }   }
+
+    *smallerPtr = *largerPtr = 0;
+
+    assert(matchEndIdx > current+8);
+    ms->nextToUpdate = matchEndIdx - 8;  /* skip repetitive patterns */
+    return mnum;
+}
+
+
+FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches (
+                        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+                        const BYTE* ip, const BYTE* const iHighLimit, int const extDict,
+                        U32 rep[ZSTD_REP_NUM], U32 const ll0,
+                        ZSTD_match_t* matches, U32 const lengthToBeat)
+{
+    U32 const matchLengthSearch = cParams->searchLength;
+    DEBUGLOG(7, "ZSTD_BtGetAllMatches");
+    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
+    ZSTD_updateTree_internal(ms, cParams, ip, iHighLimit, matchLengthSearch, extDict);
+    switch(matchLengthSearch)
+    {
+    case 3 : return ZSTD_insertBtAndGetAllMatches(ms, cParams, ip, iHighLimit, extDict, rep, ll0, matches, lengthToBeat, 3);
+    default :
+    case 4 : return ZSTD_insertBtAndGetAllMatches(ms, cParams, ip, iHighLimit, extDict, rep, ll0, matches, lengthToBeat, 4);
+    case 5 : return ZSTD_insertBtAndGetAllMatches(ms, cParams, ip, iHighLimit, extDict, rep, ll0, matches, lengthToBeat, 5);
+    case 7 :
+    case 6 : return ZSTD_insertBtAndGetAllMatches(ms, cParams, ip, iHighLimit, extDict, rep, ll0, matches, lengthToBeat, 6);
+    }
+}
+
+
+/*-*******************************
+*  Optimal parser
+*********************************/
+typedef struct repcodes_s {
+    U32 rep[3];
+} repcodes_t;
+
+repcodes_t ZSTD_updateRep(U32 const rep[3], U32 const offset, U32 const ll0)
+{
+    repcodes_t newReps;
+    if (offset >= ZSTD_REP_NUM) {  /* full offset */
+        newReps.rep[2] = rep[1];
+        newReps.rep[1] = rep[0];
+        newReps.rep[0] = offset - ZSTD_REP_MOVE;
+    } else {   /* repcode */
+        U32 const repCode = offset + ll0;
+        if (repCode > 0) {  /* note : if repCode==0, no change */
+            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
+            newReps.rep[2] = (repCode >= 2) ? rep[1] : rep[2];
+            newReps.rep[1] = rep[0];
+            newReps.rep[0] = currentOffset;
+        } else {   /* repCode == 0 */
+            memcpy(&newReps, rep, sizeof(newReps));
+        }
+    }
+    return newReps;
+}
+
+
+typedef struct {
+    const BYTE* anchor;
+    U32 litlen;
+    U32 rawLitCost;
+} cachedLiteralPrice_t;
+
+static U32 ZSTD_rawLiteralsCost_cached(
+                            cachedLiteralPrice_t* const cachedLitPrice,
+                            const BYTE* const anchor, U32 const litlen,
+                            const optState_t* const optStatePtr)
+{
+    U32 startCost;
+    U32 remainingLength;
+    const BYTE* startPosition;
+
+    if (anchor == cachedLitPrice->anchor) {
+        startCost = cachedLitPrice->rawLitCost;
+        startPosition = anchor + cachedLitPrice->litlen;
+        assert(litlen >= cachedLitPrice->litlen);
+        remainingLength = litlen - cachedLitPrice->litlen;
+    } else {
+        startCost = 0;
+        startPosition = anchor;
+        remainingLength = litlen;
+    }
+
+    {   U32 const rawLitCost = startCost + ZSTD_rawLiteralsCost(startPosition, remainingLength, optStatePtr);
+        cachedLitPrice->anchor = anchor;
+        cachedLitPrice->litlen = litlen;
+        cachedLitPrice->rawLitCost = rawLitCost;
+        return rawLitCost;
+    }
+}
+
+static U32 ZSTD_fullLiteralsCost_cached(
+                            cachedLiteralPrice_t* const cachedLitPrice,
+                            const BYTE* const anchor, U32 const litlen,
+                            const optState_t* const optStatePtr)
+{
+    return ZSTD_rawLiteralsCost_cached(cachedLitPrice, anchor, litlen, optStatePtr)
+         + ZSTD_litLengthPrice(litlen, optStatePtr);
+}
+
+static int ZSTD_literalsContribution_cached(
+                            cachedLiteralPrice_t* const cachedLitPrice,
+                            const BYTE* const anchor, U32 const litlen,
+                            const optState_t* const optStatePtr)
+{
+    int const contribution = ZSTD_rawLiteralsCost_cached(cachedLitPrice, anchor, litlen, optStatePtr)
+                           + ZSTD_litLengthContribution(litlen, optStatePtr);
+    return contribution;
+}
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,seqStore_t* seqStore,
+                                      U32 rep[ZSTD_REP_NUM],
+                                      ZSTD_compressionParameters const* cParams,
+                                      const void* src, size_t srcSize,
+                                      const int optLevel, const int extDict)
+{
+    optState_t* const optStatePtr = &ms->opt;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const prefixStart = base + ms->window.dictLimit;
+
+    U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
+    U32 const minMatch = (cParams->searchLength == 3) ? 3 : 4;
+
+    ZSTD_optimal_t* const opt = optStatePtr->priceTable;
+    ZSTD_match_t* const matches = optStatePtr->matchTable;
+    cachedLiteralPrice_t cachedLitPrice;
+
+    /* init */
+    DEBUGLOG(5, "ZSTD_compressBlock_opt_generic");
+    ms->nextToUpdate3 = ms->nextToUpdate;
+    ZSTD_rescaleFreqs(optStatePtr, (const BYTE*)src, srcSize);
+    ip += (ip==prefixStart);
+    memset(&cachedLitPrice, 0, sizeof(cachedLitPrice));
+
+    /* Match Loop */
+    while (ip < ilimit) {
+        U32 cur, last_pos = 0;
+        U32 best_mlen, best_off;
+
+        /* find first match */
+        {   U32 const litlen = (U32)(ip - anchor);
+            U32 const ll0 = !litlen;
+            U32 const nbMatches = ZSTD_BtGetAllMatches(ms, cParams, ip, iend, extDict, rep, ll0, matches, minMatch);
+            if (!nbMatches) { ip++; continue; }
+
+            /* initialize opt[0] */
+            { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; }
+            opt[0].mlen = 1;
+            opt[0].litlen = litlen;
+
+            /* large match -> immediate encoding */
+            {   U32 const maxML = matches[nbMatches-1].len;
+                DEBUGLOG(7, "found %u matches of maxLength=%u and offset=%u at cPos=%u => start new serie",
+                            nbMatches, maxML, matches[nbMatches-1].off, (U32)(ip-prefixStart));
+
+                if (maxML > sufficient_len) {
+                    best_mlen = maxML;
+                    best_off = matches[nbMatches-1].off;
+                    DEBUGLOG(7, "large match (%u>%u), immediate encoding",
+                                best_mlen, sufficient_len);
+                    cur = 0;
+                    last_pos = 1;
+                    goto _shortestPath;
+            }   }
+
+            /* set prices for first matches starting position == 0 */
+            {   U32 const literalsPrice = ZSTD_fullLiteralsCost_cached(&cachedLitPrice, anchor, litlen, optStatePtr);
+                U32 pos;
+                U32 matchNb;
+                for (pos = 0; pos < minMatch; pos++) {
+                    opt[pos].mlen = 1;
+                    opt[pos].price = ZSTD_MAX_PRICE;
+                }
+                for (matchNb = 0; matchNb < nbMatches; matchNb++) {
+                    U32 const offset = matches[matchNb].off;
+                    U32 const end = matches[matchNb].len;
+                    repcodes_t const repHistory = ZSTD_updateRep(rep, offset, ll0);
+                    for ( ; pos <= end ; pos++ ) {
+                        U32 const matchPrice = literalsPrice + ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel);
+                        DEBUGLOG(7, "rPos:%u => set initial price : %u",
+                                    pos, matchPrice);
+                        opt[pos].mlen = pos;
+                        opt[pos].off = offset;
+                        opt[pos].litlen = litlen;
+                        opt[pos].price = matchPrice;
+                        memcpy(opt[pos].rep, &repHistory, sizeof(repHistory));
+                }   }
+                last_pos = pos-1;
+            }
+        }
+
+        /* check further positions */
+        for (cur = 1; cur <= last_pos; cur++) {
+            const BYTE* const inr = ip + cur;
+            assert(cur < ZSTD_OPT_NUM);
+
+            /* Fix current position with one literal if cheaper */
+            {   U32 const litlen = (opt[cur-1].mlen == 1) ? opt[cur-1].litlen + 1 : 1;
+                int price;  /* note : contribution can be negative */
+                if (cur > litlen) {
+                    price = opt[cur - litlen].price + ZSTD_literalsContribution(inr-litlen, litlen, optStatePtr);
+                } else {
+                    price = ZSTD_literalsContribution_cached(&cachedLitPrice, anchor, litlen, optStatePtr);
+                }
+                assert(price < 1000000000); /* overflow check */
+                if (price <= opt[cur].price) {
+                    DEBUGLOG(7, "rPos:%u : better price (%u<%u) using literal",
+                                cur, price, opt[cur].price);
+                    opt[cur].mlen = 1;
+                    opt[cur].off = 0;
+                    opt[cur].litlen = litlen;
+                    opt[cur].price = price;
+                    memcpy(opt[cur].rep, opt[cur-1].rep, sizeof(opt[cur].rep));
+            }   }
+
+            /* last match must start at a minimum distance of 8 from oend */
+            if (inr > ilimit) continue;
+
+            if (cur == last_pos) break;
+
+             if ( (optLevel==0) /*static*/
+               && (opt[cur+1].price <= opt[cur].price) )
+                continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
+
+            {   U32 const ll0 = (opt[cur].mlen != 1);
+                U32 const litlen = (opt[cur].mlen == 1) ? opt[cur].litlen : 0;
+                U32 const previousPrice = (cur > litlen) ? opt[cur-litlen].price : 0;
+                U32 const basePrice = previousPrice + ZSTD_fullLiteralsCost(inr-litlen, litlen, optStatePtr);
+                U32 const nbMatches = ZSTD_BtGetAllMatches(ms, cParams, inr, iend, extDict, opt[cur].rep, ll0, matches, minMatch);
+                U32 matchNb;
+                if (!nbMatches) continue;
+
+                {   U32 const maxML = matches[nbMatches-1].len;
+                    DEBUGLOG(7, "rPos:%u, found %u matches, of maxLength=%u",
+                                cur, nbMatches, maxML);
+
+                    if ( (maxML > sufficient_len)
+                       | (cur + maxML >= ZSTD_OPT_NUM) ) {
+                        best_mlen = maxML;
+                        best_off = matches[nbMatches-1].off;
+                        last_pos = cur + 1;
+                        goto _shortestPath;
+                    }
+                }
+
+                /* set prices using matches found at position == cur */
+                for (matchNb = 0; matchNb < nbMatches; matchNb++) {
+                    U32 const offset = matches[matchNb].off;
+                    repcodes_t const repHistory = ZSTD_updateRep(opt[cur].rep, offset, ll0);
+                    U32 const lastML = matches[matchNb].len;
+                    U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
+                    U32 mlen;
+
+                    DEBUGLOG(7, "testing match %u => offCode=%u, mlen=%u, llen=%u",
+                                matchNb, matches[matchNb].off, lastML, litlen);
+
+                    for (mlen = lastML; mlen >= startML; mlen--) {
+                        U32 const pos = cur + mlen;
+                        int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
+
+                        if ((pos > last_pos) || (price < opt[pos].price)) {
+                            DEBUGLOG(7, "rPos:%u => new better price (%u<%u)",
+                                        pos, price, opt[pos].price);
+                            while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; }
+                            opt[pos].mlen = mlen;
+                            opt[pos].off = offset;
+                            opt[pos].litlen = litlen;
+                            opt[pos].price = price;
+                            memcpy(opt[pos].rep, &repHistory, sizeof(repHistory));
+                        } else {
+                            if (optLevel==0) break;  /* gets ~+10% speed for about -0.01 ratio loss */
+                        }
+            }   }   }
+        }  /* for (cur = 1; cur <= last_pos; cur++) */
+
+        best_mlen = opt[last_pos].mlen;
+        best_off = opt[last_pos].off;
+        cur = last_pos - best_mlen;
+
+_shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
+        assert(opt[0].mlen == 1);
+
+        /* reverse traversal */
+        DEBUGLOG(7, "start reverse traversal (last_pos:%u, cur:%u)",
+                    last_pos, cur);
+        {   U32 selectedMatchLength = best_mlen;
+            U32 selectedOffset = best_off;
+            U32 pos = cur;
+            while (1) {
+                U32 const mlen = opt[pos].mlen;
+                U32 const off = opt[pos].off;
+                opt[pos].mlen = selectedMatchLength;
+                opt[pos].off = selectedOffset;
+                selectedMatchLength = mlen;
+                selectedOffset = off;
+                if (mlen > pos) break;
+                pos -= mlen;
+        }   }
+
+        /* save sequences */
+        {   U32 pos;
+            for (pos=0; pos < last_pos; ) {
+                U32 const llen = (U32)(ip - anchor);
+                U32 const mlen = opt[pos].mlen;
+                U32 const offset = opt[pos].off;
+                if (mlen == 1) { ip++; pos++; continue; }  /* literal position => move on */
+                pos += mlen; ip += mlen;
+
+                /* repcodes update : like ZSTD_updateRep(), but update in place */
+                if (offset >= ZSTD_REP_NUM) {  /* full offset */
+                    rep[2] = rep[1];
+                    rep[1] = rep[0];
+                    rep[0] = offset - ZSTD_REP_MOVE;
+                } else {   /* repcode */
+                    U32 const repCode = offset + (llen==0);
+                    if (repCode) {  /* note : if repCode==0, no change */
+                        U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
+                        if (repCode >= 2) rep[2] = rep[1];
+                        rep[1] = rep[0];
+                        rep[0] = currentOffset;
+                    }
+                }
+
+                ZSTD_updateStats(optStatePtr, llen, anchor, offset, mlen);
+                ZSTD_storeSeq(seqStore, llen, anchor, offset, mlen-MINMATCH);
+                anchor = ip;
+        }   }
+        ZSTD_setLog2Prices(optStatePtr);
+    }   /* while (ip < ilimit) */
+
+    /* Return the last literals size */
+    return iend - anchor;
+}
+
+
+size_t ZSTD_compressBlock_btopt(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    DEBUGLOG(5, "ZSTD_compressBlock_btopt");
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, cParams, src, srcSize, 0 /*optLevel*/, 0 /*extDict*/);
+}
+
+size_t ZSTD_compressBlock_btultra(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, cParams, src, srcSize, 2 /*optLevel*/, 0 /*extDict*/);
+}
+
+size_t ZSTD_compressBlock_btopt_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, cParams, src, srcSize, 0 /*optLevel*/, 1 /*extDict*/);
+}
+
+size_t ZSTD_compressBlock_btultra_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, cParams, src, srcSize, 2 /*optLevel*/, 1 /*extDict*/);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_opt.h b/vendor/github.com/DataDog/zstd/zstd_opt.h
new file mode 100644
index 0000000..b8dc389
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_opt.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_OPT_H
+#define ZSTD_OPT_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include "zstd_compress_internal.h"
+
+void ZSTD_updateTree(
+        ZSTD_matchState_t* ms, ZSTD_compressionParameters const* cParams,
+        const BYTE* ip, const BYTE* iend);  /* used in ZSTD_loadDictionaryContent() */
+
+size_t ZSTD_compressBlock_btopt(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_btultra(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_btopt_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_btultra_extDict(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_compressionParameters const* cParams, void const* src, size_t srcSize);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_OPT_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_stream.go b/vendor/github.com/DataDog/zstd/zstd_stream.go
new file mode 100644
index 0000000..d5d1336
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_stream.go
@@ -0,0 +1,291 @@
+package zstd
+
+/*
+#define ZSTD_STATIC_LINKING_ONLY
+#define ZBUFF_DISABLE_DEPRECATE_WARNINGS
+#include "zstd.h"
+#include "zbuff.h"
+*/
+import "C"
+import (
+	"errors"
+	"fmt"
+	"io"
+	"unsafe"
+)
+
+var errShortRead = errors.New("short read")
+
+// Writer is an io.WriteCloser that zstd-compresses its input.
+type Writer struct {
+	CompressionLevel int
+
+	ctx              *C.ZSTD_CCtx
+	dict             []byte
+	dstBuffer        []byte
+	firstError       error
+	underlyingWriter io.Writer
+}
+
+func resize(in []byte, newSize int) []byte {
+	if in == nil {
+		return make([]byte, newSize)
+	}
+	if newSize <= cap(in) {
+		return in[:newSize]
+	}
+	toAdd := newSize - len(in)
+	return append(in, make([]byte, toAdd)...)
+}
+
+// NewWriter creates a new Writer with default compression options.  Writes to
+// the writer will be written in compressed form to w.
+func NewWriter(w io.Writer) *Writer {
+	return NewWriterLevelDict(w, DefaultCompression, nil)
+}
+
+// NewWriterLevel is like NewWriter but specifies the compression level instead
+// of assuming default compression.
+//
+// The level can be DefaultCompression or any integer value between BestSpeed
+// and BestCompression inclusive.
+func NewWriterLevel(w io.Writer, level int) *Writer {
+	return NewWriterLevelDict(w, level, nil)
+
+}
+
+// NewWriterLevelDict is like NewWriterLevel but specifies a dictionary to
+// compress with.  If the dictionary is empty or nil it is ignored. The dictionary
+// should not be modified until the writer is closed.
+func NewWriterLevelDict(w io.Writer, level int, dict []byte) *Writer {
+	var err error
+	ctx := C.ZSTD_createCCtx()
+
+	if dict == nil {
+		err = getError(int(C.ZSTD_compressBegin(ctx,
+			C.int(level))))
+	} else {
+		err = getError(int(C.ZSTD_compressBegin_usingDict(
+			ctx,
+			unsafe.Pointer(&dict[0]),
+			C.size_t(len(dict)),
+			C.int(level))))
+	}
+
+	return &Writer{
+		CompressionLevel: level,
+		ctx:              ctx,
+		dict:             dict,
+		dstBuffer:        make([]byte, CompressBound(1024)),
+		firstError:       err,
+		underlyingWriter: w,
+	}
+}
+
+// Write writes a compressed form of p to the underlying io.Writer.
+func (w *Writer) Write(p []byte) (int, error) {
+	if w.firstError != nil {
+		return 0, w.firstError
+	}
+	if len(p) == 0 {
+		return 0, nil
+	}
+	// Check if dstBuffer is enough
+	if len(w.dstBuffer) < CompressBound(len(p)) {
+		w.dstBuffer = make([]byte, CompressBound(len(p)))
+	}
+
+	retCode := C.ZSTD_compressContinue(
+		w.ctx,
+		unsafe.Pointer(&w.dstBuffer[0]),
+		C.size_t(len(w.dstBuffer)),
+		unsafe.Pointer(&p[0]),
+		C.size_t(len(p)))
+
+	if err := getError(int(retCode)); err != nil {
+		return 0, err
+	}
+	written := int(retCode)
+
+	// Write to underlying buffer
+	_, err := w.underlyingWriter.Write(w.dstBuffer[:written])
+
+	// Same behaviour as zlib, we can't know how much data we wrote, only
+	// if there was an error
+	if err != nil {
+		return 0, err
+	}
+	return len(p), err
+}
+
+// Close closes the Writer, flushing any unwritten data to the underlying
+// io.Writer and freeing objects, but does not close the underlying io.Writer.
+func (w *Writer) Close() error {
+	retCode := C.ZSTD_compressEnd(
+		w.ctx,
+		unsafe.Pointer(&w.dstBuffer[0]),
+		C.size_t(len(w.dstBuffer)),
+		unsafe.Pointer(nil),
+		C.size_t(0))
+
+	if err := getError(int(retCode)); err != nil {
+		return err
+	}
+	written := int(retCode)
+	retCode = C.ZSTD_freeCCtx(w.ctx) // Safely close buffer before writing the end
+
+	if err := getError(int(retCode)); err != nil {
+		return err
+	}
+
+	_, err := w.underlyingWriter.Write(w.dstBuffer[:written])
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+// reader is an io.ReadCloser that decompresses when read from.
+type reader struct {
+	ctx                 *C.ZBUFF_DCtx
+	compressionBuffer   []byte
+	compressionLeft     int
+	decompressionBuffer []byte
+	decompOff           int
+	decompSize          int
+	dict                []byte
+	firstError          error
+	recommendedSrcSize  int
+	underlyingReader    io.Reader
+}
+
+// NewReader creates a new io.ReadCloser.  Reads from the returned ReadCloser
+// read and decompress data from r.  It is the caller's responsibility to call
+// Close on the ReadCloser when done.  If this is not done, underlying objects
+// in the zstd library will not be freed.
+func NewReader(r io.Reader) io.ReadCloser {
+	return NewReaderDict(r, nil)
+}
+
+// NewReaderDict is like NewReader but uses a preset dictionary.  NewReaderDict
+// ignores the dictionary if it is nil.
+func NewReaderDict(r io.Reader, dict []byte) io.ReadCloser {
+	var err error
+	ctx := C.ZBUFF_createDCtx()
+	if len(dict) == 0 {
+		err = getError(int(C.ZBUFF_decompressInit(ctx)))
+	} else {
+		err = getError(int(C.ZBUFF_decompressInitDictionary(
+			ctx,
+			unsafe.Pointer(&dict[0]),
+			C.size_t(len(dict)))))
+	}
+	cSize := int(C.ZBUFF_recommendedDInSize())
+	dSize := int(C.ZBUFF_recommendedDOutSize())
+	if cSize <= 0 {
+		panic(fmt.Errorf("ZBUFF_recommendedDInSize() returned invalid size: %v", cSize))
+	}
+	if dSize <= 0 {
+		panic(fmt.Errorf("ZBUFF_recommendedDOutSize() returned invalid size: %v", dSize))
+	}
+
+	compressionBuffer := make([]byte, cSize)
+	decompressionBuffer := make([]byte, dSize)
+	return &reader{
+		ctx:                 ctx,
+		dict:                dict,
+		compressionBuffer:   compressionBuffer,
+		decompressionBuffer: decompressionBuffer,
+		firstError:          err,
+		recommendedSrcSize:  cSize,
+		underlyingReader:    r,
+	}
+}
+
+// Close frees the allocated C objects
+func (r *reader) Close() error {
+	return getError(int(C.ZBUFF_freeDCtx(r.ctx)))
+}
+
+func (r *reader) Read(p []byte) (int, error) {
+
+	// If we already have enough bytes, return
+	if r.decompSize-r.decompOff >= len(p) {
+		copy(p, r.decompressionBuffer[r.decompOff:])
+		r.decompOff += len(p)
+		return len(p), nil
+	}
+
+	copy(p, r.decompressionBuffer[r.decompOff:r.decompSize])
+	got := r.decompSize - r.decompOff
+	r.decompSize = 0
+	r.decompOff = 0
+
+	for got < len(p) {
+		// Populate src
+		src := r.compressionBuffer
+		reader := r.underlyingReader
+		n, err := TryReadFull(reader, src[r.compressionLeft:])
+		if err != nil && err != errShortRead { // Handle underlying reader errors first
+			return 0, fmt.Errorf("failed to read from underlying reader: %s", err)
+		} else if n == 0 && r.compressionLeft == 0 {
+			return got, io.EOF
+		}
+		src = src[:r.compressionLeft+n]
+
+		// C code
+		cSrcSize := C.size_t(len(src))
+		cDstSize := C.size_t(len(r.decompressionBuffer))
+		retCode := int(C.ZBUFF_decompressContinue(
+			r.ctx,
+			unsafe.Pointer(&r.decompressionBuffer[0]),
+			&cDstSize,
+			unsafe.Pointer(&src[0]),
+			&cSrcSize))
+
+		if err = getError(retCode); err != nil {
+			return 0, fmt.Errorf("failed to decompress: %s", err)
+		}
+
+		// Put everything in buffer
+		if int(cSrcSize) < len(src) {
+			left := src[int(cSrcSize):]
+			copy(r.compressionBuffer, left)
+		}
+		r.compressionLeft = len(src) - int(cSrcSize)
+		r.decompSize = int(cDstSize)
+		r.decompOff = copy(p[got:], r.decompressionBuffer[:r.decompSize])
+		got += r.decompOff
+
+		// Resize buffers
+		nsize := retCode // Hint for next src buffer size
+		if nsize <= 0 {
+			// Reset to recommended size
+			nsize = r.recommendedSrcSize
+		}
+		if nsize < r.compressionLeft {
+			nsize = r.compressionLeft
+		}
+		r.compressionBuffer = resize(r.compressionBuffer, nsize)
+	}
+	return got, nil
+}
+
+// TryReadFull reads buffer just as ReadFull does
+// Here we expect that buffer may end and we do not return ErrUnexpectedEOF as ReadAtLeast does.
+// We return errShortRead instead to distinguish short reads and failures.
+// We cannot use ReadFull/ReadAtLeast because it masks Reader errors, such as network failures
+// and causes panic instead of error.
+func TryReadFull(r io.Reader, buf []byte) (n int, err error) {
+	for n < len(buf) && err == nil {
+		var nn int
+		nn, err = r.Read(buf[n:])
+		n += nn
+	}
+	if n == len(buf) && err == io.EOF {
+		err = nil // EOF at the end is somewhat expected
+	} else if err == io.EOF {
+		err = errShortRead
+	}
+	return
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_v01.c b/vendor/github.com/DataDog/zstd/zstd_v01.c
new file mode 100644
index 0000000..ae1cb2c
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v01.c
@@ -0,0 +1,2127 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/******************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include "zstd_v01.h"
+#include "error_private.h"
+
+
+/******************************************
+*  Static allocation
+******************************************/
+/* You can statically allocate FSE CTable/DTable as a table of unsigned using below macro */
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+/* You can statically allocate Huff0 DTable as a table of unsigned short using below macro */
+#define HUF_DTABLE_SIZE_U16(maxTableLog)   (1 + (1<<maxTableLog))
+#define HUF_CREATE_STATIC_DTABLE(DTable, maxTableLog) \
+        unsigned short DTable[HUF_DTABLE_SIZE_U16(maxTableLog)] = { maxTableLog }
+
+
+/******************************************
+*  Error Management
+******************************************/
+#define FSE_LIST_ERRORS(ITEM) \
+        ITEM(FSE_OK_NoError) ITEM(FSE_ERROR_GENERIC) \
+        ITEM(FSE_ERROR_tableLog_tooLarge) ITEM(FSE_ERROR_maxSymbolValue_tooLarge) ITEM(FSE_ERROR_maxSymbolValue_tooSmall) \
+        ITEM(FSE_ERROR_dstSize_tooSmall) ITEM(FSE_ERROR_srcSize_wrong)\
+        ITEM(FSE_ERROR_corruptionDetected) \
+        ITEM(FSE_ERROR_maxCode)
+
+#define FSE_GENERATE_ENUM(ENUM) ENUM,
+typedef enum { FSE_LIST_ERRORS(FSE_GENERATE_ENUM) } FSE_errorCodes;  /* enum is exposed, to detect & handle specific errors; compare function result to -enum value */
+
+
+/******************************************
+*  FSE symbol compression API
+******************************************/
+/*
+   This API consists of small unitary functions, which highly benefit from being inlined.
+   You will want to enable link-time-optimization to ensure these functions are properly inlined in your binary.
+   Visual seems to do it automatically.
+   For gcc or clang, you'll need to add -flto flag at compilation and linking stages.
+   If none of these solutions is applicable, include "fse.c" directly.
+*/
+
+typedef unsigned FSE_CTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+
+typedef struct
+{
+    size_t bitContainer;
+    int    bitPos;
+    char*  startPtr;
+    char*  ptr;
+    char*  endPtr;
+} FSE_CStream_t;
+
+typedef struct
+{
+    ptrdiff_t   value;
+    const void* stateTable;
+    const void* symbolTT;
+    unsigned    stateLog;
+} FSE_CState_t;
+
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} FSE_DStream_t;
+
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+typedef enum { FSE_DStream_unfinished = 0,
+               FSE_DStream_endOfBuffer = 1,
+               FSE_DStream_completed = 2,
+               FSE_DStream_tooFar = 3 } FSE_DStream_status;  /* result of FSE_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... ?! */
+
+
+/****************************************************************
+*  Tuning parameters
+****************************************************************/
+/* MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSE_MAX_MEMORY_USAGE 14
+#define FSE_DEFAULT_MEMORY_USAGE 13
+
+/* FSE_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSE_MAX_SYMBOL_VALUE 255
+
+
+/****************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+
+
+/****************************************************************
+*  Byte symbol type
+****************************************************************/
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSE_decode_t;   /* size == U32 */
+
+
+
+/****************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  define GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/****************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+
+#ifndef MEM_ACCESS_MODULE
+#define MEM_ACCESS_MODULE
+/****************************************************************
+*  Basic Types
+*****************************************************************/
+#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+# include <stdint.h>
+typedef  uint8_t BYTE;
+typedef uint16_t U16;
+typedef  int16_t S16;
+typedef uint32_t U32;
+typedef  int32_t S32;
+typedef uint64_t U64;
+typedef  int64_t S64;
+#else
+typedef unsigned char       BYTE;
+typedef unsigned short      U16;
+typedef   signed short      S16;
+typedef unsigned int        U32;
+typedef   signed int        S32;
+typedef unsigned long long  U64;
+typedef   signed long long  S64;
+#endif
+
+#endif   /* MEM_ACCESS_MODULE */
+
+/****************************************************************
+*  Memory I/O
+*****************************************************************/
+/* FSE_FORCE_MEMORY_ACCESS
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets generating assembly depending on alignment.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef FSE_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define FSE_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define FSE_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+
+static unsigned FSE_32bits(void)
+{
+    return sizeof(void*)==4;
+}
+
+static unsigned FSE_isLittleEndian(void)
+{
+    const union { U32 i; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(FSE_FORCE_MEMORY_ACCESS) && (FSE_FORCE_MEMORY_ACCESS==2)
+
+static U16 FSE_read16(const void* memPtr) { return *(const U16*) memPtr; }
+static U32 FSE_read32(const void* memPtr) { return *(const U32*) memPtr; }
+static U64 FSE_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+#elif defined(FSE_FORCE_MEMORY_ACCESS) && (FSE_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
+
+static U16 FSE_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+static U32 FSE_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+static U64 FSE_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+#else
+
+static U16 FSE_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+static U32 FSE_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+static U64 FSE_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+#endif // FSE_FORCE_MEMORY_ACCESS
+
+static U16 FSE_readLE16(const void* memPtr)
+{
+    if (FSE_isLittleEndian())
+        return FSE_read16(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+static U32 FSE_readLE32(const void* memPtr)
+{
+    if (FSE_isLittleEndian())
+        return FSE_read32(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+
+static U64 FSE_readLE64(const void* memPtr)
+{
+    if (FSE_isLittleEndian())
+        return FSE_read64(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U64)((U64)p[0] + ((U64)p[1]<<8) + ((U64)p[2]<<16) + ((U64)p[3]<<24)
+                     + ((U64)p[4]<<32) + ((U64)p[5]<<40) + ((U64)p[6]<<48) + ((U64)p[7]<<56));
+    }
+}
+
+static size_t FSE_readLEST(const void* memPtr)
+{
+    if (FSE_32bits())
+        return (size_t)FSE_readLE32(memPtr);
+    else
+        return (size_t)FSE_readLE64(memPtr);
+}
+
+
+
+/****************************************************************
+*  Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+#error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+
+/****************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/****************************************************************
+*  Complex types
+****************************************************************/
+typedef struct
+{
+    int deltaFindState;
+    U32 deltaNbBits;
+} FSE_symbolCompressionTransform; /* total 8 bytes */
+
+typedef U32 DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+
+/****************************************************************
+*  Internal functions
+****************************************************************/
+FORCE_INLINE unsigned FSE_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (GCC_VERSION >= 304)   /* GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+/****************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+
+
+static U32 FSE_tableStep(U32 tableSize) { return (tableSize>>1) + (tableSize>>3) + 3; }
+
+#define FSE_DECODE_TYPE FSE_decode_t
+
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSE_DTableHeader;   /* sizeof U32 */
+
+static size_t FSE_buildDTable
+(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*)(ptr) + 1;   /* because dt is unsigned, 32-bits aligned on 32-bits */
+    const U32 tableSize = 1 << tableLog;
+    const U32 tableMask = tableSize-1;
+    const U32 step = FSE_tableStep(tableSize);
+    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U32 position = 0;
+    U32 highThreshold = tableSize-1;
+    const S16 largeLimit= (S16)(1 << (tableLog-1));
+    U32 noLarge = 1;
+    U32 s;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return (size_t)-FSE_ERROR_maxSymbolValue_tooLarge;
+    if (tableLog > FSE_MAX_TABLELOG) return (size_t)-FSE_ERROR_tableLog_tooLarge;
+
+    /* Init, lay down lowprob symbols */
+    DTableH[0].tableLog = (U16)tableLog;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        if (normalizedCounter[s]==-1)
+        {
+            tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+            symbolNext[s] = 1;
+        }
+        else
+        {
+            if (normalizedCounter[s] >= largeLimit) noLarge=0;
+            symbolNext[s] = normalizedCounter[s];
+        }
+    }
+
+    /* Spread symbols */
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        int i;
+        for (i=0; i<normalizedCounter[s]; i++)
+        {
+            tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+            position = (position + step) & tableMask;
+            while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }
+    }
+
+    if (position!=0) return (size_t)-FSE_ERROR_GENERIC;   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+
+    /* Build Decoding table */
+    {
+        U32 i;
+        for (i=0; i<tableSize; i++)
+        {
+            FSE_FUNCTION_TYPE symbol = (FSE_FUNCTION_TYPE)(tableDecode[i].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[i].nbBits = (BYTE) (tableLog - FSE_highbit32 ((U32)nextState) );
+            tableDecode[i].newState = (U16) ( (nextState << tableDecode[i].nbBits) - tableSize);
+        }
+    }
+
+    DTableH->fastMode = (U16)noLarge;
+    return 0;
+}
+
+
+/******************************************
+*  FSE byte symbol
+******************************************/
+#ifndef FSE_COMMONDEFS_ONLY
+
+static unsigned FSE_isError(size_t code) { return (code > (size_t)(-FSE_ERROR_maxCode)); }
+
+static short FSE_abs(short a)
+{
+    return a<0? -a : a;
+}
+
+
+/****************************************************************
+*  Header bitstream management
+****************************************************************/
+static size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return (size_t)-FSE_ERROR_srcSize_wrong;
+    bitStream = FSE_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return (size_t)-FSE_ERROR_tableLog_tooLarge;
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr))
+    {
+        if (previous0)
+        {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF)
+            {
+                n0+=24;
+                if (ip < iend-5)
+                {
+                    ip+=2;
+                    bitStream = FSE_readLE32(ip) >> bitCount;
+                }
+                else
+                {
+                    bitStream >>= 16;
+                    bitCount+=16;
+                }
+            }
+            while ((bitStream & 3) == 3)
+            {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return (size_t)-FSE_ERROR_maxSymbolValue_tooSmall;
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+            {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = FSE_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {
+            const short max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max)
+            {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            }
+            else
+            {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSE_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold)
+            {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            {
+                if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+                {
+                    ip += bitCount>>3;
+                    bitCount &= 7;
+                }
+                else
+                {
+                    bitCount -= (int)(8 * (iend - 4 - ip));
+                    ip = iend - 4;
+                }
+                bitStream = FSE_readLE32(ip) >> (bitCount & 31);
+            }
+        }
+    }
+    if (remaining != 1) return (size_t)-FSE_ERROR_GENERIC;
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+    return ip-istart;
+}
+
+
+/*********************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const cell = (FSE_decode_t*)(ptr) + 1;   /* because dt is unsigned */
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const dinfo = (FSE_decode_t*)(ptr) + 1;   /* because dt is unsigned */
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return (size_t)-FSE_ERROR_GENERIC;             /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+
+/* FSE_initDStream
+ * Initialize a FSE_DStream_t.
+ * srcBuffer must point at the beginning of an FSE block.
+ * The function result is the size of the FSE_block (== srcSize).
+ * If srcSize is too small, the function will return an errorCode;
+ */
+static size_t FSE_initDStream(FSE_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) return (size_t)-FSE_ERROR_srcSize_wrong;
+
+    if (srcSize >=  sizeof(size_t))
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(size_t);
+        bitD->bitContainer = FSE_readLEST(bitD->ptr);
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return (size_t)-FSE_ERROR_GENERIC;   /* stop bit not present */
+        bitD->bitsConsumed = 8 - FSE_highbit32(contain32);
+    }
+    else
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[6]) << (sizeof(size_t)*8 - 16);
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[5]) << (sizeof(size_t)*8 - 24);
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[4]) << (sizeof(size_t)*8 - 32);
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[3]) << 24;
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[2]) << 16;
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[1]) <<  8;
+            default:;
+        }
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return (size_t)-FSE_ERROR_GENERIC;   /* stop bit not present */
+        bitD->bitsConsumed = 8 - FSE_highbit32(contain32);
+        bitD->bitsConsumed += (U32)(sizeof(size_t) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+
+/*!FSE_lookBits
+ * Provides next n bits from the bitContainer.
+ * bitContainer is not modified (bits are still present for next read/look)
+ * On 32-bits, maxNbBits==25
+ * On 64-bits, maxNbBits==57
+ * return : value extracted.
+ */
+static size_t FSE_lookBits(FSE_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+static size_t FSE_lookBitsFast(FSE_DStream_t* bitD, U32 nbBits)   /* only if nbBits >= 1 !! */
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+static void FSE_skipBits(FSE_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+
+/*!FSE_readBits
+ * Read next n bits from the bitContainer.
+ * On 32-bits, don't read more than maxNbBits==25
+ * On 64-bits, don't read more than maxNbBits==57
+ * Use the fast variant *only* if n >= 1.
+ * return : value extracted.
+ */
+static size_t FSE_readBits(FSE_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = FSE_lookBits(bitD, nbBits);
+    FSE_skipBits(bitD, nbBits);
+    return value;
+}
+
+static size_t FSE_readBitsFast(FSE_DStream_t* bitD, U32 nbBits)   /* only if nbBits >= 1 !! */
+{
+    size_t value = FSE_lookBitsFast(bitD, nbBits);
+    FSE_skipBits(bitD, nbBits);
+    return value;
+}
+
+static unsigned FSE_reloadDStream(FSE_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return FSE_DStream_tooFar;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer))
+    {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = FSE_readLEST(bitD->ptr);
+        return FSE_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start)
+    {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return FSE_DStream_endOfBuffer;
+        return FSE_DStream_completed;
+    }
+    {
+        U32 nbBytes = bitD->bitsConsumed >> 3;
+        U32 result = FSE_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start)
+        {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = FSE_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = FSE_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+
+static void FSE_initDState(FSE_DState_t* DStatePtr, FSE_DStream_t* bitD, const FSE_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;
+    DStatePtr->state = FSE_readBits(bitD, DTableH->tableLog);
+    FSE_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+static BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, FSE_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32  nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = FSE_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+static BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, FSE_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32 nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = FSE_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+/* FSE_endOfDStream
+   Tells if bitD has reached end of bitStream or not */
+
+static unsigned FSE_endOfDStream(const FSE_DStream_t* bitD)
+{
+    return ((bitD->ptr == bitD->start) && (bitD->bitsConsumed == sizeof(bitD->bitContainer)*8));
+}
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+FORCE_INLINE size_t FSE_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    FSE_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+    size_t errorCode;
+
+    /* Init */
+    errorCode = FSE_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+    if (FSE_isError(errorCode)) return errorCode;
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (FSE_reloadDStream(&bitD)==FSE_DStream_unfinished) && (op<olimit) ; op+=4)
+    {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            FSE_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (FSE_reloadDStream(&bitD) > FSE_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            FSE_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : FSE_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly FSE_DStream_completed */
+    while (1)
+    {
+        if ( (FSE_reloadDStream(&bitD)>FSE_DStream_completed) || (op==omax) || (FSE_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state1))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state1);
+
+        if ( (FSE_reloadDStream(&bitD)>FSE_DStream_completed) || (op==omax) || (FSE_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state2))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state2);
+    }
+
+    /* end ? */
+    if (FSE_endOfDStream(&bitD) && FSE_endOfDState(&state1) && FSE_endOfDState(&state2))
+        return op-ostart;
+
+    if (op==omax) return (size_t)-FSE_ERROR_dstSize_tooSmall;   /* dst buffer is full, but cSrc unfinished */
+
+    return (size_t)-FSE_ERROR_corruptionDetected;
+}
+
+
+static size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));   /* memcpy() into local variable, to avoid strict aliasing warning */
+
+    /* select fast mode (static) */
+    if (DTableH.fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSE_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    size_t errorCode;
+
+    if (cSrcSize<2) return (size_t)-FSE_ERROR_srcSize_wrong;   /* too small input size */
+
+    /* normal FSE decoding mode */
+    errorCode = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSE_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return (size_t)-FSE_ERROR_srcSize_wrong;   /* too small input size */
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    errorCode = FSE_buildDTable (dt, counting, maxSymbolValue, tableLog);
+    if (FSE_isError(errorCode)) return errorCode;
+
+    /* always return, even if it is an error code */
+    return FSE_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);
+}
+
+
+
+/* *******************************************************
+*  Huff0 : Huffman block compression
+*********************************************************/
+#define HUF_MAX_SYMBOL_VALUE 255
+#define HUF_DEFAULT_TABLELOG  12       /* used by default, when not specified */
+#define HUF_MAX_TABLELOG  12           /* max possible tableLog; for allocation purpose; can be modified */
+#define HUF_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#if (HUF_MAX_TABLELOG > HUF_ABSOLUTEMAX_TABLELOG)
+#  error "HUF_MAX_TABLELOG is too large !"
+#endif
+
+typedef struct HUF_CElt_s {
+  U16  val;
+  BYTE nbBits;
+} HUF_CElt ;
+
+typedef struct nodeElt_s {
+    U32 count;
+    U16 parent;
+    BYTE byte;
+    BYTE nbBits;
+} nodeElt;
+
+
+/* *******************************************************
+*  Huff0 : Huffman block decompression
+*********************************************************/
+typedef struct {
+    BYTE byte;
+    BYTE nbBits;
+} HUF_DElt;
+
+static size_t HUF_readDTable (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];  /* large enough for values from 0 to 16 */
+    U32 weightTotal;
+    U32 maxBits;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+    U32 n;
+    U32 nextRankStart;
+    void* ptr = DTable+1;
+    HUF_DElt* const dt = (HUF_DElt*)ptr;
+
+    if (!srcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+    iSize = ip[0];
+
+    FSE_STATIC_ASSERT(sizeof(HUF_DElt) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* should not be necessary, but some analyzer complain ... */
+    if (iSize >= 128)  /* special header */
+    {
+        if (iSize >= (242))   /* RLE */
+        {
+            static int l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, sizeof(huffWeight));
+            iSize = 0;
+        }
+        else   /* Incompressible */
+        {
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+            ip += 1;
+            for (n=0; n<oSize; n+=2)
+            {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+            }
+        }
+    }
+    else  /* header compressed with FSE (normal case) */
+    {
+        if (iSize+1 > srcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+        oSize = FSE_decompress(huffWeight, HUF_MAX_SYMBOL_VALUE, ip+1, iSize);   /* max 255 values decoded, last one is implied */
+        if (FSE_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankVal, 0, sizeof(rankVal));
+    weightTotal = 0;
+    for (n=0; n<oSize; n++)
+    {
+        if (huffWeight[n] >= HUF_ABSOLUTEMAX_TABLELOG) return (size_t)-FSE_ERROR_corruptionDetected;
+        rankVal[huffWeight[n]]++;
+        weightTotal += (1 << huffWeight[n]) >> 1;
+    }
+    if (weightTotal == 0) return (size_t)-FSE_ERROR_corruptionDetected;
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    maxBits = FSE_highbit32(weightTotal) + 1;
+    if (maxBits > DTable[0]) return (size_t)-FSE_ERROR_tableLog_tooLarge;   /* DTable is too small */
+    DTable[0] = (U16)maxBits;
+    {
+        U32 total = 1 << maxBits;
+        U32 rest = total - weightTotal;
+        U32 verif = 1 << FSE_highbit32(rest);
+        U32 lastWeight = FSE_highbit32(rest) + 1;
+        if (verif != rest) return (size_t)-FSE_ERROR_corruptionDetected;    /* last value must be a clean power of 2 */
+        huffWeight[oSize] = (BYTE)lastWeight;
+        rankVal[lastWeight]++;
+    }
+
+    /* check tree construction validity */
+    if ((rankVal[1] < 2) || (rankVal[1] & 1)) return (size_t)-FSE_ERROR_corruptionDetected;   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<=maxBits; n++)
+    {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<=oSize; n++)
+    {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUF_DElt D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(maxBits + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize+1;
+}
+
+
+static BYTE HUF_decodeSymbol(FSE_DStream_t* Dstream, const HUF_DElt* dt, const U32 dtLog)
+{
+        const size_t val = FSE_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+        const BYTE c = dt[val].byte;
+        FSE_skipBits(Dstream, dt[val].nbBits);
+        return c;
+}
+
+static size_t HUF_decompress_usingDTable(   /* -3% slower when non static */
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-15;
+
+    const void* ptr = DTable;
+    const HUF_DElt* const dt = (const HUF_DElt*)(ptr)+1;
+    const U32 dtLog = DTable[0];
+    size_t errorCode;
+    U32 reloadStatus;
+
+    /* Init */
+
+    const U16* jumpTable = (const U16*)cSrc;
+    const size_t length1 = FSE_readLE16(jumpTable);
+    const size_t length2 = FSE_readLE16(jumpTable+1);
+    const size_t length3 = FSE_readLE16(jumpTable+2);
+    const size_t length4 = cSrcSize - 6 - length1 - length2 - length3;   // check coherency !!
+    const char* const start1 = (const char*)(cSrc) + 6;
+    const char* const start2 = start1 + length1;
+    const char* const start3 = start2 + length2;
+    const char* const start4 = start3 + length3;
+    FSE_DStream_t bitD1, bitD2, bitD3, bitD4;
+
+    if (length1+length2+length3+6 >= cSrcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+
+    errorCode = FSE_initDStream(&bitD1, start1, length1);
+    if (FSE_isError(errorCode)) return errorCode;
+    errorCode = FSE_initDStream(&bitD2, start2, length2);
+    if (FSE_isError(errorCode)) return errorCode;
+    errorCode = FSE_initDStream(&bitD3, start3, length3);
+    if (FSE_isError(errorCode)) return errorCode;
+    errorCode = FSE_initDStream(&bitD4, start4, length4);
+    if (FSE_isError(errorCode)) return errorCode;
+
+    reloadStatus=FSE_reloadDStream(&bitD2);
+
+    /* 16 symbols per loop */
+    for ( ; (reloadStatus<FSE_DStream_completed) && (op<olimit);  /* D2-3-4 are supposed to be synchronized and finish together */
+        op+=16, reloadStatus = FSE_reloadDStream(&bitD2) | FSE_reloadDStream(&bitD3) | FSE_reloadDStream(&bitD4), FSE_reloadDStream(&bitD1))
+    {
+#define HUF_DECODE_SYMBOL_0(n, Dstream) \
+        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog);
+
+#define HUF_DECODE_SYMBOL_1(n, Dstream) \
+        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog); \
+        if (FSE_32bits() && (HUF_MAX_TABLELOG>12)) FSE_reloadDStream(&Dstream)
+
+#define HUF_DECODE_SYMBOL_2(n, Dstream) \
+        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog); \
+        if (FSE_32bits()) FSE_reloadDStream(&Dstream)
+
+        HUF_DECODE_SYMBOL_1( 0, bitD1);
+        HUF_DECODE_SYMBOL_1( 1, bitD2);
+        HUF_DECODE_SYMBOL_1( 2, bitD3);
+        HUF_DECODE_SYMBOL_1( 3, bitD4);
+        HUF_DECODE_SYMBOL_2( 4, bitD1);
+        HUF_DECODE_SYMBOL_2( 5, bitD2);
+        HUF_DECODE_SYMBOL_2( 6, bitD3);
+        HUF_DECODE_SYMBOL_2( 7, bitD4);
+        HUF_DECODE_SYMBOL_1( 8, bitD1);
+        HUF_DECODE_SYMBOL_1( 9, bitD2);
+        HUF_DECODE_SYMBOL_1(10, bitD3);
+        HUF_DECODE_SYMBOL_1(11, bitD4);
+        HUF_DECODE_SYMBOL_0(12, bitD1);
+        HUF_DECODE_SYMBOL_0(13, bitD2);
+        HUF_DECODE_SYMBOL_0(14, bitD3);
+        HUF_DECODE_SYMBOL_0(15, bitD4);
+    }
+
+    if (reloadStatus!=FSE_DStream_completed)   /* not complete : some bitStream might be FSE_DStream_unfinished */
+        return (size_t)-FSE_ERROR_corruptionDetected;
+
+    /* tail */
+    {
+        // bitTail = bitD1;   // *much* slower : -20% !??!
+        FSE_DStream_t bitTail;
+        bitTail.ptr = bitD1.ptr;
+        bitTail.bitsConsumed = bitD1.bitsConsumed;
+        bitTail.bitContainer = bitD1.bitContainer;   // required in case of FSE_DStream_endOfBuffer
+        bitTail.start = start1;
+        for ( ; (FSE_reloadDStream(&bitTail) < FSE_DStream_completed) && (op<omax) ; op++)
+        {
+            HUF_DECODE_SYMBOL_0(0, bitTail);
+        }
+
+        if (FSE_endOfDStream(&bitTail))
+            return op-ostart;
+    }
+
+    if (op==omax) return (size_t)-FSE_ERROR_dstSize_tooSmall;   /* dst buffer is full, but cSrc unfinished */
+
+    return (size_t)-FSE_ERROR_corruptionDetected;
+}
+
+
+static size_t HUF_decompress (void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLE(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUF_readDTable (DTable, cSrc, cSrcSize);
+    if (FSE_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUF_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, DTable);
+}
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
+
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/****************************************************************
+*  Tuning parameters
+*****************************************************************/
+/* MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect */
+#define ZSTD_MEMORY_USAGE 17
+
+
+/**************************************
+   CPU Feature Detection
+**************************************/
+/*
+ * Automated efficient unaligned memory access detection
+ * Based on known hardware architectures
+ * This list will be updated thanks to feedbacks
+ */
+#if defined(CPU_HAS_EFFICIENT_UNALIGNED_MEMORY_ACCESS) \
+    || defined(__ARM_FEATURE_UNALIGNED) \
+    || defined(__i386__) || defined(__x86_64__) \
+    || defined(_M_IX86) || defined(_M_X64) \
+    || defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_8__) \
+    || (defined(_M_ARM) && (_M_ARM >= 7))
+#  define ZSTD_UNALIGNED_ACCESS 1
+#else
+#  define ZSTD_UNALIGNED_ACCESS 0
+#endif
+
+
+/********************************************************
+*  Includes
+*********************************************************/
+#include <stdlib.h>      /* calloc */
+#include <string.h>      /* memcpy, memmove */
+#include <stdio.h>       /* debug : printf */
+
+
+/********************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef __AVX2__
+#  include <immintrin.h>   /* AVX2 intrinsics */
+#endif
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+
+#ifndef MEM_ACCESS_MODULE
+#define MEM_ACCESS_MODULE
+/********************************************************
+*  Basic Types
+*********************************************************/
+#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+# include <stdint.h>
+typedef  uint8_t BYTE;
+typedef uint16_t U16;
+typedef  int16_t S16;
+typedef uint32_t U32;
+typedef  int32_t S32;
+typedef uint64_t U64;
+#else
+typedef unsigned char       BYTE;
+typedef unsigned short      U16;
+typedef   signed short      S16;
+typedef unsigned int        U32;
+typedef   signed int        S32;
+typedef unsigned long long  U64;
+#endif
+
+#endif   /* MEM_ACCESS_MODULE */
+
+
+/********************************************************
+*  Constants
+*********************************************************/
+static const U32 ZSTD_magicNumber = 0xFD2FB51E;   /* 3rd version : seqNb header */
+
+#define HASH_LOG (ZSTD_MEMORY_USAGE - 2)
+#define HASH_TABLESIZE (1 << HASH_LOG)
+#define HASH_MASK (HASH_TABLESIZE - 1)
+
+#define KNUTH 2654435761
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BLOCKSIZE (128 KB)                 /* define, for static allocation */
+
+#define WORKPLACESIZE (BLOCKSIZE*3)
+#define MINMATCH 4
+#define MLbits   7
+#define LLbits   6
+#define Offbits  5
+#define MaxML  ((1<<MLbits )-1)
+#define MaxLL  ((1<<LLbits )-1)
+#define MaxOff ((1<<Offbits)-1)
+#define LitFSELog  11
+#define MLFSELog   10
+#define LLFSELog   10
+#define OffFSELog   9
+#define MAX(a,b) ((a)<(b)?(b):(a))
+#define MaxSeq MAX(MaxLL, MaxML)
+
+#define LITERAL_NOENTROPY 63
+#define COMMAND_NOENTROPY 7   /* to remove */
+
+static const size_t ZSTD_blockHeaderSize = 3;
+static const size_t ZSTD_frameHeaderSize = 4;
+
+
+/********************************************************
+*  Memory operations
+*********************************************************/
+static unsigned ZSTD_32bits(void) { return sizeof(void*)==4; }
+
+static unsigned ZSTD_isLittleEndian(void)
+{
+    const union { U32 i; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+static U16    ZSTD_read16(const void* p) { U16 r; memcpy(&r, p, sizeof(r)); return r; }
+
+static U32    ZSTD_read32(const void* p) { U32 r; memcpy(&r, p, sizeof(r)); return r; }
+
+static void   ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+static void   ZSTD_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+
+#define COPY8(d,s)    { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+static void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    while (op < oend) COPY8(op, ip);
+}
+
+static U16 ZSTD_readLE16(const void* memPtr)
+{
+    if (ZSTD_isLittleEndian()) return ZSTD_read16(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)((U16)p[0] + ((U16)p[1]<<8));
+    }
+}
+
+
+static U32 ZSTD_readLE32(const void* memPtr)
+{
+    if (ZSTD_isLittleEndian())
+        return ZSTD_read32(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+static U32 ZSTD_readBE32(const void* memPtr)
+{
+    const BYTE* p = (const BYTE*)memPtr;
+    return (U32)(((U32)p[0]<<24) + ((U32)p[1]<<16) + ((U32)p[2]<<8) + ((U32)p[3]<<0));
+}
+
+
+/**************************************
+*  Local structures
+***************************************/
+typedef struct ZSTD_Cctx_s ZSTD_Cctx;
+
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* offCode;
+    BYTE* litStart;
+    BYTE* lit;
+    BYTE* litLengthStart;
+    BYTE* litLength;
+    BYTE* matchLengthStart;
+    BYTE* matchLength;
+    BYTE* dumpsStart;
+    BYTE* dumps;
+} seqStore_t;
+
+
+typedef struct ZSTD_Cctx_s
+{
+    const BYTE* base;
+    U32 current;
+    U32 nextUpdate;
+    seqStore_t seqStore;
+#ifdef __AVX2__
+    __m256i hashTable[HASH_TABLESIZE>>3];
+#else
+    U32 hashTable[HASH_TABLESIZE];
+#endif
+    BYTE buffer[WORKPLACESIZE];
+} cctxi_t;
+
+
+
+
+/**************************************
+*  Error Management
+**************************************/
+/* published entry point */
+unsigned ZSTDv01_isError(size_t code) { return ERR_isError(code); }
+
+
+/**************************************
+*  Tool functions
+**************************************/
+#define ZSTD_VERSION_MAJOR    0    /* for breaking interface changes  */
+#define ZSTD_VERSION_MINOR    1    /* for new (non-breaking) interface capabilities */
+#define ZSTD_VERSION_RELEASE  3    /* for tweaks, bug-fixes, or development */
+#define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
+
+/**************************************************************
+*   Decompression code
+**************************************************************/
+
+size_t ZSTDv01_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    BYTE headerFlags;
+    U32 cSize;
+
+    if (srcSize < 3) return ERROR(srcSize_wrong);
+
+    headerFlags = *in;
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+
+    bpPtr->blockType = (blockType_t)(headerFlags >> 6);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+
+static size_t ZSTD_copyUncompressedBlock(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    if (srcSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+static size_t ZSTD_decompressLiterals(void* ctx,
+                                      void* dst, size_t maxDstSize,
+                                const void* src, size_t srcSize)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + maxDstSize;
+    const BYTE* ip = (const BYTE*)src;
+    size_t errorCode;
+    size_t litSize;
+
+    /* check : minimum 2, for litSize, +1, for content */
+    if (srcSize <= 3) return ERROR(corruption_detected);
+
+    litSize = ip[1] + (ip[0]<<8);
+    litSize += ((ip[-3] >> 3) & 7) << 16;   // mmmmh....
+    op = oend - litSize;
+
+    (void)ctx;
+    if (litSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    errorCode = HUF_decompress(op, litSize, ip+2, srcSize-2);
+    if (FSE_isError(errorCode)) return ERROR(GENERIC);
+    return litSize;
+}
+
+
+size_t ZSTDv01_decodeLiteralsBlock(void* ctx,
+                                void* dst, size_t maxDstSize,
+                          const BYTE** litStart, size_t* litSize,
+                          const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + maxDstSize;
+    blockProperties_t litbp;
+
+    size_t litcSize = ZSTDv01_getcBlockSize(src, srcSize, &litbp);
+    if (ZSTDv01_isError(litcSize)) return litcSize;
+    if (litcSize > srcSize - ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    ip += ZSTD_blockHeaderSize;
+
+    switch(litbp.blockType)
+    {
+    case bt_raw:
+        *litStart = ip;
+        ip += litcSize;
+        *litSize = litcSize;
+        break;
+    case bt_rle:
+        {
+            size_t rleSize = litbp.origSize;
+            if (rleSize>maxDstSize) return ERROR(dstSize_tooSmall);
+            if (!srcSize) return ERROR(srcSize_wrong);
+            memset(oend - rleSize, *ip, rleSize);
+            *litStart = oend - rleSize;
+            *litSize = rleSize;
+            ip++;
+            break;
+        }
+    case bt_compressed:
+        {
+            size_t decodedLitSize = ZSTD_decompressLiterals(ctx, dst, maxDstSize, ip, litcSize);
+            if (ZSTDv01_isError(decodedLitSize)) return decodedLitSize;
+            *litStart = oend - decodedLitSize;
+            *litSize = decodedLitSize;
+            ip += litcSize;
+            break;
+        }
+    case bt_end:
+    default:
+        return ERROR(GENERIC);
+    }
+
+    return ip-istart;
+}
+
+
+size_t ZSTDv01_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t* dumpsLengthPtr,
+                         FSE_DTable* DTableLL, FSE_DTable* DTableML, FSE_DTable* DTableOffb,
+                         const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    const BYTE* const iend = istart + srcSize;
+    U32 LLtype, Offtype, MLtype;
+    U32 LLlog, Offlog, MLlog;
+    size_t dumpsLength;
+
+    /* check */
+    if (srcSize < 5) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    *nbSeq = ZSTD_readLE16(ip); ip+=2;
+    LLtype  = *ip >> 6;
+    Offtype = (*ip >> 4) & 3;
+    MLtype  = (*ip >> 2) & 3;
+    if (*ip & 2)
+    {
+        dumpsLength  = ip[2];
+        dumpsLength += ip[1] << 8;
+        ip += 3;
+    }
+    else
+    {
+        dumpsLength  = ip[1];
+        dumpsLength += (ip[0] & 1) << 8;
+        ip += 2;
+    }
+    *dumpsPtr = ip;
+    ip += dumpsLength;
+    *dumpsLengthPtr = dumpsLength;
+
+    /* check */
+    if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+    /* sequences */
+    {
+        S16 norm[MaxML+1];    /* assumption : MaxML >= MaxLL and MaxOff */
+        size_t headerSize;
+
+        /* Build DTables */
+        switch(LLtype)
+        {
+        case bt_rle :
+            LLlog = 0;
+            FSE_buildDTable_rle(DTableLL, *ip++); break;
+        case bt_raw :
+            LLlog = LLbits;
+            FSE_buildDTable_raw(DTableLL, LLbits); break;
+        default :
+            {   U32 max = MaxLL;
+                headerSize = FSE_readNCount(norm, &max, &LLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (LLlog > LLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableLL, norm, max, LLlog);
+        }   }
+
+        switch(Offtype)
+        {
+        case bt_rle :
+            Offlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableOffb, *ip++); break;
+        case bt_raw :
+            Offlog = Offbits;
+            FSE_buildDTable_raw(DTableOffb, Offbits); break;
+        default :
+            {   U32 max = MaxOff;
+                headerSize = FSE_readNCount(norm, &max, &Offlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (Offlog > OffFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableOffb, norm, max, Offlog);
+        }   }
+
+        switch(MLtype)
+        {
+        case bt_rle :
+            MLlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableML, *ip++); break;
+        case bt_raw :
+            MLlog = MLbits;
+            FSE_buildDTable_raw(DTableML, MLbits); break;
+        default :
+            {   U32 max = MaxML;
+                headerSize = FSE_readNCount(norm, &max, &MLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (MLlog > MLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableML, norm, max, MLlog);
+    }   }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t offset;
+    size_t matchLength;
+} seq_t;
+
+typedef struct {
+    FSE_DStream_t DStream;
+    FSE_DState_t stateLL;
+    FSE_DState_t stateOffb;
+    FSE_DState_t stateML;
+    size_t prevOffset;
+    const BYTE* dumps;
+    const BYTE* dumpsEnd;
+} seqState_t;
+
+
+static void ZSTD_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    size_t litLength;
+    size_t prevOffset;
+    size_t offset;
+    size_t matchLength;
+    const BYTE* dumps = seqState->dumps;
+    const BYTE* const de = seqState->dumpsEnd;
+
+    /* Literal length */
+    litLength = FSE_decodeSymbol(&(seqState->stateLL), &(seqState->DStream));
+    prevOffset = litLength ? seq->offset : seqState->prevOffset;
+    seqState->prevOffset = seq->offset;
+    if (litLength == MaxLL)
+    {
+        U32 add = dumps<de ? *dumps++ : 0;
+        if (add < 255) litLength += add;
+        else
+        {
+            if (dumps<=(de-3))
+            {
+                litLength = ZSTD_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+                dumps += 3;
+            }
+        }
+    }
+
+    /* Offset */
+    {
+        U32 offsetCode, nbBits;
+        offsetCode = FSE_decodeSymbol(&(seqState->stateOffb), &(seqState->DStream));
+        if (ZSTD_32bits()) FSE_reloadDStream(&(seqState->DStream));
+        nbBits = offsetCode - 1;
+        if (offsetCode==0) nbBits = 0;   /* cmove */
+        offset = ((size_t)1 << (nbBits & ((sizeof(offset)*8)-1))) + FSE_readBits(&(seqState->DStream), nbBits);
+        if (ZSTD_32bits()) FSE_reloadDStream(&(seqState->DStream));
+        if (offsetCode==0) offset = prevOffset;
+    }
+
+    /* MatchLength */
+    matchLength = FSE_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
+    if (matchLength == MaxML)
+    {
+        U32 add = dumps<de ? *dumps++ : 0;
+        if (add < 255) matchLength += add;
+        else
+        {
+            if (dumps<=(de-3))
+            {
+                matchLength = ZSTD_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+                dumps += 3;
+            }
+        }
+    }
+    matchLength += MINMATCH;
+
+    /* save result */
+    seq->litLength = litLength;
+    seq->offset = offset;
+    seq->matchLength = matchLength;
+    seqState->dumps = dumps;
+}
+
+
+static size_t ZSTD_execSequence(BYTE* op,
+                                seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                BYTE* const base, BYTE* const oend)
+{
+    static const int dec32table[] = {0, 1, 2, 1, 4, 4, 4, 4};   /* added */
+    static const int dec64table[] = {8, 8, 8, 7, 8, 9,10,11};   /* substracted */
+    const BYTE* const ostart = op;
+    const size_t litLength = sequence.litLength;
+    BYTE* const endMatch = op + litLength + sequence.matchLength;    /* risk : address space overflow (32-bits) */
+    const BYTE* const litEnd = *litPtr + litLength;
+
+    /* check */
+    if (endMatch > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);
+    if (sequence.matchLength > (size_t)(*litPtr-op))  return ERROR(dstSize_tooSmall);    /* overwrite literal segment */
+
+    /* copy Literals */
+    if (((size_t)(*litPtr - op) < 8) || ((size_t)(oend-litEnd) < 8) || (op+litLength > oend-8))
+        memmove(op, *litPtr, litLength);   /* overwrite risk */
+    else
+        ZSTD_wildcopy(op, *litPtr, litLength);
+    op += litLength;
+    *litPtr = litEnd;   /* update for next sequence */
+
+    /* check : last match must be at a minimum distance of 8 from end of dest buffer */
+    if (oend-op < 8) return ERROR(dstSize_tooSmall);
+
+    /* copy Match */
+    {
+        const U32 overlapRisk = (((size_t)(litEnd - endMatch)) < 12);
+        const BYTE* match = op - sequence.offset;            /* possible underflow at op - offset ? */
+        size_t qutt = 12;
+        U64 saved[2];
+
+        /* check */
+        if (match < base) return ERROR(corruption_detected);
+        if (sequence.offset > (size_t)base) return ERROR(corruption_detected);
+
+        /* save beginning of literal sequence, in case of write overlap */
+        if (overlapRisk)
+        {
+            if ((endMatch + qutt) > oend) qutt = oend-endMatch;
+            memcpy(saved, endMatch, qutt);
+        }
+
+        if (sequence.offset < 8)
+        {
+            const int dec64 = dec64table[sequence.offset];
+            op[0] = match[0];
+            op[1] = match[1];
+            op[2] = match[2];
+            op[3] = match[3];
+            match += dec32table[sequence.offset];
+            ZSTD_copy4(op+4, match);
+            match -= dec64;
+        } else { ZSTD_copy8(op, match); }
+        op += 8; match += 8;
+
+        if (endMatch > oend-(16-MINMATCH))
+        {
+            if (op < oend-8)
+            {
+                ZSTD_wildcopy(op, match, (oend-8) - op);
+                match += (oend-8) - op;
+                op = oend-8;
+            }
+            while (op<endMatch) *op++ = *match++;
+        }
+        else
+            ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+
+        /* restore, in case of overlap */
+        if (overlapRisk) memcpy(endMatch, saved, qutt);
+    }
+
+    return endMatch-ostart;
+}
+
+typedef struct ZSTDv01_Dctx_s
+{
+    U32 LLTable[FSE_DTABLE_SIZE_U32(LLFSELog)];
+    U32 OffTable[FSE_DTABLE_SIZE_U32(OffFSELog)];
+    U32 MLTable[FSE_DTABLE_SIZE_U32(MLFSELog)];
+    void* previousDstEnd;
+    void* base;
+    size_t expected;
+    blockType_t bType;
+    U32 phase;
+} dctx_t;
+
+
+static size_t ZSTD_decompressSequences(
+                               void* ctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize,
+                         const BYTE* litStart, size_t litSize)
+{
+    dctx_t* dctx = (dctx_t*)ctx;
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t errorCode, dumpsLength;
+    const BYTE* litPtr = litStart;
+    const BYTE* const litEnd = litStart + litSize;
+    int nbSeq;
+    const BYTE* dumps;
+    U32* DTableLL = dctx->LLTable;
+    U32* DTableML = dctx->MLTable;
+    U32* DTableOffb = dctx->OffTable;
+    BYTE* const base = (BYTE*) (dctx->base);
+
+    /* Build Decoding Tables */
+    errorCode = ZSTDv01_decodeSeqHeaders(&nbSeq, &dumps, &dumpsLength,
+                                      DTableLL, DTableML, DTableOffb,
+                                      ip, iend-ip);
+    if (ZSTDv01_isError(errorCode)) return errorCode;
+    ip += errorCode;
+
+    /* Regen sequences */
+    {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        seqState.dumps = dumps;
+        seqState.dumpsEnd = dumps + dumpsLength;
+        seqState.prevOffset = 1;
+        errorCode = FSE_initDStream(&(seqState.DStream), ip, iend-ip);
+        if (FSE_isError(errorCode)) return ERROR(corruption_detected);
+        FSE_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSE_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSE_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (FSE_reloadDStream(&(seqState.DStream)) <= FSE_DStream_completed) && (nbSeq>0) ; )
+        {
+            size_t oneSeqSize;
+            nbSeq--;
+            ZSTD_decodeSequence(&sequence, &seqState);
+            oneSeqSize = ZSTD_execSequence(op, sequence, &litPtr, litEnd, base, oend);
+            if (ZSTDv01_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* check if reached exact end */
+        if ( !FSE_endOfDStream(&(seqState.DStream)) ) return ERROR(corruption_detected);   /* requested too much : data is corrupted */
+        if (nbSeq<0) return ERROR(corruption_detected);   /* requested too many sequences : data is corrupted */
+
+        /* last literal segment */
+        {
+            size_t lastLLSize = litEnd - litPtr;
+            if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+            if (op != litPtr) memmove(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+    }
+
+    return op-ostart;
+}
+
+
+static size_t ZSTD_decompressBlock(
+                            void* ctx,
+                            void* dst, size_t maxDstSize,
+                      const void* src, size_t srcSize)
+{
+    /* blockType == blockCompressed, srcSize is trusted */
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* litPtr = NULL;
+    size_t litSize = 0;
+    size_t errorCode;
+
+    /* Decode literals sub-block */
+    errorCode = ZSTDv01_decodeLiteralsBlock(ctx, dst, maxDstSize, &litPtr, &litSize, src, srcSize);
+    if (ZSTDv01_isError(errorCode)) return errorCode;
+    ip += errorCode;
+    srcSize -= errorCode;
+
+    return ZSTD_decompressSequences(ctx, dst, maxDstSize, ip, srcSize, litPtr, litSize);
+}
+
+
+size_t ZSTDv01_decompressDCtx(void* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    size_t errorCode=0;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = ZSTD_readBE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t blockSize = ZSTDv01_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTDv01_isError(blockSize)) return blockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (blockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            errorCode = ZSTD_decompressBlock(ctx, op, oend-op, ip, blockSize);
+            break;
+        case bt_raw :
+            errorCode = ZSTD_copyUncompressedBlock(op, oend-op, ip, blockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);
+        }
+        if (blockSize == 0) break;   /* bt_end */
+
+        if (ZSTDv01_isError(errorCode)) return errorCode;
+        op += errorCode;
+        ip += blockSize;
+        remainingSize -= blockSize;
+    }
+
+    return op-ostart;
+}
+
+size_t ZSTDv01_decompress(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    dctx_t ctx;
+    ctx.base = dst;
+    return ZSTDv01_decompressDCtx(&ctx, dst, maxDstSize, src, srcSize);
+}
+
+size_t ZSTDv01_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = ZSTD_readBE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t blockSize = ZSTDv01_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTDv01_isError(blockSize)) return blockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (blockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (blockSize == 0) break;   /* bt_end */
+
+        ip += blockSize;
+        remainingSize -= blockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/*******************************
+*  Streaming Decompression API
+*******************************/
+
+size_t ZSTDv01_resetDCtx(ZSTDv01_Dctx* dctx)
+{
+    dctx->expected = ZSTD_frameHeaderSize;
+    dctx->phase = 0;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    return 0;
+}
+
+ZSTDv01_Dctx* ZSTDv01_createDCtx(void)
+{
+    ZSTDv01_Dctx* dctx = (ZSTDv01_Dctx*)malloc(sizeof(ZSTDv01_Dctx));
+    if (dctx==NULL) return NULL;
+    ZSTDv01_resetDCtx(dctx);
+    return dctx;
+}
+
+size_t ZSTDv01_freeDCtx(ZSTDv01_Dctx* dctx)
+{
+    free(dctx);
+    return 0;
+}
+
+size_t ZSTDv01_nextSrcSizeToDecompress(ZSTDv01_Dctx* dctx)
+{
+    return ((dctx_t*)dctx)->expected;
+}
+
+size_t ZSTDv01_decompressContinue(ZSTDv01_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    dctx_t* ctx = (dctx_t*)dctx;
+
+    /* Sanity check */
+    if (srcSize != ctx->expected) return ERROR(srcSize_wrong);
+    if (dst != ctx->previousDstEnd)  /* not contiguous */
+        ctx->base = dst;
+
+    /* Decompress : frame header */
+    if (ctx->phase == 0)
+    {
+        /* Check frame magic header */
+        U32 magicNumber = ZSTD_readBE32(src);
+        if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        return 0;
+    }
+
+    /* Decompress : block header */
+    if (ctx->phase == 1)
+    {
+        blockProperties_t bp;
+        size_t blockSize = ZSTDv01_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
+        if (ZSTDv01_isError(blockSize)) return blockSize;
+        if (bp.blockType == bt_end)
+        {
+            ctx->expected = 0;
+            ctx->phase = 0;
+        }
+        else
+        {
+            ctx->expected = blockSize;
+            ctx->bType = bp.blockType;
+            ctx->phase = 2;
+        }
+
+        return 0;
+    }
+
+    /* Decompress : block content */
+    {
+        size_t rSize;
+        switch(ctx->bType)
+        {
+        case bt_compressed:
+            rSize = ZSTD_decompressBlock(ctx, dst, maxDstSize, src, srcSize);
+            break;
+        case bt_raw :
+            rSize = ZSTD_copyUncompressedBlock(dst, maxDstSize, src, srcSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet handled */
+            break;
+        case bt_end :   /* should never happen (filtered at phase 1) */
+            rSize = 0;
+            break;
+        default:
+            return ERROR(GENERIC);
+        }
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        ctx->previousDstEnd = (void*)( ((char*)dst) + rSize);
+        return rSize;
+    }
+
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_v01.h b/vendor/github.com/DataDog/zstd/zstd_v01.h
new file mode 100644
index 0000000..42f0897
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v01.h
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_V01_H_28739879432
+#define ZSTD_V01_H_28739879432
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Simple one-step function
+***************************************/
+/**
+ZSTDv01_decompress() : decompress ZSTD frames compliant with v0.1.x format
+    compressedSize : is the exact source size
+    maxOriginalSize : is the size of the 'dst' buffer, which must be already allocated.
+                      It must be equal or larger than originalSize, otherwise decompression will fail.
+    return : the number of bytes decompressed into destination buffer (originalSize)
+             or an errorCode if it fails (which can be tested using ZSTDv01_isError())
+*/
+size_t ZSTDv01_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize);
+
+/**
+ZSTDv01_getFrameSrcSize() : get the source length of a ZSTD frame compliant with v0.1.x format
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv01_isError())
+*/
+size_t ZSTDv01_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/**
+ZSTDv01_isError() : tells if the result of ZSTDv01_decompress() is an error
+*/
+unsigned ZSTDv01_isError(size_t code);
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTDv01_Dctx_s ZSTDv01_Dctx;
+ZSTDv01_Dctx* ZSTDv01_createDCtx(void);
+size_t ZSTDv01_freeDCtx(ZSTDv01_Dctx* dctx);
+
+size_t ZSTDv01_decompressDCtx(void* ctx,
+                              void* dst, size_t maxOriginalSize,
+                        const void* src, size_t compressedSize);
+
+/* *************************************
+*  Streaming functions
+***************************************/
+size_t ZSTDv01_resetDCtx(ZSTDv01_Dctx* dctx);
+
+size_t ZSTDv01_nextSrcSizeToDecompress(ZSTDv01_Dctx* dctx);
+size_t ZSTDv01_decompressContinue(ZSTDv01_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
+/**
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTDv01_magicNumber   0xFD2FB51E   /* Big Endian version */
+#define ZSTDv01_magicNumberLE 0x1EB52FFD   /* Little Endian version */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_V01_H_28739879432 */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v02.c b/vendor/github.com/DataDog/zstd/zstd_v02.c
new file mode 100644
index 0000000..8bc0ece
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v02.c
@@ -0,0 +1,3483 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include "zstd_v02.h"
+#include "error_private.h"
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+
+
+/* ******************************************************************
+   mem.h
+   low-level memory access routines
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/******************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include <string.h>    /* memcpy */
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/****************************************************************
+*  Basic Types
+*****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/****************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets generating assembly depending on alignment.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif // MEM_FORCE_MEMORY_ACCESS
+
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian())
+    {
+        MEM_write16(memPtr, val);
+    }
+    else
+    {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U64)((U64)p[0] + ((U64)p[1]<<8) + ((U64)p[2]<<16) + ((U64)p[3]<<24)
+                     + ((U64)p[4]<<32) + ((U64)p[5]<<40) + ((U64)p[6]<<48) + ((U64)p[7]<<56));
+    }
+}
+
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+
+
+/* ******************************************************************
+   bitstream
+   Part of NewGen Entropy library
+   header file (to include)
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which highly benefit from being inlined.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+
+/**********************************************
+*  bitStream decompression API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BIT_DStream_t;
+
+typedef enum { BIT_DStream_unfinished = 0,
+               BIT_DStream_endOfBuffer = 1,
+               BIT_DStream_completed = 2,
+               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
+
+
+/******************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/****************************************************************
+*  Helper functions
+****************************************************************/
+MEM_STATIC unsigned BIT_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+
+/**********************************************************
+* bitStream decoding
+**********************************************************/
+
+/*!BIT_initDStream
+*  Initialize a BIT_DStream_t.
+*  @bitD : a pointer to an already allocated BIT_DStream_t structure
+*  @srcBuffer must point at the beginning of a bitStream
+*  @srcSize must be the exact size of the bitStream
+*  @result : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(size_t))   /* normal case */
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(size_t);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+    }
+    else
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[6]) << (sizeof(size_t)*8 - 16);
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[5]) << (sizeof(size_t)*8 - 24);
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[4]) << (sizeof(size_t)*8 - 32);
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[3]) << 24;
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[2]) << 16;
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[1]) <<  8;
+            default:;
+        }
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+        bitD->bitsConsumed += (U32)(sizeof(size_t) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BIT_lookBitsFast :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBits(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*!BIT_readBitsFast :
+*  unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBitsFast(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return BIT_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer))
+    {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BIT_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start)
+    {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
+        return BIT_DStream_completed;
+    }
+    {
+        U32 nbBytes = bitD->bitsConsumed >> 3;
+        BIT_DStream_status result = BIT_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start)
+        {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BIT_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BIT_endOfDStream
+*   @return Tells if DStream has reached its exact end
+*/
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
+/* ******************************************************************
+   Error codes and messages
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef ERROR_H_MODULE
+#define ERROR_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define ERR_STATIC static inline
+#elif defined(_MSC_VER)
+#  define ERR_STATIC static __inline
+#elif defined(__GNUC__)
+#  define ERR_STATIC static __attribute__((unused))
+#else
+#  define ERR_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/******************************************
+*  Error Management
+******************************************/
+#define PREFIX(name) ZSTD_error_##name
+
+#define ERROR(name) (size_t)-PREFIX(name)
+
+#define ERROR_LIST(ITEM) \
+        ITEM(PREFIX(No_Error)) ITEM(PREFIX(GENERIC)) \
+        ITEM(PREFIX(dstSize_tooSmall)) ITEM(PREFIX(srcSize_wrong)) \
+        ITEM(PREFIX(prefix_unknown)) ITEM(PREFIX(corruption_detected)) \
+        ITEM(PREFIX(tableLog_tooLarge)) ITEM(PREFIX(maxSymbolValue_tooLarge)) ITEM(PREFIX(maxSymbolValue_tooSmall)) \
+        ITEM(PREFIX(maxCode))
+
+#define ERROR_GENERATE_ENUM(ENUM) ENUM,
+typedef enum { ERROR_LIST(ERROR_GENERATE_ENUM) } ERR_codes;  /* enum is exposed, to detect & handle specific errors; compare function result to -enum value */
+
+#define ERROR_CONVERTTOSTRING(STRING) #STRING,
+#define ERROR_GENERATE_STRING(EXPR) ERROR_CONVERTTOSTRING(EXPR)
+static const char* ERR_strings[] = { ERROR_LIST(ERROR_GENERATE_STRING) };
+
+ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }
+
+ERR_STATIC const char* ERR_getErrorName(size_t code)
+{
+    static const char* codeError = "Unspecified error code";
+    if (ERR_isError(code)) return ERR_strings[-(int)(code)];
+    return codeError;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ERROR_H_MODULE */
+/*
+Constructor and Destructor of type FSE_CTable
+    Note that its size depends on 'tableLog' and 'maxSymbolValue' */
+typedef unsigned FSE_CTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+
+
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/******************************************
+*  Static allocation
+******************************************/
+/* FSE buffer bounds */
+#define FSE_NCOUNTBOUND 512
+#define FSE_BLOCKBOUND(size) (size + (size>>7))
+#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* You can statically allocate FSE CTable/DTable as a table of unsigned using below macro */
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/******************************************
+*  FSE advanced API
+******************************************/
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
+/* build a fake FSE_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
+/* build a fake FSE_DTable, designed to always generate the same symbolValue */
+
+
+/******************************************
+*  FSE symbol decompression API
+******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+
+static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
+
+static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
+
+
+/******************************************
+*  FSE unsafe API
+******************************************/
+static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/******************************************
+*  Implementation of inline functions
+******************************************/
+
+/* decompression */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSE_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSE_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));
+    DStatePtr->state = BIT_readBits(bitD, DTableH.tableLog);
+    BIT_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32  nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32 nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/******************************************
+*  Static allocation macros
+******************************************/
+/* Huff0 buffer bounds */
+#define HUF_CTABLEBOUND 129
+#define HUF_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true if incompressible pre-filtered with fast heuristic */
+#define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* static allocation of Huff0's DTable */
+#define HUF_DTABLE_SIZE(maxTableLog)   (1 + (1<<maxTableLog))  /* nb Cells; use unsigned short for X2, unsigned int for X4 */
+#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        unsigned short DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX6(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog) * 3 / 2] = { maxTableLog }
+
+
+/******************************************
+*  Advanced functions
+******************************************/
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbols decoder */
+static size_t HUF_decompress4X6 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* quad-symbols decoder */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+/*
+    zstd - standard compression library
+    Header File
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Version
+***************************************/
+#define ZSTD_VERSION_MAJOR    0    /* for breaking interface changes  */
+#define ZSTD_VERSION_MINOR    2    /* for new (non-breaking) interface capabilities */
+#define ZSTD_VERSION_RELEASE  2    /* for tweaks, bug-fixes, or development */
+#define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTD_CCtx_s ZSTD_CCtx;   /* incomplete type */
+
+#if defined (__cplusplus)
+}
+#endif
+/*
+    zstd - standard compression library
+    Header File for static linking only
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* The objects defined into this file should be considered experimental.
+ * They are not labelled stable, as their prototype may change in the future.
+ * You can use them for tests, provide feedback, or if you can endure risk of future changes.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Streaming functions
+***************************************/
+
+typedef struct ZSTD_DCtx_s ZSTD_DCtx;
+
+/*
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTD_magicNumber 0xFD2FB522   /* v0.2 (current)*/
+
+
+#if defined (__cplusplus)
+}
+#endif
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/****************************************************************
+*  Tuning parameters
+****************************************************************/
+/* MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSE_MAX_MEMORY_USAGE 14
+#define FSE_DEFAULT_MEMORY_USAGE 13
+
+/* FSE_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSE_MAX_SYMBOL_VALUE 255
+
+
+/****************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+
+
+/****************************************************************
+*  Byte symbol type
+****************************************************************/
+#endif   /* !FSE_COMMONDEFS_ONLY */
+
+
+/****************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/****************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+/****************************************************************
+*  Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+#error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+
+/****************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/****************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+
+
+/****************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+
+/* Function templates */
+
+#define FSE_DECODE_TYPE FSE_decode_t
+
+static U32 FSE_tableStep(U32 tableSize) { return (tableSize>>1) + (tableSize>>3) + 3; }
+
+static size_t FSE_buildDTable
+(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* ptr = dt+1;
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*)ptr;
+    FSE_DTableHeader DTableH;
+    const U32 tableSize = 1 << tableLog;
+    const U32 tableMask = tableSize-1;
+    const U32 step = FSE_tableStep(tableSize);
+    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U32 position = 0;
+    U32 highThreshold = tableSize-1;
+    const S16 largeLimit= (S16)(1 << (tableLog-1));
+    U32 noLarge = 1;
+    U32 s;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    DTableH.tableLog = (U16)tableLog;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        if (normalizedCounter[s]==-1)
+        {
+            tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+            symbolNext[s] = 1;
+        }
+        else
+        {
+            if (normalizedCounter[s] >= largeLimit) noLarge=0;
+            symbolNext[s] = normalizedCounter[s];
+        }
+    }
+
+    /* Spread symbols */
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        int i;
+        for (i=0; i<normalizedCounter[s]; i++)
+        {
+            tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+            position = (position + step) & tableMask;
+            while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }
+    }
+
+    if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+
+    /* Build Decoding table */
+    {
+        U32 i;
+        for (i=0; i<tableSize; i++)
+        {
+            FSE_FUNCTION_TYPE symbol = (FSE_FUNCTION_TYPE)(tableDecode[i].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[i].nbBits = (BYTE) (tableLog - BIT_highbit32 ((U32)nextState) );
+            tableDecode[i].newState = (U16) ( (nextState << tableDecode[i].nbBits) - tableSize);
+        }
+    }
+
+    DTableH.fastMode = (U16)noLarge;
+    memcpy(dt, &DTableH, sizeof(DTableH));   /* memcpy(), to avoid strict aliasing warnings */
+    return 0;
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+/******************************************
+*  FSE helper functions
+******************************************/
+static unsigned FSE_isError(size_t code) { return ERR_isError(code); }
+
+
+/****************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+static short FSE_abs(short a)
+{
+    return (short)(a<0 ? -a : a);
+}
+
+static size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr))
+    {
+        if (previous0)
+        {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF)
+            {
+                n0+=24;
+                if (ip < iend-5)
+                {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                }
+                else
+                {
+                    bitStream >>= 16;
+                    bitCount+=16;
+                }
+            }
+            while ((bitStream & 3) == 3)
+            {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+            {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {
+            const short max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max)
+            {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            }
+            else
+            {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSE_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold)
+            {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            {
+                if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+                {
+                    ip += bitCount>>3;
+                    bitCount &= 7;
+                }
+                else
+                {
+                    bitCount -= (int)(8 * (iend - 4 - ip));
+                    ip = iend - 4;
+                }
+                bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+            }
+        }
+    }
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+
+
+/*********************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const cell = (FSE_decode_t*)(ptr) + 1;   /* because dt is unsigned */
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const dinfo = (FSE_decode_t*)(ptr) + 1;   /* because dt is unsigned */
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSE_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BIT_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+    size_t errorCode;
+
+    /* Init */
+    errorCode = BIT_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+    if (FSE_isError(errorCode)) return errorCode;
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) && (op<olimit) ; op+=4)
+    {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+    while (1)
+    {
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state1))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state1);
+
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state2))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state2);
+    }
+
+    /* end ? */
+    if (BIT_endOfDStream(&bitD) && FSE_endOfDState(&state1) && FSE_endOfDState(&state2))
+        return op-ostart;
+
+    if (op==omax) return ERROR(dstSize_tooSmall);   /* dst buffer is full, but cSrc unfinished */
+
+    return ERROR(corruption_detected);
+}
+
+
+static size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));
+
+    /* select fast mode (static) */
+    if (DTableH.fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSE_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    size_t errorCode;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSE decoding mode */
+    errorCode = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSE_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    errorCode = FSE_buildDTable (dt, counting, maxSymbolValue, tableLog);
+    if (FSE_isError(errorCode)) return errorCode;
+
+    /* always return, even if it is an error code */
+    return FSE_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);
+}
+
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+Huff0 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/****************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/****************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+/****************************************************************
+*  Error Management
+****************************************************************/
+#define HUF_STATIC_ASSERT(c) { enum { HUF_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/******************************************
+*  Helper functions
+******************************************/
+static unsigned HUF_isError(size_t code) { return ERR_isError(code); }
+
+#define HUF_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_MAX_TABLELOG  12           /* max configured tableLog (for static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_DEFAULT_TABLELOG  HUF_MAX_TABLELOG   /* tableLog by default, when not specified */
+#define HUF_MAX_SYMBOL_VALUE 255
+#if (HUF_MAX_TABLELOG > HUF_ABSOLUTEMAX_TABLELOG)
+#  error "HUF_MAX_TABLELOG is too large !"
+#endif
+
+
+
+/*********************************************************
+*  Huff0 : Huffman block decompression
+*********************************************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUF_DEltX2;   /* single-symbol decoding */
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+/*! HUF_readStats
+    Read compact Huffman tree, saved by HUF_writeCTable
+    @huffWeight : destination buffer
+    @return : size read from `src`
+*/
+static size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                            U32* nbSymbolsPtr, U32* tableLogPtr,
+                            const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    U32 tableLog;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+    U32 n;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  /* special header */
+    {
+        if (iSize >= (242))   /* RLE */
+        {
+            static int l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else   /* Incompressible */
+        {
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            for (n=0; n<oSize; n+=2)
+            {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+            }
+        }
+    }
+    else  /* header compressed with FSE (normal case) */
+    {
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSE_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSE_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUF_ABSOLUTEMAX_TABLELOG + 1) * sizeof(U32));
+    weightTotal = 0;
+    for (n=0; n<oSize; n++)
+    {
+        if (huffWeight[n] >= HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+        rankStats[huffWeight[n]]++;
+        weightTotal += (1 << huffWeight[n]) >> 1;
+    }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    tableLog = BIT_highbit32(weightTotal) + 1;
+    if (tableLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+    {
+        U32 total = 1 << tableLog;
+        U32 rest = total - weightTotal;
+        U32 verif = 1 << BIT_highbit32(rest);
+        U32 lastWeight = BIT_highbit32(rest) + 1;
+        if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+        huffWeight[oSize] = (BYTE)lastWeight;
+        rankStats[lastWeight]++;
+    }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    *tableLogPtr = tableLog;
+    return iSize+1;
+}
+
+
+/**************************/
+/* single-symbol decoding */
+/**************************/
+
+static size_t HUF_readDTableX2 (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize = ip[0];
+    U32 nbSymbols = 0;
+    U32 n;
+    U32 nextRankStart;
+    void* ptr = DTable+1;
+    HUF_DEltX2* const dt = (HUF_DEltX2*)ptr;
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(huffWeight, HUF_MAX_SYMBOL_VALUE + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > DTable[0]) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+    DTable[0] = (U16)tableLog;   /* maybe should separate sizeof DTable, as allocated, from used size of DTable, in case of DTable re-use */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<=tableLog; n++)
+    {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<nbSymbols; n++)
+    {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUF_DEltX2 D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize;
+}
+
+static BYTE HUF_decodeSymbolX2(BIT_DStream_t* Dstream, const HUF_DEltX2* dt, const U32 dtLog)
+{
+        const size_t val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+        const BYTE c = dt[val].byte;
+        BIT_skipBits(Dstream, dt[val].nbBits);
+        return c;
+}
+
+#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUF_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-4))
+    {
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd))
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+
+static size_t HUF_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+
+        const void* ptr = DTable;
+        const HUF_DEltX2* const dt = ((const HUF_DEltX2*)ptr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUF_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUF_decompress4X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/***************************/
+/* double-symbols decoding */
+/***************************/
+
+static void HUF_fillDTableX4Level2(HUF_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUF_DEltX4 DElt;
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    U32 s;
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1)
+    {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)   /* note : sortedSymbols already skipped */
+    {
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }
+}
+
+typedef U32 rankVal_t[HUF_ABSOLUTEMAX_TABLELOG][HUF_ABSOLUTEMAX_TABLELOG + 1];
+
+static void HUF_fillDTableX4(HUF_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)
+    {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits)   /* enough room for a second symbol */
+        {
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUF_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        }
+        else
+        {
+            U32 i;
+            const U32 end = start + length;
+            HUF_DEltX4 DElt;
+
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits   = (BYTE)(nbBits);
+            DElt.length   = 1;
+            for (i = start; i < end; i++)
+                DTable[i] = DElt;
+        }
+        rankVal[weight] += length;
+    }
+}
+
+static size_t HUF_readDTableX4 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUF_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUF_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUF_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    const U32 memLog = DTable[0];
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize = ip[0];
+    void* ptr = DTable;
+    HUF_DEltX4* const dt = ((HUF_DEltX4*)ptr) + 1;
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX4) == sizeof(U32));   /* if compilation fails here, assertion is false */
+    if (memLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(weightList, HUF_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--)
+        {if (!maxW) return ERROR(GENERIC); }  /* necessarily finds a solution before maxW==0 */
+
+    /* Get start index of each weight */
+    {
+        U32 w, nextRankStart = 0;
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {
+        U32 s;
+        for (s=0; s<nbSymbols; s++)
+        {
+            U32 w = weightList[s];
+            U32 r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {
+        const U32 minBits = tableLog+1 - maxW;
+        U32 nextRankVal = 0;
+        U32 w, consumed;
+        const int rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+        U32* rankVal0 = rankVal[0];
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankVal;
+            nextRankVal += rankStats[w] << (w+rescale);
+            rankVal0[w] = current;
+        }
+        for (consumed = minBits; consumed <= memLog - minBits; consumed++)
+        {
+            U32* rankValPtr = rankVal[consumed];
+            for (w = 1; w <= maxW; w++)
+            {
+                rankValPtr[w] = rankVal0[w] >> consumed;
+            }
+        }
+    }
+
+    HUF_fillDTableX4(dt, memLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    return iSize;
+}
+
+
+static U32 HUF_decodeSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BIT_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUF_decodeLastSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BIT_skipBits(DStream, dt[val].nbBits);
+    else
+    {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8))
+        {
+            BIT_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+        }
+    }
+    return 1;
+}
+
+
+#define HUF_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUF_decodeStreamX4(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd, const HUF_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd-7))
+    {
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-2))
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUF_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+
+
+static size_t HUF_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+
+        const void* ptr = DTable;
+        const HUF_DEltX4* const dt = ((const HUF_DEltX4*)ptr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX4(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUF_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUF_decompress4X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/**********************************/
+/* quad-symbol decoding           */
+/**********************************/
+typedef struct { BYTE nbBits; BYTE nbBytes; } HUF_DDescX6;
+typedef union { BYTE byte[4]; U32 sequence; } HUF_DSeqX6;
+
+/* recursive, up to level 3; may benefit from <template>-like strategy to nest each level inline */
+static void HUF_fillDTableX6LevelN(HUF_DDescX6* DDescription, HUF_DSeqX6* DSequence, int sizeLog,
+                           const rankVal_t rankValOrigin, const U32 consumed, const int minWeight, const U32 maxWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize, const U32* rankStart,
+                           const U32 nbBitsBaseline, HUF_DSeqX6 baseSeq, HUF_DDescX6 DDesc)
+{
+    const int scaleLog = nbBitsBaseline - sizeLog;   /* note : targetLog >= (nbBitsBaseline-1), hence scaleLog <= 1 */
+    const int minBits  = nbBitsBaseline - maxWeight;
+    const U32 level = DDesc.nbBytes;
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    U32 symbolStartPos, s;
+
+    /* local rankVal, will be modified */
+    memcpy(rankVal, rankValOrigin[consumed], sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1)
+    {
+        U32 i;
+        const U32 skipSize = rankVal[minWeight];
+        for (i = 0; i < skipSize; i++)
+        {
+            DSequence[i] = baseSeq;
+            DDescription[i] = DDesc;
+        }
+    }
+
+    /* fill DTable */
+    DDesc.nbBytes++;
+    symbolStartPos = rankStart[minWeight];
+    for (s=symbolStartPos; s<sortedListSize; s++)
+    {
+        const BYTE symbol = sortedSymbols[s].symbol;
+        const U32  weight = sortedSymbols[s].weight;   /* >= 1 (sorted) */
+        const int  nbBits = nbBitsBaseline - weight;   /* >= 1 (by construction) */
+        const int  totalBits = consumed+nbBits;
+        const U32  start  = rankVal[weight];
+        const U32  length = 1 << (sizeLog-nbBits);
+        baseSeq.byte[level] = symbol;
+        DDesc.nbBits = (BYTE)totalBits;
+
+        if ((level<3) && (sizeLog-totalBits >= minBits))   /* enough room for another symbol */
+        {
+            int nextMinWeight = totalBits + scaleLog;
+            if (nextMinWeight < 1) nextMinWeight = 1;
+            HUF_fillDTableX6LevelN(DDescription+start, DSequence+start, sizeLog-nbBits,
+                           rankValOrigin, totalBits, nextMinWeight, maxWeight,
+                           sortedSymbols, sortedListSize, rankStart,
+                           nbBitsBaseline, baseSeq, DDesc);   /* recursive (max : level 3) */
+        }
+        else
+        {
+            U32 i;
+            const U32 end = start + length;
+            for (i = start; i < end; i++)
+            {
+                DDescription[i] = DDesc;
+                DSequence[i] = baseSeq;
+            }
+        }
+        rankVal[weight] += length;
+    }
+}
+
+
+/* note : same preparation as X4 */
+static size_t HUF_readDTableX6 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUF_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUF_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUF_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    rankVal_t rankVal;
+    const U32 memLog = DTable[0];
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize = ip[0];
+
+    if (memLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(weightList, HUF_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--)
+        { if (!maxW) return ERROR(GENERIC); }  /* necessarily finds a solution before maxW==0 */
+
+
+    /* Get start index of each weight */
+    {
+        U32 w, nextRankStart = 0;
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {
+        U32 s;
+        for (s=0; s<nbSymbols; s++)
+        {
+            U32 w = weightList[s];
+            U32 r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {
+        const U32 minBits = tableLog+1 - maxW;
+        U32 nextRankVal = 0;
+        U32 w, consumed;
+        const int rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+        U32* rankVal0 = rankVal[0];
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankVal;
+            nextRankVal += rankStats[w] << (w+rescale);
+            rankVal0[w] = current;
+        }
+        for (consumed = minBits; consumed <= memLog - minBits; consumed++)
+        {
+            U32* rankValPtr = rankVal[consumed];
+            for (w = 1; w <= maxW; w++)
+            {
+                rankValPtr[w] = rankVal0[w] >> consumed;
+            }
+        }
+    }
+
+
+    /* fill tables */
+    {
+        void* ptr = DTable+1;
+        HUF_DDescX6* DDescription = (HUF_DDescX6*)(ptr);
+        void* dSeqStart = DTable + 1 + ((size_t)1<<(memLog-1));
+        HUF_DSeqX6* DSequence = (HUF_DSeqX6*)(dSeqStart);
+        HUF_DSeqX6 DSeq;
+        HUF_DDescX6 DDesc;
+        DSeq.sequence = 0;
+        DDesc.nbBits = 0;
+        DDesc.nbBytes = 0;
+        HUF_fillDTableX6LevelN(DDescription, DSequence, memLog,
+                       (const U32 (*)[HUF_ABSOLUTEMAX_TABLELOG + 1])rankVal, 0, 1, maxW,
+                       sortedSymbol, sizeOfSort, rankStart0,
+                       tableLog+1, DSeq, DDesc);
+    }
+
+    return iSize;
+}
+
+
+static U32 HUF_decodeSymbolX6(void* op, BIT_DStream_t* DStream, const HUF_DDescX6* dd, const HUF_DSeqX6* ds, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, ds+val, sizeof(HUF_DSeqX6));
+    BIT_skipBits(DStream, dd[val].nbBits);
+    return dd[val].nbBytes;
+}
+
+static U32 HUF_decodeLastSymbolsX6(void* op, const U32 maxL, BIT_DStream_t* DStream,
+                                  const HUF_DDescX6* dd, const HUF_DSeqX6* ds, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    U32 length = dd[val].nbBytes;
+    if (length <= maxL)
+    {
+        memcpy(op, ds+val, length);
+        BIT_skipBits(DStream, dd[val].nbBits);
+        return length;
+    }
+    memcpy(op, ds+val, maxL);
+    if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8))
+    {
+        BIT_skipBits(DStream, dd[val].nbBits);
+        if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+            DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+    }
+    return maxL;
+}
+
+
+#define HUF_DECODE_SYMBOLX6_0(ptr, DStreamPtr) \
+    ptr += HUF_decodeSymbolX6(ptr, DStreamPtr, dd, ds, dtLog)
+
+#define HUF_DECODE_SYMBOLX6_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        HUF_DECODE_SYMBOLX6_0(ptr, DStreamPtr)
+
+#define HUF_DECODE_SYMBOLX6_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUF_DECODE_SYMBOLX6_0(ptr, DStreamPtr)
+
+static inline size_t HUF_decodeStreamX6(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd, const U32* DTable, const U32 dtLog)
+{
+    const void* ddPtr = DTable+1;
+    const HUF_DDescX6* dd = (const HUF_DDescX6*)(ddPtr);
+    const void* dsPtr = DTable + 1 + ((size_t)1<<(dtLog-1));
+    const HUF_DSeqX6* ds = (const HUF_DSeqX6*)(dsPtr);
+    BYTE* const pStart = p;
+
+    /* up to 16 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-16))
+    {
+        HUF_DECODE_SYMBOLX6_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX6_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX6_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX6_0(p, bitDPtr);
+    }
+
+    /* closer to the end, up to 4 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-4))
+        HUF_DECODE_SYMBOLX6_0(p, bitDPtr);
+
+    while (p <= pEnd-4)
+        HUF_DECODE_SYMBOLX6_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    while (p < pEnd)
+        p += HUF_decodeLastSymbolsX6(p, (U32)(pEnd-p), bitDPtr, dd, ds, dtLog);
+
+    return p-pStart;
+}
+
+
+
+static size_t HUF_decompress4X6_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+
+        const U32 dtLog = DTable[0];
+        const void* ddPtr = DTable+1;
+        const HUF_DDescX6* dd = (const HUF_DDescX6*)(ddPtr);
+        const void* dsPtr = DTable + 1 + ((size_t)1<<(dtLog-1));
+        const HUF_DSeqX6* ds = (const HUF_DSeqX6*)(dsPtr);
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-64 symbols per loop (4-16 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (op3 <= opStart4) && (endSignal==BIT_DStream_unfinished) && (op4<=(oend-16)) ; )
+        {
+            HUF_DECODE_SYMBOLX6_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX6_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX6_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX6_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX6_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX6_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX6_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX6_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX6_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX6_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX6_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX6_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX6_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX6_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX6_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX6_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX6(op1, &bitD1, opStart2, DTable, dtLog);
+        HUF_decodeStreamX6(op2, &bitD2, opStart3, DTable, dtLog);
+        HUF_decodeStreamX6(op3, &bitD3, opStart4, DTable, dtLog);
+        HUF_decodeStreamX6(op4, &bitD4, oend,     DTable, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X6 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX6(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUF_readDTableX6 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUF_decompress4X6_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/**********************************/
+/* Generic decompression selector */
+/**********************************/
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[3] = { HUF_decompress4X2, HUF_decompress4X4, HUF_decompress4X6 };
+    /* estimate decompression time */
+    U32 Q;
+    const U32 D256 = (U32)(dstSize >> 8);
+    U32 Dtime[3];
+    U32 algoNb = 0;
+    int n;
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    /* decoder timing evaluation */
+    Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+    for (n=0; n<3; n++)
+        Dtime[n] = algoTime[Q][n].tableTime + (algoTime[Q][n].decode256Time * D256);
+
+    Dtime[1] += Dtime[1] >> 4; Dtime[2] += Dtime[2] >> 3; /* advantage to algorithms using less memory, for cache eviction */
+
+    if (Dtime[1] < Dtime[0]) algoNb = 1;
+    if (Dtime[2] < Dtime[algoNb]) algoNb = 2;
+
+    return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+
+    //return HUF_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUF_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+    //return HUF_decompress4X6(dst, dstSize, cSrc, cSrcSize);   /* multi-streams quad-symbols decoding */
+}
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+*  MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*/
+#define ZSTD_MEMORY_USAGE 17
+
+/*!
+ * HEAPMODE :
+ * Select how default compression functions will allocate memory for their hash table,
+ * in memory stack (0, fastest), or in memory heap (1, requires malloc())
+ * Note that compression context is fairly large, as a consequence heap memory is recommended.
+ */
+#ifndef ZSTD_HEAPMODE
+#  define ZSTD_HEAPMODE 1
+#endif /* ZSTD_HEAPMODE */
+
+/*!
+*  LEGACY_SUPPORT :
+*  decompressor can decode older formats (starting from Zstd 0.1+)
+*/
+#ifndef ZSTD_LEGACY_SUPPORT
+#  define ZSTD_LEGACY_SUPPORT 1
+#endif
+
+
+/* *******************************************************
+*  Includes
+*********************************************************/
+#include <stdlib.h>      /* calloc */
+#include <string.h>      /* memcpy, memmove */
+#include <stdio.h>       /* debug : printf */
+
+
+/* *******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef __AVX2__
+#  include <immintrin.h>   /* AVX2 intrinsics */
+#endif
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+
+/* *******************************************************
+*  Constants
+*********************************************************/
+#define HASH_LOG (ZSTD_MEMORY_USAGE - 2)
+#define HASH_TABLESIZE (1 << HASH_LOG)
+#define HASH_MASK (HASH_TABLESIZE - 1)
+
+#define KNUTH 2654435761
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BLOCKSIZE (128 KB)                 /* define, for static allocation */
+#define MIN_SEQUENCES_SIZE (2 /*seqNb*/ + 2 /*dumps*/ + 3 /*seqTables*/ + 1 /*bitStream*/)
+#define MIN_CBLOCK_SIZE (3 /*litCSize*/ + MIN_SEQUENCES_SIZE)
+#define IS_RAW BIT0
+#define IS_RLE BIT1
+
+#define WORKPLACESIZE (BLOCKSIZE*3)
+#define MINMATCH 4
+#define MLbits   7
+#define LLbits   6
+#define Offbits  5
+#define MaxML  ((1<<MLbits )-1)
+#define MaxLL  ((1<<LLbits )-1)
+#define MaxOff   31
+#define LitFSELog  11
+#define MLFSELog   10
+#define LLFSELog   10
+#define OffFSELog   9
+#define MAX(a,b) ((a)<(b)?(b):(a))
+#define MaxSeq MAX(MaxLL, MaxML)
+
+#define LITERAL_NOENTROPY 63
+#define COMMAND_NOENTROPY 7   /* to remove */
+
+static const size_t ZSTD_blockHeaderSize = 3;
+static const size_t ZSTD_frameHeaderSize = 4;
+
+
+/* *******************************************************
+*  Memory operations
+**********************************************************/
+static void   ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+static void   ZSTD_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+
+#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTD_wildcopy : custom version of memcpy(), can copy up to 7-8 bytes too many */
+static void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do COPY8(op, ip) while (op < oend);
+}
+
+
+/* **************************************
+*  Local structures
+****************************************/
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* offCode;
+    BYTE* litStart;
+    BYTE* lit;
+    BYTE* litLengthStart;
+    BYTE* litLength;
+    BYTE* matchLengthStart;
+    BYTE* matchLength;
+    BYTE* dumpsStart;
+    BYTE* dumps;
+} seqStore_t;
+
+
+/* *************************************
+*  Error Management
+***************************************/
+/*! ZSTD_isError
+*   tells if a return value is an error code */
+static unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }
+
+
+
+/* *************************************************************
+*   Decompression section
+***************************************************************/
+struct ZSTD_DCtx_s
+{
+    U32 LLTable[FSE_DTABLE_SIZE_U32(LLFSELog)];
+    U32 OffTable[FSE_DTABLE_SIZE_U32(OffFSELog)];
+    U32 MLTable[FSE_DTABLE_SIZE_U32(MLFSELog)];
+    void* previousDstEnd;
+    void* base;
+    size_t expected;
+    blockType_t bType;
+    U32 phase;
+    const BYTE* litPtr;
+    size_t litSize;
+    BYTE litBuffer[BLOCKSIZE + 8 /* margin for wildcopy */];
+};   /* typedef'd to ZSTD_Dctx within "zstd_static.h" */
+
+
+static size_t ZSTD_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    BYTE headerFlags;
+    U32 cSize;
+
+    if (srcSize < 3) return ERROR(srcSize_wrong);
+
+    headerFlags = *in;
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+
+    bpPtr->blockType = (blockType_t)(headerFlags >> 6);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+static size_t ZSTD_copyUncompressedBlock(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    if (srcSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/** ZSTD_decompressLiterals
+    @return : nb of bytes read from src, or an error code*/
+static size_t ZSTD_decompressLiterals(void* dst, size_t* maxDstSizePtr,
+                                const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+
+    const size_t litSize = (MEM_readLE32(src) & 0x1FFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+    const size_t litCSize = (MEM_readLE32(ip+2) & 0xFFFFFF) >> 5;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+
+    if (litSize > *maxDstSizePtr) return ERROR(corruption_detected);
+    if (litCSize + 5 > srcSize) return ERROR(corruption_detected);
+
+    if (HUF_isError(HUF_decompress(dst, litSize, ip+5, litCSize))) return ERROR(corruption_detected);
+
+    *maxDstSizePtr = litSize;
+    return litCSize + 5;
+}
+
+
+/** ZSTD_decodeLiteralsBlock
+    @return : nb of bytes read from src (< srcSize )*/
+static size_t ZSTD_decodeLiteralsBlock(void* ctx,
+                          const void* src, size_t srcSize)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)ctx;
+    const BYTE* const istart = (const BYTE* const)src;
+
+    /* any compressed block with literals segment must be at least this size */
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch(*istart & 3)
+    {
+    default:
+    case 0:
+        {
+            size_t litSize = BLOCKSIZE;
+            const size_t readSize = ZSTD_decompressLiterals(dctx->litBuffer, &litSize, src, srcSize);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, 8);
+            return readSize;   /* works if it's an error too */
+        }
+    case IS_RAW:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > srcSize-11)   /* risk of reading too far with wildcopy */
+            {
+                if (litSize > srcSize-3) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, 8);
+                return litSize+3;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+3;
+            dctx->litSize = litSize;
+            return litSize+3;
+        }
+    case IS_RLE:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > BLOCKSIZE) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[3], litSize + 8);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return 4;
+        }
+    }
+}
+
+
+static size_t ZSTD_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t* dumpsLengthPtr,
+                         FSE_DTable* DTableLL, FSE_DTable* DTableML, FSE_DTable* DTableOffb,
+                         const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    const BYTE* const iend = istart + srcSize;
+    U32 LLtype, Offtype, MLtype;
+    U32 LLlog, Offlog, MLlog;
+    size_t dumpsLength;
+
+    /* check */
+    if (srcSize < 5) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    *nbSeq = MEM_readLE16(ip); ip+=2;
+    LLtype  = *ip >> 6;
+    Offtype = (*ip >> 4) & 3;
+    MLtype  = (*ip >> 2) & 3;
+    if (*ip & 2)
+    {
+        dumpsLength  = ip[2];
+        dumpsLength += ip[1] << 8;
+        ip += 3;
+    }
+    else
+    {
+        dumpsLength  = ip[1];
+        dumpsLength += (ip[0] & 1) << 8;
+        ip += 2;
+    }
+    *dumpsPtr = ip;
+    ip += dumpsLength;
+    *dumpsLengthPtr = dumpsLength;
+
+    /* check */
+    if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+    /* sequences */
+    {
+        S16 norm[MaxML+1];    /* assumption : MaxML >= MaxLL and MaxOff */
+        size_t headerSize;
+
+        /* Build DTables */
+        switch(LLtype)
+        {
+        case bt_rle :
+            LLlog = 0;
+            FSE_buildDTable_rle(DTableLL, *ip++); break;
+        case bt_raw :
+            LLlog = LLbits;
+            FSE_buildDTable_raw(DTableLL, LLbits); break;
+        default :
+            {   U32 max = MaxLL;
+                headerSize = FSE_readNCount(norm, &max, &LLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (LLlog > LLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableLL, norm, max, LLlog);
+        }   }
+
+        switch(Offtype)
+        {
+        case bt_rle :
+            Offlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong);   /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableOffb, *ip++ & MaxOff); /* if *ip > MaxOff, data is corrupted */
+            break;
+        case bt_raw :
+            Offlog = Offbits;
+            FSE_buildDTable_raw(DTableOffb, Offbits); break;
+        default :
+            {   U32 max = MaxOff;
+                headerSize = FSE_readNCount(norm, &max, &Offlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (Offlog > OffFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableOffb, norm, max, Offlog);
+        }   }
+
+        switch(MLtype)
+        {
+        case bt_rle :
+            MLlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableML, *ip++); break;
+        case bt_raw :
+            MLlog = MLbits;
+            FSE_buildDTable_raw(DTableML, MLbits); break;
+        default :
+            {   U32 max = MaxML;
+                headerSize = FSE_readNCount(norm, &max, &MLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (MLlog > MLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableML, norm, max, MLlog);
+    }   }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t offset;
+    size_t matchLength;
+} seq_t;
+
+typedef struct {
+    BIT_DStream_t DStream;
+    FSE_DState_t stateLL;
+    FSE_DState_t stateOffb;
+    FSE_DState_t stateML;
+    size_t prevOffset;
+    const BYTE* dumps;
+    const BYTE* dumpsEnd;
+} seqState_t;
+
+
+static void ZSTD_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    size_t litLength;
+    size_t prevOffset;
+    size_t offset;
+    size_t matchLength;
+    const BYTE* dumps = seqState->dumps;
+    const BYTE* const de = seqState->dumpsEnd;
+
+    /* Literal length */
+    litLength = FSE_decodeSymbol(&(seqState->stateLL), &(seqState->DStream));
+    prevOffset = litLength ? seq->offset : seqState->prevOffset;
+    seqState->prevOffset = seq->offset;
+    if (litLength == MaxLL)
+    {
+        U32 add = *dumps++;
+        if (add < 255) litLength += add;
+        else
+        {
+            litLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+            dumps += 3;
+        }
+        if (dumps >= de) dumps = de-1;   /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+
+    /* Offset */
+    {
+        static const size_t offsetPrefix[MaxOff+1] = {  /* note : size_t faster than U32 */
+                1 /*fake*/, 1, 2, 4, 8, 16, 32, 64, 128, 256,
+                512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144,
+                524288, 1048576, 2097152, 4194304, 8388608, 16777216, 33554432, /*fake*/ 1, 1, 1, 1, 1 };
+        U32 offsetCode, nbBits;
+        offsetCode = FSE_decodeSymbol(&(seqState->stateOffb), &(seqState->DStream));   /* <= maxOff, by table construction */
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        nbBits = offsetCode - 1;
+        if (offsetCode==0) nbBits = 0;   /* cmove */
+        offset = offsetPrefix[offsetCode] + BIT_readBits(&(seqState->DStream), nbBits);
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        if (offsetCode==0) offset = prevOffset;   /* cmove */
+    }
+
+    /* MatchLength */
+    matchLength = FSE_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
+    if (matchLength == MaxML)
+    {
+        U32 add = *dumps++;
+        if (add < 255) matchLength += add;
+        else
+        {
+            matchLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+            dumps += 3;
+        }
+        if (dumps >= de) dumps = de-1;   /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+    matchLength += MINMATCH;
+
+    /* save result */
+    seq->litLength = litLength;
+    seq->offset = offset;
+    seq->matchLength = matchLength;
+    seqState->dumps = dumps;
+}
+
+
+static size_t ZSTD_execSequence(BYTE* op,
+                                seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                BYTE* const base, BYTE* const oend)
+{
+    static const int dec32table[] = {0, 1, 2, 1, 4, 4, 4, 4};   /* added */
+    static const int dec64table[] = {8, 8, 8, 7, 8, 9,10,11};   /* substracted */
+    const BYTE* const ostart = op;
+    BYTE* const oLitEnd = op + sequence.litLength;
+    BYTE* const oMatchEnd = op + sequence.litLength + sequence.matchLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_8 = oend-8;
+    const BYTE* const litEnd = *litPtr + sequence.litLength;
+
+    /* checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
+
+    /* copy Literals */
+    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = litEnd;   /* update for next sequence */
+
+    /* copy Match */
+    {
+        const BYTE* match = op - sequence.offset;
+
+        /* check */
+        if (sequence.offset > (size_t)op) return ERROR(corruption_detected);   /* address space overflow test (this test seems kept by clang optimizer) */
+        //if (match > op) return ERROR(corruption_detected);   /* address space overflow test (is clang optimizer removing this test ?) */
+        if (match < base) return ERROR(corruption_detected);
+
+        /* close range match, overlap */
+        if (sequence.offset < 8)
+        {
+            const int dec64 = dec64table[sequence.offset];
+            op[0] = match[0];
+            op[1] = match[1];
+            op[2] = match[2];
+            op[3] = match[3];
+            match += dec32table[sequence.offset];
+            ZSTD_copy4(op+4, match);
+            match -= dec64;
+        }
+        else
+        {
+            ZSTD_copy8(op, match);
+        }
+        op += 8; match += 8;
+
+        if (oMatchEnd > oend-(16-MINMATCH))
+        {
+            if (op < oend_8)
+            {
+                ZSTD_wildcopy(op, match, oend_8 - op);
+                match += oend_8 - op;
+                op = oend_8;
+            }
+            while (op < oMatchEnd) *op++ = *match++;
+        }
+        else
+        {
+            ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+        }
+    }
+
+    return oMatchEnd - ostart;
+}
+
+static size_t ZSTD_decompressSequences(
+                               void* ctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)ctx;
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t errorCode, dumpsLength;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    int nbSeq;
+    const BYTE* dumps;
+    U32* DTableLL = dctx->LLTable;
+    U32* DTableML = dctx->MLTable;
+    U32* DTableOffb = dctx->OffTable;
+    BYTE* const base = (BYTE*) (dctx->base);
+
+    /* Build Decoding Tables */
+    errorCode = ZSTD_decodeSeqHeaders(&nbSeq, &dumps, &dumpsLength,
+                                      DTableLL, DTableML, DTableOffb,
+                                      ip, iend-ip);
+    if (ZSTD_isError(errorCode)) return errorCode;
+    ip += errorCode;
+
+    /* Regen sequences */
+    {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        seqState.dumps = dumps;
+        seqState.dumpsEnd = dumps + dumpsLength;
+        seqState.prevOffset = 1;
+        errorCode = BIT_initDStream(&(seqState.DStream), ip, iend-ip);
+        if (ERR_isError(errorCode)) return ERROR(corruption_detected);
+        FSE_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSE_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSE_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (nbSeq>0) ; )
+        {
+            size_t oneSeqSize;
+            nbSeq--;
+            ZSTD_decodeSequence(&sequence, &seqState);
+            oneSeqSize = ZSTD_execSequence(op, sequence, &litPtr, litEnd, base, oend);
+            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* check if reached exact end */
+        if ( !BIT_endOfDStream(&(seqState.DStream)) ) return ERROR(corruption_detected);   /* requested too much : data is corrupted */
+        if (nbSeq<0) return ERROR(corruption_detected);   /* requested too many sequences : data is corrupted */
+
+        /* last literal segment */
+        {
+            size_t lastLLSize = litEnd - litPtr;
+            if (litPtr > litEnd) return ERROR(corruption_detected);
+            if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+            if (op != litPtr) memmove(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+    }
+
+    return op-ostart;
+}
+
+
+static size_t ZSTD_decompressBlock(
+                            void* ctx,
+                            void* dst, size_t maxDstSize,
+                      const void* src, size_t srcSize)
+{
+    /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+
+    /* Decode literals sub-block */
+    size_t litCSize = ZSTD_decodeLiteralsBlock(ctx, src, srcSize);
+    if (ZSTD_isError(litCSize)) return litCSize;
+    ip += litCSize;
+    srcSize -= litCSize;
+
+    return ZSTD_decompressSequences(ctx, dst, maxDstSize, ip, srcSize);
+}
+
+
+static size_t ZSTD_decompressDCtx(void* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t decodedSize=0;
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTD_decompressBlock(ctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTD_copyUncompressedBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        if (ZSTD_isError(decodedSize)) return decodedSize;
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+static size_t ZSTD_decompress(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    ZSTD_DCtx ctx;
+    ctx.base = dst;
+    return ZSTD_decompressDCtx(&ctx, dst, maxDstSize, src, srcSize);
+}
+
+static size_t ZSTD_findFrameCompressedSize(const void *src, size_t srcSize)
+{
+
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/*******************************
+*  Streaming Decompression API
+*******************************/
+
+static size_t ZSTD_resetDCtx(ZSTD_DCtx* dctx)
+{
+    dctx->expected = ZSTD_frameHeaderSize;
+    dctx->phase = 0;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    return 0;
+}
+
+static ZSTD_DCtx* ZSTD_createDCtx(void)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)malloc(sizeof(ZSTD_DCtx));
+    if (dctx==NULL) return NULL;
+    ZSTD_resetDCtx(dctx);
+    return dctx;
+}
+
+static size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
+{
+    free(dctx);
+    return 0;
+}
+
+static size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+static size_t ZSTD_decompressContinue(ZSTD_DCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != ctx->expected) return ERROR(srcSize_wrong);
+    if (dst != ctx->previousDstEnd)  /* not contiguous */
+        ctx->base = dst;
+
+    /* Decompress : frame header */
+    if (ctx->phase == 0)
+    {
+        /* Check frame magic header */
+        U32 magicNumber = MEM_readLE32(src);
+        if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        return 0;
+    }
+
+    /* Decompress : block header */
+    if (ctx->phase == 1)
+    {
+        blockProperties_t bp;
+        size_t blockSize = ZSTD_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
+        if (ZSTD_isError(blockSize)) return blockSize;
+        if (bp.blockType == bt_end)
+        {
+            ctx->expected = 0;
+            ctx->phase = 0;
+        }
+        else
+        {
+            ctx->expected = blockSize;
+            ctx->bType = bp.blockType;
+            ctx->phase = 2;
+        }
+
+        return 0;
+    }
+
+    /* Decompress : block content */
+    {
+        size_t rSize;
+        switch(ctx->bType)
+        {
+        case bt_compressed:
+            rSize = ZSTD_decompressBlock(ctx, dst, maxDstSize, src, srcSize);
+            break;
+        case bt_raw :
+            rSize = ZSTD_copyUncompressedBlock(dst, maxDstSize, src, srcSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet handled */
+            break;
+        case bt_end :   /* should never happen (filtered at phase 1) */
+            rSize = 0;
+            break;
+        default:
+            return ERROR(GENERIC);
+        }
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        ctx->previousDstEnd = (void*)( ((char*)dst) + rSize);
+        return rSize;
+    }
+
+}
+
+
+/* wrapper layer */
+
+unsigned ZSTDv02_isError(size_t code)
+{
+    return ZSTD_isError(code);
+}
+
+size_t ZSTDv02_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize)
+{
+    return ZSTD_decompress(dst, maxOriginalSize, src, compressedSize);
+}
+
+size_t ZSTDv02_findFrameCompressedSize(const void *src, size_t compressedSize)
+{
+    return ZSTD_findFrameCompressedSize(src, compressedSize);
+}
+
+ZSTDv02_Dctx* ZSTDv02_createDCtx(void)
+{
+    return (ZSTDv02_Dctx*)ZSTD_createDCtx();
+}
+
+size_t ZSTDv02_freeDCtx(ZSTDv02_Dctx* dctx)
+{
+    return ZSTD_freeDCtx((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv02_resetDCtx(ZSTDv02_Dctx* dctx)
+{
+    return ZSTD_resetDCtx((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv02_nextSrcSizeToDecompress(ZSTDv02_Dctx* dctx)
+{
+    return ZSTD_nextSrcSizeToDecompress((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv02_decompressContinue(ZSTDv02_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return ZSTD_decompressContinue((ZSTD_DCtx*)dctx, dst, maxDstSize, src, srcSize);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_v02.h b/vendor/github.com/DataDog/zstd/zstd_v02.h
new file mode 100644
index 0000000..0dde7a6
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v02.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_V02_H_4174539423
+#define ZSTD_V02_H_4174539423
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Simple one-step function
+***************************************/
+/**
+ZSTDv02_decompress() : decompress ZSTD frames compliant with v0.2.x format
+    compressedSize : is the exact source size
+    maxOriginalSize : is the size of the 'dst' buffer, which must be already allocated.
+                      It must be equal or larger than originalSize, otherwise decompression will fail.
+    return : the number of bytes decompressed into destination buffer (originalSize)
+             or an errorCode if it fails (which can be tested using ZSTDv01_isError())
+*/
+size_t ZSTDv02_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize);
+
+/**
+ZSTDv02_getFrameSrcSize() : get the source length of a ZSTD frame compliant with v0.2.x format
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv02_isError())
+*/
+size_t ZSTDv02_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/**
+ZSTDv02_isError() : tells if the result of ZSTDv02_decompress() is an error
+*/
+unsigned ZSTDv02_isError(size_t code);
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTDv02_Dctx_s ZSTDv02_Dctx;
+ZSTDv02_Dctx* ZSTDv02_createDCtx(void);
+size_t ZSTDv02_freeDCtx(ZSTDv02_Dctx* dctx);
+
+size_t ZSTDv02_decompressDCtx(void* ctx,
+                              void* dst, size_t maxOriginalSize,
+                        const void* src, size_t compressedSize);
+
+/* *************************************
+*  Streaming functions
+***************************************/
+size_t ZSTDv02_resetDCtx(ZSTDv02_Dctx* dctx);
+
+size_t ZSTDv02_nextSrcSizeToDecompress(ZSTDv02_Dctx* dctx);
+size_t ZSTDv02_decompressContinue(ZSTDv02_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
+/**
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTDv02_magicNumber 0xFD2FB522   /* v0.2 */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_V02_H_4174539423 */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v03.c b/vendor/github.com/DataDog/zstd/zstd_v03.c
new file mode 100644
index 0000000..54445af
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v03.c
@@ -0,0 +1,3124 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include "zstd_v03.h"
+#include "error_private.h"
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+
+
+
+/* ******************************************************************
+   mem.h
+   low-level memory access routines
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/******************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include <string.h>    /* memcpy */
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/****************************************************************
+*  Basic Types
+*****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/****************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets generating assembly depending on alignment.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+
+#endif // MEM_FORCE_MEMORY_ACCESS
+
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian())
+    {
+        MEM_write16(memPtr, val);
+    }
+    else
+    {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U64)((U64)p[0] + ((U64)p[1]<<8) + ((U64)p[2]<<16) + ((U64)p[3]<<24)
+                     + ((U64)p[4]<<32) + ((U64)p[5]<<40) + ((U64)p[6]<<48) + ((U64)p[7]<<56));
+    }
+}
+
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+
+
+/* ******************************************************************
+   bitstream
+   Part of NewGen Entropy library
+   header file (to include)
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which highly benefit from being inlined.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+
+/**********************************************
+*  bitStream decompression API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BIT_DStream_t;
+
+typedef enum { BIT_DStream_unfinished = 0,
+               BIT_DStream_endOfBuffer = 1,
+               BIT_DStream_completed = 2,
+               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
+
+
+
+/******************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/****************************************************************
+*  Helper functions
+****************************************************************/
+MEM_STATIC unsigned BIT_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+
+/**********************************************************
+* bitStream decoding
+**********************************************************/
+
+/*!BIT_initDStream
+*  Initialize a BIT_DStream_t.
+*  @bitD : a pointer to an already allocated BIT_DStream_t structure
+*  @srcBuffer must point at the beginning of a bitStream
+*  @srcSize must be the exact size of the bitStream
+*  @result : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(size_t))   /* normal case */
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(size_t);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+    }
+    else
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[6]) << (sizeof(size_t)*8 - 16);
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[5]) << (sizeof(size_t)*8 - 24);
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[4]) << (sizeof(size_t)*8 - 32);
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[3]) << 24;
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[2]) << 16;
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[1]) <<  8;
+            default:;
+        }
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+        bitD->bitsConsumed += (U32)(sizeof(size_t) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BIT_lookBitsFast :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBits(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*!BIT_readBitsFast :
+*  unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBitsFast(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return BIT_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer))
+    {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BIT_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start)
+    {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
+        return BIT_DStream_completed;
+    }
+    {
+        U32 nbBytes = bitD->bitsConsumed >> 3;
+        BIT_DStream_status result = BIT_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start)
+        {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BIT_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BIT_endOfDStream
+*   @return Tells if DStream has reached its exact end
+*/
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
+/* ******************************************************************
+   Error codes and messages
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef ERROR_H_MODULE
+#define ERROR_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define ERR_STATIC static inline
+#elif defined(_MSC_VER)
+#  define ERR_STATIC static __inline
+#elif defined(__GNUC__)
+#  define ERR_STATIC static __attribute__((unused))
+#else
+#  define ERR_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/******************************************
+*  Error Management
+******************************************/
+#define PREFIX(name) ZSTD_error_##name
+
+#define ERROR(name) (size_t)-PREFIX(name)
+
+#define ERROR_LIST(ITEM) \
+        ITEM(PREFIX(No_Error)) ITEM(PREFIX(GENERIC)) \
+        ITEM(PREFIX(dstSize_tooSmall)) ITEM(PREFIX(srcSize_wrong)) \
+        ITEM(PREFIX(prefix_unknown)) ITEM(PREFIX(corruption_detected)) \
+        ITEM(PREFIX(tableLog_tooLarge)) ITEM(PREFIX(maxSymbolValue_tooLarge)) ITEM(PREFIX(maxSymbolValue_tooSmall)) \
+        ITEM(PREFIX(maxCode))
+
+#define ERROR_GENERATE_ENUM(ENUM) ENUM,
+typedef enum { ERROR_LIST(ERROR_GENERATE_ENUM) } ERR_codes;  /* enum is exposed, to detect & handle specific errors; compare function result to -enum value */
+
+#define ERROR_CONVERTTOSTRING(STRING) #STRING,
+#define ERROR_GENERATE_STRING(EXPR) ERROR_CONVERTTOSTRING(EXPR)
+static const char* ERR_strings[] = { ERROR_LIST(ERROR_GENERATE_STRING) };
+
+ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }
+
+ERR_STATIC const char* ERR_getErrorName(size_t code)
+{
+    static const char* codeError = "Unspecified error code";
+    if (ERR_isError(code)) return ERR_strings[-(int)(code)];
+    return codeError;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ERROR_H_MODULE */
+/*
+Constructor and Destructor of type FSE_CTable
+    Note that its size depends on 'tableLog' and 'maxSymbolValue' */
+typedef unsigned FSE_CTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+
+
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/******************************************
+*  Static allocation
+******************************************/
+/* FSE buffer bounds */
+#define FSE_NCOUNTBOUND 512
+#define FSE_BLOCKBOUND(size) (size + (size>>7))
+#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* You can statically allocate FSE CTable/DTable as a table of unsigned using below macro */
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/******************************************
+*  FSE advanced API
+******************************************/
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
+/* build a fake FSE_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
+/* build a fake FSE_DTable, designed to always generate the same symbolValue */
+
+
+/******************************************
+*  FSE symbol decompression API
+******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+
+static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
+
+static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
+
+
+/******************************************
+*  FSE unsafe API
+******************************************/
+static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/******************************************
+*  Implementation of inline functions
+******************************************/
+
+/* decompression */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSE_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSE_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));
+    DStatePtr->state = BIT_readBits(bitD, DTableH.tableLog);
+    BIT_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32  nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32 nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/******************************************
+*  Static allocation macros
+******************************************/
+/* Huff0 buffer bounds */
+#define HUF_CTABLEBOUND 129
+#define HUF_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true if incompressible pre-filtered with fast heuristic */
+#define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* static allocation of Huff0's DTable */
+#define HUF_DTABLE_SIZE(maxTableLog)   (1 + (1<<maxTableLog))  /* nb Cells; use unsigned short for X2, unsigned int for X4 */
+#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        unsigned short DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX6(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog) * 3 / 2] = { maxTableLog }
+
+
+/******************************************
+*  Advanced functions
+******************************************/
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbols decoder */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+/*
+    zstd - standard compression library
+    Header File
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Version
+***************************************/
+#define ZSTD_VERSION_MAJOR    0    /* for breaking interface changes  */
+#define ZSTD_VERSION_MINOR    2    /* for new (non-breaking) interface capabilities */
+#define ZSTD_VERSION_RELEASE  2    /* for tweaks, bug-fixes, or development */
+#define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTD_CCtx_s ZSTD_CCtx;   /* incomplete type */
+
+#if defined (__cplusplus)
+}
+#endif
+/*
+    zstd - standard compression library
+    Header File for static linking only
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* The objects defined into this file should be considered experimental.
+ * They are not labelled stable, as their prototype may change in the future.
+ * You can use them for tests, provide feedback, or if you can endure risk of future changes.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Streaming functions
+***************************************/
+
+typedef struct ZSTD_DCtx_s ZSTD_DCtx;
+
+/*
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTD_magicNumber 0xFD2FB523   /* v0.3 */
+
+
+#if defined (__cplusplus)
+}
+#endif
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/****************************************************************
+*  Tuning parameters
+****************************************************************/
+/* MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSE_MAX_MEMORY_USAGE 14
+#define FSE_DEFAULT_MEMORY_USAGE 13
+
+/* FSE_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSE_MAX_SYMBOL_VALUE 255
+
+
+/****************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+
+
+/****************************************************************
+*  Byte symbol type
+****************************************************************/
+#endif   /* !FSE_COMMONDEFS_ONLY */
+
+
+/****************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/****************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+/****************************************************************
+*  Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+#error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+
+/****************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/****************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+
+
+/****************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+
+/* Function templates */
+
+#define FSE_DECODE_TYPE FSE_decode_t
+
+static U32 FSE_tableStep(U32 tableSize) { return (tableSize>>1) + (tableSize>>3) + 3; }
+
+static size_t FSE_buildDTable
+(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* ptr = dt+1;
+    FSE_DTableHeader DTableH;
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*)ptr;
+    const U32 tableSize = 1 << tableLog;
+    const U32 tableMask = tableSize-1;
+    const U32 step = FSE_tableStep(tableSize);
+    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U32 position = 0;
+    U32 highThreshold = tableSize-1;
+    const S16 largeLimit= (S16)(1 << (tableLog-1));
+    U32 noLarge = 1;
+    U32 s;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    DTableH.tableLog = (U16)tableLog;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        if (normalizedCounter[s]==-1)
+        {
+            tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+            symbolNext[s] = 1;
+        }
+        else
+        {
+            if (normalizedCounter[s] >= largeLimit) noLarge=0;
+            symbolNext[s] = normalizedCounter[s];
+        }
+    }
+
+    /* Spread symbols */
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        int i;
+        for (i=0; i<normalizedCounter[s]; i++)
+        {
+            tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+            position = (position + step) & tableMask;
+            while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }
+    }
+
+    if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+
+    /* Build Decoding table */
+    {
+        U32 i;
+        for (i=0; i<tableSize; i++)
+        {
+            FSE_FUNCTION_TYPE symbol = (FSE_FUNCTION_TYPE)(tableDecode[i].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[i].nbBits = (BYTE) (tableLog - BIT_highbit32 ((U32)nextState) );
+            tableDecode[i].newState = (U16) ( (nextState << tableDecode[i].nbBits) - tableSize);
+        }
+    }
+
+    DTableH.fastMode = (U16)noLarge;
+    memcpy(dt, &DTableH, sizeof(DTableH));
+    return 0;
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+/******************************************
+*  FSE helper functions
+******************************************/
+static unsigned FSE_isError(size_t code) { return ERR_isError(code); }
+
+
+/****************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+static short FSE_abs(short a)
+{
+    return a<0 ? -a : a;
+}
+
+static size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr))
+    {
+        if (previous0)
+        {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF)
+            {
+                n0+=24;
+                if (ip < iend-5)
+                {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                }
+                else
+                {
+                    bitStream >>= 16;
+                    bitCount+=16;
+                }
+            }
+            while ((bitStream & 3) == 3)
+            {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+            {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {
+            const short max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max)
+            {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            }
+            else
+            {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSE_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold)
+            {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            {
+                if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+                {
+                    ip += bitCount>>3;
+                    bitCount &= 7;
+                }
+                else
+                {
+                    bitCount -= (int)(8 * (iend - 4 - ip));
+                    ip = iend - 4;
+                }
+                bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+            }
+        }
+    }
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+
+
+/*********************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const cell = (FSE_decode_t*)(ptr) + 1;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    FSE_decode_t* const dinfo = (FSE_decode_t*)(ptr) + 1;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSE_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BIT_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+    size_t errorCode;
+
+    /* Init */
+    errorCode = BIT_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+    if (FSE_isError(errorCode)) return errorCode;
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) && (op<olimit) ; op+=4)
+    {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+    while (1)
+    {
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state1))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state1);
+
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state2))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state2);
+    }
+
+    /* end ? */
+    if (BIT_endOfDStream(&bitD) && FSE_endOfDState(&state1) && FSE_endOfDState(&state2))
+        return op-ostart;
+
+    if (op==omax) return ERROR(dstSize_tooSmall);   /* dst buffer is full, but cSrc unfinished */
+
+    return ERROR(corruption_detected);
+}
+
+
+static size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));
+
+    /* select fast mode (static) */
+    if (DTableH.fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSE_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    size_t errorCode;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSE decoding mode */
+    errorCode = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSE_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    errorCode = FSE_buildDTable (dt, counting, maxSymbolValue, tableLog);
+    if (FSE_isError(errorCode)) return errorCode;
+
+    /* always return, even if it is an error code */
+    return FSE_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);
+}
+
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+Huff0 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/****************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+/****************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+/****************************************************************
+*  Error Management
+****************************************************************/
+#define HUF_STATIC_ASSERT(c) { enum { HUF_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/******************************************
+*  Helper functions
+******************************************/
+static unsigned HUF_isError(size_t code) { return ERR_isError(code); }
+
+#define HUF_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_MAX_TABLELOG  12           /* max configured tableLog (for static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_DEFAULT_TABLELOG  HUF_MAX_TABLELOG   /* tableLog by default, when not specified */
+#define HUF_MAX_SYMBOL_VALUE 255
+#if (HUF_MAX_TABLELOG > HUF_ABSOLUTEMAX_TABLELOG)
+#  error "HUF_MAX_TABLELOG is too large !"
+#endif
+
+
+
+/*********************************************************
+*  Huff0 : Huffman block decompression
+*********************************************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUF_DEltX2;   /* single-symbol decoding */
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+/*! HUF_readStats
+    Read compact Huffman tree, saved by HUF_writeCTable
+    @huffWeight : destination buffer
+    @return : size read from `src`
+*/
+static size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                            U32* nbSymbolsPtr, U32* tableLogPtr,
+                            const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    U32 tableLog;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+    U32 n;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  /* special header */
+    {
+        if (iSize >= (242))   /* RLE */
+        {
+            static int l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else   /* Incompressible */
+        {
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            for (n=0; n<oSize; n+=2)
+            {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+            }
+        }
+    }
+    else  /* header compressed with FSE (normal case) */
+    {
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSE_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSE_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUF_ABSOLUTEMAX_TABLELOG + 1) * sizeof(U32));
+    weightTotal = 0;
+    for (n=0; n<oSize; n++)
+    {
+        if (huffWeight[n] >= HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+        rankStats[huffWeight[n]]++;
+        weightTotal += (1 << huffWeight[n]) >> 1;
+    }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    tableLog = BIT_highbit32(weightTotal) + 1;
+    if (tableLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+    {
+        U32 total = 1 << tableLog;
+        U32 rest = total - weightTotal;
+        U32 verif = 1 << BIT_highbit32(rest);
+        U32 lastWeight = BIT_highbit32(rest) + 1;
+        if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+        huffWeight[oSize] = (BYTE)lastWeight;
+        rankStats[lastWeight]++;
+    }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    *tableLogPtr = tableLog;
+    return iSize+1;
+}
+
+
+/**************************/
+/* single-symbol decoding */
+/**************************/
+
+static size_t HUF_readDTableX2 (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize = ip[0];
+    U32 nbSymbols = 0;
+    U32 n;
+    U32 nextRankStart;
+    void* ptr = DTable+1;
+    HUF_DEltX2* const dt = (HUF_DEltX2*)(ptr);
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(huffWeight, HUF_MAX_SYMBOL_VALUE + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > DTable[0]) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+    DTable[0] = (U16)tableLog;   /* maybe should separate sizeof DTable, as allocated, from used size of DTable, in case of DTable re-use */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<=tableLog; n++)
+    {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<nbSymbols; n++)
+    {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUF_DEltX2 D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize;
+}
+
+static BYTE HUF_decodeSymbolX2(BIT_DStream_t* Dstream, const HUF_DEltX2* dt, const U32 dtLog)
+{
+        const size_t val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+        const BYTE c = dt[val].byte;
+        BIT_skipBits(Dstream, dt[val].nbBits);
+        return c;
+}
+
+#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUF_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-4))
+    {
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd))
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+
+static size_t HUF_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+
+        const void* ptr = DTable;
+        const HUF_DEltX2* const dt = ((const HUF_DEltX2*)ptr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUF_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUF_decompress4X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/***************************/
+/* double-symbols decoding */
+/***************************/
+
+static void HUF_fillDTableX4Level2(HUF_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUF_DEltX4 DElt;
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    U32 s;
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1)
+    {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)   /* note : sortedSymbols already skipped */
+    {
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }
+}
+
+typedef U32 rankVal_t[HUF_ABSOLUTEMAX_TABLELOG][HUF_ABSOLUTEMAX_TABLELOG + 1];
+
+static void HUF_fillDTableX4(HUF_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)
+    {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits)   /* enough room for a second symbol */
+        {
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUF_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        }
+        else
+        {
+            U32 i;
+            const U32 end = start + length;
+            HUF_DEltX4 DElt;
+
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits   = (BYTE)(nbBits);
+            DElt.length   = 1;
+            for (i = start; i < end; i++)
+                DTable[i] = DElt;
+        }
+        rankVal[weight] += length;
+    }
+}
+
+static size_t HUF_readDTableX4 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUF_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUF_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUF_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    const U32 memLog = DTable[0];
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize = ip[0];
+    void* ptr = DTable;
+    HUF_DEltX4* const dt = ((HUF_DEltX4*)ptr) + 1;
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX4) == sizeof(U32));   /* if compilation fails here, assertion is false */
+    if (memLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(weightList, HUF_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--)
+        { if (!maxW) return ERROR(GENERIC); }  /* necessarily finds a solution before maxW==0 */
+
+    /* Get start index of each weight */
+    {
+        U32 w, nextRankStart = 0;
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {
+        U32 s;
+        for (s=0; s<nbSymbols; s++)
+        {
+            U32 w = weightList[s];
+            U32 r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {
+        const U32 minBits = tableLog+1 - maxW;
+        U32 nextRankVal = 0;
+        U32 w, consumed;
+        const int rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+        U32* rankVal0 = rankVal[0];
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankVal;
+            nextRankVal += rankStats[w] << (w+rescale);
+            rankVal0[w] = current;
+        }
+        for (consumed = minBits; consumed <= memLog - minBits; consumed++)
+        {
+            U32* rankValPtr = rankVal[consumed];
+            for (w = 1; w <= maxW; w++)
+            {
+                rankValPtr[w] = rankVal0[w] >> consumed;
+            }
+        }
+    }
+
+    HUF_fillDTableX4(dt, memLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    return iSize;
+}
+
+
+static U32 HUF_decodeSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BIT_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUF_decodeLastSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BIT_skipBits(DStream, dt[val].nbBits);
+    else
+    {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8))
+        {
+            BIT_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+        }
+    }
+    return 1;
+}
+
+
+#define HUF_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUF_decodeStreamX4(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd, const HUF_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd-7))
+    {
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-2))
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUF_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+
+
+static size_t HUF_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+
+        const void* ptr = DTable;
+        const HUF_DEltX4* const dt = ((const HUF_DEltX4*)ptr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX4(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUF_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUF_decompress4X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/**********************************/
+/* Generic decompression selector */
+/**********************************/
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[3] = { HUF_decompress4X2, HUF_decompress4X4, NULL };
+    /* estimate decompression time */
+    U32 Q;
+    const U32 D256 = (U32)(dstSize >> 8);
+    U32 Dtime[3];
+    U32 algoNb = 0;
+    int n;
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    /* decoder timing evaluation */
+    Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+    for (n=0; n<3; n++)
+        Dtime[n] = algoTime[Q][n].tableTime + (algoTime[Q][n].decode256Time * D256);
+
+    Dtime[1] += Dtime[1] >> 4; Dtime[2] += Dtime[2] >> 3; /* advantage to algorithms using less memory, for cache eviction */
+
+    if (Dtime[1] < Dtime[0]) algoNb = 1;
+
+    return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+
+    //return HUF_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUF_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+    //return HUF_decompress4X6(dst, dstSize, cSrc, cSrcSize);   /* multi-streams quad-symbols decoding */
+}
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+*  MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*/
+#define ZSTD_MEMORY_USAGE 17
+
+/*!
+ * HEAPMODE :
+ * Select how default compression functions will allocate memory for their hash table,
+ * in memory stack (0, fastest), or in memory heap (1, requires malloc())
+ * Note that compression context is fairly large, as a consequence heap memory is recommended.
+ */
+#ifndef ZSTD_HEAPMODE
+#  define ZSTD_HEAPMODE 1
+#endif /* ZSTD_HEAPMODE */
+
+/*!
+*  LEGACY_SUPPORT :
+*  decompressor can decode older formats (starting from Zstd 0.1+)
+*/
+#ifndef ZSTD_LEGACY_SUPPORT
+#  define ZSTD_LEGACY_SUPPORT 1
+#endif
+
+
+/* *******************************************************
+*  Includes
+*********************************************************/
+#include <stdlib.h>      /* calloc */
+#include <string.h>      /* memcpy, memmove */
+#include <stdio.h>       /* debug : printf */
+
+
+/* *******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef __AVX2__
+#  include <immintrin.h>   /* AVX2 intrinsics */
+#endif
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#else
+#  define GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#endif
+
+
+/* *******************************************************
+*  Constants
+*********************************************************/
+#define HASH_LOG (ZSTD_MEMORY_USAGE - 2)
+#define HASH_TABLESIZE (1 << HASH_LOG)
+#define HASH_MASK (HASH_TABLESIZE - 1)
+
+#define KNUTH 2654435761
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BLOCKSIZE (128 KB)                 /* define, for static allocation */
+#define MIN_SEQUENCES_SIZE (2 /*seqNb*/ + 2 /*dumps*/ + 3 /*seqTables*/ + 1 /*bitStream*/)
+#define MIN_CBLOCK_SIZE (3 /*litCSize*/ + MIN_SEQUENCES_SIZE)
+#define IS_RAW BIT0
+#define IS_RLE BIT1
+
+#define WORKPLACESIZE (BLOCKSIZE*3)
+#define MINMATCH 4
+#define MLbits   7
+#define LLbits   6
+#define Offbits  5
+#define MaxML  ((1<<MLbits )-1)
+#define MaxLL  ((1<<LLbits )-1)
+#define MaxOff   31
+#define LitFSELog  11
+#define MLFSELog   10
+#define LLFSELog   10
+#define OffFSELog   9
+#define MAX(a,b) ((a)<(b)?(b):(a))
+#define MaxSeq MAX(MaxLL, MaxML)
+
+#define LITERAL_NOENTROPY 63
+#define COMMAND_NOENTROPY 7   /* to remove */
+
+static const size_t ZSTD_blockHeaderSize = 3;
+static const size_t ZSTD_frameHeaderSize = 4;
+
+
+/* *******************************************************
+*  Memory operations
+**********************************************************/
+static void   ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+static void   ZSTD_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+
+#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTD_wildcopy : custom version of memcpy(), can copy up to 7-8 bytes too many */
+static void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do COPY8(op, ip) while (op < oend);
+}
+
+
+/* **************************************
+*  Local structures
+****************************************/
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* offCode;
+    BYTE* litStart;
+    BYTE* lit;
+    BYTE* litLengthStart;
+    BYTE* litLength;
+    BYTE* matchLengthStart;
+    BYTE* matchLength;
+    BYTE* dumpsStart;
+    BYTE* dumps;
+} seqStore_t;
+
+
+/* *************************************
+*  Error Management
+***************************************/
+/*! ZSTD_isError
+*   tells if a return value is an error code */
+static unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }
+
+
+
+/* *************************************************************
+*   Decompression section
+***************************************************************/
+struct ZSTD_DCtx_s
+{
+    U32 LLTable[FSE_DTABLE_SIZE_U32(LLFSELog)];
+    U32 OffTable[FSE_DTABLE_SIZE_U32(OffFSELog)];
+    U32 MLTable[FSE_DTABLE_SIZE_U32(MLFSELog)];
+    void* previousDstEnd;
+    void* base;
+    size_t expected;
+    blockType_t bType;
+    U32 phase;
+    const BYTE* litPtr;
+    size_t litSize;
+    BYTE litBuffer[BLOCKSIZE + 8 /* margin for wildcopy */];
+};   /* typedef'd to ZSTD_Dctx within "zstd_static.h" */
+
+
+static size_t ZSTD_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    BYTE headerFlags;
+    U32 cSize;
+
+    if (srcSize < 3) return ERROR(srcSize_wrong);
+
+    headerFlags = *in;
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+
+    bpPtr->blockType = (blockType_t)(headerFlags >> 6);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+static size_t ZSTD_copyUncompressedBlock(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    if (srcSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/** ZSTD_decompressLiterals
+    @return : nb of bytes read from src, or an error code*/
+static size_t ZSTD_decompressLiterals(void* dst, size_t* maxDstSizePtr,
+                                const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+
+    const size_t litSize = (MEM_readLE32(src) & 0x1FFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+    const size_t litCSize = (MEM_readLE32(ip+2) & 0xFFFFFF) >> 5;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+
+    if (litSize > *maxDstSizePtr) return ERROR(corruption_detected);
+    if (litCSize + 5 > srcSize) return ERROR(corruption_detected);
+
+    if (HUF_isError(HUF_decompress(dst, litSize, ip+5, litCSize))) return ERROR(corruption_detected);
+
+    *maxDstSizePtr = litSize;
+    return litCSize + 5;
+}
+
+
+/** ZSTD_decodeLiteralsBlock
+    @return : nb of bytes read from src (< srcSize )*/
+static size_t ZSTD_decodeLiteralsBlock(void* ctx,
+                          const void* src, size_t srcSize)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)ctx;
+    const BYTE* const istart = (const BYTE* const)src;
+
+    /* any compressed block with literals segment must be at least this size */
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch(*istart & 3)
+    {
+    default:
+    case 0:
+        {
+            size_t litSize = BLOCKSIZE;
+            const size_t readSize = ZSTD_decompressLiterals(dctx->litBuffer, &litSize, src, srcSize);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, 8);
+            return readSize;   /* works if it's an error too */
+        }
+    case IS_RAW:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > srcSize-11)   /* risk of reading too far with wildcopy */
+            {
+                if (litSize > srcSize-3) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, 8);
+                return litSize+3;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+3;
+            dctx->litSize = litSize;
+            return litSize+3;
+        }
+    case IS_RLE:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > BLOCKSIZE) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[3], litSize + 8);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return 4;
+        }
+    }
+}
+
+
+static size_t ZSTD_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t* dumpsLengthPtr,
+                         FSE_DTable* DTableLL, FSE_DTable* DTableML, FSE_DTable* DTableOffb,
+                         const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    const BYTE* const iend = istart + srcSize;
+    U32 LLtype, Offtype, MLtype;
+    U32 LLlog, Offlog, MLlog;
+    size_t dumpsLength;
+
+    /* check */
+    if (srcSize < 5) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    *nbSeq = MEM_readLE16(ip); ip+=2;
+    LLtype  = *ip >> 6;
+    Offtype = (*ip >> 4) & 3;
+    MLtype  = (*ip >> 2) & 3;
+    if (*ip & 2)
+    {
+        dumpsLength  = ip[2];
+        dumpsLength += ip[1] << 8;
+        ip += 3;
+    }
+    else
+    {
+        dumpsLength  = ip[1];
+        dumpsLength += (ip[0] & 1) << 8;
+        ip += 2;
+    }
+    *dumpsPtr = ip;
+    ip += dumpsLength;
+    *dumpsLengthPtr = dumpsLength;
+
+    /* check */
+    if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+    /* sequences */
+    {
+        S16 norm[MaxML+1];    /* assumption : MaxML >= MaxLL and MaxOff */
+        size_t headerSize;
+
+        /* Build DTables */
+        switch(LLtype)
+        {
+        case bt_rle :
+            LLlog = 0;
+            FSE_buildDTable_rle(DTableLL, *ip++); break;
+        case bt_raw :
+            LLlog = LLbits;
+            FSE_buildDTable_raw(DTableLL, LLbits); break;
+        default :
+            {   U32 max = MaxLL;
+                headerSize = FSE_readNCount(norm, &max, &LLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (LLlog > LLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableLL, norm, max, LLlog);
+        }   }
+
+        switch(Offtype)
+        {
+        case bt_rle :
+            Offlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong);   /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableOffb, *ip++ & MaxOff); /* if *ip > MaxOff, data is corrupted */
+            break;
+        case bt_raw :
+            Offlog = Offbits;
+            FSE_buildDTable_raw(DTableOffb, Offbits); break;
+        default :
+            {   U32 max = MaxOff;
+                headerSize = FSE_readNCount(norm, &max, &Offlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (Offlog > OffFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableOffb, norm, max, Offlog);
+        }   }
+
+        switch(MLtype)
+        {
+        case bt_rle :
+            MLlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableML, *ip++); break;
+        case bt_raw :
+            MLlog = MLbits;
+            FSE_buildDTable_raw(DTableML, MLbits); break;
+        default :
+            {   U32 max = MaxML;
+                headerSize = FSE_readNCount(norm, &max, &MLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (MLlog > MLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableML, norm, max, MLlog);
+    }   }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t offset;
+    size_t matchLength;
+} seq_t;
+
+typedef struct {
+    BIT_DStream_t DStream;
+    FSE_DState_t stateLL;
+    FSE_DState_t stateOffb;
+    FSE_DState_t stateML;
+    size_t prevOffset;
+    const BYTE* dumps;
+    const BYTE* dumpsEnd;
+} seqState_t;
+
+
+static void ZSTD_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    size_t litLength;
+    size_t prevOffset;
+    size_t offset;
+    size_t matchLength;
+    const BYTE* dumps = seqState->dumps;
+    const BYTE* const de = seqState->dumpsEnd;
+
+    /* Literal length */
+    litLength = FSE_decodeSymbol(&(seqState->stateLL), &(seqState->DStream));
+    prevOffset = litLength ? seq->offset : seqState->prevOffset;
+    seqState->prevOffset = seq->offset;
+    if (litLength == MaxLL)
+    {
+        U32 add = *dumps++;
+        if (add < 255) litLength += add;
+        else
+        {
+            litLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+            dumps += 3;
+        }
+        if (dumps >= de) dumps = de-1;   /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+
+    /* Offset */
+    {
+        static const size_t offsetPrefix[MaxOff+1] = {  /* note : size_t faster than U32 */
+                1 /*fake*/, 1, 2, 4, 8, 16, 32, 64, 128, 256,
+                512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144,
+                524288, 1048576, 2097152, 4194304, 8388608, 16777216, 33554432, /*fake*/ 1, 1, 1, 1, 1 };
+        U32 offsetCode, nbBits;
+        offsetCode = FSE_decodeSymbol(&(seqState->stateOffb), &(seqState->DStream));   /* <= maxOff, by table construction */
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        nbBits = offsetCode - 1;
+        if (offsetCode==0) nbBits = 0;   /* cmove */
+        offset = offsetPrefix[offsetCode] + BIT_readBits(&(seqState->DStream), nbBits);
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        if (offsetCode==0) offset = prevOffset;   /* cmove */
+    }
+
+    /* MatchLength */
+    matchLength = FSE_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
+    if (matchLength == MaxML)
+    {
+        U32 add = *dumps++;
+        if (add < 255) matchLength += add;
+        else
+        {
+            matchLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+            dumps += 3;
+        }
+        if (dumps >= de) dumps = de-1;   /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+    matchLength += MINMATCH;
+
+    /* save result */
+    seq->litLength = litLength;
+    seq->offset = offset;
+    seq->matchLength = matchLength;
+    seqState->dumps = dumps;
+}
+
+
+static size_t ZSTD_execSequence(BYTE* op,
+                                seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                BYTE* const base, BYTE* const oend)
+{
+    static const int dec32table[] = {0, 1, 2, 1, 4, 4, 4, 4};   /* added */
+    static const int dec64table[] = {8, 8, 8, 7, 8, 9,10,11};   /* substracted */
+    const BYTE* const ostart = op;
+    BYTE* const oLitEnd = op + sequence.litLength;
+    BYTE* const oMatchEnd = op + sequence.litLength + sequence.matchLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_8 = oend-8;
+    const BYTE* const litEnd = *litPtr + sequence.litLength;
+
+    /* checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
+
+    /* copy Literals */
+    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = litEnd;   /* update for next sequence */
+
+    /* copy Match */
+    {
+        const BYTE* match = op - sequence.offset;
+
+        /* check */
+        if (sequence.offset > (size_t)op) return ERROR(corruption_detected);   /* address space overflow test (this test seems kept by clang optimizer) */
+        //if (match > op) return ERROR(corruption_detected);   /* address space overflow test (is clang optimizer removing this test ?) */
+        if (match < base) return ERROR(corruption_detected);
+
+        /* close range match, overlap */
+        if (sequence.offset < 8)
+        {
+            const int dec64 = dec64table[sequence.offset];
+            op[0] = match[0];
+            op[1] = match[1];
+            op[2] = match[2];
+            op[3] = match[3];
+            match += dec32table[sequence.offset];
+            ZSTD_copy4(op+4, match);
+            match -= dec64;
+        }
+        else
+        {
+            ZSTD_copy8(op, match);
+        }
+        op += 8; match += 8;
+
+        if (oMatchEnd > oend-(16-MINMATCH))
+        {
+            if (op < oend_8)
+            {
+                ZSTD_wildcopy(op, match, oend_8 - op);
+                match += oend_8 - op;
+                op = oend_8;
+            }
+            while (op < oMatchEnd) *op++ = *match++;
+        }
+        else
+        {
+            ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+        }
+    }
+
+    return oMatchEnd - ostart;
+}
+
+static size_t ZSTD_decompressSequences(
+                               void* ctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)ctx;
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t errorCode, dumpsLength;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    int nbSeq;
+    const BYTE* dumps;
+    U32* DTableLL = dctx->LLTable;
+    U32* DTableML = dctx->MLTable;
+    U32* DTableOffb = dctx->OffTable;
+    BYTE* const base = (BYTE*) (dctx->base);
+
+    /* Build Decoding Tables */
+    errorCode = ZSTD_decodeSeqHeaders(&nbSeq, &dumps, &dumpsLength,
+                                      DTableLL, DTableML, DTableOffb,
+                                      ip, iend-ip);
+    if (ZSTD_isError(errorCode)) return errorCode;
+    ip += errorCode;
+
+    /* Regen sequences */
+    {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        seqState.dumps = dumps;
+        seqState.dumpsEnd = dumps + dumpsLength;
+        seqState.prevOffset = sequence.offset = 4;
+        errorCode = BIT_initDStream(&(seqState.DStream), ip, iend-ip);
+        if (ERR_isError(errorCode)) return ERROR(corruption_detected);
+        FSE_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSE_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSE_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (nbSeq>0) ; )
+        {
+            size_t oneSeqSize;
+            nbSeq--;
+            ZSTD_decodeSequence(&sequence, &seqState);
+            oneSeqSize = ZSTD_execSequence(op, sequence, &litPtr, litEnd, base, oend);
+            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* check if reached exact end */
+        if ( !BIT_endOfDStream(&(seqState.DStream)) ) return ERROR(corruption_detected);   /* requested too much : data is corrupted */
+        if (nbSeq<0) return ERROR(corruption_detected);   /* requested too many sequences : data is corrupted */
+
+        /* last literal segment */
+        {
+            size_t lastLLSize = litEnd - litPtr;
+            if (litPtr > litEnd) return ERROR(corruption_detected);
+            if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+            if (op != litPtr) memmove(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+    }
+
+    return op-ostart;
+}
+
+
+static size_t ZSTD_decompressBlock(
+                            void* ctx,
+                            void* dst, size_t maxDstSize,
+                      const void* src, size_t srcSize)
+{
+    /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+
+    /* Decode literals sub-block */
+    size_t litCSize = ZSTD_decodeLiteralsBlock(ctx, src, srcSize);
+    if (ZSTD_isError(litCSize)) return litCSize;
+    ip += litCSize;
+    srcSize -= litCSize;
+
+    return ZSTD_decompressSequences(ctx, dst, maxDstSize, ip, srcSize);
+}
+
+
+static size_t ZSTD_decompressDCtx(void* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t decodedSize=0;
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTD_decompressBlock(ctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTD_copyUncompressedBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        if (ZSTD_isError(decodedSize)) return decodedSize;
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+static size_t ZSTD_decompress(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    ZSTD_DCtx ctx;
+    ctx.base = dst;
+    return ZSTD_decompressDCtx(&ctx, dst, maxDstSize, src, srcSize);
+}
+
+static size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    U32 magicNumber;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize; remainingSize -= ZSTD_frameHeaderSize;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+
+/*******************************
+*  Streaming Decompression API
+*******************************/
+
+static size_t ZSTD_resetDCtx(ZSTD_DCtx* dctx)
+{
+    dctx->expected = ZSTD_frameHeaderSize;
+    dctx->phase = 0;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    return 0;
+}
+
+static ZSTD_DCtx* ZSTD_createDCtx(void)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)malloc(sizeof(ZSTD_DCtx));
+    if (dctx==NULL) return NULL;
+    ZSTD_resetDCtx(dctx);
+    return dctx;
+}
+
+static size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
+{
+    free(dctx);
+    return 0;
+}
+
+static size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+static size_t ZSTD_decompressContinue(ZSTD_DCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != ctx->expected) return ERROR(srcSize_wrong);
+    if (dst != ctx->previousDstEnd)  /* not contiguous */
+        ctx->base = dst;
+
+    /* Decompress : frame header */
+    if (ctx->phase == 0)
+    {
+        /* Check frame magic header */
+        U32 magicNumber = MEM_readLE32(src);
+        if (magicNumber != ZSTD_magicNumber) return ERROR(prefix_unknown);
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        return 0;
+    }
+
+    /* Decompress : block header */
+    if (ctx->phase == 1)
+    {
+        blockProperties_t bp;
+        size_t blockSize = ZSTD_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
+        if (ZSTD_isError(blockSize)) return blockSize;
+        if (bp.blockType == bt_end)
+        {
+            ctx->expected = 0;
+            ctx->phase = 0;
+        }
+        else
+        {
+            ctx->expected = blockSize;
+            ctx->bType = bp.blockType;
+            ctx->phase = 2;
+        }
+
+        return 0;
+    }
+
+    /* Decompress : block content */
+    {
+        size_t rSize;
+        switch(ctx->bType)
+        {
+        case bt_compressed:
+            rSize = ZSTD_decompressBlock(ctx, dst, maxDstSize, src, srcSize);
+            break;
+        case bt_raw :
+            rSize = ZSTD_copyUncompressedBlock(dst, maxDstSize, src, srcSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet handled */
+            break;
+        case bt_end :   /* should never happen (filtered at phase 1) */
+            rSize = 0;
+            break;
+        default:
+            return ERROR(GENERIC);
+        }
+        ctx->phase = 1;
+        ctx->expected = ZSTD_blockHeaderSize;
+        ctx->previousDstEnd = (void*)( ((char*)dst) + rSize);
+        return rSize;
+    }
+
+}
+
+
+/* wrapper layer */
+
+unsigned ZSTDv03_isError(size_t code)
+{
+    return ZSTD_isError(code);
+}
+
+size_t ZSTDv03_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize)
+{
+    return ZSTD_decompress(dst, maxOriginalSize, src, compressedSize);
+}
+
+size_t ZSTDv03_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    return ZSTD_findFrameCompressedSize(src, srcSize);
+}
+
+ZSTDv03_Dctx* ZSTDv03_createDCtx(void)
+{
+    return (ZSTDv03_Dctx*)ZSTD_createDCtx();
+}
+
+size_t ZSTDv03_freeDCtx(ZSTDv03_Dctx* dctx)
+{
+    return ZSTD_freeDCtx((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv03_resetDCtx(ZSTDv03_Dctx* dctx)
+{
+    return ZSTD_resetDCtx((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv03_nextSrcSizeToDecompress(ZSTDv03_Dctx* dctx)
+{
+    return ZSTD_nextSrcSizeToDecompress((ZSTD_DCtx*)dctx);
+}
+
+size_t ZSTDv03_decompressContinue(ZSTDv03_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return ZSTD_decompressContinue((ZSTD_DCtx*)dctx, dst, maxDstSize, src, srcSize);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_v03.h b/vendor/github.com/DataDog/zstd/zstd_v03.h
new file mode 100644
index 0000000..b4449e2
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v03.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_V03_H_298734209782
+#define ZSTD_V03_H_298734209782
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Simple one-step function
+***************************************/
+/**
+ZSTDv03_decompress() : decompress ZSTD frames compliant with v0.3.x format
+    compressedSize : is the exact source size
+    maxOriginalSize : is the size of the 'dst' buffer, which must be already allocated.
+                      It must be equal or larger than originalSize, otherwise decompression will fail.
+    return : the number of bytes decompressed into destination buffer (originalSize)
+             or an errorCode if it fails (which can be tested using ZSTDv01_isError())
+*/
+size_t ZSTDv03_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize);
+
+/**
+ZSTDv03_getFrameSrcSize() : get the source length of a ZSTD frame compliant with v0.3.x format
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv03_isError())
+*/
+size_t ZSTDv03_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+    /**
+ZSTDv03_isError() : tells if the result of ZSTDv03_decompress() is an error
+*/
+unsigned ZSTDv03_isError(size_t code);
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTDv03_Dctx_s ZSTDv03_Dctx;
+ZSTDv03_Dctx* ZSTDv03_createDCtx(void);
+size_t ZSTDv03_freeDCtx(ZSTDv03_Dctx* dctx);
+
+size_t ZSTDv03_decompressDCtx(void* ctx,
+                              void* dst, size_t maxOriginalSize,
+                        const void* src, size_t compressedSize);
+
+/* *************************************
+*  Streaming functions
+***************************************/
+size_t ZSTDv03_resetDCtx(ZSTDv03_Dctx* dctx);
+
+size_t ZSTDv03_nextSrcSizeToDecompress(ZSTDv03_Dctx* dctx);
+size_t ZSTDv03_decompressContinue(ZSTDv03_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
+/**
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTDv03_magicNumber 0xFD2FB523   /* v0.3 */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_V03_H_298734209782 */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v04.c b/vendor/github.com/DataDog/zstd/zstd_v04.c
new file mode 100644
index 0000000..fb6d1d4
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v04.c
@@ -0,0 +1,3677 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*- Dependencies -*/
+#include "zstd_v04.h"
+#include "error_private.h"
+
+
+/* ******************************************************************
+   mem.h
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/******************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include <string.h>    /* memcpy */
+
+
+/******************************************
+*  Compiler-specific
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/****************************************************************
+*  Basic Types
+*****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/*-*************************************
+*  Debug
+***************************************/
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
+#define ZSTD_STATIC_ASSERT(c) { enum { ZSTD_static_assert = 1/(int)(!!(c)) }; }
+
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
+#  include <stdio.h>
+extern int g_debuglog_enable;
+/* recommended values for ZSTD_DEBUG display levels :
+ * 1 : no display, enables assert() only
+ * 2 : reserved for currently active debug path
+ * 3 : events once per object lifetime (CCtx, CDict, etc.)
+ * 4 : events once per frame
+ * 5 : events once per block
+ * 6 : events once per sequence (*very* verbose) */
+#  define RAWLOG(l, ...) {                                      \
+                if ((g_debuglog_enable) & (l<=ZSTD_DEBUG)) {    \
+                    fprintf(stderr, __VA_ARGS__);               \
+            }   }
+#  define DEBUGLOG(l, ...) {                                    \
+                if ((g_debuglog_enable) & (l<=ZSTD_DEBUG)) {    \
+                    fprintf(stderr, __FILE__ ": " __VA_ARGS__); \
+                    fprintf(stderr, " \n");                     \
+            }   }
+#else
+#  define RAWLOG(l, ...)      {}    /* disabled */
+#  define DEBUGLOG(l, ...)    {}    /* disabled */
+#endif
+
+
+/****************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets generating assembly depending on alignment.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif // MEM_FORCE_MEMORY_ACCESS
+
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian())
+    {
+        MEM_write16(memPtr, val);
+    }
+    else
+    {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+    {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U64)((U64)p[0] + ((U64)p[1]<<8) + ((U64)p[2]<<16) + ((U64)p[3]<<24)
+                     + ((U64)p[4]<<32) + ((U64)p[5]<<40) + ((U64)p[6]<<48) + ((U64)p[7]<<56));
+    }
+}
+
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+
+/*
+    zstd - standard compression library
+    Header File for static linking only
+*/
+#ifndef ZSTD_STATIC_H
+#define ZSTD_STATIC_H
+
+/* The objects defined into this file shall be considered experimental.
+ * They are not considered stable, as their prototype may change in the future.
+ * You can use them for tests, provide feedback, or if you can endure risks of future changes.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Types
+***************************************/
+#define ZSTD_WINDOWLOG_MAX 26
+#define ZSTD_WINDOWLOG_MIN 18
+#define ZSTD_WINDOWLOG_ABSOLUTEMIN 11
+#define ZSTD_CONTENTLOG_MAX (ZSTD_WINDOWLOG_MAX+1)
+#define ZSTD_CONTENTLOG_MIN 4
+#define ZSTD_HASHLOG_MAX 28
+#define ZSTD_HASHLOG_MIN 4
+#define ZSTD_SEARCHLOG_MAX (ZSTD_CONTENTLOG_MAX-1)
+#define ZSTD_SEARCHLOG_MIN 1
+#define ZSTD_SEARCHLENGTH_MAX 7
+#define ZSTD_SEARCHLENGTH_MIN 4
+
+/** from faster to stronger */
+typedef enum { ZSTD_fast, ZSTD_greedy, ZSTD_lazy, ZSTD_lazy2, ZSTD_btlazy2 } ZSTD_strategy;
+
+typedef struct
+{
+    U64 srcSize;       /* optional : tells how much bytes are present in the frame. Use 0 if not known. */
+    U32 windowLog;     /* largest match distance : larger == more compression, more memory needed during decompression */
+    U32 contentLog;    /* full search segment : larger == more compression, slower, more memory (useless for fast) */
+    U32 hashLog;       /* dispatch table : larger == more memory, faster */
+    U32 searchLog;     /* nb of searches : larger == more compression, slower */
+    U32 searchLength;  /* size of matches : larger == faster decompression, sometimes less compression */
+    ZSTD_strategy strategy;
+} ZSTD_parameters;
+
+typedef ZSTDv04_Dctx ZSTD_DCtx;
+
+/* *************************************
+*  Advanced functions
+***************************************/
+/** ZSTD_decompress_usingDict
+*   Same as ZSTD_decompressDCtx, using a Dictionary content as prefix
+*   Note : dict can be NULL, in which case, it's equivalent to ZSTD_decompressDCtx() */
+static size_t ZSTD_decompress_usingDict(ZSTD_DCtx* ctx,
+                                             void* dst, size_t maxDstSize,
+                                       const void* src, size_t srcSize,
+                                       const void* dict,size_t dictSize);
+
+
+/* **************************************
+*  Streaming functions (direct mode)
+****************************************/
+static size_t ZSTD_resetDCtx(ZSTD_DCtx* dctx);
+static size_t ZSTD_getFrameParams(ZSTD_parameters* params, const void* src, size_t srcSize);
+static void   ZSTD_decompress_insertDictionary(ZSTD_DCtx* ctx, const void* src, size_t srcSize);
+
+static size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
+static size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
+
+/**
+  Streaming decompression, bufferless mode
+
+  A ZSTD_DCtx object is required to track streaming operations.
+  Use ZSTD_createDCtx() / ZSTD_freeDCtx() to manage it.
+  A ZSTD_DCtx object can be re-used multiple times. Use ZSTD_resetDCtx() to return to fresh status.
+
+  First operation is to retrieve frame parameters, using ZSTD_getFrameParams().
+  This function doesn't consume its input. It needs enough input data to properly decode the frame header.
+  Objective is to retrieve *params.windowlog, to know minimum amount of memory required during decoding.
+  Result : 0 when successful, it means the ZSTD_parameters structure has been filled.
+           >0 : means there is not enough data into src. Provides the expected size to successfully decode header.
+           errorCode, which can be tested using ZSTD_isError() (For example, if it's not a ZSTD header)
+
+  Then, you can optionally insert a dictionary.
+  This operation must mimic the compressor behavior, otherwise decompression will fail or be corrupted.
+
+  Then it's possible to start decompression.
+  Use ZSTD_nextSrcSizeToDecompress() and ZSTD_decompressContinue() alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() requires this exact amount of bytes, or it will fail.
+  ZSTD_decompressContinue() needs previous data blocks during decompression, up to (1 << windowlog).
+  They should preferably be located contiguously, prior to current block. Alternatively, a round buffer is also possible.
+
+  @result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+
+  A frame is fully decoded when ZSTD_nextSrcSizeToDecompress() returns zero.
+  Context can then be reset to start a new decompression.
+*/
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+
+#endif  /* ZSTD_STATIC_H */
+
+
+/*
+    zstd_internal - common functions to include
+    Header File for include
+*/
+#ifndef ZSTD_CCOMMON_H_MODULE
+#define ZSTD_CCOMMON_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Common macros
+***************************************/
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+
+
+/* *************************************
+*  Common constants
+***************************************/
+#define ZSTD_MAGICNUMBER 0xFD2FB524   /* v0.4 */
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BLOCKSIZE (128 KB)                 /* define, for static allocation */
+
+static const size_t ZSTD_blockHeaderSize = 3;
+static const size_t ZSTD_frameHeaderSize_min = 5;
+#define ZSTD_frameHeaderSize_max 5         /* define, for static allocation */
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define IS_RAW BIT0
+#define IS_RLE BIT1
+
+#define MINMATCH 4
+#define REPCODE_STARTVALUE 4
+
+#define MLbits   7
+#define LLbits   6
+#define Offbits  5
+#define MaxML  ((1<<MLbits) - 1)
+#define MaxLL  ((1<<LLbits) - 1)
+#define MaxOff ((1<<Offbits)- 1)
+#define MLFSELog   10
+#define LLFSELog   10
+#define OffFSELog   9
+#define MaxSeq MAX(MaxLL, MaxML)
+
+#define MIN_SEQUENCES_SIZE (2 /*seqNb*/ + 2 /*dumps*/ + 3 /*seqTables*/ + 1 /*bitStream*/)
+#define MIN_CBLOCK_SIZE (3 /*litCSize*/ + MIN_SEQUENCES_SIZE)
+
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+
+/* ******************************************
+*  Shared functions to include for inlining
+********************************************/
+static void ZSTD_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+
+#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTD_wildcopy : custom version of memcpy(), can copy up to 7-8 bytes too many */
+static void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   header file
+****************************************************************** */
+#ifndef FSE_H
+#define FSE_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* *****************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+
+
+/* *****************************************
+*  FSE simple functions
+******************************************/
+static size_t FSE_decompress(void* dst,  size_t maxDstSize,
+                const void* cSrc, size_t cSrcSize);
+/*!
+FSE_decompress():
+    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'maxDstSize'.
+    return : size of regenerated data (<= maxDstSize)
+             or an error code, which can be tested using FSE_isError()
+
+    ** Important ** : FSE_decompress() doesn't decompress non-compressible nor RLE data !!!
+    Why ? : making this distinction requires a header.
+    Header management is intentionally delegated to the user layer, which can better manage special cases.
+*/
+
+
+/* *****************************************
+*  Tool functions
+******************************************/
+/* Error Management */
+static unsigned    FSE_isError(size_t code);        /* tells if a return value is an error code */
+
+
+
+/* *****************************************
+*  FSE detailed API
+******************************************/
+/*!
+FSE_compress() does the following:
+1. count symbol occurrence from source[] into table count[]
+2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)
+3. save normalized counters to memory buffer using writeNCount()
+4. build encoding table 'CTable' from normalized counters
+5. encode the data stream using encoding table 'CTable'
+
+FSE_decompress() does the following:
+1. read normalized counters with readNCount()
+2. build decoding table 'DTable' from normalized counters
+3. decode the data stream using decoding table 'DTable'
+
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and provide normalized distribution using external method.
+*/
+
+
+/* *** DECOMPRESSION *** */
+
+/*!
+FSE_readNCount():
+   Read compactly saved 'normalizedCounter' from 'rBuffer'.
+   return : size read from 'rBuffer'
+            or an errorCode, which can be tested using FSE_isError()
+            maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+static  size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, const void* rBuffer, size_t rBuffSize);
+
+/*!
+Constructor and Destructor of type FSE_DTable
+    Note that its size depends on 'tableLog' */
+typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+
+/*!
+FSE_buildDTable():
+   Builds 'dt', which must be already allocated, using FSE_createDTable()
+   return : 0,
+            or an errorCode, which can be tested using FSE_isError() */
+static size_t FSE_buildDTable ( FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*!
+FSE_decompress_usingDTable():
+   Decompress compressed source 'cSrc' of size 'cSrcSize' using 'dt'
+   into 'dst' which must be already allocated.
+   return : size of regenerated data (necessarily <= maxDstSize)
+            or an errorCode, which can be tested using FSE_isError() */
+static  size_t FSE_decompress_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
+
+/*!
+Tutorial :
+----------
+(Note : these functions only decompress FSE-compressed blocks.
+ If block is uncompressed, use memcpy() instead
+ If block is a single repeated byte, use memset() instead )
+
+The first step is to obtain the normalized frequencies of symbols.
+This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
+In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
+or size the table to handle worst case situations (typically 256).
+FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
+The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.
+Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.
+This is performed by the function FSE_buildDTable().
+The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+'FSE_DTable' can then be used to decompress 'cSrc', with FSE_decompress_usingDTable().
+'cSrcSize' must be strictly correct, otherwise decompression will fail.
+FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=maxDstSize).
+If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)
+*/
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSE_H */
+
+
+/* ******************************************************************
+   bitstream
+   Part of NewGen Entropy library
+   header file (to include)
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which highly benefit from being inlined.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+/**********************************************
+*  bitStream decompression API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BIT_DStream_t;
+
+typedef enum { BIT_DStream_unfinished = 0,
+               BIT_DStream_endOfBuffer = 1,
+               BIT_DStream_completed = 2,
+               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
+
+
+
+
+/******************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/****************************************************************
+*  Helper functions
+****************************************************************/
+MEM_STATIC unsigned BIT_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+/**********************************************************
+* bitStream decoding
+**********************************************************/
+
+/*!BIT_initDStream
+*  Initialize a BIT_DStream_t.
+*  @bitD : a pointer to an already allocated BIT_DStream_t structure
+*  @srcBuffer must point at the beginning of a bitStream
+*  @srcSize must be the exact size of the bitStream
+*  @result : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(size_t))   /* normal case */
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(size_t);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+    }
+    else
+    {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[6]) << (sizeof(size_t)*8 - 16);/* fall-through */
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[5]) << (sizeof(size_t)*8 - 24);/* fall-through */
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[4]) << (sizeof(size_t)*8 - 32);/* fall-through */
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[3]) << 24; /* fall-through */
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[2]) << 16; /* fall-through */
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[1]) <<  8; /* fall-through */
+            default: break;
+        }
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BIT_highbit32(contain32);
+        bitD->bitsConsumed += (U32)(sizeof(size_t) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BIT_lookBitsFast :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBits(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*!BIT_readBitsFast :
+*  unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BIT_lookBitsFast(bitD, nbBits);
+    BIT_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return BIT_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer))
+    {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BIT_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start)
+    {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
+        return BIT_DStream_completed;
+    }
+    {
+        U32 nbBytes = bitD->bitsConsumed >> 3;
+        BIT_DStream_status result = BIT_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start)
+        {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BIT_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BIT_endOfDStream
+*   @return Tells if DStream has reached its exact end
+*/
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
+
+
+
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef FSE_STATIC_H
+#define FSE_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* *****************************************
+*  Static allocation
+*******************************************/
+/* FSE buffer bounds */
+#define FSE_NCOUNTBOUND 512
+#define FSE_BLOCKBOUND(size) (size + (size>>7))
+#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* It is possible to statically allocate FSE CTable/DTable as a table of unsigned using below macros */
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/* *****************************************
+*  FSE advanced API
+*******************************************/
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
+/* build a fake FSE_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
+/* build a fake FSE_DTable, designed to always generate the same symbolValue */
+
+
+
+/* *****************************************
+*  FSE symbol decompression API
+*******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+
+static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
+
+static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
+
+
+/* *****************************************
+*  FSE unsafe API
+*******************************************/
+static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* *****************************************
+*  Implementation of inlined functions
+*******************************************/
+/* decompression */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSE_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSE_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    memcpy(&DTableH, dt, sizeof(DTableH));
+    DStatePtr->state = BIT_readBits(bitD, DTableH.tableLog);
+    BIT_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32  nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+    const FSE_decode_t DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32 nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BIT_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSE_STATIC_H */
+
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/* **************************************************************
+*  Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSE_MAX_MEMORY_USAGE 14
+#define FSE_DEFAULT_MEMORY_USAGE 13
+
+/*!FSE_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSE_MAX_SYMBOL_VALUE 255
+
+
+/* **************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+#define FSE_DECODE_TYPE FSE_decode_t
+
+
+#endif   /* !FSE_COMMONDEFS_ONLY */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/* **************************************************************
+*  Dependencies
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+
+/* ***************************************************************
+*  Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+#error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_STATIC_ASSERT(c) { enum { FSE_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+
+
+/*-**************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+static U32 FSE_tableStep(U32 tableSize) { return (tableSize>>1) + (tableSize>>3) + 3; }
+
+
+static size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    FSE_DTableHeader DTableH;
+    void* const tdPtr = dt+1;   /* because dt is unsigned, 32-bits aligned on 32-bits */
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
+    const U32 tableSize = 1 << tableLog;
+    const U32 tableMask = tableSize-1;
+    const U32 step = FSE_tableStep(tableSize);
+    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U32 position = 0;
+    U32 highThreshold = tableSize-1;
+    const S16 largeLimit= (S16)(1 << (tableLog-1));
+    U32 noLarge = 1;
+    U32 s;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    DTableH.tableLog = (U16)tableLog;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        if (normalizedCounter[s]==-1)
+        {
+            tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+            symbolNext[s] = 1;
+        }
+        else
+        {
+            if (normalizedCounter[s] >= largeLimit) noLarge=0;
+            symbolNext[s] = normalizedCounter[s];
+        }
+    }
+
+    /* Spread symbols */
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        int i;
+        for (i=0; i<normalizedCounter[s]; i++)
+        {
+            tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+            position = (position + step) & tableMask;
+            while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }
+    }
+
+    if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+
+    /* Build Decoding table */
+    {
+        U32 i;
+        for (i=0; i<tableSize; i++)
+        {
+            FSE_FUNCTION_TYPE symbol = (FSE_FUNCTION_TYPE)(tableDecode[i].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[i].nbBits = (BYTE) (tableLog - BIT_highbit32 ((U32)nextState) );
+            tableDecode[i].newState = (U16) ( (nextState << tableDecode[i].nbBits) - tableSize);
+        }
+    }
+
+    DTableH.fastMode = (U16)noLarge;
+    memcpy(dt, &DTableH, sizeof(DTableH));
+    return 0;
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+/******************************************
+*  FSE helper functions
+******************************************/
+static unsigned FSE_isError(size_t code) { return ERR_isError(code); }
+
+
+/****************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+static short FSE_abs(short a)
+{
+    return a<0 ? -a : a;
+}
+
+static size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr))
+    {
+        if (previous0)
+        {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF)
+            {
+                n0+=24;
+                if (ip < iend-5)
+                {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                }
+                else
+                {
+                    bitStream >>= 16;
+                    bitCount+=16;
+                }
+            }
+            while ((bitStream & 3) == 3)
+            {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+            {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {
+            const short max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max)
+            {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            }
+            else
+            {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSE_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold)
+            {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            {
+                if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4))
+                {
+                    ip += bitCount>>3;
+                    bitCount &= 7;
+                }
+                else
+                {
+                    bitCount -= (int)(8 * (iend - 4 - ip));
+                    ip = iend - 4;
+                }
+                bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+            }
+        }
+    }
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+
+
+/*********************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+static size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+static size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<=maxSymbolValue; s++)
+    {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSE_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BIT_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+    size_t errorCode;
+
+    /* Init */
+    errorCode = BIT_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+    if (FSE_isError(errorCode)) return errorCode;
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) && (op<olimit) ; op+=4)
+    {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+    while (1)
+    {
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state1))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state1);
+
+        if ( (BIT_reloadDStream(&bitD)>BIT_DStream_completed) || (op==omax) || (BIT_endOfDStream(&bitD) && (fast || FSE_endOfDState(&state2))) )
+            break;
+
+        *op++ = FSE_GETSYMBOL(&state2);
+    }
+
+    /* end ? */
+    if (BIT_endOfDStream(&bitD) && FSE_endOfDState(&state1) && FSE_endOfDState(&state2))
+        return op-ostart;
+
+    if (op==omax) return ERROR(dstSize_tooSmall);   /* dst buffer is full, but cSrc unfinished */
+
+    return ERROR(corruption_detected);
+}
+
+
+static size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSE_DTable* dt)
+{
+    FSE_DTableHeader DTableH;
+    U32 fastMode;
+
+    memcpy(&DTableH, dt, sizeof(DTableH));
+    fastMode = DTableH.fastMode;
+
+    /* select fast mode (static) */
+    if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSE_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    size_t errorCode;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSE decoding mode */
+    errorCode = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSE_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    errorCode = FSE_buildDTable (dt, counting, maxSymbolValue, tableLog);
+    if (FSE_isError(errorCode)) return errorCode;
+
+    /* always return, even if it is an error code */
+    return FSE_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);
+}
+
+
+
+#endif   /* FSE_COMMONDEFS_ONLY */
+
+
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   header file
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef HUFF0_H
+#define HUFF0_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************************
+*  Dependency
+******************************************/
+#include <stddef.h>    /* size_t */
+
+
+/* ****************************************
+*  Huff0 simple functions
+******************************************/
+static size_t HUF_decompress(void* dst,  size_t dstSize,
+                const void* cSrc, size_t cSrcSize);
+/*!
+HUF_decompress():
+    Decompress Huff0 data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstSize'.
+    'dstSize' must be the exact size of original (uncompressed) data.
+    Note : in contrast with FSE, HUF_decompress can regenerate RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data, because it knows size to regenerate.
+    @return : size of regenerated data (== dstSize)
+              or an error code, which can be tested using HUF_isError()
+*/
+
+
+/* ****************************************
+*  Tool functions
+******************************************/
+/* Error Management */
+static unsigned    HUF_isError(size_t code);        /* tells if a return value is an error code */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* HUFF0_H */
+
+
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef HUFF0_STATIC_H
+#define HUFF0_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/* ****************************************
+*  Static allocation macros
+******************************************/
+/* static allocation of Huff0's DTable */
+#define HUF_DTABLE_SIZE(maxTableLog)   (1 + (1<<maxTableLog))  /* nb Cells; use unsigned short for X2, unsigned int for X4 */
+#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        unsigned short DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUF_CREATE_STATIC_DTABLEX6(DTable, maxTableLog) \
+        unsigned int DTable[HUF_DTABLE_SIZE(maxTableLog) * 3 / 2] = { maxTableLog }
+
+
+/* ****************************************
+*  Advanced decompression functions
+******************************************/
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbols decoder */
+
+
+/* ****************************************
+*  Huff0 detailed API
+******************************************/
+/*!
+HUF_decompress() does the following:
+1. select the decompression algorithm (X2, X4, X6) based on pre-computed heuristics
+2. build Huffman table from save, using HUF_readDTableXn()
+3. decode 1 or 4 segments in parallel using HUF_decompressSXn_usingDTable
+
+*/
+static size_t HUF_readDTableX2 (unsigned short* DTable, const void* src, size_t srcSize);
+static size_t HUF_readDTableX4 (unsigned* DTable, const void* src, size_t srcSize);
+
+static size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned short* DTable);
+static size_t HUF_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned* DTable);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* HUFF0_STATIC_H */
+
+
+
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+Huff0 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+
+/* **************************************************************
+*  Constants
+****************************************************************/
+#define HUF_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_MAX_TABLELOG  12           /* max configured tableLog (for static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_DEFAULT_TABLELOG  HUF_MAX_TABLELOG   /* tableLog by default, when not specified */
+#define HUF_MAX_SYMBOL_VALUE 255
+#if (HUF_MAX_TABLELOG > HUF_ABSOLUTEMAX_TABLELOG)
+#  error "HUF_MAX_TABLELOG is too large !"
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+static unsigned HUF_isError(size_t code) { return ERR_isError(code); }
+#define HUF_STATIC_ASSERT(c) { enum { HUF_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+
+/*-*******************************************************
+*  Huff0 : Huffman block decompression
+*********************************************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUF_DEltX2;   /* single-symbol decoding */
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+/*! HUF_readStats
+    Read compact Huffman tree, saved by HUF_writeCTable
+    @huffWeight : destination buffer
+    @return : size read from `src`
+*/
+static size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                            U32* nbSymbolsPtr, U32* tableLogPtr,
+                            const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    U32 tableLog;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+    U32 n;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  /* special header */
+    {
+        if (iSize >= (242))   /* RLE */
+        {
+            static int l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else   /* Incompressible */
+        {
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            for (n=0; n<oSize; n+=2)
+            {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+            }
+        }
+    }
+    else  /* header compressed with FSE (normal case) */
+    {
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSE_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSE_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUF_ABSOLUTEMAX_TABLELOG + 1) * sizeof(U32));
+    weightTotal = 0;
+    for (n=0; n<oSize; n++)
+    {
+        if (huffWeight[n] >= HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+        rankStats[huffWeight[n]]++;
+        weightTotal += (1 << huffWeight[n]) >> 1;
+    }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    tableLog = BIT_highbit32(weightTotal) + 1;
+    if (tableLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+    {
+        U32 total = 1 << tableLog;
+        U32 rest = total - weightTotal;
+        U32 verif = 1 << BIT_highbit32(rest);
+        U32 lastWeight = BIT_highbit32(rest) + 1;
+        if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+        huffWeight[oSize] = (BYTE)lastWeight;
+        rankStats[lastWeight]++;
+    }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    *tableLogPtr = tableLog;
+    return iSize+1;
+}
+
+
+/**************************/
+/* single-symbol decoding */
+/**************************/
+
+static size_t HUF_readDTableX2 (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    size_t iSize;
+    U32 nbSymbols = 0;
+    U32 n;
+    U32 nextRankStart;
+    void* const dtPtr = DTable + 1;
+    HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(huffWeight, HUF_MAX_SYMBOL_VALUE + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > DTable[0]) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+    DTable[0] = (U16)tableLog;   /* maybe should separate sizeof DTable, as allocated, from used size of DTable, in case of DTable re-use */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<=tableLog; n++)
+    {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<nbSymbols; n++)
+    {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUF_DEltX2 D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize;
+}
+
+static BYTE HUF_decodeSymbolX2(BIT_DStream_t* Dstream, const HUF_DEltX2* dt, const U32 dtLog)
+{
+        const size_t val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+        const BYTE c = dt[val].byte;
+        BIT_skipBits(Dstream, dt[val].nbBits);
+        return c;
+}
+
+#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUF_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-4))
+    {
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd))
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+
+static size_t HUF_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable;
+        const HUF_DEltX2* const dt = ((const HUF_DEltX2*)dtPtr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUF_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUF_decompress4X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/***************************/
+/* double-symbols decoding */
+/***************************/
+
+static void HUF_fillDTableX4Level2(HUF_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUF_DEltX4 DElt;
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    U32 s;
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1)
+    {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)   /* note : sortedSymbols already skipped */
+    {
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }
+}
+
+typedef U32 rankVal_t[HUF_ABSOLUTEMAX_TABLELOG][HUF_ABSOLUTEMAX_TABLELOG + 1];
+
+static void HUF_fillDTableX4(HUF_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUF_ABSOLUTEMAX_TABLELOG + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++)
+    {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits)   /* enough room for a second symbol */
+        {
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUF_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        }
+        else
+        {
+            U32 i;
+            const U32 end = start + length;
+            HUF_DEltX4 DElt;
+
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits   = (BYTE)(nbBits);
+            DElt.length   = 1;
+            for (i = start; i < end; i++)
+                DTable[i] = DElt;
+        }
+        rankVal[weight] += length;
+    }
+}
+
+static size_t HUF_readDTableX4 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUF_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUF_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUF_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUF_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    const U32 memLog = DTable[0];
+    size_t iSize;
+    void* dtPtr = DTable;
+    HUF_DEltX4* const dt = ((HUF_DEltX4*)dtPtr) + 1;
+
+    HUF_STATIC_ASSERT(sizeof(HUF_DEltX4) == sizeof(U32));   /* if compilation fails here, assertion is false */
+    if (memLog > HUF_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUF_readStats(weightList, HUF_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUF_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--)
+        { if (!maxW) return ERROR(GENERIC); }  /* necessarily finds a solution before maxW==0 */
+
+    /* Get start index of each weight */
+    {
+        U32 w, nextRankStart = 0;
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {
+        U32 s;
+        for (s=0; s<nbSymbols; s++)
+        {
+            U32 w = weightList[s];
+            U32 r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {
+        const U32 minBits = tableLog+1 - maxW;
+        U32 nextRankVal = 0;
+        U32 w, consumed;
+        const int rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+        U32* rankVal0 = rankVal[0];
+        for (w=1; w<=maxW; w++)
+        {
+            U32 current = nextRankVal;
+            nextRankVal += rankStats[w] << (w+rescale);
+            rankVal0[w] = current;
+        }
+        for (consumed = minBits; consumed <= memLog - minBits; consumed++)
+        {
+            U32* rankValPtr = rankVal[consumed];
+            for (w = 1; w <= maxW; w++)
+            {
+                rankValPtr[w] = rankVal0[w] >> consumed;
+            }
+        }
+    }
+
+    HUF_fillDTableX4(dt, memLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    return iSize;
+}
+
+
+static U32 HUF_decodeSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BIT_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUF_decodeLastSymbolX4(void* op, BIT_DStream_t* DStream, const HUF_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BIT_skipBits(DStream, dt[val].nbBits);
+    else
+    {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8))
+        {
+            BIT_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+        }
+    }
+    return 1;
+}
+
+
+#define HUF_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUF_MAX_TABLELOG<=12)) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUF_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUF_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUF_decodeStreamX4(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd, const HUF_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p < pEnd-7))
+    {
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) && (p <= pEnd-2))
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUF_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUF_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+static size_t HUF_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable;
+        const HUF_DEltX4* const dt = ((const HUF_DEltX4*)dtPtr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BIT_DStream_t bitD1;
+        BIT_DStream_t bitD2;
+        BIT_DStream_t bitD3;
+        BIT_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BIT_initDStream(&bitD1, istart1, length1);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD2, istart2, length2);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD3, istart3, length3);
+        if (HUF_isError(errorCode)) return errorCode;
+        errorCode = BIT_initDStream(&bitD4, istart4, length4);
+        if (HUF_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        for ( ; (endSignal==BIT_DStream_unfinished) && (op4<(oend-7)) ; )
+        {
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUF_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUF_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUF_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUF_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BIT_reloadDStream(&bitD1) | BIT_reloadDStream(&bitD2) | BIT_reloadDStream(&bitD3) | BIT_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUF_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUF_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUF_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUF_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUF_CREATE_STATIC_DTABLEX4(DTable, HUF_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUF_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUF_decompress4X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/**********************************/
+/* Generic decompression selector */
+/**********************************/
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[3] = { HUF_decompress4X2, HUF_decompress4X4, NULL };
+    /* estimate decompression time */
+    U32 Q;
+    const U32 D256 = (U32)(dstSize >> 8);
+    U32 Dtime[3];
+    U32 algoNb = 0;
+    int n;
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    /* decoder timing evaluation */
+    Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+    for (n=0; n<3; n++)
+        Dtime[n] = algoTime[Q][n].tableTime + (algoTime[Q][n].decode256Time * D256);
+
+    Dtime[1] += Dtime[1] >> 4; Dtime[2] += Dtime[2] >> 3; /* advantage to algorithms using less memory, for cache eviction */
+
+    if (Dtime[1] < Dtime[0]) algoNb = 1;
+
+    return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+
+    //return HUF_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUF_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+    //return HUF_decompress4X6(dst, dstSize, cSrc, cSrcSize);   /* multi-streams quad-symbols decoding */
+}
+
+
+
+#endif   /* ZSTD_CCOMMON_H_MODULE */
+
+
+/*
+    zstd - decompression module fo v0.4 legacy format
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * HEAPMODE :
+ * Select how default decompression function ZSTD_decompress() will allocate memory,
+ * in memory stack (0), or in memory heap (1, requires malloc())
+ */
+#ifndef ZSTD_HEAPMODE
+#  define ZSTD_HEAPMODE 1
+#endif
+
+
+/* *******************************************************
+*  Includes
+*********************************************************/
+#include <stdlib.h>      /* calloc */
+#include <string.h>      /* memcpy, memmove */
+#include <stdio.h>       /* debug : printf */
+
+
+/* *******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+
+/* *************************************
+*  Local types
+***************************************/
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+
+/* *******************************************************
+*  Memory operations
+**********************************************************/
+static void ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+
+/* *************************************
+*  Error Management
+***************************************/
+
+/*! ZSTD_isError
+*   tells if a return value is an error code */
+static unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }
+
+
+/* *************************************************************
+*   Context management
+***************************************************************/
+typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
+               ZSTDds_decodeBlockHeader, ZSTDds_decompressBlock } ZSTD_dStage;
+
+struct ZSTDv04_Dctx_s
+{
+    U32 LLTable[FSE_DTABLE_SIZE_U32(LLFSELog)];
+    U32 OffTable[FSE_DTABLE_SIZE_U32(OffFSELog)];
+    U32 MLTable[FSE_DTABLE_SIZE_U32(MLFSELog)];
+    const void* previousDstEnd;
+    const void* base;
+    const void* vBase;
+    const void* dictEnd;
+    size_t expected;
+    size_t headerSize;
+    ZSTD_parameters params;
+    blockType_t bType;
+    ZSTD_dStage stage;
+    const BYTE* litPtr;
+    size_t litSize;
+    BYTE litBuffer[BLOCKSIZE + 8 /* margin for wildcopy */];
+    BYTE headerBuffer[ZSTD_frameHeaderSize_max];
+};  /* typedef'd to ZSTD_DCtx within "zstd_static.h" */
+
+static size_t ZSTD_resetDCtx(ZSTD_DCtx* dctx)
+{
+    dctx->expected = ZSTD_frameHeaderSize_min;
+    dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    dctx->vBase = NULL;
+    dctx->dictEnd = NULL;
+    return 0;
+}
+
+static ZSTD_DCtx* ZSTD_createDCtx(void)
+{
+    ZSTD_DCtx* dctx = (ZSTD_DCtx*)malloc(sizeof(ZSTD_DCtx));
+    if (dctx==NULL) return NULL;
+    ZSTD_resetDCtx(dctx);
+    return dctx;
+}
+
+static size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
+{
+    free(dctx);
+    return 0;
+}
+
+
+/* *************************************************************
+*   Decompression section
+***************************************************************/
+/** ZSTD_decodeFrameHeader_Part1
+*   decode the 1st part of the Frame Header, which tells Frame Header size.
+*   srcSize must be == ZSTD_frameHeaderSize_min
+*   @return : the full size of the Frame Header */
+static size_t ZSTD_decodeFrameHeader_Part1(ZSTD_DCtx* zc, const void* src, size_t srcSize)
+{
+    U32 magicNumber;
+    if (srcSize != ZSTD_frameHeaderSize_min) return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_MAGICNUMBER) return ERROR(prefix_unknown);
+    zc->headerSize = ZSTD_frameHeaderSize_min;
+    return zc->headerSize;
+}
+
+
+static size_t ZSTD_getFrameParams(ZSTD_parameters* params, const void* src, size_t srcSize)
+{
+    U32 magicNumber;
+    if (srcSize < ZSTD_frameHeaderSize_min) return ZSTD_frameHeaderSize_max;
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTD_MAGICNUMBER) return ERROR(prefix_unknown);
+    memset(params, 0, sizeof(*params));
+    params->windowLog = (((const BYTE*)src)[4] & 15) + ZSTD_WINDOWLOG_ABSOLUTEMIN;
+    if ((((const BYTE*)src)[4] >> 4) != 0) return ERROR(frameParameter_unsupported);   /* reserved bits */
+    return 0;
+}
+
+/** ZSTD_decodeFrameHeader_Part2
+*   decode the full Frame Header
+*   srcSize must be the size provided by ZSTD_decodeFrameHeader_Part1
+*   @return : 0, or an error code, which can be tested using ZSTD_isError() */
+static size_t ZSTD_decodeFrameHeader_Part2(ZSTD_DCtx* zc, const void* src, size_t srcSize)
+{
+    size_t result;
+    if (srcSize != zc->headerSize) return ERROR(srcSize_wrong);
+    result = ZSTD_getFrameParams(&(zc->params), src, srcSize);
+    if ((MEM_32bits()) && (zc->params.windowLog > 25)) return ERROR(frameParameter_unsupported);
+    return result;
+}
+
+
+static size_t ZSTD_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    BYTE headerFlags;
+    U32 cSize;
+
+    if (srcSize < 3) return ERROR(srcSize_wrong);
+
+    headerFlags = *in;
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+
+    bpPtr->blockType = (blockType_t)(headerFlags >> 6);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+static size_t ZSTD_copyRawBlock(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    if (srcSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/** ZSTD_decompressLiterals
+    @return : nb of bytes read from src, or an error code*/
+static size_t ZSTD_decompressLiterals(void* dst, size_t* maxDstSizePtr,
+                                const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+
+    const size_t litSize = (MEM_readLE32(src) & 0x1FFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+    const size_t litCSize = (MEM_readLE32(ip+2) & 0xFFFFFF) >> 5;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+
+    if (litSize > *maxDstSizePtr) return ERROR(corruption_detected);
+    if (litCSize + 5 > srcSize) return ERROR(corruption_detected);
+
+    if (HUF_isError(HUF_decompress(dst, litSize, ip+5, litCSize))) return ERROR(corruption_detected);
+
+    *maxDstSizePtr = litSize;
+    return litCSize + 5;
+}
+
+
+/** ZSTD_decodeLiteralsBlock
+    @return : nb of bytes read from src (< srcSize ) */
+static size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
+                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+{
+    const BYTE* const istart = (const BYTE*) src;
+
+    /* any compressed block with literals segment must be at least this size */
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch(*istart & 3)
+    {
+    /* compressed */
+    case 0:
+        {
+            size_t litSize = BLOCKSIZE;
+            const size_t readSize = ZSTD_decompressLiterals(dctx->litBuffer, &litSize, src, srcSize);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, 8);
+            return readSize;   /* works if it's an error too */
+        }
+    case IS_RAW:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > srcSize-11)   /* risk of reading too far with wildcopy */
+            {
+                if (litSize > srcSize-3) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, 8);
+                return litSize+3;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+3;
+            dctx->litSize = litSize;
+            return litSize+3;        }
+    case IS_RLE:
+        {
+            const size_t litSize = (MEM_readLE32(istart) & 0xFFFFFF) >> 2;   /* no buffer issue : srcSize >= MIN_CBLOCK_SIZE */
+            if (litSize > BLOCKSIZE) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[3], litSize + 8);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return 4;
+        }
+    default:
+        return ERROR(corruption_detected);   /* forbidden nominal case */
+    }
+}
+
+
+static size_t ZSTD_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t* dumpsLengthPtr,
+                         FSE_DTable* DTableLL, FSE_DTable* DTableML, FSE_DTable* DTableOffb,
+                         const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    const BYTE* const iend = istart + srcSize;
+    U32 LLtype, Offtype, MLtype;
+    U32 LLlog, Offlog, MLlog;
+    size_t dumpsLength;
+
+    /* check */
+    if (srcSize < 5) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    *nbSeq = MEM_readLE16(ip); ip+=2;
+    LLtype  = *ip >> 6;
+    Offtype = (*ip >> 4) & 3;
+    MLtype  = (*ip >> 2) & 3;
+    if (*ip & 2)
+    {
+        dumpsLength  = ip[2];
+        dumpsLength += ip[1] << 8;
+        ip += 3;
+    }
+    else
+    {
+        dumpsLength  = ip[1];
+        dumpsLength += (ip[0] & 1) << 8;
+        ip += 2;
+    }
+    *dumpsPtr = ip;
+    ip += dumpsLength;
+    *dumpsLengthPtr = dumpsLength;
+
+    /* check */
+    if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+    /* sequences */
+    {
+        S16 norm[MaxML+1];    /* assumption : MaxML >= MaxLL >= MaxOff */
+        size_t headerSize;
+
+        /* Build DTables */
+        switch(LLtype)
+        {
+        case bt_rle :
+            LLlog = 0;
+            FSE_buildDTable_rle(DTableLL, *ip++); break;
+        case bt_raw :
+            LLlog = LLbits;
+            FSE_buildDTable_raw(DTableLL, LLbits); break;
+        default :
+            {   U32 max = MaxLL;
+                headerSize = FSE_readNCount(norm, &max, &LLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (LLlog > LLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableLL, norm, max, LLlog);
+        }   }
+
+        switch(Offtype)
+        {
+        case bt_rle :
+            Offlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong);   /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableOffb, *ip++ & MaxOff); /* if *ip > MaxOff, data is corrupted */
+            break;
+        case bt_raw :
+            Offlog = Offbits;
+            FSE_buildDTable_raw(DTableOffb, Offbits); break;
+        default :
+            {   U32 max = MaxOff;
+                headerSize = FSE_readNCount(norm, &max, &Offlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (Offlog > OffFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableOffb, norm, max, Offlog);
+        }   }
+
+        switch(MLtype)
+        {
+        case bt_rle :
+            MLlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSE_buildDTable_rle(DTableML, *ip++); break;
+        case bt_raw :
+            MLlog = MLbits;
+            FSE_buildDTable_raw(DTableML, MLbits); break;
+        default :
+            {   U32 max = MaxML;
+                headerSize = FSE_readNCount(norm, &max, &MLlog, ip, iend-ip);
+                if (FSE_isError(headerSize)) return ERROR(GENERIC);
+                if (MLlog > MLFSELog) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSE_buildDTable(DTableML, norm, max, MLlog);
+    }   }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t offset;
+    size_t matchLength;
+} seq_t;
+
+typedef struct {
+    BIT_DStream_t DStream;
+    FSE_DState_t stateLL;
+    FSE_DState_t stateOffb;
+    FSE_DState_t stateML;
+    size_t prevOffset;
+    const BYTE* dumps;
+    const BYTE* dumpsEnd;
+} seqState_t;
+
+
+static void ZSTD_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    size_t litLength;
+    size_t prevOffset;
+    size_t offset;
+    size_t matchLength;
+    const BYTE* dumps = seqState->dumps;
+    const BYTE* const de = seqState->dumpsEnd;
+
+    /* Literal length */
+    litLength = FSE_decodeSymbol(&(seqState->stateLL), &(seqState->DStream));
+    prevOffset = litLength ? seq->offset : seqState->prevOffset;
+    if (litLength == MaxLL) {
+        U32 add = *dumps++;
+        if (add < 255) litLength += add;
+        else {
+            litLength = dumps[0] + (dumps[1]<<8) + (dumps[2]<<16);
+            dumps += 3;
+        }
+        if (dumps > de) { litLength = MaxLL+255; }  /* late correction, to avoid using uninitialized memory */
+        if (dumps >= de) { dumps = de-1; }  /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+
+    /* Offset */
+    {   static const U32 offsetPrefix[MaxOff+1] = {
+                1 /*fake*/, 1, 2, 4, 8, 16, 32, 64, 128, 256,
+                512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144,
+                524288, 1048576, 2097152, 4194304, 8388608, 16777216, 33554432, /*fake*/ 1, 1, 1, 1, 1 };
+        U32 offsetCode, nbBits;
+        offsetCode = FSE_decodeSymbol(&(seqState->stateOffb), &(seqState->DStream));   /* <= maxOff, by table construction */
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        nbBits = offsetCode - 1;
+        if (offsetCode==0) nbBits = 0;   /* cmove */
+        offset = offsetPrefix[offsetCode] + BIT_readBits(&(seqState->DStream), nbBits);
+        if (MEM_32bits()) BIT_reloadDStream(&(seqState->DStream));
+        if (offsetCode==0) offset = prevOffset;   /* cmove */
+        if (offsetCode | !litLength) seqState->prevOffset = seq->offset;   /* cmove */
+    }
+
+    /* MatchLength */
+    matchLength = FSE_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
+    if (matchLength == MaxML) {
+        U32 add = *dumps++;
+        if (add < 255) matchLength += add;
+        else {
+            matchLength = dumps[0] + (dumps[1]<<8) + (dumps[2]<<16);
+            dumps += 3;
+        }
+        if (dumps > de) { matchLength = MaxML+255; }  /* late correction, to avoid using uninitialized memory */
+        if (dumps >= de) { dumps = de-1; }  /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+    matchLength += MINMATCH;
+
+    /* save result */
+    seq->litLength = litLength;
+    seq->offset = offset;
+    seq->matchLength = matchLength;
+    seqState->dumps = dumps;
+}
+
+
+static size_t ZSTD_execSequence(BYTE* op,
+                                BYTE* const oend, seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    static const int dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+    static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* substracted */
+    BYTE* const oLitEnd = op + sequence.litLength;
+    const size_t sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_8 = oend-8;
+    const BYTE* const litEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* risk read beyond lit buffer */
+
+    /* copy Literals */
+    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = litEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base))
+    {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - vBase))
+            return ERROR(corruption_detected);
+        match = dictEnd - (base-match);
+        if (match + sequence.matchLength <= dictEnd)
+        {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {
+            size_t length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+            if (op > oend_8 || sequence.matchLength < MINMATCH) {
+              while (op < oMatchEnd) *op++ = *match++;
+              return sequenceLength;
+            }
+        }
+    }
+    /* Requirement: op <= oend_8 */
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        const int sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTD_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTD_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH))
+    {
+        if (op < oend_8)
+        {
+            ZSTD_wildcopy(op, match, oend_8 - op);
+            match += oend_8 - op;
+            op = oend_8;
+        }
+        while (op < oMatchEnd) *op++ = *match++;
+    }
+    else
+    {
+        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+
+static size_t ZSTD_decompressSequences(
+                               ZSTD_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t errorCode, dumpsLength;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    int nbSeq;
+    const BYTE* dumps;
+    U32* DTableLL = dctx->LLTable;
+    U32* DTableML = dctx->MLTable;
+    U32* DTableOffb = dctx->OffTable;
+    const BYTE* const base = (const BYTE*) (dctx->base);
+    const BYTE* const vBase = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+
+    /* Build Decoding Tables */
+    errorCode = ZSTD_decodeSeqHeaders(&nbSeq, &dumps, &dumpsLength,
+                                      DTableLL, DTableML, DTableOffb,
+                                      ip, iend-ip);
+    if (ZSTD_isError(errorCode)) return errorCode;
+    ip += errorCode;
+
+    /* Regen sequences */
+    {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        sequence.offset = 4;
+        seqState.dumps = dumps;
+        seqState.dumpsEnd = dumps + dumpsLength;
+        seqState.prevOffset = 4;
+        errorCode = BIT_initDStream(&(seqState.DStream), ip, iend-ip);
+        if (ERR_isError(errorCode)) return ERROR(corruption_detected);
+        FSE_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSE_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSE_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && nbSeq ; )
+        {
+            size_t oneSeqSize;
+            nbSeq--;
+            ZSTD_decodeSequence(&sequence, &seqState);
+            oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, base, vBase, dictEnd);
+            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* check if reached exact end */
+        if ( !BIT_endOfDStream(&(seqState.DStream)) ) return ERROR(corruption_detected);   /* DStream should be entirely and exactly consumed; otherwise data is corrupted */
+
+        /* last literal segment */
+        {
+            size_t lastLLSize = litEnd - litPtr;
+            if (litPtr > litEnd) return ERROR(corruption_detected);
+            if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+            if (op != litPtr) memcpy(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+    }
+
+    return op-ostart;
+}
+
+
+static void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst)
+{
+    if (dst != dctx->previousDstEnd)   /* not contiguous */
+    {
+        dctx->dictEnd = dctx->previousDstEnd;
+        dctx->vBase = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+        dctx->base = dst;
+        dctx->previousDstEnd = dst;
+    }
+}
+
+
+static size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
+                            void* dst, size_t maxDstSize,
+                      const void* src, size_t srcSize)
+{
+    /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+
+    /* Decode literals sub-block */
+    size_t litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize);
+    if (ZSTD_isError(litCSize)) return litCSize;
+    ip += litCSize;
+    srcSize -= litCSize;
+
+    return ZSTD_decompressSequences(dctx, dst, maxDstSize, ip, srcSize);
+}
+
+
+static size_t ZSTD_decompress_usingDict(ZSTD_DCtx* ctx,
+                                 void* dst, size_t maxDstSize,
+                                 const void* src, size_t srcSize,
+                                 const void* dict, size_t dictSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties;
+
+    /* init */
+    ZSTD_resetDCtx(ctx);
+    if (dict)
+    {
+        ZSTD_decompress_insertDictionary(ctx, dict, dictSize);
+        ctx->dictEnd = ctx->previousDstEnd;
+        ctx->vBase = (const char*)dst - ((const char*)(ctx->previousDstEnd) - (const char*)(ctx->base));
+        ctx->base = dst;
+    }
+    else
+    {
+        ctx->vBase = ctx->base = ctx->dictEnd = dst;
+    }
+
+    /* Frame Header */
+    {
+        size_t frameHeaderSize;
+        if (srcSize < ZSTD_frameHeaderSize_min+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+        frameHeaderSize = ZSTD_decodeFrameHeader_Part1(ctx, src, ZSTD_frameHeaderSize_min);
+        if (ZSTD_isError(frameHeaderSize)) return frameHeaderSize;
+        if (srcSize < frameHeaderSize+ZSTD_blockHeaderSize) return ERROR(srcSize_wrong);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+        frameHeaderSize = ZSTD_decodeFrameHeader_Part2(ctx, src, frameHeaderSize);
+        if (ZSTD_isError(frameHeaderSize)) return frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t decodedSize=0;
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTD_decompressBlock_internal(ctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTD_copyRawBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        if (ZSTD_isError(decodedSize)) return decodedSize;
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+static size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTD_frameHeaderSize_min) return ERROR(srcSize_wrong);
+    if (MEM_readLE32(src) != ZSTD_MAGICNUMBER) return ERROR(prefix_unknown);
+    ip += ZSTD_frameHeaderSize_min; remainingSize -= ZSTD_frameHeaderSize_min;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTD_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTD_blockHeaderSize;
+        remainingSize -= ZSTD_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/* ******************************
+*  Streaming Decompression API
+********************************/
+static size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+static size_t ZSTD_decompressContinue(ZSTD_DCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != ctx->expected) return ERROR(srcSize_wrong);
+    ZSTD_checkContinuity(ctx, dst);
+
+    /* Decompress : frame header; part 1 */
+    switch (ctx->stage)
+    {
+    case ZSTDds_getFrameHeaderSize :
+        /* get frame header size */
+        if (srcSize != ZSTD_frameHeaderSize_min) return ERROR(srcSize_wrong);   /* impossible */
+        ctx->headerSize = ZSTD_decodeFrameHeader_Part1(ctx, src, ZSTD_frameHeaderSize_min);
+        if (ZSTD_isError(ctx->headerSize)) return ctx->headerSize;
+        memcpy(ctx->headerBuffer, src, ZSTD_frameHeaderSize_min);
+        if (ctx->headerSize > ZSTD_frameHeaderSize_min) return ERROR(GENERIC);   /* impossible */
+        ctx->expected = 0;   /* not necessary to copy more */
+        /* fallthrough */
+    case ZSTDds_decodeFrameHeader:
+        /* get frame header */
+        {   size_t const result = ZSTD_decodeFrameHeader_Part2(ctx, ctx->headerBuffer, ctx->headerSize);
+            if (ZSTD_isError(result)) return result;
+            ctx->expected = ZSTD_blockHeaderSize;
+            ctx->stage = ZSTDds_decodeBlockHeader;
+            return 0;
+        }
+    case ZSTDds_decodeBlockHeader:
+        /* Decode block header */
+        {   blockProperties_t bp;
+            size_t const blockSize = ZSTD_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
+            if (ZSTD_isError(blockSize)) return blockSize;
+            if (bp.blockType == bt_end)
+            {
+                ctx->expected = 0;
+                ctx->stage = ZSTDds_getFrameHeaderSize;
+            }
+            else
+            {
+                ctx->expected = blockSize;
+                ctx->bType = bp.blockType;
+                ctx->stage = ZSTDds_decompressBlock;
+            }
+            return 0;
+        }
+    case ZSTDds_decompressBlock:
+        {
+            /* Decompress : block content */
+            size_t rSize;
+            switch(ctx->bType)
+            {
+            case bt_compressed:
+                rSize = ZSTD_decompressBlock_internal(ctx, dst, maxDstSize, src, srcSize);
+                break;
+            case bt_raw :
+                rSize = ZSTD_copyRawBlock(dst, maxDstSize, src, srcSize);
+                break;
+            case bt_rle :
+                return ERROR(GENERIC);   /* not yet handled */
+                break;
+            case bt_end :   /* should never happen (filtered at phase 1) */
+                rSize = 0;
+                break;
+            default:
+                return ERROR(GENERIC);
+            }
+            ctx->stage = ZSTDds_decodeBlockHeader;
+            ctx->expected = ZSTD_blockHeaderSize;
+            ctx->previousDstEnd = (char*)dst + rSize;
+            return rSize;
+        }
+    default:
+        return ERROR(GENERIC);   /* impossible */
+    }
+}
+
+
+static void ZSTD_decompress_insertDictionary(ZSTD_DCtx* ctx, const void* dict, size_t dictSize)
+{
+    ctx->dictEnd = ctx->previousDstEnd;
+    ctx->vBase = (const char*)dict - ((const char*)(ctx->previousDstEnd) - (const char*)(ctx->base));
+    ctx->base = dict;
+    ctx->previousDstEnd = (const char*)dict + dictSize;
+}
+
+
+
+/*
+    Buffered version of Zstd compression library
+    Copyright (C) 2015, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* The objects defined into this file should be considered experimental.
+ * They are not labelled stable, as their prototype may change in the future.
+ * You can use them for tests, provide feedback, or if you can endure risk of future changes.
+ */
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stdlib.h>
+
+
+/** ************************************************
+*  Streaming decompression
+*
+*  A ZBUFF_DCtx object is required to track streaming operation.
+*  Use ZBUFF_createDCtx() and ZBUFF_freeDCtx() to create/release resources.
+*  Use ZBUFF_decompressInit() to start a new decompression operation.
+*  ZBUFF_DCtx objects can be reused multiple times.
+*
+*  Use ZBUFF_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *maxDstSizePtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *maxDstSizePtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to call again the function with remaining input.
+*  The content of dst will be overwritten (up to *maxDstSizePtr) at each function call, so save its content if it matters or change dst .
+*  return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to improve latency)
+*            or 0 when a frame is completely decoded
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory)
+*  output : 128 KB block size is the internal unit, it ensures it's always possible to write a full block when it's decoded.
+*  input : just follow indications from ZBUFF_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* **************************************************/
+
+typedef enum { ZBUFFds_init, ZBUFFds_readHeader, ZBUFFds_loadHeader, ZBUFFds_decodeHeader,
+               ZBUFFds_read, ZBUFFds_load, ZBUFFds_flush } ZBUFF_dStage;
+
+/* *** Resource management *** */
+
+#define ZSTD_frameHeaderSize_max 5   /* too magical, should come from reference */
+struct ZBUFFv04_DCtx_s {
+    ZSTD_DCtx* zc;
+    ZSTD_parameters params;
+    char* inBuff;
+    size_t inBuffSize;
+    size_t inPos;
+    char* outBuff;
+    size_t outBuffSize;
+    size_t outStart;
+    size_t outEnd;
+    size_t hPos;
+    const char* dict;
+    size_t dictSize;
+    ZBUFF_dStage stage;
+    unsigned char headerBuffer[ZSTD_frameHeaderSize_max];
+};   /* typedef'd to ZBUFF_DCtx within "zstd_buffered.h" */
+
+typedef ZBUFFv04_DCtx ZBUFF_DCtx;
+
+
+static ZBUFF_DCtx* ZBUFF_createDCtx(void)
+{
+    ZBUFF_DCtx* zbc = (ZBUFF_DCtx*)malloc(sizeof(ZBUFF_DCtx));
+    if (zbc==NULL) return NULL;
+    memset(zbc, 0, sizeof(*zbc));
+    zbc->zc = ZSTD_createDCtx();
+    zbc->stage = ZBUFFds_init;
+    return zbc;
+}
+
+static size_t ZBUFF_freeDCtx(ZBUFF_DCtx* zbc)
+{
+    if (zbc==NULL) return 0;   /* support free on null */
+    ZSTD_freeDCtx(zbc->zc);
+    free(zbc->inBuff);
+    free(zbc->outBuff);
+    free(zbc);
+    return 0;
+}
+
+
+/* *** Initialization *** */
+
+static size_t ZBUFF_decompressInit(ZBUFF_DCtx* zbc)
+{
+    zbc->stage = ZBUFFds_readHeader;
+    zbc->hPos = zbc->inPos = zbc->outStart = zbc->outEnd = zbc->dictSize = 0;
+    return ZSTD_resetDCtx(zbc->zc);
+}
+
+
+static size_t ZBUFF_decompressWithDictionary(ZBUFF_DCtx* zbc, const void* src, size_t srcSize)
+{
+    zbc->dict = (const char*)src;
+    zbc->dictSize = srcSize;
+    return 0;
+}
+
+static size_t ZBUFF_limitCopy(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    size_t length = MIN(maxDstSize, srcSize);
+    memcpy(dst, src, length);
+    return length;
+}
+
+/* *** Decompression *** */
+
+static size_t ZBUFF_decompressContinue(ZBUFF_DCtx* zbc, void* dst, size_t* maxDstSizePtr, const void* src, size_t* srcSizePtr)
+{
+    const char* const istart = (const char*)src;
+    const char* ip = istart;
+    const char* const iend = istart + *srcSizePtr;
+    char* const ostart = (char*)dst;
+    char* op = ostart;
+    char* const oend = ostart + *maxDstSizePtr;
+    U32 notDone = 1;
+
+    DEBUGLOG(5, "ZBUFF_decompressContinue");
+    while (notDone)
+    {
+        switch(zbc->stage)
+        {
+
+        case ZBUFFds_init :
+            DEBUGLOG(5, "ZBUFF_decompressContinue: stage==ZBUFFds_init => ERROR(init_missing)");
+            return ERROR(init_missing);
+
+        case ZBUFFds_readHeader :
+            /* read header from src */
+            {   size_t const headerSize = ZSTD_getFrameParams(&(zbc->params), src, *srcSizePtr);
+                if (ZSTD_isError(headerSize)) return headerSize;
+                if (headerSize) {
+                    /* not enough input to decode header : tell how many bytes would be necessary */
+                    memcpy(zbc->headerBuffer+zbc->hPos, src, *srcSizePtr);
+                    zbc->hPos += *srcSizePtr;
+                    *maxDstSizePtr = 0;
+                    zbc->stage = ZBUFFds_loadHeader;
+                    return headerSize - zbc->hPos;
+                }
+                zbc->stage = ZBUFFds_decodeHeader;
+                break;
+            }
+
+        case ZBUFFds_loadHeader:
+            /* complete header from src */
+            {   size_t headerSize = ZBUFF_limitCopy(
+                    zbc->headerBuffer + zbc->hPos, ZSTD_frameHeaderSize_max - zbc->hPos,
+                    src, *srcSizePtr);
+                zbc->hPos += headerSize;
+                ip += headerSize;
+                headerSize = ZSTD_getFrameParams(&(zbc->params), zbc->headerBuffer, zbc->hPos);
+                if (ZSTD_isError(headerSize)) return headerSize;
+                if (headerSize) {
+                    /* not enough input to decode header : tell how many bytes would be necessary */
+                    *maxDstSizePtr = 0;
+                    return headerSize - zbc->hPos;
+            }   }
+            /* intentional fallthrough */
+
+        case ZBUFFds_decodeHeader:
+                /* apply header to create / resize buffers */
+                {   size_t const neededOutSize = (size_t)1 << zbc->params.windowLog;
+                    size_t const neededInSize = BLOCKSIZE;   /* a block is never > BLOCKSIZE */
+                    if (zbc->inBuffSize < neededInSize) {
+                        free(zbc->inBuff);
+                        zbc->inBuffSize = neededInSize;
+                        zbc->inBuff = (char*)malloc(neededInSize);
+                        if (zbc->inBuff == NULL) return ERROR(memory_allocation);
+                    }
+                    if (zbc->outBuffSize < neededOutSize) {
+                        free(zbc->outBuff);
+                        zbc->outBuffSize = neededOutSize;
+                        zbc->outBuff = (char*)malloc(neededOutSize);
+                        if (zbc->outBuff == NULL) return ERROR(memory_allocation);
+                }   }
+                if (zbc->dictSize)
+                    ZSTD_decompress_insertDictionary(zbc->zc, zbc->dict, zbc->dictSize);
+                if (zbc->hPos) {
+                    /* some data already loaded into headerBuffer : transfer into inBuff */
+                    memcpy(zbc->inBuff, zbc->headerBuffer, zbc->hPos);
+                    zbc->inPos = zbc->hPos;
+                    zbc->hPos = 0;
+                    zbc->stage = ZBUFFds_load;
+                    break;
+                }
+                zbc->stage = ZBUFFds_read;
+		/* fall-through */
+        case ZBUFFds_read:
+            {
+                size_t neededInSize = ZSTD_nextSrcSizeToDecompress(zbc->zc);
+                if (neededInSize==0)   /* end of frame */
+                {
+                    zbc->stage = ZBUFFds_init;
+                    notDone = 0;
+                    break;
+                }
+                if ((size_t)(iend-ip) >= neededInSize)
+                {
+                    /* directly decode from src */
+                    size_t decodedSize = ZSTD_decompressContinue(zbc->zc,
+                        zbc->outBuff + zbc->outStart, zbc->outBuffSize - zbc->outStart,
+                        ip, neededInSize);
+                    if (ZSTD_isError(decodedSize)) return decodedSize;
+                    ip += neededInSize;
+                    if (!decodedSize) break;   /* this was just a header */
+                    zbc->outEnd = zbc->outStart +  decodedSize;
+                    zbc->stage = ZBUFFds_flush;
+                    break;
+                }
+                if (ip==iend) { notDone = 0; break; }   /* no more input */
+                zbc->stage = ZBUFFds_load;
+            }
+	    /* fall-through */
+        case ZBUFFds_load:
+            {
+                size_t neededInSize = ZSTD_nextSrcSizeToDecompress(zbc->zc);
+                size_t toLoad = neededInSize - zbc->inPos;   /* should always be <= remaining space within inBuff */
+                size_t loadedSize;
+                if (toLoad > zbc->inBuffSize - zbc->inPos) return ERROR(corruption_detected);   /* should never happen */
+                loadedSize = ZBUFF_limitCopy(zbc->inBuff + zbc->inPos, toLoad, ip, iend-ip);
+                ip += loadedSize;
+                zbc->inPos += loadedSize;
+                if (loadedSize < toLoad) { notDone = 0; break; }   /* not enough input, wait for more */
+                {
+                    size_t decodedSize = ZSTD_decompressContinue(zbc->zc,
+                        zbc->outBuff + zbc->outStart, zbc->outBuffSize - zbc->outStart,
+                        zbc->inBuff, neededInSize);
+                    if (ZSTD_isError(decodedSize)) return decodedSize;
+                    zbc->inPos = 0;   /* input is consumed */
+                    if (!decodedSize) { zbc->stage = ZBUFFds_read; break; }   /* this was just a header */
+                    zbc->outEnd = zbc->outStart +  decodedSize;
+                    zbc->stage = ZBUFFds_flush;
+                    /* ZBUFFds_flush follows */
+                }
+            }
+	    /* fall-through */
+        case ZBUFFds_flush:
+            {
+                size_t toFlushSize = zbc->outEnd - zbc->outStart;
+                size_t flushedSize = ZBUFF_limitCopy(op, oend-op, zbc->outBuff + zbc->outStart, toFlushSize);
+                op += flushedSize;
+                zbc->outStart += flushedSize;
+                if (flushedSize == toFlushSize)
+                {
+                    zbc->stage = ZBUFFds_read;
+                    if (zbc->outStart + BLOCKSIZE > zbc->outBuffSize)
+                        zbc->outStart = zbc->outEnd = 0;
+                    break;
+                }
+                /* cannot flush everything */
+                notDone = 0;
+                break;
+            }
+        default: return ERROR(GENERIC);   /* impossible */
+        }
+    }
+
+    *srcSizePtr = ip-istart;
+    *maxDstSizePtr = op-ostart;
+
+    {
+        size_t nextSrcSizeHint = ZSTD_nextSrcSizeToDecompress(zbc->zc);
+        if (nextSrcSizeHint > 3) nextSrcSizeHint+= 3;   /* get the next block header while at it */
+        nextSrcSizeHint -= zbc->inPos;   /* already loaded*/
+        return nextSrcSizeHint;
+    }
+}
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+unsigned ZBUFFv04_isError(size_t errorCode) { return ERR_isError(errorCode); }
+const char* ZBUFFv04_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
+
+size_t ZBUFFv04_recommendedDInSize()  { return BLOCKSIZE + 3; }
+size_t ZBUFFv04_recommendedDOutSize() { return BLOCKSIZE; }
+
+
+
+/*- ========================================================================= -*/
+
+/* final wrapping stage */
+
+size_t ZSTDv04_decompressDCtx(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return ZSTD_decompress_usingDict(dctx, dst, maxDstSize, src, srcSize, NULL, 0);
+}
+
+size_t ZSTDv04_decompress(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+#if defined(ZSTD_HEAPMODE) && (ZSTD_HEAPMODE==1)
+    size_t regenSize;
+    ZSTD_DCtx* dctx = ZSTD_createDCtx();
+    if (dctx==NULL) return ERROR(memory_allocation);
+    regenSize = ZSTDv04_decompressDCtx(dctx, dst, maxDstSize, src, srcSize);
+    ZSTD_freeDCtx(dctx);
+    return regenSize;
+#else
+    ZSTD_DCtx dctx;
+    return ZSTDv04_decompressDCtx(&dctx, dst, maxDstSize, src, srcSize);
+#endif
+}
+
+size_t ZSTDv04_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    return ZSTD_findFrameCompressedSize(src, srcSize);
+}
+
+size_t ZSTDv04_resetDCtx(ZSTDv04_Dctx* dctx) { return ZSTD_resetDCtx(dctx); }
+
+size_t ZSTDv04_nextSrcSizeToDecompress(ZSTDv04_Dctx* dctx)
+{
+    return ZSTD_nextSrcSizeToDecompress(dctx);
+}
+
+size_t ZSTDv04_decompressContinue(ZSTDv04_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return ZSTD_decompressContinue(dctx, dst, maxDstSize, src, srcSize);
+}
+
+
+
+ZBUFFv04_DCtx* ZBUFFv04_createDCtx(void) { return ZBUFF_createDCtx(); }
+size_t ZBUFFv04_freeDCtx(ZBUFFv04_DCtx* dctx) { return ZBUFF_freeDCtx(dctx); }
+
+size_t ZBUFFv04_decompressInit(ZBUFFv04_DCtx* dctx) { return ZBUFF_decompressInit(dctx); }
+size_t ZBUFFv04_decompressWithDictionary(ZBUFFv04_DCtx* dctx, const void* src, size_t srcSize)
+{ return ZBUFF_decompressWithDictionary(dctx, src, srcSize); }
+
+size_t ZBUFFv04_decompressContinue(ZBUFFv04_DCtx* dctx, void* dst, size_t* maxDstSizePtr, const void* src, size_t* srcSizePtr)
+{
+    DEBUGLOG(5, "ZBUFFv04_decompressContinue");
+    return ZBUFF_decompressContinue(dctx, dst, maxDstSizePtr, src, srcSizePtr);
+}
+
+ZSTD_DCtx* ZSTDv04_createDCtx(void) { return ZSTD_createDCtx(); }
+size_t ZSTDv04_freeDCtx(ZSTD_DCtx* dctx) { return ZSTD_freeDCtx(dctx); }
+
+size_t ZSTDv04_getFrameParams(ZSTD_parameters* params, const void* src, size_t srcSize)
+{
+    return ZSTD_getFrameParams(params, src, srcSize);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstd_v04.h b/vendor/github.com/DataDog/zstd/zstd_v04.h
new file mode 100644
index 0000000..6391631
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v04.h
@@ -0,0 +1,137 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_V04_H_91868324769238
+#define ZSTD_V04_H_91868324769238
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* *************************************
+*  Includes
+***************************************/
+#include <stddef.h>   /* size_t */
+
+
+/* *************************************
+*  Simple one-step function
+***************************************/
+/**
+ZSTDv04_decompress() : decompress ZSTD frames compliant with v0.4.x format
+    compressedSize : is the exact source size
+    maxOriginalSize : is the size of the 'dst' buffer, which must be already allocated.
+                      It must be equal or larger than originalSize, otherwise decompression will fail.
+    return : the number of bytes decompressed into destination buffer (originalSize)
+             or an errorCode if it fails (which can be tested using ZSTDv01_isError())
+*/
+size_t ZSTDv04_decompress( void* dst, size_t maxOriginalSize,
+                     const void* src, size_t compressedSize);
+
+/**
+ZSTDv04_getFrameSrcSize() : get the source length of a ZSTD frame compliant with v0.4.x format
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv04_isError())
+*/
+size_t ZSTDv04_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/**
+ZSTDv04_isError() : tells if the result of ZSTDv04_decompress() is an error
+*/
+unsigned ZSTDv04_isError(size_t code);
+
+
+/* *************************************
+*  Advanced functions
+***************************************/
+typedef struct ZSTDv04_Dctx_s ZSTDv04_Dctx;
+ZSTDv04_Dctx* ZSTDv04_createDCtx(void);
+size_t ZSTDv04_freeDCtx(ZSTDv04_Dctx* dctx);
+
+size_t ZSTDv04_decompressDCtx(ZSTDv04_Dctx* dctx,
+                              void* dst, size_t maxOriginalSize,
+                        const void* src, size_t compressedSize);
+
+
+/* *************************************
+*  Direct Streaming
+***************************************/
+size_t ZSTDv04_resetDCtx(ZSTDv04_Dctx* dctx);
+
+size_t ZSTDv04_nextSrcSizeToDecompress(ZSTDv04_Dctx* dctx);
+size_t ZSTDv04_decompressContinue(ZSTDv04_Dctx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
+/**
+  Use above functions alternatively.
+  ZSTD_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTD_decompressContinue().
+  ZSTD_decompressContinue() will use previous data blocks to improve compression if they are located prior to current block.
+  Result is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTD_decompressContinue() has decoded some header.
+*/
+
+
+/* *************************************
+*  Buffered Streaming
+***************************************/
+typedef struct ZBUFFv04_DCtx_s ZBUFFv04_DCtx;
+ZBUFFv04_DCtx* ZBUFFv04_createDCtx(void);
+size_t         ZBUFFv04_freeDCtx(ZBUFFv04_DCtx* dctx);
+
+size_t ZBUFFv04_decompressInit(ZBUFFv04_DCtx* dctx);
+size_t ZBUFFv04_decompressWithDictionary(ZBUFFv04_DCtx* dctx, const void* dict, size_t dictSize);
+
+size_t ZBUFFv04_decompressContinue(ZBUFFv04_DCtx* dctx, void* dst, size_t* maxDstSizePtr, const void* src, size_t* srcSizePtr);
+
+/** ************************************************
+*  Streaming decompression
+*
+*  A ZBUFF_DCtx object is required to track streaming operation.
+*  Use ZBUFF_createDCtx() and ZBUFF_freeDCtx() to create/release resources.
+*  Use ZBUFF_decompressInit() to start a new decompression operation.
+*  ZBUFF_DCtx objects can be reused multiple times.
+*
+*  Optionally, a reference to a static dictionary can be set, using ZBUFF_decompressWithDictionary()
+*  It must be the same content as the one set during compression phase.
+*  Dictionary content must remain accessible during the decompression process.
+*
+*  Use ZBUFF_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *maxDstSizePtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *maxDstSizePtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of dst will be overwritten (up to *maxDstSizePtr) at each function call, so save its content if it matters or change dst.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to improve latency)
+*            or 0 when a frame is completely decoded
+*            or an error code, which can be tested using ZBUFF_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFF_recommendedDInSize / ZBUFF_recommendedDOutSize
+*  output : ZBUFF_recommendedDOutSize==128 KB block size is the internal unit, it ensures it's always possible to write a full block when it's decoded.
+*  input : ZBUFF_recommendedDInSize==128Kb+3; just follow indications from ZBUFF_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* **************************************************/
+unsigned ZBUFFv04_isError(size_t errorCode);
+const char* ZBUFFv04_getErrorName(size_t errorCode);
+
+
+/** The below functions provide recommended buffer sizes for Compression or Decompression operations.
+*   These sizes are not compulsory, they just tend to offer better latency */
+size_t ZBUFFv04_recommendedDInSize(void);
+size_t ZBUFFv04_recommendedDOutSize(void);
+
+
+/* *************************************
+*  Prefix - version detection
+***************************************/
+#define ZSTDv04_magicNumber 0xFD2FB524   /* v0.4 */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_V04_H_91868324769238 */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v05.c b/vendor/github.com/DataDog/zstd/zstd_v05.c
new file mode 100644
index 0000000..a5e1b1f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v05.c
@@ -0,0 +1,4011 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*- Dependencies -*/
+#include "zstd_v05.h"
+#include "error_private.h"
+
+
+/* ******************************************************************
+   mem.h
+   low-level memory access routines
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSEv05 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-****************************************
+*  Dependencies
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include <string.h>    /* memcpy */
+
+
+/*-****************************************
+*  Compiler specifics
+******************************************/
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/*-**************************************************************
+*  Basic Types
+*****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/*-**************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets depending on alignment.
+ *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard, by lying on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(U32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign*)memPtr)->u32 = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign*)memPtr)->u64 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write32(void* memPtr, U32 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write64(void* memPtr, U64 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif /* MEM_FORCE_MEMORY_ACCESS */
+
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian()) {
+        MEM_write16(memPtr, val);
+    } else {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U32)((U32)p[0] + ((U32)p[1]<<8) + ((U32)p[2]<<16) + ((U32)p[3]<<24));
+    }
+}
+
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U64)((U64)p[0] + ((U64)p[1]<<8) + ((U64)p[2]<<16) + ((U64)p[3]<<24)
+                     + ((U64)p[4]<<32) + ((U64)p[5]<<40) + ((U64)p[6]<<48) + ((U64)p[7]<<56));
+    }
+}
+
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+
+/*
+    zstd - standard compression library
+    Header File for static linking only
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net
+*/
+#ifndef ZSTD_STATIC_H
+#define ZSTD_STATIC_H
+
+/* The prototypes defined within this file are considered experimental.
+ * They should not be used in the context DLL as they may change in the future.
+ * Prefer static linking if you need them, to control breaking version changes issues.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/*-*************************************
+*  Types
+***************************************/
+#define ZSTDv05_WINDOWLOG_ABSOLUTEMIN 11
+
+
+/*-*************************************
+*  Advanced functions
+***************************************/
+/*- Advanced Decompression functions -*/
+
+/*! ZSTDv05_decompress_usingPreparedDCtx() :
+*   Same as ZSTDv05_decompress_usingDict, but using a reference context `preparedDCtx`, where dictionary has been loaded.
+*   It avoids reloading the dictionary each time.
+*   `preparedDCtx` must have been properly initialized using ZSTDv05_decompressBegin_usingDict().
+*   Requires 2 contexts : 1 for reference, which will not be modified, and 1 to run the decompression operation */
+size_t ZSTDv05_decompress_usingPreparedDCtx(
+                                             ZSTDv05_DCtx* dctx, const ZSTDv05_DCtx* preparedDCtx,
+                                             void* dst, size_t dstCapacity,
+                                       const void* src, size_t srcSize);
+
+
+/* **************************************
+*  Streaming functions (direct mode)
+****************************************/
+size_t ZSTDv05_decompressBegin(ZSTDv05_DCtx* dctx);
+
+/*
+  Streaming decompression, direct mode (bufferless)
+
+  A ZSTDv05_DCtx object is required to track streaming operations.
+  Use ZSTDv05_createDCtx() / ZSTDv05_freeDCtx() to manage it.
+  A ZSTDv05_DCtx object can be re-used multiple times.
+
+  First typical operation is to retrieve frame parameters, using ZSTDv05_getFrameParams().
+  This operation is independent, and just needs enough input data to properly decode the frame header.
+  Objective is to retrieve *params.windowlog, to know minimum amount of memory required during decoding.
+  Result : 0 when successful, it means the ZSTDv05_parameters structure has been filled.
+           >0 : means there is not enough data into src. Provides the expected size to successfully decode header.
+           errorCode, which can be tested using ZSTDv05_isError()
+
+  Start decompression, with ZSTDv05_decompressBegin() or ZSTDv05_decompressBegin_usingDict()
+  Alternatively, you can copy a prepared context, using ZSTDv05_copyDCtx()
+
+  Then use ZSTDv05_nextSrcSizeToDecompress() and ZSTDv05_decompressContinue() alternatively.
+  ZSTDv05_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTDv05_decompressContinue().
+  ZSTDv05_decompressContinue() requires this exact amount of bytes, or it will fail.
+  ZSTDv05_decompressContinue() needs previous data blocks during decompression, up to (1 << windowlog).
+  They should preferably be located contiguously, prior to current block. Alternatively, a round buffer is also possible.
+
+  @result of ZSTDv05_decompressContinue() is the number of bytes regenerated within 'dst'.
+  It can be zero, which is not an error; it just means ZSTDv05_decompressContinue() has decoded some header.
+
+  A frame is fully decoded when ZSTDv05_nextSrcSizeToDecompress() returns zero.
+  Context can then be reset to start a new decompression.
+*/
+
+
+/* **************************************
+*  Block functions
+****************************************/
+/*! Block functions produce and decode raw zstd blocks, without frame metadata.
+    User will have to take in charge required information to regenerate data, such as block sizes.
+
+    A few rules to respect :
+    - Uncompressed block size must be <= 128 KB
+    - Compressing or decompressing requires a context structure
+      + Use ZSTDv05_createCCtx() and ZSTDv05_createDCtx()
+    - It is necessary to init context before starting
+      + compression : ZSTDv05_compressBegin()
+      + decompression : ZSTDv05_decompressBegin()
+      + variants _usingDict() are also allowed
+      + copyCCtx() and copyDCtx() work too
+    - When a block is considered not compressible enough, ZSTDv05_compressBlock() result will be zero.
+      In which case, nothing is produced into `dst`.
+      + User must test for such outcome and deal directly with uncompressed data
+      + ZSTDv05_decompressBlock() doesn't accept uncompressed data as input !!
+*/
+
+size_t ZSTDv05_decompressBlock(ZSTDv05_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* ZSTDv05_STATIC_H */
+
+
+/*
+    zstd_internal - common functions to include
+    Header File for include
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+*/
+#ifndef ZSTD_CCOMMON_H_MODULE
+#define ZSTD_CCOMMON_H_MODULE
+
+
+
+/*-*************************************
+*  Common macros
+***************************************/
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+
+
+/*-*************************************
+*  Common constants
+***************************************/
+#define ZSTDv05_DICT_MAGIC  0xEC30A435
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BLOCKSIZE (128 KB)                 /* define, for static allocation */
+
+static const size_t ZSTDv05_blockHeaderSize = 3;
+static const size_t ZSTDv05_frameHeaderSize_min = 5;
+#define ZSTDv05_frameHeaderSize_max 5         /* define, for static allocation */
+
+#define BITv057 128
+#define BITv056  64
+#define BITv055  32
+#define BITv054  16
+#define BITv051   2
+#define BITv050   1
+
+#define IS_HUFv05 0
+#define IS_PCH 1
+#define IS_RAW 2
+#define IS_RLE 3
+
+#define MINMATCH 4
+#define REPCODE_STARTVALUE 1
+
+#define Litbits  8
+#define MLbits   7
+#define LLbits   6
+#define Offbits  5
+#define MaxLit ((1<<Litbits) - 1)
+#define MaxML  ((1<<MLbits) - 1)
+#define MaxLL  ((1<<LLbits) - 1)
+#define MaxOff ((1<<Offbits)- 1)
+#define MLFSEv05Log   10
+#define LLFSEv05Log   10
+#define OffFSEv05Log   9
+#define MaxSeq MAX(MaxLL, MaxML)
+
+#define FSEv05_ENCODING_RAW     0
+#define FSEv05_ENCODING_RLE     1
+#define FSEv05_ENCODING_STATIC  2
+#define FSEv05_ENCODING_DYNAMIC 3
+
+
+#define HufLog 12
+
+#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+
+#define WILDCOPY_OVERLENGTH 8
+
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+
+/*-*******************************************
+*  Shared functions to include for inlining
+*********************************************/
+static void ZSTDv05_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+
+#define COPY8(d,s) { ZSTDv05_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTDv05_wildcopy() :
+*   custom version of memcpy(), can copy up to 7 bytes too many (8 bytes if length==0) */
+MEM_STATIC void ZSTDv05_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+
+/*-*******************************************
+*  Private interfaces
+*********************************************/
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* offCode;
+    BYTE* litStart;
+    BYTE* lit;
+    BYTE* litLengthStart;
+    BYTE* litLength;
+    BYTE* matchLengthStart;
+    BYTE* matchLength;
+    BYTE* dumpsStart;
+    BYTE* dumps;
+    /* opt */
+    U32* matchLengthFreq;
+    U32* litLengthFreq;
+    U32* litFreq;
+    U32* offCodeFreq;
+    U32  matchLengthSum;
+    U32  litLengthSum;
+    U32  litSum;
+    U32  offCodeSum;
+} seqStore_t;
+
+
+
+#endif   /* ZSTDv05_CCOMMON_H_MODULE */
+/* ******************************************************************
+   FSEv05 : Finite State Entropy coder
+   header file
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef FSEv05_H
+#define FSEv05_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* *****************************************
+*  Includes
+******************************************/
+#include <stddef.h>    /* size_t, ptrdiff_t */
+
+
+/*-****************************************
+*  FSEv05 simple functions
+******************************************/
+size_t FSEv05_decompress(void* dst,  size_t maxDstSize,
+                const void* cSrc, size_t cSrcSize);
+/*!
+FSEv05_decompress():
+    Decompress FSEv05 data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'maxDstSize'.
+    return : size of regenerated data (<= maxDstSize)
+             or an error code, which can be tested using FSEv05_isError()
+
+    ** Important ** : FSEv05_decompress() doesn't decompress non-compressible nor RLE data !!!
+    Why ? : making this distinction requires a header.
+    Header management is intentionally delegated to the user layer, which can better manage special cases.
+*/
+
+
+/* *****************************************
+*  Tool functions
+******************************************/
+/* Error Management */
+unsigned    FSEv05_isError(size_t code);        /* tells if a return value is an error code */
+const char* FSEv05_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
+
+
+
+
+/* *****************************************
+*  FSEv05 detailed API
+******************************************/
+/* *** DECOMPRESSION *** */
+
+/*!
+FSEv05_readNCount():
+   Read compactly saved 'normalizedCounter' from 'rBuffer'.
+   return : size read from 'rBuffer'
+            or an errorCode, which can be tested using FSEv05_isError()
+            maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+size_t FSEv05_readNCount (short* normalizedCounter, unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, const void* rBuffer, size_t rBuffSize);
+
+/*!
+Constructor and Destructor of type FSEv05_DTable
+    Note that its size depends on 'tableLog' */
+typedef unsigned FSEv05_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+FSEv05_DTable* FSEv05_createDTable(unsigned tableLog);
+void        FSEv05_freeDTable(FSEv05_DTable* dt);
+
+/*!
+FSEv05_buildDTable():
+   Builds 'dt', which must be already allocated, using FSEv05_createDTable()
+   @return : 0,
+             or an errorCode, which can be tested using FSEv05_isError() */
+size_t FSEv05_buildDTable (FSEv05_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*!
+FSEv05_decompress_usingDTable():
+   Decompress compressed source @cSrc of size @cSrcSize using `dt`
+   into `dst` which must be already allocated.
+   @return : size of regenerated data (necessarily <= @dstCapacity)
+             or an errorCode, which can be tested using FSEv05_isError() */
+size_t FSEv05_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSEv05_DTable* dt);
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSEv05_H */
+/* ******************************************************************
+   bitstream
+   Part of FSEv05 library
+   header file (to include)
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef BITv05STREAM_H_MODULE
+#define BITv05STREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which highly benefit from being inlined.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+
+
+/*-********************************************
+*  bitStream decoding API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BITv05_DStream_t;
+
+typedef enum { BITv05_DStream_unfinished = 0,
+               BITv05_DStream_endOfBuffer = 1,
+               BITv05_DStream_completed = 2,
+               BITv05_DStream_overflow = 3 } BITv05_DStream_status;  /* result of BITv05_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BITv05_initDStream(BITv05_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BITv05_readBits(BITv05_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BITv05_DStream_status BITv05_reloadDStream(BITv05_DStream_t* bitD);
+MEM_STATIC unsigned BITv05_endOfDStream(const BITv05_DStream_t* bitD);
+
+
+/*-****************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BITv05_readBitsFast(BITv05_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/*-**************************************************************
+*  Helper functions
+****************************************************************/
+MEM_STATIC unsigned BITv05_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+
+/*-********************************************************
+* bitStream decoding
+**********************************************************/
+/*!BITv05_initDStream
+*  Initialize a BITv05_DStream_t.
+*  @bitD : a pointer to an already allocated BITv05_DStream_t structure
+*  @srcBuffer must point at the beginning of a bitStream
+*  @srcSize must be the exact size of the bitStream
+*  @result : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BITv05_initDStream(BITv05_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(size_t)) {  /* normal case */
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(size_t);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BITv05_highbit32(contain32);
+    } else {
+        U32 contain32;
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[6]) << (sizeof(size_t)*8 - 16);/* fall-through */
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[5]) << (sizeof(size_t)*8 - 24);/* fall-through */
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[4]) << (sizeof(size_t)*8 - 32);/* fall-through */
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[3]) << 24; /* fall-through */
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[2]) << 16; /* fall-through */
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(bitD->start))[1]) <<  8; /* fall-through */
+            default: break;
+        }
+        contain32 = ((const BYTE*)srcBuffer)[srcSize-1];
+        if (contain32 == 0) return ERROR(GENERIC);   /* endMark not present */
+        bitD->bitsConsumed = 8 - BITv05_highbit32(contain32);
+        bitD->bitsConsumed += (U32)(sizeof(size_t) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+MEM_STATIC size_t BITv05_lookBits(BITv05_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BITv05_lookBitsFast :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv05_lookBitsFast(BITv05_DStream_t* bitD, U32 nbBits)
+{
+    const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BITv05_skipBits(BITv05_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BITv05_readBits(BITv05_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BITv05_lookBits(bitD, nbBits);
+    BITv05_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*!BITv05_readBitsFast :
+*  unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv05_readBitsFast(BITv05_DStream_t* bitD, U32 nbBits)
+{
+    size_t value = BITv05_lookBitsFast(bitD, nbBits);
+    BITv05_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BITv05_DStream_status BITv05_reloadDStream(BITv05_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return BITv05_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer)) {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BITv05_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start) {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BITv05_DStream_endOfBuffer;
+        return BITv05_DStream_completed;
+    }
+    {
+        U32 nbBytes = bitD->bitsConsumed >> 3;
+        BITv05_DStream_status result = BITv05_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start) {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BITv05_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BITv05_endOfDStream
+*   @return Tells if DStream has reached its exact end
+*/
+MEM_STATIC unsigned BITv05_endOfDStream(const BITv05_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITv05STREAM_H_MODULE */
+/* ******************************************************************
+   FSEv05 : Finite State Entropy coder
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef FSEv05_STATIC_H
+#define FSEv05_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/* *****************************************
+*  Static allocation
+*******************************************/
+/* It is possible to statically allocate FSEv05 CTable/DTable as a table of unsigned using below macros */
+#define FSEv05_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/* *****************************************
+*  FSEv05 advanced API
+*******************************************/
+size_t FSEv05_buildDTable_raw (FSEv05_DTable* dt, unsigned nbBits);
+/* build a fake FSEv05_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+size_t FSEv05_buildDTable_rle (FSEv05_DTable* dt, unsigned char symbolValue);
+/* build a fake FSEv05_DTable, designed to always generate the same symbolValue */
+
+
+
+/* *****************************************
+*  FSEv05 symbol decompression API
+*******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSEv05_DState_t;
+
+
+static void     FSEv05_initDState(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD, const FSEv05_DTable* dt);
+
+static unsigned char FSEv05_decodeSymbol(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD);
+
+static unsigned FSEv05_endOfDState(const FSEv05_DState_t* DStatePtr);
+
+
+
+/* *****************************************
+*  FSEv05 unsafe API
+*******************************************/
+static unsigned char FSEv05_decodeSymbolFast(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* *****************************************
+*  Implementation of inlined functions
+*******************************************/
+/* decompression */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSEv05_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSEv05_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSEv05_initDState(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD, const FSEv05_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv05_DTableHeader* const DTableH = (const FSEv05_DTableHeader*)ptr;
+    DStatePtr->state = BITv05_readBits(bitD, DTableH->tableLog);
+    BITv05_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSEv05_peakSymbol(FSEv05_DState_t* DStatePtr)
+{
+    const FSEv05_decode_t DInfo = ((const FSEv05_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    return DInfo.symbol;
+}
+
+MEM_STATIC BYTE FSEv05_decodeSymbol(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD)
+{
+    const FSEv05_decode_t DInfo = ((const FSEv05_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32  nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BITv05_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC BYTE FSEv05_decodeSymbolFast(FSEv05_DState_t* DStatePtr, BITv05_DStream_t* bitD)
+{
+    const FSEv05_decode_t DInfo = ((const FSEv05_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    const U32 nbBits = DInfo.nbBits;
+    BYTE symbol = DInfo.symbol;
+    size_t lowBits = BITv05_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+MEM_STATIC unsigned FSEv05_endOfDState(const FSEv05_DState_t* DStatePtr)
+{
+    return DStatePtr->state == 0;
+}
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSEv05_STATIC_H */
+/* ******************************************************************
+   FSEv05 : Finite State Entropy coder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSEv05 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+#ifndef FSEv05_COMMONDEFS_ONLY
+
+/* **************************************************************
+*  Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSEv05_MAX_MEMORY_USAGE 14
+#define FSEv05_DEFAULT_MEMORY_USAGE 13
+
+/*!FSEv05_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSEv05_MAX_SYMBOL_VALUE 255
+
+
+/* **************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSEv05_FUNCTION_TYPE BYTE
+#define FSEv05_FUNCTION_EXTENSION
+#define FSEv05_DECODE_TYPE FSEv05_decode_t
+
+
+#endif   /* !FSEv05_COMMONDEFS_ONLY */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+
+
+/* ***************************************************************
+*  Constants
+*****************************************************************/
+#define FSEv05_MAX_TABLELOG  (FSEv05_MAX_MEMORY_USAGE-2)
+#define FSEv05_MAX_TABLESIZE (1U<<FSEv05_MAX_TABLELOG)
+#define FSEv05_MAXTABLESIZE_MASK (FSEv05_MAX_TABLESIZE-1)
+#define FSEv05_DEFAULT_TABLELOG (FSEv05_DEFAULT_MEMORY_USAGE-2)
+#define FSEv05_MIN_TABLELOG 5
+
+#define FSEv05_TABLELOG_ABSOLUTE_MAX 15
+#if FSEv05_MAX_TABLELOG > FSEv05_TABLELOG_ABSOLUTE_MAX
+#error "FSEv05_MAX_TABLELOG > FSEv05_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSEv05_STATIC_ASSERT(c) { enum { FSEv05_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSEv05_DTABLE_SIZE_U32(FSEv05_MAX_TABLELOG)];
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSEv05_FUNCTION_EXTENSION
+#  error "FSEv05_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSEv05_FUNCTION_TYPE
+#  error "FSEv05_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSEv05_CAT(X,Y) X##Y
+#define FSEv05_FUNCTION_NAME(X,Y) FSEv05_CAT(X,Y)
+#define FSEv05_TYPE_NAME(X,Y) FSEv05_CAT(X,Y)
+
+
+/* Function templates */
+static U32 FSEv05_tableStep(U32 tableSize) { return (tableSize>>1) + (tableSize>>3) + 3; }
+
+
+
+FSEv05_DTable* FSEv05_createDTable (unsigned tableLog)
+{
+    if (tableLog > FSEv05_TABLELOG_ABSOLUTE_MAX) tableLog = FSEv05_TABLELOG_ABSOLUTE_MAX;
+    return (FSEv05_DTable*)malloc( FSEv05_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
+}
+
+void FSEv05_freeDTable (FSEv05_DTable* dt)
+{
+    free(dt);
+}
+
+size_t FSEv05_buildDTable(FSEv05_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    FSEv05_DTableHeader DTableH;
+    void* const tdPtr = dt+1;   /* because dt is unsigned, 32-bits aligned on 32-bits */
+    FSEv05_DECODE_TYPE* const tableDecode = (FSEv05_DECODE_TYPE*) (tdPtr);
+    const U32 tableSize = 1 << tableLog;
+    const U32 tableMask = tableSize-1;
+    const U32 step = FSEv05_tableStep(tableSize);
+    U16 symbolNext[FSEv05_MAX_SYMBOL_VALUE+1];
+    U32 position = 0;
+    U32 highThreshold = tableSize-1;
+    const S16 largeLimit= (S16)(1 << (tableLog-1));
+    U32 noLarge = 1;
+    U32 s;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSEv05_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSEv05_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    DTableH.tableLog = (U16)tableLog;
+    for (s=0; s<=maxSymbolValue; s++) {
+        if (normalizedCounter[s]==-1) {
+            tableDecode[highThreshold--].symbol = (FSEv05_FUNCTION_TYPE)s;
+            symbolNext[s] = 1;
+        } else {
+            if (normalizedCounter[s] >= largeLimit) noLarge=0;
+            symbolNext[s] = normalizedCounter[s];
+    }   }
+
+    /* Spread symbols */
+    for (s=0; s<=maxSymbolValue; s++) {
+        int i;
+        for (i=0; i<normalizedCounter[s]; i++) {
+            tableDecode[position].symbol = (FSEv05_FUNCTION_TYPE)s;
+            position = (position + step) & tableMask;
+            while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+    }   }
+
+    if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+
+    /* Build Decoding table */
+    {
+        U32 i;
+        for (i=0; i<tableSize; i++) {
+            FSEv05_FUNCTION_TYPE symbol = (FSEv05_FUNCTION_TYPE)(tableDecode[i].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[i].nbBits = (BYTE) (tableLog - BITv05_highbit32 ((U32)nextState) );
+            tableDecode[i].newState = (U16) ( (nextState << tableDecode[i].nbBits) - tableSize);
+    }   }
+
+    DTableH.fastMode = (U16)noLarge;
+    memcpy(dt, &DTableH, sizeof(DTableH));
+    return 0;
+}
+
+
+#ifndef FSEv05_COMMONDEFS_ONLY
+/*-****************************************
+*  FSEv05 helper functions
+******************************************/
+unsigned FSEv05_isError(size_t code) { return ERR_isError(code); }
+
+const char* FSEv05_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/*-**************************************************************
+*  FSEv05 NCount encoding-decoding
+****************************************************************/
+static short FSEv05_abs(short a) { return a<0 ? -a : a; }
+
+
+size_t FSEv05_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSEv05_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSEv05_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr)) {
+        if (previous0) {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF) {
+                n0+=24;
+                if (ip < iend-5) {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                } else {
+                    bitStream >>= 16;
+                    bitCount+=16;
+            }   }
+            while ((bitStream & 3) == 3) {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {
+            const short max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max) {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            } else {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSEv05_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold) {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+            } else {
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+    }   }
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+
+
+
+/*-*******************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+size_t FSEv05_buildDTable_rle (FSEv05_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSEv05_DTableHeader* const DTableH = (FSEv05_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv05_decode_t* const cell = (FSEv05_decode_t*)dPtr;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+size_t FSEv05_buildDTable_raw (FSEv05_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSEv05_DTableHeader* const DTableH = (FSEv05_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv05_decode_t* const dinfo = (FSEv05_decode_t*)dPtr;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSymbolValue = tableMask;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<=maxSymbolValue; s++) {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSEv05_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSEv05_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BITv05_DStream_t bitD;
+    FSEv05_DState_t state1;
+    FSEv05_DState_t state2;
+    size_t errorCode;
+
+    /* Init */
+    errorCode = BITv05_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+    if (FSEv05_isError(errorCode)) return errorCode;
+
+    FSEv05_initDState(&state1, &bitD, dt);
+    FSEv05_initDState(&state2, &bitD, dt);
+
+#define FSEv05_GETSYMBOL(statePtr) fast ? FSEv05_decodeSymbolFast(statePtr, &bitD) : FSEv05_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BITv05_reloadDStream(&bitD)==BITv05_DStream_unfinished) && (op<olimit) ; op+=4) {
+        op[0] = FSEv05_GETSYMBOL(&state1);
+
+        if (FSEv05_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv05_reloadDStream(&bitD);
+
+        op[1] = FSEv05_GETSYMBOL(&state2);
+
+        if (FSEv05_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BITv05_reloadDStream(&bitD) > BITv05_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSEv05_GETSYMBOL(&state1);
+
+        if (FSEv05_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv05_reloadDStream(&bitD);
+
+        op[3] = FSEv05_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BITv05_reloadDStream(&bitD) >= FSEv05_DStream_partiallyFilled; Ends at exactly BITv05_DStream_completed */
+    while (1) {
+        if ( (BITv05_reloadDStream(&bitD)>BITv05_DStream_completed) || (op==omax) || (BITv05_endOfDStream(&bitD) && (fast || FSEv05_endOfDState(&state1))) )
+            break;
+
+        *op++ = FSEv05_GETSYMBOL(&state1);
+
+        if ( (BITv05_reloadDStream(&bitD)>BITv05_DStream_completed) || (op==omax) || (BITv05_endOfDStream(&bitD) && (fast || FSEv05_endOfDState(&state2))) )
+            break;
+
+        *op++ = FSEv05_GETSYMBOL(&state2);
+    }
+
+    /* end ? */
+    if (BITv05_endOfDStream(&bitD) && FSEv05_endOfDState(&state1) && FSEv05_endOfDState(&state2))
+        return op-ostart;
+
+    if (op==omax) return ERROR(dstSize_tooSmall);   /* dst buffer is full, but cSrc unfinished */
+
+    return ERROR(corruption_detected);
+}
+
+
+size_t FSEv05_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSEv05_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv05_DTableHeader* DTableH = (const FSEv05_DTableHeader*)ptr;
+    const U32 fastMode = DTableH->fastMode;
+
+    /* select fast mode (static) */
+    if (fastMode) return FSEv05_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSEv05_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+size_t FSEv05_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSEv05_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSEv05_MAX_SYMBOL_VALUE;
+    size_t errorCode;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSEv05 decoding mode */
+    errorCode = FSEv05_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+    if (FSEv05_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    errorCode = FSEv05_buildDTable (dt, counting, maxSymbolValue, tableLog);
+    if (FSEv05_isError(errorCode)) return errorCode;
+
+    /* always return, even if it is an error code */
+    return FSEv05_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);
+}
+
+
+
+#endif   /* FSEv05_COMMONDEFS_ONLY */
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   header file
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef HUFF0_H
+#define HUFF0_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/* ****************************************
+*  Huff0 simple functions
+******************************************/
+size_t HUFv05_decompress(void* dst,  size_t dstSize,
+                const void* cSrc, size_t cSrcSize);
+/*!
+HUFv05_decompress():
+    Decompress Huff0 data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstSize'.
+    @dstSize : must be the **exact** size of original (uncompressed) data.
+    Note : in contrast with FSEv05, HUFv05_decompress can regenerate
+           RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
+           because it knows size to regenerate.
+    @return : size of regenerated data (== dstSize)
+              or an error code, which can be tested using HUFv05_isError()
+*/
+
+
+/* ****************************************
+*  Tool functions
+******************************************/
+/* Error Management */
+unsigned    HUFv05_isError(size_t code);        /* tells if a return value is an error code */
+const char* HUFv05_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* HUF0_H */
+/* ******************************************************************
+   Huff0 : Huffman codec, part of New Generation Entropy library
+   header file, for static linking only
+   Copyright (C) 2013-2016, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef HUF0_STATIC_H
+#define HUF0_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/* ****************************************
+*  Static allocation
+******************************************/
+/* static allocation of Huff0's DTable */
+#define HUFv05_DTABLE_SIZE(maxTableLog)   (1 + (1<<maxTableLog))
+#define HUFv05_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        unsigned short DTable[HUFv05_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUFv05_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        unsigned int DTable[HUFv05_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUFv05_CREATE_STATIC_DTABLEX6(DTable, maxTableLog) \
+        unsigned int DTable[HUFv05_DTABLE_SIZE(maxTableLog) * 3 / 2] = { maxTableLog }
+
+
+/* ****************************************
+*  Advanced decompression functions
+******************************************/
+size_t HUFv05_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUFv05_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbols decoder */
+
+
+/* ****************************************
+*  Huff0 detailed API
+******************************************/
+/*!
+HUFv05_decompress() does the following:
+1. select the decompression algorithm (X2, X4, X6) based on pre-computed heuristics
+2. build Huffman table from save, using HUFv05_readDTableXn()
+3. decode 1 or 4 segments in parallel using HUFv05_decompressSXn_usingDTable
+*/
+size_t HUFv05_readDTableX2 (unsigned short* DTable, const void* src, size_t srcSize);
+size_t HUFv05_readDTableX4 (unsigned* DTable, const void* src, size_t srcSize);
+
+size_t HUFv05_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned short* DTable);
+size_t HUFv05_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned* DTable);
+
+
+/* single stream variants */
+
+size_t HUFv05_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUFv05_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+
+size_t HUFv05_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned short* DTable);
+size_t HUFv05_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned* DTable);
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* HUF0_STATIC_H */
+/* ******************************************************************
+   Huff0 : Huffman coder, part of New Generation Entropy library
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSEv05+Huff0 source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include <stdlib.h>     /* malloc, free, qsort */
+#include <string.h>     /* memcpy, memset */
+#include <stdio.h>      /* printf (debug) */
+
+
+/* **************************************************************
+*  Constants
+****************************************************************/
+#define HUFv05_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUFv05_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUFv05_MAX_TABLELOG  12           /* max configured tableLog (for static allocation); can be modified up to HUFv05_ABSOLUTEMAX_TABLELOG */
+#define HUFv05_DEFAULT_TABLELOG  HUFv05_MAX_TABLELOG   /* tableLog by default, when not specified */
+#define HUFv05_MAX_SYMBOL_VALUE 255
+#if (HUFv05_MAX_TABLELOG > HUFv05_ABSOLUTEMAX_TABLELOG)
+#  error "HUFv05_MAX_TABLELOG is too large !"
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+unsigned HUFv05_isError(size_t code) { return ERR_isError(code); }
+const char* HUFv05_getErrorName(size_t code) { return ERR_getErrorName(code); }
+#define HUFv05_STATIC_ASSERT(c) { enum { HUFv05_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* *******************************************************
+*  Huff0 : Huffman block decompression
+*********************************************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUFv05_DEltX2;   /* single-symbol decoding */
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUFv05_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+/*! HUFv05_readStats
+    Read compact Huffman tree, saved by HUFv05_writeCTable
+    @huffWeight : destination buffer
+    @return : size read from `src`
+*/
+static size_t HUFv05_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                            U32* nbSymbolsPtr, U32* tableLogPtr,
+                            const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    U32 tableLog;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+    U32 n;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  { /* special header */
+        if (iSize >= (242)) {  /* RLE */
+            static int l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else {   /* Incompressible */
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            for (n=0; n<oSize; n+=2) {
+                huffWeight[n]   = ip[n/2] >> 4;
+                huffWeight[n+1] = ip[n/2] & 15;
+    }   }   }
+    else  {   /* header compressed with FSEv05 (normal case) */
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSEv05_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSEv05_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUFv05_ABSOLUTEMAX_TABLELOG + 1) * sizeof(U32));
+    weightTotal = 0;
+    for (n=0; n<oSize; n++) {
+        if (huffWeight[n] >= HUFv05_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+        rankStats[huffWeight[n]]++;
+        weightTotal += (1 << huffWeight[n]) >> 1;
+    }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    tableLog = BITv05_highbit32(weightTotal) + 1;
+    if (tableLog > HUFv05_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+    {   /* determine last weight */
+        U32 total = 1 << tableLog;
+        U32 rest = total - weightTotal;
+        U32 verif = 1 << BITv05_highbit32(rest);
+        U32 lastWeight = BITv05_highbit32(rest) + 1;
+        if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+        huffWeight[oSize] = (BYTE)lastWeight;
+        rankStats[lastWeight]++;
+    }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    *tableLogPtr = tableLog;
+    return iSize+1;
+}
+
+
+/*-***************************/
+/*  single-symbol decoding   */
+/*-***************************/
+
+size_t HUFv05_readDTableX2 (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUFv05_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUFv05_ABSOLUTEMAX_TABLELOG + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    size_t iSize;
+    U32 nbSymbols = 0;
+    U32 n;
+    U32 nextRankStart;
+    void* const dtPtr = DTable + 1;
+    HUFv05_DEltX2* const dt = (HUFv05_DEltX2*)dtPtr;
+
+    HUFv05_STATIC_ASSERT(sizeof(HUFv05_DEltX2) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv05_readStats(huffWeight, HUFv05_MAX_SYMBOL_VALUE + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv05_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > DTable[0]) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+    DTable[0] = (U16)tableLog;   /* maybe should separate sizeof allocated DTable, from used size of DTable, in case of re-use */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<=tableLog; n++) {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<nbSymbols; n++) {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUFv05_DEltX2 D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize;
+}
+
+static BYTE HUFv05_decodeSymbolX2(BITv05_DStream_t* Dstream, const HUFv05_DEltX2* dt, const U32 dtLog)
+{
+        const size_t val = BITv05_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+        const BYTE c = dt[val].byte;
+        BITv05_skipBits(Dstream, dt[val].nbBits);
+        return c;
+}
+
+#define HUFv05_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUFv05_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUFv05_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv05_MAX_TABLELOG<=12)) \
+        HUFv05_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUFv05_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUFv05_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUFv05_decodeStreamX2(BYTE* p, BITv05_DStream_t* const bitDPtr, BYTE* const pEnd, const HUFv05_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BITv05_reloadDStream(bitDPtr) == BITv05_DStream_unfinished) && (p <= pEnd-4)) {
+        HUFv05_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BITv05_reloadDStream(bitDPtr) == BITv05_DStream_unfinished) && (p < pEnd))
+        HUFv05_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUFv05_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+size_t HUFv05_decompress1X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + dstSize;
+    const U32 dtLog = DTable[0];
+    const void* dtPtr = DTable;
+    const HUFv05_DEltX2* const dt = ((const HUFv05_DEltX2*)dtPtr)+1;
+    BITv05_DStream_t bitD;
+
+    if (dstSize <= cSrcSize) return ERROR(dstSize_tooSmall);
+    { size_t const errorCode = BITv05_initDStream(&bitD, cSrc, cSrcSize);
+      if (HUFv05_isError(errorCode)) return errorCode; }
+
+    HUFv05_decodeStreamX2(op, &bitD, oend, dt, dtLog);
+
+    /* check */
+    if (!BITv05_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    return dstSize;
+}
+
+size_t HUFv05_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv05_CREATE_STATIC_DTABLEX2(DTable, HUFv05_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUFv05_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUFv05_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUFv05_decompress1X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+size_t HUFv05_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    const BYTE* const istart = (const BYTE*) cSrc;
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* const oend = ostart + dstSize;
+    const void* const dtPtr = DTable;
+    const HUFv05_DEltX2* const dt = ((const HUFv05_DEltX2*)dtPtr) +1;
+    const U32 dtLog = DTable[0];
+    size_t errorCode;
+
+    /* Init */
+    BITv05_DStream_t bitD1;
+    BITv05_DStream_t bitD2;
+    BITv05_DStream_t bitD3;
+    BITv05_DStream_t bitD4;
+    const size_t length1 = MEM_readLE16(istart);
+    const size_t length2 = MEM_readLE16(istart+2);
+    const size_t length3 = MEM_readLE16(istart+4);
+    size_t length4;
+    const BYTE* const istart1 = istart + 6;  /* jumpTable */
+    const BYTE* const istart2 = istart1 + length1;
+    const BYTE* const istart3 = istart2 + length2;
+    const BYTE* const istart4 = istart3 + length3;
+    const size_t segmentSize = (dstSize+3) / 4;
+    BYTE* const opStart2 = ostart + segmentSize;
+    BYTE* const opStart3 = opStart2 + segmentSize;
+    BYTE* const opStart4 = opStart3 + segmentSize;
+    BYTE* op1 = ostart;
+    BYTE* op2 = opStart2;
+    BYTE* op3 = opStart3;
+    BYTE* op4 = opStart4;
+    U32 endSignal;
+
+    /* Check */
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    length4 = cSrcSize - (length1 + length2 + length3 + 6);
+    if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+    errorCode = BITv05_initDStream(&bitD1, istart1, length1);
+    if (HUFv05_isError(errorCode)) return errorCode;
+    errorCode = BITv05_initDStream(&bitD2, istart2, length2);
+    if (HUFv05_isError(errorCode)) return errorCode;
+    errorCode = BITv05_initDStream(&bitD3, istart3, length3);
+    if (HUFv05_isError(errorCode)) return errorCode;
+    errorCode = BITv05_initDStream(&bitD4, istart4, length4);
+    if (HUFv05_isError(errorCode)) return errorCode;
+
+    /* 16-32 symbols per loop (4-8 symbols per stream) */
+    endSignal = BITv05_reloadDStream(&bitD1) | BITv05_reloadDStream(&bitD2) | BITv05_reloadDStream(&bitD3) | BITv05_reloadDStream(&bitD4);
+    for ( ; (endSignal==BITv05_DStream_unfinished) && (op4<(oend-7)) ; ) {
+        HUFv05_DECODE_SYMBOLX2_2(op1, &bitD1);
+        HUFv05_DECODE_SYMBOLX2_2(op2, &bitD2);
+        HUFv05_DECODE_SYMBOLX2_2(op3, &bitD3);
+        HUFv05_DECODE_SYMBOLX2_2(op4, &bitD4);
+        HUFv05_DECODE_SYMBOLX2_1(op1, &bitD1);
+        HUFv05_DECODE_SYMBOLX2_1(op2, &bitD2);
+        HUFv05_DECODE_SYMBOLX2_1(op3, &bitD3);
+        HUFv05_DECODE_SYMBOLX2_1(op4, &bitD4);
+        HUFv05_DECODE_SYMBOLX2_2(op1, &bitD1);
+        HUFv05_DECODE_SYMBOLX2_2(op2, &bitD2);
+        HUFv05_DECODE_SYMBOLX2_2(op3, &bitD3);
+        HUFv05_DECODE_SYMBOLX2_2(op4, &bitD4);
+        HUFv05_DECODE_SYMBOLX2_0(op1, &bitD1);
+        HUFv05_DECODE_SYMBOLX2_0(op2, &bitD2);
+        HUFv05_DECODE_SYMBOLX2_0(op3, &bitD3);
+        HUFv05_DECODE_SYMBOLX2_0(op4, &bitD4);
+        endSignal = BITv05_reloadDStream(&bitD1) | BITv05_reloadDStream(&bitD2) | BITv05_reloadDStream(&bitD3) | BITv05_reloadDStream(&bitD4);
+    }
+
+    /* check corruption */
+    if (op1 > opStart2) return ERROR(corruption_detected);
+    if (op2 > opStart3) return ERROR(corruption_detected);
+    if (op3 > opStart4) return ERROR(corruption_detected);
+    /* note : op4 supposed already verified within main loop */
+
+    /* finish bitStreams one by one */
+    HUFv05_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+    HUFv05_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+    HUFv05_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+    HUFv05_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+    /* check */
+    endSignal = BITv05_endOfDStream(&bitD1) & BITv05_endOfDStream(&bitD2) & BITv05_endOfDStream(&bitD3) & BITv05_endOfDStream(&bitD4);
+    if (!endSignal) return ERROR(corruption_detected);
+
+    /* decoded size */
+    return dstSize;
+}
+
+
+size_t HUFv05_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv05_CREATE_STATIC_DTABLEX2(DTable, HUFv05_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+    size_t errorCode;
+
+    errorCode = HUFv05_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUFv05_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUFv05_decompress4X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/* *************************/
+/* double-symbols decoding */
+/* *************************/
+
+static void HUFv05_fillDTableX4Level2(HUFv05_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUFv05_DEltX4 DElt;
+    U32 rankVal[HUFv05_ABSOLUTEMAX_TABLELOG + 1];
+    U32 s;
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1) {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++) {   /* note : sortedSymbols already skipped */
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }
+}
+
+typedef U32 rankVal_t[HUFv05_ABSOLUTEMAX_TABLELOG][HUFv05_ABSOLUTEMAX_TABLELOG + 1];
+
+static void HUFv05_fillDTableX4(HUFv05_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUFv05_ABSOLUTEMAX_TABLELOG + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++) {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits) {   /* enough room for a second symbol */
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUFv05_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        } else {
+            U32 i;
+            const U32 end = start + length;
+            HUFv05_DEltX4 DElt;
+
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits   = (BYTE)(nbBits);
+            DElt.length   = 1;
+            for (i = start; i < end; i++)
+                DTable[i] = DElt;
+        }
+        rankVal[weight] += length;
+    }
+}
+
+size_t HUFv05_readDTableX4 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUFv05_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUFv05_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUFv05_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUFv05_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    const U32 memLog = DTable[0];
+    size_t iSize;
+    void* dtPtr = DTable;
+    HUFv05_DEltX4* const dt = ((HUFv05_DEltX4*)dtPtr) + 1;
+
+    HUFv05_STATIC_ASSERT(sizeof(HUFv05_DEltX4) == sizeof(U32));   /* if compilation fails here, assertion is false */
+    if (memLog > HUFv05_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv05_readStats(weightList, HUFv05_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv05_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+
+    /* Get start index of each weight */
+    {
+        U32 w, nextRankStart = 0;
+        for (w=1; w<=maxW; w++) {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {
+        U32 s;
+        for (s=0; s<nbSymbols; s++) {
+            U32 w = weightList[s];
+            U32 r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {
+        const U32 minBits = tableLog+1 - maxW;
+        U32 nextRankVal = 0;
+        U32 w, consumed;
+        const int rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+        U32* rankVal0 = rankVal[0];
+        for (w=1; w<=maxW; w++) {
+            U32 current = nextRankVal;
+            nextRankVal += rankStats[w] << (w+rescale);
+            rankVal0[w] = current;
+        }
+        for (consumed = minBits; consumed <= memLog - minBits; consumed++) {
+            U32* rankValPtr = rankVal[consumed];
+            for (w = 1; w <= maxW; w++) {
+                rankValPtr[w] = rankVal0[w] >> consumed;
+    }   }   }
+
+    HUFv05_fillDTableX4(dt, memLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    return iSize;
+}
+
+
+static U32 HUFv05_decodeSymbolX4(void* op, BITv05_DStream_t* DStream, const HUFv05_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv05_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BITv05_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUFv05_decodeLastSymbolX4(void* op, BITv05_DStream_t* DStream, const HUFv05_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv05_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BITv05_skipBits(DStream, dt[val].nbBits);
+    else {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
+            BITv05_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+    }   }
+    return 1;
+}
+
+
+#define HUFv05_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUFv05_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv05_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv05_MAX_TABLELOG<=12)) \
+        ptr += HUFv05_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv05_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUFv05_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUFv05_decodeStreamX4(BYTE* p, BITv05_DStream_t* bitDPtr, BYTE* const pEnd, const HUFv05_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BITv05_reloadDStream(bitDPtr) == BITv05_DStream_unfinished) && (p < pEnd-7)) {
+        HUFv05_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv05_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BITv05_reloadDStream(bitDPtr) == BITv05_DStream_unfinished) && (p <= pEnd-2))
+        HUFv05_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUFv05_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUFv05_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+
+size_t HUFv05_decompress1X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    const BYTE* const istart = (const BYTE*) cSrc;
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* const oend = ostart + dstSize;
+
+    const U32 dtLog = DTable[0];
+    const void* const dtPtr = DTable;
+    const HUFv05_DEltX4* const dt = ((const HUFv05_DEltX4*)dtPtr) +1;
+    size_t errorCode;
+
+    /* Init */
+    BITv05_DStream_t bitD;
+    errorCode = BITv05_initDStream(&bitD, istart, cSrcSize);
+    if (HUFv05_isError(errorCode)) return errorCode;
+
+    /* finish bitStreams one by one */
+    HUFv05_decodeStreamX4(ostart, &bitD, oend,     dt, dtLog);
+
+    /* check */
+    if (!BITv05_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    /* decoded size */
+    return dstSize;
+}
+
+size_t HUFv05_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv05_CREATE_STATIC_DTABLEX4(DTable, HUFv05_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUFv05_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUFv05_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUFv05_decompress1X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+size_t HUFv05_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {
+        const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable;
+        const HUFv05_DEltX4* const dt = ((const HUFv05_DEltX4*)dtPtr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BITv05_DStream_t bitD1;
+        BITv05_DStream_t bitD2;
+        BITv05_DStream_t bitD3;
+        BITv05_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BITv05_initDStream(&bitD1, istart1, length1);
+        if (HUFv05_isError(errorCode)) return errorCode;
+        errorCode = BITv05_initDStream(&bitD2, istart2, length2);
+        if (HUFv05_isError(errorCode)) return errorCode;
+        errorCode = BITv05_initDStream(&bitD3, istart3, length3);
+        if (HUFv05_isError(errorCode)) return errorCode;
+        errorCode = BITv05_initDStream(&bitD4, istart4, length4);
+        if (HUFv05_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BITv05_reloadDStream(&bitD1) | BITv05_reloadDStream(&bitD2) | BITv05_reloadDStream(&bitD3) | BITv05_reloadDStream(&bitD4);
+        for ( ; (endSignal==BITv05_DStream_unfinished) && (op4<(oend-7)) ; ) {
+            HUFv05_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv05_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv05_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv05_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv05_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUFv05_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUFv05_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUFv05_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUFv05_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv05_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv05_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv05_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv05_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUFv05_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUFv05_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUFv05_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BITv05_reloadDStream(&bitD1) | BITv05_reloadDStream(&bitD2) | BITv05_reloadDStream(&bitD3) | BITv05_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUFv05_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUFv05_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUFv05_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUFv05_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BITv05_endOfDStream(&bitD1) & BITv05_endOfDStream(&bitD2) & BITv05_endOfDStream(&bitD3) & BITv05_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+size_t HUFv05_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv05_CREATE_STATIC_DTABLEX4(DTable, HUFv05_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUFv05_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUFv05_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUFv05_decompress4X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/* ********************************/
+/* Generic decompression selector */
+/* ********************************/
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+size_t HUFv05_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[3] = { HUFv05_decompress4X2, HUFv05_decompress4X4, NULL };
+    /* estimate decompression time */
+    U32 Q;
+    const U32 D256 = (U32)(dstSize >> 8);
+    U32 Dtime[3];
+    U32 algoNb = 0;
+    int n;
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize >= dstSize) return ERROR(corruption_detected);   /* invalid, or not compressed, but not compressed already dealt with */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    /* decoder timing evaluation */
+    Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+    for (n=0; n<3; n++)
+        Dtime[n] = algoTime[Q][n].tableTime + (algoTime[Q][n].decode256Time * D256);
+
+    Dtime[1] += Dtime[1] >> 4; Dtime[2] += Dtime[2] >> 3; /* advantage to algorithms using less memory, for cache eviction */
+
+    if (Dtime[1] < Dtime[0]) algoNb = 1;
+
+    return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+
+    //return HUFv05_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUFv05_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+    //return HUFv05_decompress4X6(dst, dstSize, cSrc, cSrcSize);   /* multi-streams quad-symbols decoding */
+}
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * HEAPMODE :
+ * Select how default decompression function ZSTDv05_decompress() will allocate memory,
+ * in memory stack (0), or in memory heap (1, requires malloc())
+ */
+#ifndef ZSTDv05_HEAPMODE
+#  define ZSTDv05_HEAPMODE 1
+#endif
+
+
+/*-*******************************************************
+*  Dependencies
+*********************************************************/
+#include <stdlib.h>      /* calloc */
+#include <string.h>      /* memcpy, memmove */
+#include <stdio.h>       /* debug only : printf */
+
+
+/*-*******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+
+/*-*************************************
+*  Local types
+***************************************/
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+
+/* *******************************************************
+*  Memory operations
+**********************************************************/
+static void ZSTDv05_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+
+/* *************************************
+*  Error Management
+***************************************/
+/*! ZSTDv05_isError() :
+*   tells if a return value is an error code */
+unsigned ZSTDv05_isError(size_t code) { return ERR_isError(code); }
+
+
+/*! ZSTDv05_getErrorName() :
+*   provides error code string (useful for debugging) */
+const char* ZSTDv05_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/* *************************************************************
+*   Context management
+***************************************************************/
+typedef enum { ZSTDv05ds_getFrameHeaderSize, ZSTDv05ds_decodeFrameHeader,
+               ZSTDv05ds_decodeBlockHeader, ZSTDv05ds_decompressBlock } ZSTDv05_dStage;
+
+struct ZSTDv05_DCtx_s
+{
+    FSEv05_DTable LLTable[FSEv05_DTABLE_SIZE_U32(LLFSEv05Log)];
+    FSEv05_DTable OffTable[FSEv05_DTABLE_SIZE_U32(OffFSEv05Log)];
+    FSEv05_DTable MLTable[FSEv05_DTABLE_SIZE_U32(MLFSEv05Log)];
+    unsigned   hufTableX4[HUFv05_DTABLE_SIZE(HufLog)];
+    const void* previousDstEnd;
+    const void* base;
+    const void* vBase;
+    const void* dictEnd;
+    size_t expected;
+    size_t headerSize;
+    ZSTDv05_parameters params;
+    blockType_t bType;   /* used in ZSTDv05_decompressContinue(), to transfer blockType between header decoding and block decoding stages */
+    ZSTDv05_dStage stage;
+    U32 flagStaticTables;
+    const BYTE* litPtr;
+    size_t litSize;
+    BYTE litBuffer[BLOCKSIZE + WILDCOPY_OVERLENGTH];
+    BYTE headerBuffer[ZSTDv05_frameHeaderSize_max];
+};  /* typedef'd to ZSTDv05_DCtx within "zstd_static.h" */
+
+size_t ZSTDv05_sizeofDCtx (void) { return sizeof(ZSTDv05_DCtx); }
+
+size_t ZSTDv05_decompressBegin(ZSTDv05_DCtx* dctx)
+{
+    dctx->expected = ZSTDv05_frameHeaderSize_min;
+    dctx->stage = ZSTDv05ds_getFrameHeaderSize;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    dctx->vBase = NULL;
+    dctx->dictEnd = NULL;
+    dctx->hufTableX4[0] = HufLog;
+    dctx->flagStaticTables = 0;
+    return 0;
+}
+
+ZSTDv05_DCtx* ZSTDv05_createDCtx(void)
+{
+    ZSTDv05_DCtx* dctx = (ZSTDv05_DCtx*)malloc(sizeof(ZSTDv05_DCtx));
+    if (dctx==NULL) return NULL;
+    ZSTDv05_decompressBegin(dctx);
+    return dctx;
+}
+
+size_t ZSTDv05_freeDCtx(ZSTDv05_DCtx* dctx)
+{
+    free(dctx);
+    return 0;   /* reserved as a potential error code in the future */
+}
+
+void ZSTDv05_copyDCtx(ZSTDv05_DCtx* dstDCtx, const ZSTDv05_DCtx* srcDCtx)
+{
+    memcpy(dstDCtx, srcDCtx,
+           sizeof(ZSTDv05_DCtx) - (BLOCKSIZE+WILDCOPY_OVERLENGTH + ZSTDv05_frameHeaderSize_max));  /* no need to copy workspace */
+}
+
+
+/* *************************************************************
+*   Decompression section
+***************************************************************/
+
+/* Frame format description
+   Frame Header -  [ Block Header - Block ] - Frame End
+   1) Frame Header
+      - 4 bytes - Magic Number : ZSTDv05_MAGICNUMBER (defined within zstd_internal.h)
+      - 1 byte  - Window Descriptor
+   2) Block Header
+      - 3 bytes, starting with a 2-bits descriptor
+                 Uncompressed, Compressed, Frame End, unused
+   3) Block
+      See Block Format Description
+   4) Frame End
+      - 3 bytes, compatible with Block Header
+*/
+
+/* Block format description
+
+   Block = Literal Section - Sequences Section
+   Prerequisite : size of (compressed) block, maximum size of regenerated data
+
+   1) Literal Section
+
+   1.1) Header : 1-5 bytes
+        flags: 2 bits
+            00 compressed by Huff0
+            01 unused
+            10 is Raw (uncompressed)
+            11 is Rle
+            Note : using 01 => Huff0 with precomputed table ?
+            Note : delta map ? => compressed ?
+
+   1.1.1) Huff0-compressed literal block : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+   1.1.2) Raw (uncompressed) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RAW<<6) + (0<<4) + size
+               12 bits: (IS_RAW<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RAW<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.3) Rle (repeated single byte) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RLE<<6) + (0<<4) + size
+               12 bits: (IS_RLE<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RLE<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.4) Huff0-compressed literal block, using precomputed CTables : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+        1- CTable available (stored into workspace ?)
+        2- Small input (fast heuristic ? Full comparison ? depend on clevel ?)
+
+
+   1.2) Literal block content
+
+   1.2.1) Huff0 block, using sizes from header
+        See Huff0 format
+
+   1.2.2) Huff0 block, using prepared table
+
+   1.2.3) Raw content
+
+   1.2.4) single byte
+
+
+   2) Sequences section
+      TO DO
+*/
+
+
+/** ZSTDv05_decodeFrameHeader_Part1() :
+*   decode the 1st part of the Frame Header, which tells Frame Header size.
+*   srcSize must be == ZSTDv05_frameHeaderSize_min.
+*   @return : the full size of the Frame Header */
+static size_t ZSTDv05_decodeFrameHeader_Part1(ZSTDv05_DCtx* zc, const void* src, size_t srcSize)
+{
+    U32 magicNumber;
+    if (srcSize != ZSTDv05_frameHeaderSize_min)
+        return ERROR(srcSize_wrong);
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTDv05_MAGICNUMBER) return ERROR(prefix_unknown);
+    zc->headerSize = ZSTDv05_frameHeaderSize_min;
+    return zc->headerSize;
+}
+
+
+size_t ZSTDv05_getFrameParams(ZSTDv05_parameters* params, const void* src, size_t srcSize)
+{
+    U32 magicNumber;
+    if (srcSize < ZSTDv05_frameHeaderSize_min) return ZSTDv05_frameHeaderSize_max;
+    magicNumber = MEM_readLE32(src);
+    if (magicNumber != ZSTDv05_MAGICNUMBER) return ERROR(prefix_unknown);
+    memset(params, 0, sizeof(*params));
+    params->windowLog = (((const BYTE*)src)[4] & 15) + ZSTDv05_WINDOWLOG_ABSOLUTEMIN;
+    if ((((const BYTE*)src)[4] >> 4) != 0) return ERROR(frameParameter_unsupported);   /* reserved bits */
+    return 0;
+}
+
+/** ZSTDv05_decodeFrameHeader_Part2() :
+*   decode the full Frame Header.
+*   srcSize must be the size provided by ZSTDv05_decodeFrameHeader_Part1().
+*   @return : 0, or an error code, which can be tested using ZSTDv05_isError() */
+static size_t ZSTDv05_decodeFrameHeader_Part2(ZSTDv05_DCtx* zc, const void* src, size_t srcSize)
+{
+    size_t result;
+    if (srcSize != zc->headerSize)
+        return ERROR(srcSize_wrong);
+    result = ZSTDv05_getFrameParams(&(zc->params), src, srcSize);
+    if ((MEM_32bits()) && (zc->params.windowLog > 25)) return ERROR(frameParameter_unsupported);
+    return result;
+}
+
+
+size_t ZSTDv05_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    BYTE headerFlags;
+    U32 cSize;
+
+    if (srcSize < 3)
+        return ERROR(srcSize_wrong);
+
+    headerFlags = *in;
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+
+    bpPtr->blockType = (blockType_t)(headerFlags >> 6);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+
+static size_t ZSTDv05_copyRawBlock(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    if (srcSize > maxDstSize) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/*! ZSTDv05_decodeLiteralsBlock() :
+    @return : nb of bytes read from src (< srcSize ) */
+size_t ZSTDv05_decodeLiteralsBlock(ZSTDv05_DCtx* dctx,
+                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+{
+    const BYTE* const istart = (const BYTE*) src;
+
+    /* any compressed block with literals segment must be at least this size */
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch(istart[0]>> 6)
+    {
+    case IS_HUFv05:
+        {
+            size_t litSize, litCSize, singleStream=0;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            if (srcSize < 5) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for case 3 */
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                /* 2 - 2 - 10 - 10 */
+                lhSize=3;
+                singleStream = istart[0] & 16;
+                litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+                litCSize = ((istart[1] &  3) << 8) + istart[2];
+                break;
+            case 2:
+                /* 2 - 2 - 14 - 14 */
+                lhSize=4;
+                litSize  = ((istart[0] & 15) << 10) + (istart[1] << 2) + (istart[2] >> 6);
+                litCSize = ((istart[2] & 63) <<  8) + istart[3];
+                break;
+            case 3:
+                /* 2 - 2 - 18 - 18 */
+                lhSize=5;
+                litSize  = ((istart[0] & 15) << 14) + (istart[1] << 6) + (istart[2] >> 2);
+                litCSize = ((istart[2] &  3) << 16) + (istart[3] << 8) + istart[4];
+                break;
+            }
+            if (litSize > BLOCKSIZE) return ERROR(corruption_detected);
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            if (HUFv05_isError(singleStream ?
+                            HUFv05_decompress1X2(dctx->litBuffer, litSize, istart+lhSize, litCSize) :
+                            HUFv05_decompress   (dctx->litBuffer, litSize, istart+lhSize, litCSize) ))
+                return ERROR(corruption_detected);
+
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case IS_PCH:
+        {
+            size_t errorCode;
+            size_t litSize, litCSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            if (lhSize != 1)  /* only case supported for now : small litSize, single stream */
+                return ERROR(corruption_detected);
+            if (!dctx->flagStaticTables)
+                return ERROR(dictionary_corrupted);
+
+            /* 2 - 2 - 10 - 10 */
+            lhSize=3;
+            litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+            litCSize = ((istart[1] &  3) << 8) + istart[2];
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            errorCode = HUFv05_decompress1X4_usingDTable(dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->hufTableX4);
+            if (HUFv05_isError(errorCode)) return ERROR(corruption_detected);
+
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case IS_RAW:
+        {
+            size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize=1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                break;
+            }
+
+            if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
+                if (litSize+lhSize > srcSize) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                return lhSize+litSize;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+lhSize;
+            dctx->litSize = litSize;
+            return lhSize+litSize;
+        }
+    case IS_RLE:
+        {
+            size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize = 1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                if (srcSize<4) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4 */
+                break;
+            }
+            if (litSize > BLOCKSIZE) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return lhSize+1;
+        }
+    default:
+        return ERROR(corruption_detected);   /* impossible */
+    }
+}
+
+
+size_t ZSTDv05_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t* dumpsLengthPtr,
+                         FSEv05_DTable* DTableLL, FSEv05_DTable* DTableML, FSEv05_DTable* DTableOffb,
+                         const void* src, size_t srcSize, U32 flagStaticTable)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* ip = istart;
+    const BYTE* const iend = istart + srcSize;
+    U32 LLtype, Offtype, MLtype;
+    U32 LLlog, Offlog, MLlog;
+    size_t dumpsLength;
+
+    /* check */
+    if (srcSize < MIN_SEQUENCES_SIZE)
+        return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    *nbSeq = *ip++;
+    if (*nbSeq==0) return 1;
+    if (*nbSeq >= 128) {
+        if (ip >= iend) return ERROR(srcSize_wrong);
+        *nbSeq = ((nbSeq[0]-128)<<8) + *ip++;
+    }
+
+    if (ip >= iend) return ERROR(srcSize_wrong);
+    LLtype  = *ip >> 6;
+    Offtype = (*ip >> 4) & 3;
+    MLtype  = (*ip >> 2) & 3;
+    if (*ip & 2) {
+        if (ip+3 > iend) return ERROR(srcSize_wrong);
+        dumpsLength  = ip[2];
+        dumpsLength += ip[1] << 8;
+        ip += 3;
+    } else {
+        if (ip+2 > iend) return ERROR(srcSize_wrong);
+        dumpsLength  = ip[1];
+        dumpsLength += (ip[0] & 1) << 8;
+        ip += 2;
+    }
+    *dumpsPtr = ip;
+    ip += dumpsLength;
+    *dumpsLengthPtr = dumpsLength;
+
+    /* check */
+    if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+    /* sequences */
+    {
+        S16 norm[MaxML+1];    /* assumption : MaxML >= MaxLL >= MaxOff */
+        size_t headerSize;
+
+        /* Build DTables */
+        switch(LLtype)
+        {
+        case FSEv05_ENCODING_RLE :
+            LLlog = 0;
+            FSEv05_buildDTable_rle(DTableLL, *ip++);
+            break;
+        case FSEv05_ENCODING_RAW :
+            LLlog = LLbits;
+            FSEv05_buildDTable_raw(DTableLL, LLbits);
+            break;
+        case FSEv05_ENCODING_STATIC:
+            if (!flagStaticTable) return ERROR(corruption_detected);
+            break;
+        case FSEv05_ENCODING_DYNAMIC :
+        default :   /* impossible */
+            {   U32 max = MaxLL;
+                headerSize = FSEv05_readNCount(norm, &max, &LLlog, ip, iend-ip);
+                if (FSEv05_isError(headerSize)) return ERROR(GENERIC);
+                if (LLlog > LLFSEv05Log) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSEv05_buildDTable(DTableLL, norm, max, LLlog);
+        }   }
+
+        switch(Offtype)
+        {
+        case FSEv05_ENCODING_RLE :
+            Offlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong);   /* min : "raw", hence no header, but at least xxLog bits */
+            FSEv05_buildDTable_rle(DTableOffb, *ip++ & MaxOff); /* if *ip > MaxOff, data is corrupted */
+            break;
+        case FSEv05_ENCODING_RAW :
+            Offlog = Offbits;
+            FSEv05_buildDTable_raw(DTableOffb, Offbits);
+            break;
+        case FSEv05_ENCODING_STATIC:
+            if (!flagStaticTable) return ERROR(corruption_detected);
+            break;
+        case FSEv05_ENCODING_DYNAMIC :
+        default :   /* impossible */
+            {   U32 max = MaxOff;
+                headerSize = FSEv05_readNCount(norm, &max, &Offlog, ip, iend-ip);
+                if (FSEv05_isError(headerSize)) return ERROR(GENERIC);
+                if (Offlog > OffFSEv05Log) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSEv05_buildDTable(DTableOffb, norm, max, Offlog);
+        }   }
+
+        switch(MLtype)
+        {
+        case FSEv05_ENCODING_RLE :
+            MLlog = 0;
+            if (ip > iend-2) return ERROR(srcSize_wrong); /* min : "raw", hence no header, but at least xxLog bits */
+            FSEv05_buildDTable_rle(DTableML, *ip++);
+            break;
+        case FSEv05_ENCODING_RAW :
+            MLlog = MLbits;
+            FSEv05_buildDTable_raw(DTableML, MLbits);
+            break;
+        case FSEv05_ENCODING_STATIC:
+            if (!flagStaticTable) return ERROR(corruption_detected);
+            break;
+        case FSEv05_ENCODING_DYNAMIC :
+        default :   /* impossible */
+            {   U32 max = MaxML;
+                headerSize = FSEv05_readNCount(norm, &max, &MLlog, ip, iend-ip);
+                if (FSEv05_isError(headerSize)) return ERROR(GENERIC);
+                if (MLlog > MLFSEv05Log) return ERROR(corruption_detected);
+                ip += headerSize;
+                FSEv05_buildDTable(DTableML, norm, max, MLlog);
+    }   }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t matchLength;
+    size_t offset;
+} seq_t;
+
+typedef struct {
+    BITv05_DStream_t DStream;
+    FSEv05_DState_t stateLL;
+    FSEv05_DState_t stateOffb;
+    FSEv05_DState_t stateML;
+    size_t prevOffset;
+    const BYTE* dumps;
+    const BYTE* dumpsEnd;
+} seqState_t;
+
+
+
+static void ZSTDv05_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    size_t litLength;
+    size_t prevOffset;
+    size_t offset;
+    size_t matchLength;
+    const BYTE* dumps = seqState->dumps;
+    const BYTE* const de = seqState->dumpsEnd;
+
+    /* Literal length */
+    litLength = FSEv05_peakSymbol(&(seqState->stateLL));
+    prevOffset = litLength ? seq->offset : seqState->prevOffset;
+    if (litLength == MaxLL) {
+        U32 add = *dumps++;
+        if (add < 255) litLength += add;
+        else {
+            litLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no risk : dumps is always followed by seq tables > 1 byte */
+            if (litLength&1) litLength>>=1, dumps += 3;
+            else litLength = (U16)(litLength)>>1, dumps += 2;
+        }
+        if (dumps > de) { litLength = MaxLL+255; }  /* late correction, to avoid using uninitialized memory */
+        if (dumps >= de) { dumps = de-1; }  /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+
+    /* Offset */
+    {
+        static const U32 offsetPrefix[MaxOff+1] = {
+                1 /*fake*/, 1, 2, 4, 8, 16, 32, 64, 128, 256,
+                512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072, 262144,
+                524288, 1048576, 2097152, 4194304, 8388608, 16777216, 33554432, /*fake*/ 1, 1, 1, 1, 1 };
+        U32 offsetCode = FSEv05_peakSymbol(&(seqState->stateOffb));   /* <= maxOff, by table construction */
+        U32 nbBits = offsetCode - 1;
+        if (offsetCode==0) nbBits = 0;   /* cmove */
+        offset = offsetPrefix[offsetCode] + BITv05_readBits(&(seqState->DStream), nbBits);
+        if (MEM_32bits()) BITv05_reloadDStream(&(seqState->DStream));
+        if (offsetCode==0) offset = prevOffset;   /* repcode, cmove */
+        if (offsetCode | !litLength) seqState->prevOffset = seq->offset;   /* cmove */
+        FSEv05_decodeSymbol(&(seqState->stateOffb), &(seqState->DStream));    /* update */
+    }
+
+    /* Literal length update */
+    FSEv05_decodeSymbol(&(seqState->stateLL), &(seqState->DStream));   /* update */
+    if (MEM_32bits()) BITv05_reloadDStream(&(seqState->DStream));
+
+    /* MatchLength */
+    matchLength = FSEv05_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
+    if (matchLength == MaxML) {
+        U32 add = *dumps++;
+        if (add < 255) matchLength += add;
+        else {
+            matchLength = MEM_readLE32(dumps) & 0xFFFFFF;  /* no pb : dumps is always followed by seq tables > 1 byte */
+            if (matchLength&1) matchLength>>=1, dumps += 3;
+            else matchLength = (U16)(matchLength)>>1, dumps += 2;
+        }
+        if (dumps > de) { matchLength = MaxML+255; }  /* late correction, to avoid using uninitialized memory */
+        if (dumps >= de) { dumps = de-1; }  /* late correction, to avoid read overflow (data is now corrupted anyway) */
+    }
+    matchLength += MINMATCH;
+
+    /* save result */
+    seq->litLength = litLength;
+    seq->offset = offset;
+    seq->matchLength = matchLength;
+    seqState->dumps = dumps;
+
+#if 0   /* debug */
+    {
+        static U64 totalDecoded = 0;
+        printf("pos %6u : %3u literals & match %3u bytes at distance %6u \n",
+           (U32)(totalDecoded), (U32)litLength, (U32)matchLength, (U32)offset);
+        totalDecoded += litLength + matchLength;
+    }
+#endif
+}
+
+
+static size_t ZSTDv05_execSequence(BYTE* op,
+                                BYTE* const oend, seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    static const int dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+    static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* substracted */
+    BYTE* const oLitEnd = op + sequence.litLength;
+    const size_t sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_8 = oend-8;
+    const BYTE* const litEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* risk read beyond lit buffer */
+
+    /* copy Literals */
+    ZSTDv05_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = litEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base)) {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - vBase))
+            return ERROR(corruption_detected);
+        match = dictEnd - (base-match);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {
+            size_t length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+            if (op > oend_8 || sequence.matchLength < MINMATCH) {
+              while (op < oMatchEnd) *op++ = *match++;
+              return sequenceLength;
+            }
+    }   }
+    /* Requirement: op <= oend_8 */
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        const int sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTDv05_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTDv05_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH)) {
+        if (op < oend_8) {
+            ZSTDv05_wildcopy(op, match, oend_8 - op);
+            match += oend_8 - op;
+            op = oend_8;
+        }
+        while (op < oMatchEnd)
+            *op++ = *match++;
+    } else {
+        ZSTDv05_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+
+static size_t ZSTDv05_decompressSequences(
+                               ZSTDv05_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t errorCode, dumpsLength;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    int nbSeq;
+    const BYTE* dumps;
+    U32* DTableLL = dctx->LLTable;
+    U32* DTableML = dctx->MLTable;
+    U32* DTableOffb = dctx->OffTable;
+    const BYTE* const base = (const BYTE*) (dctx->base);
+    const BYTE* const vBase = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+
+    /* Build Decoding Tables */
+    errorCode = ZSTDv05_decodeSeqHeaders(&nbSeq, &dumps, &dumpsLength,
+                                      DTableLL, DTableML, DTableOffb,
+                                      ip, seqSize, dctx->flagStaticTables);
+    if (ZSTDv05_isError(errorCode)) return errorCode;
+    ip += errorCode;
+
+    /* Regen sequences */
+    if (nbSeq) {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        sequence.offset = REPCODE_STARTVALUE;
+        seqState.dumps = dumps;
+        seqState.dumpsEnd = dumps + dumpsLength;
+        seqState.prevOffset = REPCODE_STARTVALUE;
+        errorCode = BITv05_initDStream(&(seqState.DStream), ip, iend-ip);
+        if (ERR_isError(errorCode)) return ERROR(corruption_detected);
+        FSEv05_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSEv05_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSEv05_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BITv05_reloadDStream(&(seqState.DStream)) <= BITv05_DStream_completed) && nbSeq ; ) {
+            size_t oneSeqSize;
+            nbSeq--;
+            ZSTDv05_decodeSequence(&sequence, &seqState);
+            oneSeqSize = ZSTDv05_execSequence(op, oend, sequence, &litPtr, litEnd, base, vBase, dictEnd);
+            if (ZSTDv05_isError(oneSeqSize)) return oneSeqSize;
+            op += oneSeqSize;
+        }
+
+        /* check if reached exact end */
+        if (nbSeq) return ERROR(corruption_detected);
+    }
+
+    /* last literal segment */
+    {
+        size_t lastLLSize = litEnd - litPtr;
+        if (litPtr > litEnd) return ERROR(corruption_detected);   /* too many literals already used */
+        if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+        memcpy(op, litPtr, lastLLSize);
+        op += lastLLSize;
+    }
+
+    return op-ostart;
+}
+
+
+static void ZSTDv05_checkContinuity(ZSTDv05_DCtx* dctx, const void* dst)
+{
+    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+        dctx->dictEnd = dctx->previousDstEnd;
+        dctx->vBase = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+        dctx->base = dst;
+        dctx->previousDstEnd = dst;
+    }
+}
+
+
+static size_t ZSTDv05_decompressBlock_internal(ZSTDv05_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{   /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+    size_t litCSize;
+
+    if (srcSize >= BLOCKSIZE) return ERROR(srcSize_wrong);
+
+    /* Decode literals sub-block */
+    litCSize = ZSTDv05_decodeLiteralsBlock(dctx, src, srcSize);
+    if (ZSTDv05_isError(litCSize)) return litCSize;
+    ip += litCSize;
+    srcSize -= litCSize;
+
+    return ZSTDv05_decompressSequences(dctx, dst, dstCapacity, ip, srcSize);
+}
+
+
+size_t ZSTDv05_decompressBlock(ZSTDv05_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{
+    ZSTDv05_checkContinuity(dctx, dst);
+    return ZSTDv05_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+/*! ZSTDv05_decompress_continueDCtx
+*   dctx must have been properly initialized */
+static size_t ZSTDv05_decompress_continueDCtx(ZSTDv05_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                                 const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + maxDstSize;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    {
+        size_t frameHeaderSize;
+        if (srcSize < ZSTDv05_frameHeaderSize_min+ZSTDv05_blockHeaderSize) return ERROR(srcSize_wrong);
+        frameHeaderSize = ZSTDv05_decodeFrameHeader_Part1(dctx, src, ZSTDv05_frameHeaderSize_min);
+        if (ZSTDv05_isError(frameHeaderSize)) return frameHeaderSize;
+        if (srcSize < frameHeaderSize+ZSTDv05_blockHeaderSize) return ERROR(srcSize_wrong);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+        frameHeaderSize = ZSTDv05_decodeFrameHeader_Part2(dctx, src, frameHeaderSize);
+        if (ZSTDv05_isError(frameHeaderSize)) return frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t decodedSize=0;
+        size_t cBlockSize = ZSTDv05_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTDv05_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv05_blockHeaderSize;
+        remainingSize -= ZSTDv05_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTDv05_decompressBlock_internal(dctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTDv05_copyRawBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        if (ZSTDv05_isError(decodedSize)) return decodedSize;
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+
+size_t ZSTDv05_decompress_usingPreparedDCtx(ZSTDv05_DCtx* dctx, const ZSTDv05_DCtx* refDCtx,
+                                         void* dst, size_t maxDstSize,
+                                   const void* src, size_t srcSize)
+{
+    ZSTDv05_copyDCtx(dctx, refDCtx);
+    ZSTDv05_checkContinuity(dctx, dst);
+    return ZSTDv05_decompress_continueDCtx(dctx, dst, maxDstSize, src, srcSize);
+}
+
+
+size_t ZSTDv05_decompress_usingDict(ZSTDv05_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                                 const void* src, size_t srcSize,
+                                 const void* dict, size_t dictSize)
+{
+    ZSTDv05_decompressBegin_usingDict(dctx, dict, dictSize);
+    ZSTDv05_checkContinuity(dctx, dst);
+    return ZSTDv05_decompress_continueDCtx(dctx, dst, maxDstSize, src, srcSize);
+}
+
+
+size_t ZSTDv05_decompressDCtx(ZSTDv05_DCtx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    return ZSTDv05_decompress_usingDict(dctx, dst, maxDstSize, src, srcSize, NULL, 0);
+}
+
+size_t ZSTDv05_decompress(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+#if defined(ZSTDv05_HEAPMODE) && (ZSTDv05_HEAPMODE==1)
+    size_t regenSize;
+    ZSTDv05_DCtx* dctx = ZSTDv05_createDCtx();
+    if (dctx==NULL) return ERROR(memory_allocation);
+    regenSize = ZSTDv05_decompressDCtx(dctx, dst, maxDstSize, src, srcSize);
+    ZSTDv05_freeDCtx(dctx);
+    return regenSize;
+#else
+    ZSTDv05_DCtx dctx;
+    return ZSTDv05_decompressDCtx(&dctx, dst, maxDstSize, src, srcSize);
+#endif
+}
+
+size_t ZSTDv05_findFrameCompressedSize(const void *src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties;
+
+    /* Frame Header */
+    if (srcSize < ZSTDv05_frameHeaderSize_min) return ERROR(srcSize_wrong);
+    if (MEM_readLE32(src) != ZSTDv05_MAGICNUMBER) return ERROR(prefix_unknown);
+    ip += ZSTDv05_frameHeaderSize_min; remainingSize -= ZSTDv05_frameHeaderSize_min;
+
+    /* Loop on each block */
+    while (1)
+    {
+        size_t cBlockSize = ZSTDv05_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTDv05_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv05_blockHeaderSize;
+        remainingSize -= ZSTDv05_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/* ******************************
+*  Streaming Decompression API
+********************************/
+size_t ZSTDv05_nextSrcSizeToDecompress(ZSTDv05_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+size_t ZSTDv05_decompressContinue(ZSTDv05_DCtx* dctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != dctx->expected) return ERROR(srcSize_wrong);
+    ZSTDv05_checkContinuity(dctx, dst);
+
+    /* Decompress : frame header; part 1 */
+    switch (dctx->stage)
+    {
+    case ZSTDv05ds_getFrameHeaderSize :
+        /* get frame header size */
+        if (srcSize != ZSTDv05_frameHeaderSize_min) return ERROR(srcSize_wrong);   /* impossible */
+        dctx->headerSize = ZSTDv05_decodeFrameHeader_Part1(dctx, src, ZSTDv05_frameHeaderSize_min);
+        if (ZSTDv05_isError(dctx->headerSize)) return dctx->headerSize;
+        memcpy(dctx->headerBuffer, src, ZSTDv05_frameHeaderSize_min);
+        if (dctx->headerSize > ZSTDv05_frameHeaderSize_min) return ERROR(GENERIC); /* should never happen */
+        dctx->expected = 0;   /* not necessary to copy more */
+        /* fallthrough */
+    case ZSTDv05ds_decodeFrameHeader:
+        /* get frame header */
+        {   size_t const result = ZSTDv05_decodeFrameHeader_Part2(dctx, dctx->headerBuffer, dctx->headerSize);
+            if (ZSTDv05_isError(result)) return result;
+            dctx->expected = ZSTDv05_blockHeaderSize;
+            dctx->stage = ZSTDv05ds_decodeBlockHeader;
+            return 0;
+        }
+    case ZSTDv05ds_decodeBlockHeader:
+        {
+            /* Decode block header */
+            blockProperties_t bp;
+            size_t blockSize = ZSTDv05_getcBlockSize(src, ZSTDv05_blockHeaderSize, &bp);
+            if (ZSTDv05_isError(blockSize)) return blockSize;
+            if (bp.blockType == bt_end) {
+                dctx->expected = 0;
+                dctx->stage = ZSTDv05ds_getFrameHeaderSize;
+            }
+            else {
+                dctx->expected = blockSize;
+                dctx->bType = bp.blockType;
+                dctx->stage = ZSTDv05ds_decompressBlock;
+            }
+            return 0;
+        }
+    case ZSTDv05ds_decompressBlock:
+        {
+            /* Decompress : block content */
+            size_t rSize;
+            switch(dctx->bType)
+            {
+            case bt_compressed:
+                rSize = ZSTDv05_decompressBlock_internal(dctx, dst, maxDstSize, src, srcSize);
+                break;
+            case bt_raw :
+                rSize = ZSTDv05_copyRawBlock(dst, maxDstSize, src, srcSize);
+                break;
+            case bt_rle :
+                return ERROR(GENERIC);   /* not yet handled */
+                break;
+            case bt_end :   /* should never happen (filtered at phase 1) */
+                rSize = 0;
+                break;
+            default:
+                return ERROR(GENERIC);   /* impossible */
+            }
+            dctx->stage = ZSTDv05ds_decodeBlockHeader;
+            dctx->expected = ZSTDv05_blockHeaderSize;
+            dctx->previousDstEnd = (char*)dst + rSize;
+            return rSize;
+        }
+    default:
+        return ERROR(GENERIC);   /* impossible */
+    }
+}
+
+
+static void ZSTDv05_refDictContent(ZSTDv05_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    dctx->dictEnd = dctx->previousDstEnd;
+    dctx->vBase = (const char*)dict - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+    dctx->base = dict;
+    dctx->previousDstEnd = (const char*)dict + dictSize;
+}
+
+static size_t ZSTDv05_loadEntropy(ZSTDv05_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    size_t hSize, offcodeHeaderSize, matchlengthHeaderSize, errorCode, litlengthHeaderSize;
+    short offcodeNCount[MaxOff+1];
+    U32 offcodeMaxValue=MaxOff, offcodeLog;
+    short matchlengthNCount[MaxML+1];
+    unsigned matchlengthMaxValue = MaxML, matchlengthLog;
+    short litlengthNCount[MaxLL+1];
+    unsigned litlengthMaxValue = MaxLL, litlengthLog;
+
+    hSize = HUFv05_readDTableX4(dctx->hufTableX4, dict, dictSize);
+    if (HUFv05_isError(hSize)) return ERROR(dictionary_corrupted);
+    dict = (const char*)dict + hSize;
+    dictSize -= hSize;
+
+    offcodeHeaderSize = FSEv05_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dict, dictSize);
+    if (FSEv05_isError(offcodeHeaderSize)) return ERROR(dictionary_corrupted);
+    if (offcodeLog > OffFSEv05Log) return ERROR(dictionary_corrupted);
+    errorCode = FSEv05_buildDTable(dctx->OffTable, offcodeNCount, offcodeMaxValue, offcodeLog);
+    if (FSEv05_isError(errorCode)) return ERROR(dictionary_corrupted);
+    dict = (const char*)dict + offcodeHeaderSize;
+    dictSize -= offcodeHeaderSize;
+
+    matchlengthHeaderSize = FSEv05_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dict, dictSize);
+    if (FSEv05_isError(matchlengthHeaderSize)) return ERROR(dictionary_corrupted);
+    if (matchlengthLog > MLFSEv05Log) return ERROR(dictionary_corrupted);
+    errorCode = FSEv05_buildDTable(dctx->MLTable, matchlengthNCount, matchlengthMaxValue, matchlengthLog);
+    if (FSEv05_isError(errorCode)) return ERROR(dictionary_corrupted);
+    dict = (const char*)dict + matchlengthHeaderSize;
+    dictSize -= matchlengthHeaderSize;
+
+    litlengthHeaderSize = FSEv05_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dict, dictSize);
+    if (litlengthLog > LLFSEv05Log) return ERROR(dictionary_corrupted);
+    if (FSEv05_isError(litlengthHeaderSize)) return ERROR(dictionary_corrupted);
+    errorCode = FSEv05_buildDTable(dctx->LLTable, litlengthNCount, litlengthMaxValue, litlengthLog);
+    if (FSEv05_isError(errorCode)) return ERROR(dictionary_corrupted);
+
+    dctx->flagStaticTables = 1;
+    return hSize + offcodeHeaderSize + matchlengthHeaderSize + litlengthHeaderSize;
+}
+
+static size_t ZSTDv05_decompress_insertDictionary(ZSTDv05_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    size_t eSize;
+    U32 magic = MEM_readLE32(dict);
+    if (magic != ZSTDv05_DICT_MAGIC) {
+        /* pure content mode */
+        ZSTDv05_refDictContent(dctx, dict, dictSize);
+        return 0;
+    }
+    /* load entropy tables */
+    dict = (const char*)dict + 4;
+    dictSize -= 4;
+    eSize = ZSTDv05_loadEntropy(dctx, dict, dictSize);
+    if (ZSTDv05_isError(eSize)) return ERROR(dictionary_corrupted);
+
+    /* reference dictionary content */
+    dict = (const char*)dict + eSize;
+    dictSize -= eSize;
+    ZSTDv05_refDictContent(dctx, dict, dictSize);
+
+    return 0;
+}
+
+
+size_t ZSTDv05_decompressBegin_usingDict(ZSTDv05_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    size_t errorCode;
+    errorCode = ZSTDv05_decompressBegin(dctx);
+    if (ZSTDv05_isError(errorCode)) return errorCode;
+
+    if (dict && dictSize) {
+        errorCode = ZSTDv05_decompress_insertDictionary(dctx, dict, dictSize);
+        if (ZSTDv05_isError(errorCode)) return ERROR(dictionary_corrupted);
+    }
+
+    return 0;
+}
+
+/*
+    Buffered version of Zstd compression library
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd source repository : https://github.com/Cyan4973/zstd
+    - ztsd public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* The objects defined into this file should be considered experimental.
+ * They are not labelled stable, as their prototype may change in the future.
+ * You can use them for tests, provide feedback, or if you can endure risk of future changes.
+ */
+
+
+
+/* *************************************
+*  Constants
+***************************************/
+static size_t ZBUFFv05_blockHeaderSize = 3;
+
+
+
+/* *** Compression *** */
+
+static size_t ZBUFFv05_limitCopy(void* dst, size_t maxDstSize, const void* src, size_t srcSize)
+{
+    size_t length = MIN(maxDstSize, srcSize);
+    memcpy(dst, src, length);
+    return length;
+}
+
+
+
+
+/** ************************************************
+*  Streaming decompression
+*
+*  A ZBUFFv05_DCtx object is required to track streaming operation.
+*  Use ZBUFFv05_createDCtx() and ZBUFFv05_freeDCtx() to create/release resources.
+*  Use ZBUFFv05_decompressInit() to start a new decompression operation.
+*  ZBUFFv05_DCtx objects can be reused multiple times.
+*
+*  Use ZBUFFv05_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *maxDstSizePtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *maxDstSizePtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to call again the function with remaining input.
+*  The content of dst will be overwritten (up to *maxDstSizePtr) at each function call, so save its content if it matters or change dst .
+*  return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to improve latency)
+*            or 0 when a frame is completely decoded
+*            or an error code, which can be tested using ZBUFFv05_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory)
+*  output : 128 KB block size is the internal unit, it ensures it's always possible to write a full block when it's decoded.
+*  input : just follow indications from ZBUFFv05_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* **************************************************/
+
+typedef enum { ZBUFFv05ds_init, ZBUFFv05ds_readHeader, ZBUFFv05ds_loadHeader, ZBUFFv05ds_decodeHeader,
+               ZBUFFv05ds_read, ZBUFFv05ds_load, ZBUFFv05ds_flush } ZBUFFv05_dStage;
+
+/* *** Resource management *** */
+
+#define ZSTDv05_frameHeaderSize_max 5   /* too magical, should come from reference */
+struct ZBUFFv05_DCtx_s {
+    ZSTDv05_DCtx* zc;
+    ZSTDv05_parameters params;
+    char* inBuff;
+    size_t inBuffSize;
+    size_t inPos;
+    char* outBuff;
+    size_t outBuffSize;
+    size_t outStart;
+    size_t outEnd;
+    size_t hPos;
+    ZBUFFv05_dStage stage;
+    unsigned char headerBuffer[ZSTDv05_frameHeaderSize_max];
+};   /* typedef'd to ZBUFFv05_DCtx within "zstd_buffered.h" */
+
+
+ZBUFFv05_DCtx* ZBUFFv05_createDCtx(void)
+{
+    ZBUFFv05_DCtx* zbc = (ZBUFFv05_DCtx*)malloc(sizeof(ZBUFFv05_DCtx));
+    if (zbc==NULL) return NULL;
+    memset(zbc, 0, sizeof(*zbc));
+    zbc->zc = ZSTDv05_createDCtx();
+    zbc->stage = ZBUFFv05ds_init;
+    return zbc;
+}
+
+size_t ZBUFFv05_freeDCtx(ZBUFFv05_DCtx* zbc)
+{
+    if (zbc==NULL) return 0;   /* support free on null */
+    ZSTDv05_freeDCtx(zbc->zc);
+    free(zbc->inBuff);
+    free(zbc->outBuff);
+    free(zbc);
+    return 0;
+}
+
+
+/* *** Initialization *** */
+
+size_t ZBUFFv05_decompressInitDictionary(ZBUFFv05_DCtx* zbc, const void* dict, size_t dictSize)
+{
+    zbc->stage = ZBUFFv05ds_readHeader;
+    zbc->hPos = zbc->inPos = zbc->outStart = zbc->outEnd = 0;
+    return ZSTDv05_decompressBegin_usingDict(zbc->zc, dict, dictSize);
+}
+
+size_t ZBUFFv05_decompressInit(ZBUFFv05_DCtx* zbc)
+{
+    return ZBUFFv05_decompressInitDictionary(zbc, NULL, 0);
+}
+
+
+/* *** Decompression *** */
+
+size_t ZBUFFv05_decompressContinue(ZBUFFv05_DCtx* zbc, void* dst, size_t* maxDstSizePtr, const void* src, size_t* srcSizePtr)
+{
+    const char* const istart = (const char*)src;
+    const char* ip = istart;
+    const char* const iend = istart + *srcSizePtr;
+    char* const ostart = (char*)dst;
+    char* op = ostart;
+    char* const oend = ostart + *maxDstSizePtr;
+    U32 notDone = 1;
+
+    while (notDone) {
+        switch(zbc->stage)
+        {
+        case ZBUFFv05ds_init :
+            return ERROR(init_missing);
+
+        case ZBUFFv05ds_readHeader :
+            /* read header from src */
+            {
+                size_t headerSize = ZSTDv05_getFrameParams(&(zbc->params), src, *srcSizePtr);
+                if (ZSTDv05_isError(headerSize)) return headerSize;
+                if (headerSize) {
+                    /* not enough input to decode header : tell how many bytes would be necessary */
+                    memcpy(zbc->headerBuffer+zbc->hPos, src, *srcSizePtr);
+                    zbc->hPos += *srcSizePtr;
+                    *maxDstSizePtr = 0;
+                    zbc->stage = ZBUFFv05ds_loadHeader;
+                    return headerSize - zbc->hPos;
+                }
+                zbc->stage = ZBUFFv05ds_decodeHeader;
+                break;
+            }
+	    /* fall-through */
+        case ZBUFFv05ds_loadHeader:
+            /* complete header from src */
+            {
+                size_t headerSize = ZBUFFv05_limitCopy(
+                    zbc->headerBuffer + zbc->hPos, ZSTDv05_frameHeaderSize_max - zbc->hPos,
+                    src, *srcSizePtr);
+                zbc->hPos += headerSize;
+                ip += headerSize;
+                headerSize = ZSTDv05_getFrameParams(&(zbc->params), zbc->headerBuffer, zbc->hPos);
+                if (ZSTDv05_isError(headerSize)) return headerSize;
+                if (headerSize) {
+                    /* not enough input to decode header : tell how many bytes would be necessary */
+                    *maxDstSizePtr = 0;
+                    return headerSize - zbc->hPos;
+                }
+                // zbc->stage = ZBUFFv05ds_decodeHeader; break;   /* useless : stage follows */
+            }
+	    /* fall-through */
+        case ZBUFFv05ds_decodeHeader:
+                /* apply header to create / resize buffers */
+                {
+                    size_t neededOutSize = (size_t)1 << zbc->params.windowLog;
+                    size_t neededInSize = BLOCKSIZE;   /* a block is never > BLOCKSIZE */
+                    if (zbc->inBuffSize < neededInSize) {
+                        free(zbc->inBuff);
+                        zbc->inBuffSize = neededInSize;
+                        zbc->inBuff = (char*)malloc(neededInSize);
+                        if (zbc->inBuff == NULL) return ERROR(memory_allocation);
+                    }
+                    if (zbc->outBuffSize < neededOutSize) {
+                        free(zbc->outBuff);
+                        zbc->outBuffSize = neededOutSize;
+                        zbc->outBuff = (char*)malloc(neededOutSize);
+                        if (zbc->outBuff == NULL) return ERROR(memory_allocation);
+                }   }
+                if (zbc->hPos) {
+                    /* some data already loaded into headerBuffer : transfer into inBuff */
+                    memcpy(zbc->inBuff, zbc->headerBuffer, zbc->hPos);
+                    zbc->inPos = zbc->hPos;
+                    zbc->hPos = 0;
+                    zbc->stage = ZBUFFv05ds_load;
+                    break;
+                }
+                zbc->stage = ZBUFFv05ds_read;
+		/* fall-through */
+        case ZBUFFv05ds_read:
+            {
+                size_t neededInSize = ZSTDv05_nextSrcSizeToDecompress(zbc->zc);
+                if (neededInSize==0) {  /* end of frame */
+                    zbc->stage = ZBUFFv05ds_init;
+                    notDone = 0;
+                    break;
+                }
+                if ((size_t)(iend-ip) >= neededInSize) {
+                    /* directly decode from src */
+                    size_t decodedSize = ZSTDv05_decompressContinue(zbc->zc,
+                        zbc->outBuff + zbc->outStart, zbc->outBuffSize - zbc->outStart,
+                        ip, neededInSize);
+                    if (ZSTDv05_isError(decodedSize)) return decodedSize;
+                    ip += neededInSize;
+                    if (!decodedSize) break;   /* this was just a header */
+                    zbc->outEnd = zbc->outStart +  decodedSize;
+                    zbc->stage = ZBUFFv05ds_flush;
+                    break;
+                }
+                if (ip==iend) { notDone = 0; break; }   /* no more input */
+                zbc->stage = ZBUFFv05ds_load;
+            }
+	    /* fall-through */
+        case ZBUFFv05ds_load:
+            {
+                size_t neededInSize = ZSTDv05_nextSrcSizeToDecompress(zbc->zc);
+                size_t toLoad = neededInSize - zbc->inPos;   /* should always be <= remaining space within inBuff */
+                size_t loadedSize;
+                if (toLoad > zbc->inBuffSize - zbc->inPos) return ERROR(corruption_detected);   /* should never happen */
+                loadedSize = ZBUFFv05_limitCopy(zbc->inBuff + zbc->inPos, toLoad, ip, iend-ip);
+                ip += loadedSize;
+                zbc->inPos += loadedSize;
+                if (loadedSize < toLoad) { notDone = 0; break; }   /* not enough input, wait for more */
+                {
+                    size_t decodedSize = ZSTDv05_decompressContinue(zbc->zc,
+                        zbc->outBuff + zbc->outStart, zbc->outBuffSize - zbc->outStart,
+                        zbc->inBuff, neededInSize);
+                    if (ZSTDv05_isError(decodedSize)) return decodedSize;
+                    zbc->inPos = 0;   /* input is consumed */
+                    if (!decodedSize) { zbc->stage = ZBUFFv05ds_read; break; }   /* this was just a header */
+                    zbc->outEnd = zbc->outStart +  decodedSize;
+                    zbc->stage = ZBUFFv05ds_flush;
+                    // break; /* ZBUFFv05ds_flush follows */
+                }
+	    }
+	    /* fall-through */
+        case ZBUFFv05ds_flush:
+            {
+                size_t toFlushSize = zbc->outEnd - zbc->outStart;
+                size_t flushedSize = ZBUFFv05_limitCopy(op, oend-op, zbc->outBuff + zbc->outStart, toFlushSize);
+                op += flushedSize;
+                zbc->outStart += flushedSize;
+                if (flushedSize == toFlushSize) {
+                    zbc->stage = ZBUFFv05ds_read;
+                    if (zbc->outStart + BLOCKSIZE > zbc->outBuffSize)
+                        zbc->outStart = zbc->outEnd = 0;
+                    break;
+                }
+                /* cannot flush everything */
+                notDone = 0;
+                break;
+            }
+        default: return ERROR(GENERIC);   /* impossible */
+    }   }
+
+    *srcSizePtr = ip-istart;
+    *maxDstSizePtr = op-ostart;
+
+    {   size_t nextSrcSizeHint = ZSTDv05_nextSrcSizeToDecompress(zbc->zc);
+        if (nextSrcSizeHint > ZBUFFv05_blockHeaderSize) nextSrcSizeHint+= ZBUFFv05_blockHeaderSize;   /* get next block header too */
+        nextSrcSizeHint -= zbc->inPos;   /* already loaded*/
+        return nextSrcSizeHint;
+    }
+}
+
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+unsigned ZBUFFv05_isError(size_t errorCode) { return ERR_isError(errorCode); }
+const char* ZBUFFv05_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
+
+size_t ZBUFFv05_recommendedDInSize(void)  { return BLOCKSIZE + ZBUFFv05_blockHeaderSize /* block header size*/ ; }
+size_t ZBUFFv05_recommendedDOutSize(void) { return BLOCKSIZE; }
diff --git a/vendor/github.com/DataDog/zstd/zstd_v05.h b/vendor/github.com/DataDog/zstd/zstd_v05.h
new file mode 100644
index 0000000..b68fd57
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v05.h
@@ -0,0 +1,157 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTDv05_H
+#define ZSTDv05_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-*************************************
+*  Dependencies
+***************************************/
+#include <stddef.h>   /* size_t */
+#include "mem.h"      /* U64, U32 */
+
+
+/* *************************************
+*  Simple functions
+***************************************/
+/*! ZSTDv05_decompress() :
+    `compressedSize` : is the _exact_ size of the compressed blob, otherwise decompression will fail.
+    `dstCapacity` must be large enough, equal or larger than originalSize.
+    @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
+              or an errorCode if it fails (which can be tested using ZSTDv05_isError()) */
+size_t ZSTDv05_decompress( void* dst, size_t dstCapacity,
+                     const void* src, size_t compressedSize);
+
+/**
+ZSTDv05_getFrameSrcSize() : get the source length of a ZSTD frame
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv05_isError())
+*/
+size_t ZSTDv05_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/* *************************************
+*  Helper functions
+***************************************/
+/* Error Management */
+unsigned    ZSTDv05_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
+const char* ZSTDv05_getErrorName(size_t code);     /*!< provides readable string for an error code */
+
+
+/* *************************************
+*  Explicit memory management
+***************************************/
+/** Decompression context */
+typedef struct ZSTDv05_DCtx_s ZSTDv05_DCtx;
+ZSTDv05_DCtx* ZSTDv05_createDCtx(void);
+size_t ZSTDv05_freeDCtx(ZSTDv05_DCtx* dctx);      /*!< @return : errorCode */
+
+/** ZSTDv05_decompressDCtx() :
+*   Same as ZSTDv05_decompress(), but requires an already allocated ZSTDv05_DCtx (see ZSTDv05_createDCtx()) */
+size_t ZSTDv05_decompressDCtx(ZSTDv05_DCtx* ctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+/*-***********************
+*  Simple Dictionary API
+*************************/
+/*! ZSTDv05_decompress_usingDict() :
+*   Decompression using a pre-defined Dictionary content (see dictBuilder).
+*   Dictionary must be identical to the one used during compression, otherwise regenerated data will be corrupted.
+*   Note : dict can be NULL, in which case, it's equivalent to ZSTDv05_decompressDCtx() */
+size_t ZSTDv05_decompress_usingDict(ZSTDv05_DCtx* dctx,
+                                            void* dst, size_t dstCapacity,
+                                      const void* src, size_t srcSize,
+                                      const void* dict,size_t dictSize);
+
+/*-************************
+*  Advanced Streaming API
+***************************/
+typedef enum { ZSTDv05_fast, ZSTDv05_greedy, ZSTDv05_lazy, ZSTDv05_lazy2, ZSTDv05_btlazy2, ZSTDv05_opt, ZSTDv05_btopt } ZSTDv05_strategy;
+typedef struct {
+    U64 srcSize;
+    U32 windowLog;     /* the only useful information to retrieve */
+    U32 contentLog; U32 hashLog; U32 searchLog; U32 searchLength; U32 targetLength; ZSTDv05_strategy strategy;
+} ZSTDv05_parameters;
+size_t ZSTDv05_getFrameParams(ZSTDv05_parameters* params, const void* src, size_t srcSize);
+
+size_t ZSTDv05_decompressBegin_usingDict(ZSTDv05_DCtx* dctx, const void* dict, size_t dictSize);
+void   ZSTDv05_copyDCtx(ZSTDv05_DCtx* dstDCtx, const ZSTDv05_DCtx* srcDCtx);
+size_t ZSTDv05_nextSrcSizeToDecompress(ZSTDv05_DCtx* dctx);
+size_t ZSTDv05_decompressContinue(ZSTDv05_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+/*-***********************
+*  ZBUFF API
+*************************/
+typedef struct ZBUFFv05_DCtx_s ZBUFFv05_DCtx;
+ZBUFFv05_DCtx* ZBUFFv05_createDCtx(void);
+size_t         ZBUFFv05_freeDCtx(ZBUFFv05_DCtx* dctx);
+
+size_t ZBUFFv05_decompressInit(ZBUFFv05_DCtx* dctx);
+size_t ZBUFFv05_decompressInitDictionary(ZBUFFv05_DCtx* dctx, const void* dict, size_t dictSize);
+
+size_t ZBUFFv05_decompressContinue(ZBUFFv05_DCtx* dctx,
+                                            void* dst, size_t* dstCapacityPtr,
+                                      const void* src, size_t* srcSizePtr);
+
+/*-***************************************************************************
+*  Streaming decompression
+*
+*  A ZBUFFv05_DCtx object is required to track streaming operations.
+*  Use ZBUFFv05_createDCtx() and ZBUFFv05_freeDCtx() to create/release resources.
+*  Use ZBUFFv05_decompressInit() to start a new decompression operation,
+*   or ZBUFFv05_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFFv05_DCtx objects can be reused multiple times.
+*
+*  Use ZBUFFv05_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of @dst will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters or change @dst.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to help latency)
+*            or 0 when a frame is completely decoded
+*            or an error code, which can be tested using ZBUFFv05_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFFv05_recommendedDInSize() / ZBUFFv05_recommendedDOutSize()
+*  output : ZBUFFv05_recommendedDOutSize==128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFFv05_recommendedDInSize==128Kb+3; just follow indications from ZBUFFv05_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+unsigned ZBUFFv05_isError(size_t errorCode);
+const char* ZBUFFv05_getErrorName(size_t errorCode);
+
+/** Functions below provide recommended buffer sizes for Compression or Decompression operations.
+*   These sizes are just hints, and tend to offer better latency */
+size_t ZBUFFv05_recommendedDInSize(void);
+size_t ZBUFFv05_recommendedDOutSize(void);
+
+
+
+/*-*************************************
+*  Constants
+***************************************/
+#define ZSTDv05_MAGICNUMBER 0xFD2FB525   /* v0.5 */
+
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* ZSTDv0505_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v06.c b/vendor/github.com/DataDog/zstd/zstd_v06.c
new file mode 100644
index 0000000..8b068b3
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v06.c
@@ -0,0 +1,4124 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*- Dependencies -*/
+#include "zstd_v06.h"
+#include <stddef.h>    /* size_t, ptrdiff_t */
+#include <string.h>    /* memcpy */
+#include <stdlib.h>    /* malloc, free, qsort */
+#include "error_private.h"
+
+
+
+/* ******************************************************************
+   mem.h
+   low-level memory access routines
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*-****************************************
+*  Compiler specifics
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/*-**************************************************************
+*  Basic Types
+*****************************************************************/
+#if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/*-**************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets depending on alignment.
+ *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(size_t)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard, by lying on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+
+#endif /* MEM_FORCE_MEMORY_ACCESS */
+
+MEM_STATIC U32 MEM_swap32(U32 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_ulong(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap32(in);
+#else
+    return  ((in << 24) & 0xff000000 ) |
+            ((in <<  8) & 0x00ff0000 ) |
+            ((in >>  8) & 0x0000ff00 ) |
+            ((in >> 24) & 0x000000ff );
+#endif
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_uint64(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap64(in);
+#else
+    return  ((in << 56) & 0xff00000000000000ULL) |
+            ((in << 40) & 0x00ff000000000000ULL) |
+            ((in << 24) & 0x0000ff0000000000ULL) |
+            ((in << 8)  & 0x000000ff00000000ULL) |
+            ((in >> 8)  & 0x00000000ff000000ULL) |
+            ((in >> 24) & 0x0000000000ff0000ULL) |
+            ((in >> 40) & 0x000000000000ff00ULL) |
+            ((in >> 56) & 0x00000000000000ffULL);
+#endif
+}
+
+
+/*=== Little endian r/w ===*/
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian()) {
+        MEM_write16(memPtr, val);
+    } else {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+        return MEM_swap32(MEM_read32(memPtr));
+}
+
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+        return MEM_swap64(MEM_read64(memPtr));
+}
+
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+
+/*
+    zstd - standard compression library
+    Header File for static linking only
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net
+*/
+#ifndef ZSTDv06_STATIC_H
+#define ZSTDv06_STATIC_H
+
+/* The prototypes defined within this file are considered experimental.
+ * They should not be used in the context DLL as they may change in the future.
+ * Prefer static linking if you need them, to control breaking version changes issues.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/*- Advanced Decompression functions -*/
+
+/*! ZSTDv06_decompress_usingPreparedDCtx() :
+*   Same as ZSTDv06_decompress_usingDict, but using a reference context `preparedDCtx`, where dictionary has been loaded.
+*   It avoids reloading the dictionary each time.
+*   `preparedDCtx` must have been properly initialized using ZSTDv06_decompressBegin_usingDict().
+*   Requires 2 contexts : 1 for reference (preparedDCtx), which will not be modified, and 1 to run the decompression operation (dctx) */
+ZSTDLIBv06_API size_t ZSTDv06_decompress_usingPreparedDCtx(
+                                           ZSTDv06_DCtx* dctx, const ZSTDv06_DCtx* preparedDCtx,
+                                           void* dst, size_t dstCapacity,
+                                     const void* src, size_t srcSize);
+
+
+
+#define ZSTDv06_FRAMEHEADERSIZE_MAX 13    /* for static allocation */
+static const size_t ZSTDv06_frameHeaderSize_min = 5;
+static const size_t ZSTDv06_frameHeaderSize_max = ZSTDv06_FRAMEHEADERSIZE_MAX;
+
+ZSTDLIBv06_API size_t ZSTDv06_decompressBegin(ZSTDv06_DCtx* dctx);
+
+/*
+  Streaming decompression, direct mode (bufferless)
+
+  A ZSTDv06_DCtx object is required to track streaming operations.
+  Use ZSTDv06_createDCtx() / ZSTDv06_freeDCtx() to manage it.
+  A ZSTDv06_DCtx object can be re-used multiple times.
+
+  First optional operation is to retrieve frame parameters, using ZSTDv06_getFrameParams(), which doesn't consume the input.
+  It can provide the minimum size of rolling buffer required to properly decompress data,
+  and optionally the final size of uncompressed content.
+  (Note : content size is an optional info that may not be present. 0 means : content size unknown)
+  Frame parameters are extracted from the beginning of compressed frame.
+  The amount of data to read is variable, from ZSTDv06_frameHeaderSize_min to ZSTDv06_frameHeaderSize_max (so if `srcSize` >= ZSTDv06_frameHeaderSize_max, it will always work)
+  If `srcSize` is too small for operation to succeed, function will return the minimum size it requires to produce a result.
+  Result : 0 when successful, it means the ZSTDv06_frameParams structure has been filled.
+          >0 : means there is not enough data into `src`. Provides the expected size to successfully decode header.
+           errorCode, which can be tested using ZSTDv06_isError()
+
+  Start decompression, with ZSTDv06_decompressBegin() or ZSTDv06_decompressBegin_usingDict().
+  Alternatively, you can copy a prepared context, using ZSTDv06_copyDCtx().
+
+  Then use ZSTDv06_nextSrcSizeToDecompress() and ZSTDv06_decompressContinue() alternatively.
+  ZSTDv06_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTDv06_decompressContinue().
+  ZSTDv06_decompressContinue() requires this exact amount of bytes, or it will fail.
+  ZSTDv06_decompressContinue() needs previous data blocks during decompression, up to (1 << windowlog).
+  They should preferably be located contiguously, prior to current block. Alternatively, a round buffer is also possible.
+
+  @result of ZSTDv06_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity)
+  It can be zero, which is not an error; it just means ZSTDv06_decompressContinue() has decoded some header.
+
+  A frame is fully decoded when ZSTDv06_nextSrcSizeToDecompress() returns zero.
+  Context can then be reset to start a new decompression.
+*/
+
+
+/* **************************************
+*  Block functions
+****************************************/
+/*! Block functions produce and decode raw zstd blocks, without frame metadata.
+    User will have to take in charge required information to regenerate data, such as compressed and content sizes.
+
+    A few rules to respect :
+    - Uncompressed block size must be <= ZSTDv06_BLOCKSIZE_MAX (128 KB)
+    - Compressing or decompressing requires a context structure
+      + Use ZSTDv06_createCCtx() and ZSTDv06_createDCtx()
+    - It is necessary to init context before starting
+      + compression : ZSTDv06_compressBegin()
+      + decompression : ZSTDv06_decompressBegin()
+      + variants _usingDict() are also allowed
+      + copyCCtx() and copyDCtx() work too
+    - When a block is considered not compressible enough, ZSTDv06_compressBlock() result will be zero.
+      In which case, nothing is produced into `dst`.
+      + User must test for such outcome and deal directly with uncompressed data
+      + ZSTDv06_decompressBlock() doesn't accept uncompressed data as input !!
+*/
+
+#define ZSTDv06_BLOCKSIZE_MAX (128 * 1024)   /* define, for static allocation */
+ZSTDLIBv06_API size_t ZSTDv06_decompressBlock(ZSTDv06_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* ZSTDv06_STATIC_H */
+/*
+    zstd_internal - common functions to include
+    Header File for include
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : https://www.zstd.net
+*/
+#ifndef ZSTDv06_CCOMMON_H_MODULE
+#define ZSTDv06_CCOMMON_H_MODULE
+
+
+/*-*************************************
+*  Common macros
+***************************************/
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+
+
+/*-*************************************
+*  Common constants
+***************************************/
+#define ZSTDv06_DICT_MAGIC  0xEC30A436
+
+#define ZSTDv06_REP_NUM    3
+#define ZSTDv06_REP_INIT   ZSTDv06_REP_NUM
+#define ZSTDv06_REP_MOVE   (ZSTDv06_REP_NUM-1)
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define ZSTDv06_WINDOWLOG_ABSOLUTEMIN 12
+static const size_t ZSTDv06_fcs_fieldSize[4] = { 0, 1, 2, 8 };
+
+#define ZSTDv06_BLOCKHEADERSIZE 3   /* because C standard does not allow a static const value to be defined using another static const value .... :( */
+static const size_t ZSTDv06_blockHeaderSize = ZSTDv06_BLOCKHEADERSIZE;
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+
+#define HufLog 12
+
+#define IS_HUF 0
+#define IS_PCH 1
+#define IS_RAW 2
+#define IS_RLE 3
+
+#define LONGNBSEQ 0x7F00
+
+#define MINMATCH 3
+#define EQUAL_READ32 4
+#define REPCODE_STARTVALUE 1
+
+#define Litbits  8
+#define MaxLit ((1<<Litbits) - 1)
+#define MaxML  52
+#define MaxLL  35
+#define MaxOff 28
+#define MaxSeq MAX(MaxLL, MaxML)   /* Assumption : MaxOff < MaxLL,MaxML */
+#define MLFSELog    9
+#define LLFSELog    9
+#define OffFSELog   8
+
+#define FSEv06_ENCODING_RAW     0
+#define FSEv06_ENCODING_RLE     1
+#define FSEv06_ENCODING_STATIC  2
+#define FSEv06_ENCODING_DYNAMIC 3
+
+static const U32 LL_bits[MaxLL+1] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3, 4, 6, 7, 8, 9,10,11,12,
+                                     13,14,15,16 };
+static const S16 LL_defaultNorm[MaxLL+1] = { 4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1,
+                                             2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1,
+                                            -1,-1,-1,-1 };
+static const U32 LL_defaultNormLog = 6;
+
+static const U32 ML_bits[MaxML+1] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3, 4, 4, 5, 7, 8, 9,10,11,
+                                     12,13,14,15,16 };
+static const S16 ML_defaultNorm[MaxML+1] = { 1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,-1,-1,
+                                            -1,-1,-1,-1,-1 };
+static const U32 ML_defaultNormLog = 6;
+
+static const S16 OF_defaultNorm[MaxOff+1] = { 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+                                              1, 1, 1, 1, 1, 1, 1, 1,-1,-1,-1,-1,-1 };
+static const U32 OF_defaultNormLog = 5;
+
+
+/*-*******************************************
+*  Shared functions to include for inlining
+*********************************************/
+static void ZSTDv06_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+#define COPY8(d,s) { ZSTDv06_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTDv06_wildcopy() :
+*   custom version of memcpy(), can copy up to 7 bytes too many (8 bytes if length==0) */
+#define WILDCOPY_OVERLENGTH 8
+MEM_STATIC void ZSTDv06_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+
+
+/*-*******************************************
+*  Private interfaces
+*********************************************/
+typedef struct {
+    U32 off;
+    U32 len;
+} ZSTDv06_match_t;
+
+typedef struct {
+    U32 price;
+    U32 off;
+    U32 mlen;
+    U32 litlen;
+    U32 rep[ZSTDv06_REP_INIT];
+} ZSTDv06_optimal_t;
+
+typedef struct { U32  unused; } ZSTDv06_stats_t;
+
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* litStart;
+    BYTE* lit;
+    U16*  litLengthStart;
+    U16*  litLength;
+    BYTE* llCodeStart;
+    U16*  matchLengthStart;
+    U16*  matchLength;
+    BYTE* mlCodeStart;
+    U32   longLengthID;   /* 0 == no longLength; 1 == Lit.longLength; 2 == Match.longLength; */
+    U32   longLengthPos;
+    /* opt */
+    ZSTDv06_optimal_t* priceTable;
+    ZSTDv06_match_t* matchTable;
+    U32* matchLengthFreq;
+    U32* litLengthFreq;
+    U32* litFreq;
+    U32* offCodeFreq;
+    U32  matchLengthSum;
+    U32  matchSum;
+    U32  litLengthSum;
+    U32  litSum;
+    U32  offCodeSum;
+    U32  log2matchLengthSum;
+    U32  log2matchSum;
+    U32  log2litLengthSum;
+    U32  log2litSum;
+    U32  log2offCodeSum;
+    U32  factor;
+    U32  cachedPrice;
+    U32  cachedLitLength;
+    const BYTE* cachedLiterals;
+    ZSTDv06_stats_t stats;
+} seqStore_t;
+
+void ZSTDv06_seqToCodes(const seqStore_t* seqStorePtr, size_t const nbSeq);
+
+
+#endif   /* ZSTDv06_CCOMMON_H_MODULE */
+/* ******************************************************************
+   FSE : Finite State Entropy codec
+   Public Prototypes declaration
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef FSEv06_H
+#define FSEv06_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/*-****************************************
+*  FSE simple functions
+******************************************/
+/*! FSEv06_decompress():
+    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstCapacity'.
+    @return : size of regenerated data (<= maxDstSize),
+              or an error code, which can be tested using FSEv06_isError() .
+
+    ** Important ** : FSEv06_decompress() does not decompress non-compressible nor RLE data !!!
+    Why ? : making this distinction requires a header.
+    Header management is intentionally delegated to the user layer, which can better manage special cases.
+*/
+size_t FSEv06_decompress(void* dst,  size_t dstCapacity,
+                const void* cSrc, size_t cSrcSize);
+
+
+/*-*****************************************
+*  Tool functions
+******************************************/
+size_t FSEv06_compressBound(size_t size);       /* maximum compressed size */
+
+/* Error Management */
+unsigned    FSEv06_isError(size_t code);        /* tells if a return value is an error code */
+const char* FSEv06_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
+
+
+
+/*-*****************************************
+*  FSE detailed API
+******************************************/
+/*!
+
+FSEv06_decompress() does the following:
+1. read normalized counters with readNCount()
+2. build decoding table 'DTable' from normalized counters
+3. decode the data stream using decoding table 'DTable'
+
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and provide normalized distribution using external method.
+*/
+
+
+/* *** DECOMPRESSION *** */
+
+/*! FSEv06_readNCount():
+    Read compactly saved 'normalizedCounter' from 'rBuffer'.
+    @return : size read from 'rBuffer',
+              or an errorCode, which can be tested using FSEv06_isError().
+              maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+size_t FSEv06_readNCount (short* normalizedCounter, unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, const void* rBuffer, size_t rBuffSize);
+
+/*! Constructor and Destructor of FSEv06_DTable.
+    Note that its size depends on 'tableLog' */
+typedef unsigned FSEv06_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+FSEv06_DTable* FSEv06_createDTable(unsigned tableLog);
+void        FSEv06_freeDTable(FSEv06_DTable* dt);
+
+/*! FSEv06_buildDTable():
+    Builds 'dt', which must be already allocated, using FSEv06_createDTable().
+    return : 0, or an errorCode, which can be tested using FSEv06_isError() */
+size_t FSEv06_buildDTable (FSEv06_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSEv06_decompress_usingDTable():
+    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
+    into `dst` which must be already allocated.
+    @return : size of regenerated data (necessarily <= `dstCapacity`),
+              or an errorCode, which can be tested using FSEv06_isError() */
+size_t FSEv06_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSEv06_DTable* dt);
+
+/*!
+Tutorial :
+----------
+(Note : these functions only decompress FSE-compressed blocks.
+ If block is uncompressed, use memcpy() instead
+ If block is a single repeated byte, use memset() instead )
+
+The first step is to obtain the normalized frequencies of symbols.
+This can be performed by FSEv06_readNCount() if it was saved using FSEv06_writeNCount().
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
+In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
+or size the table to handle worst case situations (typically 256).
+FSEv06_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
+The result of FSEv06_readNCount() is the number of bytes read from 'rBuffer'.
+Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
+If there is an error, the function will return an error code, which can be tested using FSEv06_isError().
+
+The next step is to build the decompression tables 'FSEv06_DTable' from 'normalizedCounter'.
+This is performed by the function FSEv06_buildDTable().
+The space required by 'FSEv06_DTable' must be already allocated using FSEv06_createDTable().
+If there is an error, the function will return an error code, which can be tested using FSEv06_isError().
+
+`FSEv06_DTable` can then be used to decompress `cSrc`, with FSEv06_decompress_usingDTable().
+`cSrcSize` must be strictly correct, otherwise decompression will fail.
+FSEv06_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
+If there is an error, the function will return an error code, which can be tested using FSEv06_isError(). (ex: dst buffer too small)
+*/
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSEv06_H */
+/* ******************************************************************
+   bitstream
+   Part of FSE library
+   header file (to include)
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which must be inlined for best performance.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+
+/*=========================================
+*  Target specific
+=========================================*/
+#if defined(__BMI__) && defined(__GNUC__)
+#  include <immintrin.h>   /* support for bextr (experimental) */
+#endif
+
+
+
+/*-********************************************
+*  bitStream decoding API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BITv06_DStream_t;
+
+typedef enum { BITv06_DStream_unfinished = 0,
+               BITv06_DStream_endOfBuffer = 1,
+               BITv06_DStream_completed = 2,
+               BITv06_DStream_overflow = 3 } BITv06_DStream_status;  /* result of BITv06_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BITv06_initDStream(BITv06_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BITv06_readBits(BITv06_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BITv06_DStream_status BITv06_reloadDStream(BITv06_DStream_t* bitD);
+MEM_STATIC unsigned BITv06_endOfDStream(const BITv06_DStream_t* bitD);
+
+
+
+/*-****************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BITv06_readBitsFast(BITv06_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/*-**************************************************************
+*  Internal functions
+****************************************************************/
+MEM_STATIC unsigned BITv06_highbit32 ( U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    unsigned r;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    r = DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+    return r;
+#   endif
+}
+
+
+
+/*-********************************************************
+* bitStream decoding
+**********************************************************/
+/*! BITv06_initDStream() :
+*   Initialize a BITv06_DStream_t.
+*   `bitD` : a pointer to an already allocated BITv06_DStream_t structure.
+*   `srcSize` must be the *exact* size of the bitStream, in bytes.
+*   @return : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BITv06_initDStream(BITv06_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(bitD->bitContainer)) {  /* normal case */
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+          if (lastByte == 0) return ERROR(GENERIC);   /* endMark not present */
+          bitD->bitsConsumed = 8 - BITv06_highbit32(lastByte); }
+    } else {
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);/* fall-through */
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);/* fall-through */
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);/* fall-through */
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24; /* fall-through */
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16; /* fall-through */
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8; /* fall-through */
+            default: break;
+        }
+        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+          if (lastByte == 0) return ERROR(GENERIC);   /* endMark not present */
+          bitD->bitsConsumed = 8 - BITv06_highbit32(lastByte); }
+        bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+
+ MEM_STATIC size_t BITv06_lookBits(const BITv06_DStream_t* bitD, U32 nbBits)
+{
+    U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BITv06_lookBitsFast() :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv06_lookBitsFast(const BITv06_DStream_t* bitD, U32 nbBits)
+{
+    U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BITv06_skipBits(BITv06_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BITv06_readBits(BITv06_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BITv06_lookBits(bitD, nbBits);
+    BITv06_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*! BITv06_readBitsFast() :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv06_readBitsFast(BITv06_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BITv06_lookBitsFast(bitD, nbBits);
+    BITv06_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BITv06_DStream_status BITv06_reloadDStream(BITv06_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should never happen */
+        return BITv06_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer)) {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BITv06_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start) {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BITv06_DStream_endOfBuffer;
+        return BITv06_DStream_completed;
+    }
+    {   U32 nbBytes = bitD->bitsConsumed >> 3;
+        BITv06_DStream_status result = BITv06_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start) {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BITv06_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BITv06_endOfDStream() :
+*   @return Tells if DStream has exactly reached its end (all bits consumed).
+*/
+MEM_STATIC unsigned BITv06_endOfDStream(const BITv06_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
+/* ******************************************************************
+   FSE : Finite State Entropy coder
+   header file for static linking (only)
+   Copyright (C) 2013-2015, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+   - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef FSEv06_STATIC_H
+#define FSEv06_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* *****************************************
+*  Static allocation
+*******************************************/
+/* FSE buffer bounds */
+#define FSEv06_NCOUNTBOUND 512
+#define FSEv06_BLOCKBOUND(size) (size + (size>>7))
+#define FSEv06_COMPRESSBOUND(size) (FSEv06_NCOUNTBOUND + FSEv06_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* It is possible to statically allocate FSE CTable/DTable as a table of unsigned using below macros */
+#define FSEv06_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/* *****************************************
+*  FSE advanced API
+*******************************************/
+size_t FSEv06_countFast(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+/* same as FSEv06_count(), but blindly trusts that all byte values within src are <= *maxSymbolValuePtr  */
+
+size_t FSEv06_buildDTable_raw (FSEv06_DTable* dt, unsigned nbBits);
+/* build a fake FSEv06_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+size_t FSEv06_buildDTable_rle (FSEv06_DTable* dt, unsigned char symbolValue);
+/* build a fake FSEv06_DTable, designed to always generate the same symbolValue */
+
+
+/* *****************************************
+*  FSE symbol decompression API
+*******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSEv06_DState_t;
+
+
+static void     FSEv06_initDState(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD, const FSEv06_DTable* dt);
+
+static unsigned char FSEv06_decodeSymbol(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD);
+
+
+/* *****************************************
+*  FSE unsafe API
+*******************************************/
+static unsigned char FSEv06_decodeSymbolFast(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* *****************************************
+*  Implementation of inlined functions
+*******************************************/
+
+
+/* ======    Decompression    ====== */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSEv06_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSEv06_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSEv06_initDState(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD, const FSEv06_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv06_DTableHeader* const DTableH = (const FSEv06_DTableHeader*)ptr;
+    DStatePtr->state = BITv06_readBits(bitD, DTableH->tableLog);
+    BITv06_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSEv06_peekSymbol(const FSEv06_DState_t* DStatePtr)
+{
+    FSEv06_decode_t const DInfo = ((const FSEv06_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    return DInfo.symbol;
+}
+
+MEM_STATIC void FSEv06_updateState(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD)
+{
+    FSEv06_decode_t const DInfo = ((const FSEv06_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    size_t const lowBits = BITv06_readBits(bitD, nbBits);
+    DStatePtr->state = DInfo.newState + lowBits;
+}
+
+MEM_STATIC BYTE FSEv06_decodeSymbol(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD)
+{
+    FSEv06_decode_t const DInfo = ((const FSEv06_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BITv06_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+/*! FSEv06_decodeSymbolFast() :
+    unsafe, only works if no symbol has a probability > 50% */
+MEM_STATIC BYTE FSEv06_decodeSymbolFast(FSEv06_DState_t* DStatePtr, BITv06_DStream_t* bitD)
+{
+    FSEv06_decode_t const DInfo = ((const FSEv06_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BITv06_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+
+
+#ifndef FSEv06_COMMONDEFS_ONLY
+
+/* **************************************************************
+*  Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSEv06_MAX_MEMORY_USAGE 14
+#define FSEv06_DEFAULT_MEMORY_USAGE 13
+
+/*!FSEv06_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSEv06_MAX_SYMBOL_VALUE 255
+
+
+/* **************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSEv06_FUNCTION_TYPE BYTE
+#define FSEv06_FUNCTION_EXTENSION
+#define FSEv06_DECODE_TYPE FSEv06_decode_t
+
+
+#endif   /* !FSEv06_COMMONDEFS_ONLY */
+
+
+/* ***************************************************************
+*  Constants
+*****************************************************************/
+#define FSEv06_MAX_TABLELOG  (FSEv06_MAX_MEMORY_USAGE-2)
+#define FSEv06_MAX_TABLESIZE (1U<<FSEv06_MAX_TABLELOG)
+#define FSEv06_MAXTABLESIZE_MASK (FSEv06_MAX_TABLESIZE-1)
+#define FSEv06_DEFAULT_TABLELOG (FSEv06_DEFAULT_MEMORY_USAGE-2)
+#define FSEv06_MIN_TABLELOG 5
+
+#define FSEv06_TABLELOG_ABSOLUTE_MAX 15
+#if FSEv06_MAX_TABLELOG > FSEv06_TABLELOG_ABSOLUTE_MAX
+#error "FSEv06_MAX_TABLELOG > FSEv06_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+#define FSEv06_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSEv06_STATIC_H */
+/*
+   Common functions of New Generation Entropy library
+   Copyright (C) 2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+*************************************************************************** */
+
+
+/*-****************************************
+*  FSE Error Management
+******************************************/
+unsigned FSEv06_isError(size_t code) { return ERR_isError(code); }
+
+const char* FSEv06_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/* **************************************************************
+*  HUF Error Management
+****************************************************************/
+unsigned HUFv06_isError(size_t code) { return ERR_isError(code); }
+
+const char* HUFv06_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/*-**************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+static short FSEv06_abs(short a) { return a<0 ? -a : a; }
+
+size_t FSEv06_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSEv06_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSEv06_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr)) {
+        if (previous0) {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF) {
+                n0+=24;
+                if (ip < iend-5) {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                } else {
+                    bitStream >>= 16;
+                    bitCount+=16;
+            }   }
+            while ((bitStream & 3) == 3) {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {   short const max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max) {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            } else {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSEv06_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold) {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+            } else {
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+    }   }   /* while ((remaining>1) && (charnum<=*maxSVPtr)) */
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+/* ******************************************************************
+   FSE : Finite State Entropy decoder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSEv06_isError ERR_isError
+#define FSEv06_STATIC_ASSERT(c) { enum { FSEv06_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSEv06_DTABLE_SIZE_U32(FSEv06_MAX_TABLELOG)];
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSEv06_FUNCTION_EXTENSION
+#  error "FSEv06_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSEv06_FUNCTION_TYPE
+#  error "FSEv06_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSEv06_CAT(X,Y) X##Y
+#define FSEv06_FUNCTION_NAME(X,Y) FSEv06_CAT(X,Y)
+#define FSEv06_TYPE_NAME(X,Y) FSEv06_CAT(X,Y)
+
+
+/* Function templates */
+FSEv06_DTable* FSEv06_createDTable (unsigned tableLog)
+{
+    if (tableLog > FSEv06_TABLELOG_ABSOLUTE_MAX) tableLog = FSEv06_TABLELOG_ABSOLUTE_MAX;
+    return (FSEv06_DTable*)malloc( FSEv06_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
+}
+
+void FSEv06_freeDTable (FSEv06_DTable* dt)
+{
+    free(dt);
+}
+
+size_t FSEv06_buildDTable(FSEv06_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
+    FSEv06_DECODE_TYPE* const tableDecode = (FSEv06_DECODE_TYPE*) (tdPtr);
+    U16 symbolNext[FSEv06_MAX_SYMBOL_VALUE+1];
+
+    U32 const maxSV1 = maxSymbolValue + 1;
+    U32 const tableSize = 1 << tableLog;
+    U32 highThreshold = tableSize-1;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSEv06_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSEv06_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    {   FSEv06_DTableHeader DTableH;
+        DTableH.tableLog = (U16)tableLog;
+        DTableH.fastMode = 1;
+        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
+            U32 s;
+            for (s=0; s<maxSV1; s++) {
+                if (normalizedCounter[s]==-1) {
+                    tableDecode[highThreshold--].symbol = (FSEv06_FUNCTION_TYPE)s;
+                    symbolNext[s] = 1;
+                } else {
+                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+                    symbolNext[s] = normalizedCounter[s];
+        }   }   }
+        memcpy(dt, &DTableH, sizeof(DTableH));
+    }
+
+    /* Spread symbols */
+    {   U32 const tableMask = tableSize-1;
+        U32 const step = FSEv06_TABLESTEP(tableSize);
+        U32 s, position = 0;
+        for (s=0; s<maxSV1; s++) {
+            int i;
+            for (i=0; i<normalizedCounter[s]; i++) {
+                tableDecode[position].symbol = (FSEv06_FUNCTION_TYPE)s;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }   }
+
+        if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+    }
+
+    /* Build Decoding table */
+    {   U32 u;
+        for (u=0; u<tableSize; u++) {
+            FSEv06_FUNCTION_TYPE const symbol = (FSEv06_FUNCTION_TYPE)(tableDecode[u].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[u].nbBits = (BYTE) (tableLog - BITv06_highbit32 ((U32)nextState) );
+            tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+    }   }
+
+    return 0;
+}
+
+
+
+#ifndef FSEv06_COMMONDEFS_ONLY
+
+/*-*******************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+size_t FSEv06_buildDTable_rle (FSEv06_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSEv06_DTableHeader* const DTableH = (FSEv06_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv06_decode_t* const cell = (FSEv06_decode_t*)dPtr;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+size_t FSEv06_buildDTable_raw (FSEv06_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSEv06_DTableHeader* const DTableH = (FSEv06_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv06_decode_t* const dinfo = (FSEv06_decode_t*)dPtr;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSV1 = tableMask+1;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<maxSV1; s++) {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSEv06_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSEv06_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BITv06_DStream_t bitD;
+    FSEv06_DState_t state1;
+    FSEv06_DState_t state2;
+
+    /* Init */
+    { size_t const errorCode = BITv06_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+      if (FSEv06_isError(errorCode)) return errorCode; }
+
+    FSEv06_initDState(&state1, &bitD, dt);
+    FSEv06_initDState(&state2, &bitD, dt);
+
+#define FSEv06_GETSYMBOL(statePtr) fast ? FSEv06_decodeSymbolFast(statePtr, &bitD) : FSEv06_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BITv06_reloadDStream(&bitD)==BITv06_DStream_unfinished) && (op<olimit) ; op+=4) {
+        op[0] = FSEv06_GETSYMBOL(&state1);
+
+        if (FSEv06_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv06_reloadDStream(&bitD);
+
+        op[1] = FSEv06_GETSYMBOL(&state2);
+
+        if (FSEv06_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BITv06_reloadDStream(&bitD) > BITv06_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSEv06_GETSYMBOL(&state1);
+
+        if (FSEv06_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv06_reloadDStream(&bitD);
+
+        op[3] = FSEv06_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BITv06_reloadDStream(&bitD) >= FSEv06_DStream_partiallyFilled; Ends at exactly BITv06_DStream_completed */
+    while (1) {
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+
+        *op++ = FSEv06_GETSYMBOL(&state1);
+
+        if (BITv06_reloadDStream(&bitD)==BITv06_DStream_overflow) {
+            *op++ = FSEv06_GETSYMBOL(&state2);
+            break;
+        }
+
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+
+        *op++ = FSEv06_GETSYMBOL(&state2);
+
+        if (BITv06_reloadDStream(&bitD)==BITv06_DStream_overflow) {
+            *op++ = FSEv06_GETSYMBOL(&state1);
+            break;
+    }   }
+
+    return op-ostart;
+}
+
+
+size_t FSEv06_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSEv06_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv06_DTableHeader* DTableH = (const FSEv06_DTableHeader*)ptr;
+    const U32 fastMode = DTableH->fastMode;
+
+    /* select fast mode (static) */
+    if (fastMode) return FSEv06_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSEv06_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+size_t FSEv06_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSEv06_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSEv06_MAX_SYMBOL_VALUE;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSE decoding mode */
+    {   size_t const NCountLength = FSEv06_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+        if (FSEv06_isError(NCountLength)) return NCountLength;
+        if (NCountLength >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
+
+    { size_t const errorCode = FSEv06_buildDTable (dt, counting, maxSymbolValue, tableLog);
+      if (FSEv06_isError(errorCode)) return errorCode; }
+
+    return FSEv06_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);   /* always return, even if it is an error code */
+}
+
+
+
+#endif   /* FSEv06_COMMONDEFS_ONLY */
+/* ******************************************************************
+   Huffman coder, part of New Generation Entropy library
+   header file
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef HUFv06_H
+#define HUFv06_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************************
+*  HUF simple functions
+******************************************/
+size_t HUFv06_decompress(void* dst,  size_t dstSize,
+                const void* cSrc, size_t cSrcSize);
+/*
+HUFv06_decompress() :
+    Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstSize'.
+    `dstSize` : must be the **exact** size of original (uncompressed) data.
+    Note : in contrast with FSE, HUFv06_decompress can regenerate
+           RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
+           because it knows size to regenerate.
+    @return : size of regenerated data (== dstSize)
+              or an error code, which can be tested using HUFv06_isError()
+*/
+
+
+/* ****************************************
+*  Tool functions
+******************************************/
+size_t HUFv06_compressBound(size_t size);       /**< maximum compressed size */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* HUFv06_H */
+/* ******************************************************************
+   Huffman codec, part of New Generation Entropy library
+   header file, for static linking only
+   Copyright (C) 2013-2016, Yann Collet
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef HUFv06_STATIC_H
+#define HUFv06_STATIC_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************************
+*  Static allocation
+******************************************/
+/* HUF buffer bounds */
+#define HUFv06_CTABLEBOUND 129
+#define HUFv06_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true if incompressible pre-filtered with fast heuristic */
+#define HUFv06_COMPRESSBOUND(size) (HUFv06_CTABLEBOUND + HUFv06_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
+
+/* static allocation of HUF's DTable */
+#define HUFv06_DTABLE_SIZE(maxTableLog)   (1 + (1<<maxTableLog))
+#define HUFv06_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        unsigned short DTable[HUFv06_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUFv06_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        unsigned int DTable[HUFv06_DTABLE_SIZE(maxTableLog)] = { maxTableLog }
+#define HUFv06_CREATE_STATIC_DTABLEX6(DTable, maxTableLog) \
+        unsigned int DTable[HUFv06_DTABLE_SIZE(maxTableLog) * 3 / 2] = { maxTableLog }
+
+
+/* ****************************************
+*  Advanced decompression functions
+******************************************/
+size_t HUFv06_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUFv06_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbols decoder */
+
+
+
+/*!
+HUFv06_decompress() does the following:
+1. select the decompression algorithm (X2, X4, X6) based on pre-computed heuristics
+2. build Huffman table from save, using HUFv06_readDTableXn()
+3. decode 1 or 4 segments in parallel using HUFv06_decompressSXn_usingDTable
+*/
+size_t HUFv06_readDTableX2 (unsigned short* DTable, const void* src, size_t srcSize);
+size_t HUFv06_readDTableX4 (unsigned* DTable, const void* src, size_t srcSize);
+
+size_t HUFv06_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned short* DTable);
+size_t HUFv06_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned* DTable);
+
+
+/* single stream variants */
+size_t HUFv06_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUFv06_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+
+size_t HUFv06_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned short* DTable);
+size_t HUFv06_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const unsigned* DTable);
+
+
+
+/* **************************************************************
+*  Constants
+****************************************************************/
+#define HUFv06_ABSOLUTEMAX_TABLELOG  16   /* absolute limit of HUFv06_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUFv06_MAX_TABLELOG  12           /* max configured tableLog (for static allocation); can be modified up to HUFv06_ABSOLUTEMAX_TABLELOG */
+#define HUFv06_DEFAULT_TABLELOG  HUFv06_MAX_TABLELOG   /* tableLog by default, when not specified */
+#define HUFv06_MAX_SYMBOL_VALUE 255
+#if (HUFv06_MAX_TABLELOG > HUFv06_ABSOLUTEMAX_TABLELOG)
+#  error "HUFv06_MAX_TABLELOG is too large !"
+#endif
+
+
+
+/*! HUFv06_readStats() :
+    Read compact Huffman tree, saved by HUFv06_writeCTable().
+    `huffWeight` is destination buffer.
+    @return : size read from `src`
+*/
+MEM_STATIC size_t HUFv06_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                            U32* nbSymbolsPtr, U32* tableLogPtr,
+                            const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  { /* special header */
+        if (iSize >= (242)) {  /* RLE */
+            static U32 l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else {   /* Incompressible */
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            {   U32 n;
+                for (n=0; n<oSize; n+=2) {
+                    huffWeight[n]   = ip[n/2] >> 4;
+                    huffWeight[n+1] = ip[n/2] & 15;
+    }   }   }   }
+    else  {   /* header compressed with FSE (normal case) */
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSEv06_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSEv06_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUFv06_ABSOLUTEMAX_TABLELOG + 1) * sizeof(U32));
+    weightTotal = 0;
+    {   U32 n; for (n=0; n<oSize; n++) {
+            if (huffWeight[n] >= HUFv06_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+            rankStats[huffWeight[n]]++;
+            weightTotal += (1 << huffWeight[n]) >> 1;
+    }   }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    {   U32 const tableLog = BITv06_highbit32(weightTotal) + 1;
+        if (tableLog > HUFv06_ABSOLUTEMAX_TABLELOG) return ERROR(corruption_detected);
+        *tableLogPtr = tableLog;
+        /* determine last weight */
+        {   U32 const total = 1 << tableLog;
+            U32 const rest = total - weightTotal;
+            U32 const verif = 1 << BITv06_highbit32(rest);
+            U32 const lastWeight = BITv06_highbit32(rest) + 1;
+            if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+            huffWeight[oSize] = (BYTE)lastWeight;
+            rankStats[lastWeight]++;
+    }   }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    return iSize+1;
+}
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* HUFv06_STATIC_H */
+/* ******************************************************************
+   Huffman decoder, part of New Generation Entropy library
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define HUFv06_STATIC_ASSERT(c) { enum { HUFv06_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+
+/* *******************************************************
+*  HUF : Huffman block decompression
+*********************************************************/
+typedef struct { BYTE byte; BYTE nbBits; } HUFv06_DEltX2;   /* single-symbol decoding */
+
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUFv06_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+
+
+/*-***************************/
+/*  single-symbol decoding   */
+/*-***************************/
+
+size_t HUFv06_readDTableX2 (U16* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUFv06_MAX_SYMBOL_VALUE + 1];
+    U32 rankVal[HUFv06_ABSOLUTEMAX_TABLELOG + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    size_t iSize;
+    U32 nbSymbols = 0;
+    U32 n;
+    U32 nextRankStart;
+    void* const dtPtr = DTable + 1;
+    HUFv06_DEltX2* const dt = (HUFv06_DEltX2*)dtPtr;
+
+    HUFv06_STATIC_ASSERT(sizeof(HUFv06_DEltX2) == sizeof(U16));   /* if compilation fails here, assertion is false */
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv06_readStats(huffWeight, HUFv06_MAX_SYMBOL_VALUE + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv06_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > DTable[0]) return ERROR(tableLog_tooLarge);   /* DTable is too small */
+    DTable[0] = (U16)tableLog;   /* maybe should separate sizeof allocated DTable, from used size of DTable, in case of re-use */
+
+    /* Prepare ranks */
+    nextRankStart = 0;
+    for (n=1; n<tableLog+1; n++) {
+        U32 current = nextRankStart;
+        nextRankStart += (rankVal[n] << (n-1));
+        rankVal[n] = current;
+    }
+
+    /* fill DTable */
+    for (n=0; n<nbSymbols; n++) {
+        const U32 w = huffWeight[n];
+        const U32 length = (1 << w) >> 1;
+        U32 i;
+        HUFv06_DEltX2 D;
+        D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+        for (i = rankVal[w]; i < rankVal[w] + length; i++)
+            dt[i] = D;
+        rankVal[w] += length;
+    }
+
+    return iSize;
+}
+
+
+static BYTE HUFv06_decodeSymbolX2(BITv06_DStream_t* Dstream, const HUFv06_DEltX2* dt, const U32 dtLog)
+{
+    const size_t val = BITv06_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+    const BYTE c = dt[val].byte;
+    BITv06_skipBits(Dstream, dt[val].nbBits);
+    return c;
+}
+
+#define HUFv06_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUFv06_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUFv06_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv06_MAX_TABLELOG<=12)) \
+        HUFv06_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUFv06_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUFv06_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUFv06_decodeStreamX2(BYTE* p, BITv06_DStream_t* const bitDPtr, BYTE* const pEnd, const HUFv06_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BITv06_reloadDStream(bitDPtr) == BITv06_DStream_unfinished) && (p <= pEnd-4)) {
+        HUFv06_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BITv06_reloadDStream(bitDPtr) == BITv06_DStream_unfinished) && (p < pEnd))
+        HUFv06_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUFv06_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+size_t HUFv06_decompress1X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + dstSize;
+    const U32 dtLog = DTable[0];
+    const void* dtPtr = DTable;
+    const HUFv06_DEltX2* const dt = ((const HUFv06_DEltX2*)dtPtr)+1;
+    BITv06_DStream_t bitD;
+
+    { size_t const errorCode = BITv06_initDStream(&bitD, cSrc, cSrcSize);
+      if (HUFv06_isError(errorCode)) return errorCode; }
+
+    HUFv06_decodeStreamX2(op, &bitD, oend, dt, dtLog);
+
+    /* check */
+    if (!BITv06_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    return dstSize;
+}
+
+size_t HUFv06_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv06_CREATE_STATIC_DTABLEX2(DTable, HUFv06_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const errorCode = HUFv06_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUFv06_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUFv06_decompress1X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+size_t HUFv06_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U16* DTable)
+{
+    /* Check */
+    if (cSrcSize < 10) return ERROR(corruption_detected);  /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable;
+        const HUFv06_DEltX2* const dt = ((const HUFv06_DEltX2*)dtPtr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BITv06_DStream_t bitD1;
+        BITv06_DStream_t bitD2;
+        BITv06_DStream_t bitD3;
+        BITv06_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BITv06_initDStream(&bitD1, istart1, length1);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD2, istart2, length2);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD3, istart3, length3);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD4, istart4, length4);
+        if (HUFv06_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BITv06_reloadDStream(&bitD1) | BITv06_reloadDStream(&bitD2) | BITv06_reloadDStream(&bitD3) | BITv06_reloadDStream(&bitD4);
+        for ( ; (endSignal==BITv06_DStream_unfinished) && (op4<(oend-7)) ; ) {
+            HUFv06_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX2_0(op4, &bitD4);
+            endSignal = BITv06_reloadDStream(&bitD1) | BITv06_reloadDStream(&bitD2) | BITv06_reloadDStream(&bitD3) | BITv06_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUFv06_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUFv06_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUFv06_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUFv06_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BITv06_endOfDStream(&bitD1) & BITv06_endOfDStream(&bitD2) & BITv06_endOfDStream(&bitD3) & BITv06_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+size_t HUFv06_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv06_CREATE_STATIC_DTABLEX2(DTable, HUFv06_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const errorCode = HUFv06_readDTableX2 (DTable, cSrc, cSrcSize);
+    if (HUFv06_isError(errorCode)) return errorCode;
+    if (errorCode >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += errorCode;
+    cSrcSize -= errorCode;
+
+    return HUFv06_decompress4X2_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+/* *************************/
+/* double-symbols decoding */
+/* *************************/
+
+static void HUFv06_fillDTableX4Level2(HUFv06_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUFv06_DEltX4 DElt;
+    U32 rankVal[HUFv06_ABSOLUTEMAX_TABLELOG + 1];
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1) {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    { U32 s; for (s=0; s<sortedListSize; s++) {   /* note : sortedSymbols already skipped */
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }}
+}
+
+typedef U32 rankVal_t[HUFv06_ABSOLUTEMAX_TABLELOG][HUFv06_ABSOLUTEMAX_TABLELOG + 1];
+
+static void HUFv06_fillDTableX4(HUFv06_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUFv06_ABSOLUTEMAX_TABLELOG + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++) {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits) {   /* enough room for a second symbol */
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUFv06_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        } else {
+            HUFv06_DEltX4 DElt;
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits = (BYTE)(nbBits);
+            DElt.length = 1;
+            {   U32 u;
+                const U32 end = start + length;
+                for (u = start; u < end; u++) DTable[u] = DElt;
+        }   }
+        rankVal[weight] += length;
+    }
+}
+
+size_t HUFv06_readDTableX4 (U32* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUFv06_MAX_SYMBOL_VALUE + 1];
+    sortedSymbol_t sortedSymbol[HUFv06_MAX_SYMBOL_VALUE + 1];
+    U32 rankStats[HUFv06_ABSOLUTEMAX_TABLELOG + 1] = { 0 };
+    U32 rankStart0[HUFv06_ABSOLUTEMAX_TABLELOG + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    const U32 memLog = DTable[0];
+    size_t iSize;
+    void* dtPtr = DTable;
+    HUFv06_DEltX4* const dt = ((HUFv06_DEltX4*)dtPtr) + 1;
+
+    HUFv06_STATIC_ASSERT(sizeof(HUFv06_DEltX4) == sizeof(U32));   /* if compilation fails here, assertion is false */
+    if (memLog > HUFv06_ABSOLUTEMAX_TABLELOG) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv06_readStats(weightList, HUFv06_MAX_SYMBOL_VALUE + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv06_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > memLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+
+    /* Get start index of each weight */
+    {   U32 w, nextRankStart = 0;
+        for (w=1; w<maxW+1; w++) {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {   U32 s;
+        for (s=0; s<nbSymbols; s++) {
+            U32 const w = weightList[s];
+            U32 const r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {   U32* const rankVal0 = rankVal[0];
+        {   int const rescale = (memLog-tableLog) - 1;   /* tableLog <= memLog */
+            U32 nextRankVal = 0;
+            U32 w;
+            for (w=1; w<maxW+1; w++) {
+                U32 current = nextRankVal;
+                nextRankVal += rankStats[w] << (w+rescale);
+                rankVal0[w] = current;
+        }   }
+        {   U32 const minBits = tableLog+1 - maxW;
+            U32 consumed;
+            for (consumed = minBits; consumed < memLog - minBits + 1; consumed++) {
+                U32* const rankValPtr = rankVal[consumed];
+                U32 w;
+                for (w = 1; w < maxW+1; w++) {
+                    rankValPtr[w] = rankVal0[w] >> consumed;
+    }   }   }   }
+
+    HUFv06_fillDTableX4(dt, memLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    return iSize;
+}
+
+
+static U32 HUFv06_decodeSymbolX4(void* op, BITv06_DStream_t* DStream, const HUFv06_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv06_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BITv06_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUFv06_decodeLastSymbolX4(void* op, BITv06_DStream_t* DStream, const HUFv06_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv06_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BITv06_skipBits(DStream, dt[val].nbBits);
+    else {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
+            BITv06_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+    }   }
+    return 1;
+}
+
+
+#define HUFv06_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUFv06_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv06_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv06_MAX_TABLELOG<=12)) \
+        ptr += HUFv06_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv06_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUFv06_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUFv06_decodeStreamX4(BYTE* p, BITv06_DStream_t* bitDPtr, BYTE* const pEnd, const HUFv06_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BITv06_reloadDStream(bitDPtr) == BITv06_DStream_unfinished) && (p < pEnd-7)) {
+        HUFv06_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv06_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BITv06_reloadDStream(bitDPtr) == BITv06_DStream_unfinished) && (p <= pEnd-2))
+        HUFv06_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUFv06_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUFv06_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+
+size_t HUFv06_decompress1X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    const BYTE* const istart = (const BYTE*) cSrc;
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* const oend = ostart + dstSize;
+
+    const U32 dtLog = DTable[0];
+    const void* const dtPtr = DTable;
+    const HUFv06_DEltX4* const dt = ((const HUFv06_DEltX4*)dtPtr) +1;
+
+    /* Init */
+    BITv06_DStream_t bitD;
+    { size_t const errorCode = BITv06_initDStream(&bitD, istart, cSrcSize);
+      if (HUFv06_isError(errorCode)) return errorCode; }
+
+    /* decode */
+    HUFv06_decodeStreamX4(ostart, &bitD, oend, dt, dtLog);
+
+    /* check */
+    if (!BITv06_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    /* decoded size */
+    return dstSize;
+}
+
+size_t HUFv06_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv06_CREATE_STATIC_DTABLEX4(DTable, HUFv06_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUFv06_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUFv06_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUFv06_decompress1X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+size_t HUFv06_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const U32* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable;
+        const HUFv06_DEltX4* const dt = ((const HUFv06_DEltX4*)dtPtr) +1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+
+        /* Init */
+        BITv06_DStream_t bitD1;
+        BITv06_DStream_t bitD2;
+        BITv06_DStream_t bitD3;
+        BITv06_DStream_t bitD4;
+        const size_t length1 = MEM_readLE16(istart);
+        const size_t length2 = MEM_readLE16(istart+2);
+        const size_t length3 = MEM_readLE16(istart+4);
+        size_t length4;
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+
+        length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        errorCode = BITv06_initDStream(&bitD1, istart1, length1);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD2, istart2, length2);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD3, istart3, length3);
+        if (HUFv06_isError(errorCode)) return errorCode;
+        errorCode = BITv06_initDStream(&bitD4, istart4, length4);
+        if (HUFv06_isError(errorCode)) return errorCode;
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BITv06_reloadDStream(&bitD1) | BITv06_reloadDStream(&bitD2) | BITv06_reloadDStream(&bitD3) | BITv06_reloadDStream(&bitD4);
+        for ( ; (endSignal==BITv06_DStream_unfinished) && (op4<(oend-7)) ; ) {
+            HUFv06_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv06_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUFv06_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUFv06_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUFv06_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BITv06_reloadDStream(&bitD1) | BITv06_reloadDStream(&bitD2) | BITv06_reloadDStream(&bitD3) | BITv06_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUFv06_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUFv06_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUFv06_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUFv06_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BITv06_endOfDStream(&bitD1) & BITv06_endOfDStream(&bitD2) & BITv06_endOfDStream(&bitD3) & BITv06_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+size_t HUFv06_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv06_CREATE_STATIC_DTABLEX4(DTable, HUFv06_MAX_TABLELOG);
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUFv06_readDTableX4 (DTable, cSrc, cSrcSize);
+    if (HUFv06_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize;
+    cSrcSize -= hSize;
+
+    return HUFv06_decompress4X4_usingDTable (dst, dstSize, ip, cSrcSize, DTable);
+}
+
+
+
+
+/* ********************************/
+/* Generic decompression selector */
+/* ********************************/
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+size_t HUFv06_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[3] = { HUFv06_decompress4X2, HUFv06_decompress4X4, NULL };
+    U32 Dtime[3];   /* decompression time estimation */
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    /* decoder timing evaluation */
+    {   U32 const Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+        U32 const D256 = (U32)(dstSize >> 8);
+        U32 n; for (n=0; n<3; n++)
+            Dtime[n] = algoTime[Q][n].tableTime + (algoTime[Q][n].decode256Time * D256);
+    }
+
+    Dtime[1] += Dtime[1] >> 4; Dtime[2] += Dtime[2] >> 3; /* advantage to algorithms using less memory, for cache eviction */
+
+    {   U32 algoNb = 0;
+        if (Dtime[1] < Dtime[0]) algoNb = 1;
+        // if (Dtime[2] < Dtime[algoNb]) algoNb = 2;   /* current speed of HUFv06_decompress4X6 is not good */
+        return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+    }
+
+    //return HUFv06_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUFv06_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+    //return HUFv06_decompress4X6(dst, dstSize, cSrc, cSrcSize);   /* multi-streams quad-symbols decoding */
+}
+/*
+    Common functions of Zstd compression library
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net/
+*/
+
+
+/*-****************************************
+*  Version
+******************************************/
+
+/*-****************************************
+*  ZSTD Error Management
+******************************************/
+/*! ZSTDv06_isError() :
+*   tells if a return value is an error code */
+unsigned ZSTDv06_isError(size_t code) { return ERR_isError(code); }
+
+/*! ZSTDv06_getErrorName() :
+*   provides error code string from function result (useful for debugging) */
+const char* ZSTDv06_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/* **************************************************************
+*  ZBUFF Error Management
+****************************************************************/
+unsigned ZBUFFv06_isError(size_t errorCode) { return ERR_isError(errorCode); }
+
+const char* ZBUFFv06_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * HEAPMODE :
+ * Select how default decompression function ZSTDv06_decompress() will allocate memory,
+ * in memory stack (0), or in memory heap (1, requires malloc())
+ */
+#ifndef ZSTDv06_HEAPMODE
+#  define ZSTDv06_HEAPMODE 1
+#endif
+
+
+
+/*-*******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#endif
+
+
+/*-*************************************
+*  Macros
+***************************************/
+#define ZSTDv06_isError ERR_isError   /* for inlining */
+#define FSEv06_isError  ERR_isError
+#define HUFv06_isError  ERR_isError
+
+
+/*_*******************************************************
+*  Memory operations
+**********************************************************/
+static void ZSTDv06_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+
+/*-*************************************************************
+*   Context management
+***************************************************************/
+typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
+               ZSTDds_decodeBlockHeader, ZSTDds_decompressBlock } ZSTDv06_dStage;
+
+struct ZSTDv06_DCtx_s
+{
+    FSEv06_DTable LLTable[FSEv06_DTABLE_SIZE_U32(LLFSELog)];
+    FSEv06_DTable OffTable[FSEv06_DTABLE_SIZE_U32(OffFSELog)];
+    FSEv06_DTable MLTable[FSEv06_DTABLE_SIZE_U32(MLFSELog)];
+    unsigned   hufTableX4[HUFv06_DTABLE_SIZE(HufLog)];
+    const void* previousDstEnd;
+    const void* base;
+    const void* vBase;
+    const void* dictEnd;
+    size_t expected;
+    size_t headerSize;
+    ZSTDv06_frameParams fParams;
+    blockType_t bType;   /* used in ZSTDv06_decompressContinue(), to transfer blockType between header decoding and block decoding stages */
+    ZSTDv06_dStage stage;
+    U32 flagRepeatTable;
+    const BYTE* litPtr;
+    size_t litSize;
+    BYTE litBuffer[ZSTDv06_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH];
+    BYTE headerBuffer[ZSTDv06_FRAMEHEADERSIZE_MAX];
+};  /* typedef'd to ZSTDv06_DCtx within "zstd_static.h" */
+
+size_t ZSTDv06_sizeofDCtx (void) { return sizeof(ZSTDv06_DCtx); }   /* non published interface */
+
+size_t ZSTDv06_decompressBegin(ZSTDv06_DCtx* dctx)
+{
+    dctx->expected = ZSTDv06_frameHeaderSize_min;
+    dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    dctx->vBase = NULL;
+    dctx->dictEnd = NULL;
+    dctx->hufTableX4[0] = HufLog;
+    dctx->flagRepeatTable = 0;
+    return 0;
+}
+
+ZSTDv06_DCtx* ZSTDv06_createDCtx(void)
+{
+    ZSTDv06_DCtx* dctx = (ZSTDv06_DCtx*)malloc(sizeof(ZSTDv06_DCtx));
+    if (dctx==NULL) return NULL;
+    ZSTDv06_decompressBegin(dctx);
+    return dctx;
+}
+
+size_t ZSTDv06_freeDCtx(ZSTDv06_DCtx* dctx)
+{
+    free(dctx);
+    return 0;   /* reserved as a potential error code in the future */
+}
+
+void ZSTDv06_copyDCtx(ZSTDv06_DCtx* dstDCtx, const ZSTDv06_DCtx* srcDCtx)
+{
+    memcpy(dstDCtx, srcDCtx,
+           sizeof(ZSTDv06_DCtx) - (ZSTDv06_BLOCKSIZE_MAX+WILDCOPY_OVERLENGTH + ZSTDv06_frameHeaderSize_max));  /* no need to copy workspace */
+}
+
+
+/*-*************************************************************
+*   Decompression section
+***************************************************************/
+
+/* Frame format description
+   Frame Header -  [ Block Header - Block ] - Frame End
+   1) Frame Header
+      - 4 bytes - Magic Number : ZSTDv06_MAGICNUMBER (defined within zstd_static.h)
+      - 1 byte  - Frame Descriptor
+   2) Block Header
+      - 3 bytes, starting with a 2-bits descriptor
+                 Uncompressed, Compressed, Frame End, unused
+   3) Block
+      See Block Format Description
+   4) Frame End
+      - 3 bytes, compatible with Block Header
+*/
+
+
+/* Frame descriptor
+
+   1 byte, using :
+   bit 0-3 : windowLog - ZSTDv06_WINDOWLOG_ABSOLUTEMIN   (see zstd_internal.h)
+   bit 4   : minmatch 4(0) or 3(1)
+   bit 5   : reserved (must be zero)
+   bit 6-7 : Frame content size : unknown, 1 byte, 2 bytes, 8 bytes
+
+   Optional : content size (0, 1, 2 or 8 bytes)
+   0 : unknown
+   1 : 0-255 bytes
+   2 : 256 - 65535+256
+   8 : up to 16 exa
+*/
+
+
+/* Compressed Block, format description
+
+   Block = Literal Section - Sequences Section
+   Prerequisite : size of (compressed) block, maximum size of regenerated data
+
+   1) Literal Section
+
+   1.1) Header : 1-5 bytes
+        flags: 2 bits
+            00 compressed by Huff0
+            01 unused
+            10 is Raw (uncompressed)
+            11 is Rle
+            Note : using 01 => Huff0 with precomputed table ?
+            Note : delta map ? => compressed ?
+
+   1.1.1) Huff0-compressed literal block : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+   1.1.2) Raw (uncompressed) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RAW<<6) + (0<<4) + size
+               12 bits: (IS_RAW<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RAW<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.3) Rle (repeated single byte) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RLE<<6) + (0<<4) + size
+               12 bits: (IS_RLE<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RLE<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.4) Huff0-compressed literal block, using precomputed CTables : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+        1- CTable available (stored into workspace ?)
+        2- Small input (fast heuristic ? Full comparison ? depend on clevel ?)
+
+
+   1.2) Literal block content
+
+   1.2.1) Huff0 block, using sizes from header
+        See Huff0 format
+
+   1.2.2) Huff0 block, using prepared table
+
+   1.2.3) Raw content
+
+   1.2.4) single byte
+
+
+   2) Sequences section
+      TO DO
+*/
+
+/** ZSTDv06_frameHeaderSize() :
+*   srcSize must be >= ZSTDv06_frameHeaderSize_min.
+*   @return : size of the Frame Header */
+static size_t ZSTDv06_frameHeaderSize(const void* src, size_t srcSize)
+{
+    if (srcSize < ZSTDv06_frameHeaderSize_min) return ERROR(srcSize_wrong);
+    { U32 const fcsId = (((const BYTE*)src)[4]) >> 6;
+      return ZSTDv06_frameHeaderSize_min + ZSTDv06_fcs_fieldSize[fcsId]; }
+}
+
+
+/** ZSTDv06_getFrameParams() :
+*   decode Frame Header, or provide expected `srcSize`.
+*   @return : 0, `fparamsPtr` is correctly filled,
+*            >0, `srcSize` is too small, result is expected `srcSize`,
+*             or an error code, which can be tested using ZSTDv06_isError() */
+size_t ZSTDv06_getFrameParams(ZSTDv06_frameParams* fparamsPtr, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+
+    if (srcSize < ZSTDv06_frameHeaderSize_min) return ZSTDv06_frameHeaderSize_min;
+    if (MEM_readLE32(src) != ZSTDv06_MAGICNUMBER) return ERROR(prefix_unknown);
+
+    /* ensure there is enough `srcSize` to fully read/decode frame header */
+    { size_t const fhsize = ZSTDv06_frameHeaderSize(src, srcSize);
+      if (srcSize < fhsize) return fhsize; }
+
+    memset(fparamsPtr, 0, sizeof(*fparamsPtr));
+    {   BYTE const frameDesc = ip[4];
+        fparamsPtr->windowLog = (frameDesc & 0xF) + ZSTDv06_WINDOWLOG_ABSOLUTEMIN;
+        if ((frameDesc & 0x20) != 0) return ERROR(frameParameter_unsupported);   /* reserved 1 bit */
+        switch(frameDesc >> 6)  /* fcsId */
+        {
+            default:   /* impossible */
+            case 0 : fparamsPtr->frameContentSize = 0; break;
+            case 1 : fparamsPtr->frameContentSize = ip[5]; break;
+            case 2 : fparamsPtr->frameContentSize = MEM_readLE16(ip+5)+256; break;
+            case 3 : fparamsPtr->frameContentSize = MEM_readLE64(ip+5); break;
+    }   }
+    return 0;
+}
+
+
+/** ZSTDv06_decodeFrameHeader() :
+*   `srcSize` must be the size provided by ZSTDv06_frameHeaderSize().
+*   @return : 0 if success, or an error code, which can be tested using ZSTDv06_isError() */
+static size_t ZSTDv06_decodeFrameHeader(ZSTDv06_DCtx* zc, const void* src, size_t srcSize)
+{
+    size_t const result = ZSTDv06_getFrameParams(&(zc->fParams), src, srcSize);
+    if ((MEM_32bits()) && (zc->fParams.windowLog > 25)) return ERROR(frameParameter_unsupported);
+    return result;
+}
+
+
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+/*! ZSTDv06_getcBlockSize() :
+*   Provides the size of compressed block from block header `src` */
+size_t ZSTDv06_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    U32 cSize;
+
+    if (srcSize < ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
+
+    bpPtr->blockType = (blockType_t)((*in) >> 6);
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+
+static size_t ZSTDv06_copyRawBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    if (srcSize > dstCapacity) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/*! ZSTDv06_decodeLiteralsBlock() :
+    @return : nb of bytes read from src (< srcSize ) */
+size_t ZSTDv06_decodeLiteralsBlock(ZSTDv06_DCtx* dctx,
+                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+{
+    const BYTE* const istart = (const BYTE*) src;
+
+    /* any compressed block with literals segment must be at least this size */
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch(istart[0]>> 6)
+    {
+    case IS_HUF:
+        {   size_t litSize, litCSize, singleStream=0;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            if (srcSize < 5) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for lhSize, + cSize (+nbSeq) */
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                /* 2 - 2 - 10 - 10 */
+                lhSize=3;
+                singleStream = istart[0] & 16;
+                litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+                litCSize = ((istart[1] &  3) << 8) + istart[2];
+                break;
+            case 2:
+                /* 2 - 2 - 14 - 14 */
+                lhSize=4;
+                litSize  = ((istart[0] & 15) << 10) + (istart[1] << 2) + (istart[2] >> 6);
+                litCSize = ((istart[2] & 63) <<  8) + istart[3];
+                break;
+            case 3:
+                /* 2 - 2 - 18 - 18 */
+                lhSize=5;
+                litSize  = ((istart[0] & 15) << 14) + (istart[1] << 6) + (istart[2] >> 2);
+                litCSize = ((istart[2] &  3) << 16) + (istart[3] << 8) + istart[4];
+                break;
+            }
+            if (litSize > ZSTDv06_BLOCKSIZE_MAX) return ERROR(corruption_detected);
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            if (HUFv06_isError(singleStream ?
+                            HUFv06_decompress1X2(dctx->litBuffer, litSize, istart+lhSize, litCSize) :
+                            HUFv06_decompress   (dctx->litBuffer, litSize, istart+lhSize, litCSize) ))
+                return ERROR(corruption_detected);
+
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case IS_PCH:
+        {   size_t litSize, litCSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            if (lhSize != 1)  /* only case supported for now : small litSize, single stream */
+                return ERROR(corruption_detected);
+            if (!dctx->flagRepeatTable)
+                return ERROR(dictionary_corrupted);
+
+            /* 2 - 2 - 10 - 10 */
+            lhSize=3;
+            litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+            litCSize = ((istart[1] &  3) << 8) + istart[2];
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            {   size_t const errorCode = HUFv06_decompress1X4_usingDTable(dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->hufTableX4);
+                if (HUFv06_isError(errorCode)) return ERROR(corruption_detected);
+            }
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case IS_RAW:
+        {   size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize=1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                break;
+            }
+
+            if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
+                if (litSize+lhSize > srcSize) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                return lhSize+litSize;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+lhSize;
+            dctx->litSize = litSize;
+            return lhSize+litSize;
+        }
+    case IS_RLE:
+        {   size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize = 1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                if (srcSize<4) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4 */
+                break;
+            }
+            if (litSize > ZSTDv06_BLOCKSIZE_MAX) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return lhSize+1;
+        }
+    default:
+        return ERROR(corruption_detected);   /* impossible */
+    }
+}
+
+
+/*! ZSTDv06_buildSeqTable() :
+    @return : nb bytes read from src,
+              or an error code if it fails, testable with ZSTDv06_isError()
+*/
+size_t ZSTDv06_buildSeqTable(FSEv06_DTable* DTable, U32 type, U32 max, U32 maxLog,
+                                 const void* src, size_t srcSize,
+                                 const S16* defaultNorm, U32 defaultLog, U32 flagRepeatTable)
+{
+    switch(type)
+    {
+    case FSEv06_ENCODING_RLE :
+        if (!srcSize) return ERROR(srcSize_wrong);
+        if ( (*(const BYTE*)src) > max) return ERROR(corruption_detected);
+        FSEv06_buildDTable_rle(DTable, *(const BYTE*)src);   /* if *src > max, data is corrupted */
+        return 1;
+    case FSEv06_ENCODING_RAW :
+        FSEv06_buildDTable(DTable, defaultNorm, max, defaultLog);
+        return 0;
+    case FSEv06_ENCODING_STATIC:
+        if (!flagRepeatTable) return ERROR(corruption_detected);
+        return 0;
+    default :   /* impossible */
+    case FSEv06_ENCODING_DYNAMIC :
+        {   U32 tableLog;
+            S16 norm[MaxSeq+1];
+            size_t const headerSize = FSEv06_readNCount(norm, &max, &tableLog, src, srcSize);
+            if (FSEv06_isError(headerSize)) return ERROR(corruption_detected);
+            if (tableLog > maxLog) return ERROR(corruption_detected);
+            FSEv06_buildDTable(DTable, norm, max, tableLog);
+            return headerSize;
+    }   }
+}
+
+
+size_t ZSTDv06_decodeSeqHeaders(int* nbSeqPtr,
+                             FSEv06_DTable* DTableLL, FSEv06_DTable* DTableML, FSEv06_DTable* DTableOffb, U32 flagRepeatTable,
+                             const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip = istart;
+
+    /* check */
+    if (srcSize < MIN_SEQUENCES_SIZE) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    {   int nbSeq = *ip++;
+        if (!nbSeq) { *nbSeqPtr=0; return 1; }
+        if (nbSeq > 0x7F) {
+            if (nbSeq == 0xFF) {
+                if (ip+2 > iend) return ERROR(srcSize_wrong);
+                nbSeq = MEM_readLE16(ip) + LONGNBSEQ, ip+=2;
+            } else {
+                if (ip >= iend) return ERROR(srcSize_wrong);
+                nbSeq = ((nbSeq-0x80)<<8) + *ip++;
+            }
+        }
+        *nbSeqPtr = nbSeq;
+    }
+
+    /* FSE table descriptors */
+    {   U32 const LLtype  = *ip >> 6;
+        U32 const Offtype = (*ip >> 4) & 3;
+        U32 const MLtype  = (*ip >> 2) & 3;
+        ip++;
+
+        /* check */
+        if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+        /* Build DTables */
+        {   size_t const bhSize = ZSTDv06_buildSeqTable(DTableLL, LLtype, MaxLL, LLFSELog, ip, iend-ip, LL_defaultNorm, LL_defaultNormLog, flagRepeatTable);
+            if (ZSTDv06_isError(bhSize)) return ERROR(corruption_detected);
+            ip += bhSize;
+        }
+        {   size_t const bhSize = ZSTDv06_buildSeqTable(DTableOffb, Offtype, MaxOff, OffFSELog, ip, iend-ip, OF_defaultNorm, OF_defaultNormLog, flagRepeatTable);
+            if (ZSTDv06_isError(bhSize)) return ERROR(corruption_detected);
+            ip += bhSize;
+        }
+        {   size_t const bhSize = ZSTDv06_buildSeqTable(DTableML, MLtype, MaxML, MLFSELog, ip, iend-ip, ML_defaultNorm, ML_defaultNormLog, flagRepeatTable);
+            if (ZSTDv06_isError(bhSize)) return ERROR(corruption_detected);
+            ip += bhSize;
+    }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t matchLength;
+    size_t offset;
+} seq_t;
+
+typedef struct {
+    BITv06_DStream_t DStream;
+    FSEv06_DState_t stateLL;
+    FSEv06_DState_t stateOffb;
+    FSEv06_DState_t stateML;
+    size_t prevOffset[ZSTDv06_REP_INIT];
+} seqState_t;
+
+
+
+static void ZSTDv06_decodeSequence(seq_t* seq, seqState_t* seqState)
+{
+    /* Literal length */
+    U32 const llCode = FSEv06_peekSymbol(&(seqState->stateLL));
+    U32 const mlCode = FSEv06_peekSymbol(&(seqState->stateML));
+    U32 const ofCode = FSEv06_peekSymbol(&(seqState->stateOffb));   /* <= maxOff, by table construction */
+
+    U32 const llBits = LL_bits[llCode];
+    U32 const mlBits = ML_bits[mlCode];
+    U32 const ofBits = ofCode;
+    U32 const totalBits = llBits+mlBits+ofBits;
+
+    static const U32 LL_base[MaxLL+1] = {
+                             0,  1,  2,  3,  4,  5,  6,  7,  8,  9,   10,    11,    12,    13,    14,     15,
+                            16, 18, 20, 22, 24, 28, 32, 40, 48, 64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
+                            0x2000, 0x4000, 0x8000, 0x10000 };
+
+    static const U32 ML_base[MaxML+1] = {
+                             0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10,   11,    12,    13,    14,    15,
+                            16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26,   27,    28,    29,    30,    31,
+                            32, 34, 36, 38, 40, 44, 48, 56, 64, 80, 96, 0x80, 0x100, 0x200, 0x400, 0x800,
+                            0x1000, 0x2000, 0x4000, 0x8000, 0x10000 };
+
+    static const U32 OF_base[MaxOff+1] = {
+                 0,        1,       3,       7,     0xF,     0x1F,     0x3F,     0x7F,
+                 0xFF,   0x1FF,   0x3FF,   0x7FF,   0xFFF,   0x1FFF,   0x3FFF,   0x7FFF,
+                 0xFFFF, 0x1FFFF, 0x3FFFF, 0x7FFFF, 0xFFFFF, 0x1FFFFF, 0x3FFFFF, 0x7FFFFF,
+                 0xFFFFFF, 0x1FFFFFF, 0x3FFFFFF, /*fake*/ 1, 1 };
+
+    /* sequence */
+    {   size_t offset;
+        if (!ofCode)
+            offset = 0;
+        else {
+            offset = OF_base[ofCode] + BITv06_readBits(&(seqState->DStream), ofBits);   /* <=  26 bits */
+            if (MEM_32bits()) BITv06_reloadDStream(&(seqState->DStream));
+        }
+
+        if (offset < ZSTDv06_REP_NUM) {
+            if (llCode == 0 && offset <= 1) offset = 1-offset;
+
+            if (offset != 0) {
+                size_t temp = seqState->prevOffset[offset];
+                if (offset != 1) {
+                    seqState->prevOffset[2] = seqState->prevOffset[1];
+                }
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset = temp;
+
+            } else {
+                offset = seqState->prevOffset[0];
+            }
+        } else {
+            offset -= ZSTDv06_REP_MOVE;
+            seqState->prevOffset[2] = seqState->prevOffset[1];
+            seqState->prevOffset[1] = seqState->prevOffset[0];
+            seqState->prevOffset[0] = offset;
+        }
+        seq->offset = offset;
+    }
+
+    seq->matchLength = ML_base[mlCode] + MINMATCH + ((mlCode>31) ? BITv06_readBits(&(seqState->DStream), mlBits) : 0);   /* <=  16 bits */
+    if (MEM_32bits() && (mlBits+llBits>24)) BITv06_reloadDStream(&(seqState->DStream));
+
+    seq->litLength = LL_base[llCode] + ((llCode>15) ? BITv06_readBits(&(seqState->DStream), llBits) : 0);   /* <=  16 bits */
+    if (MEM_32bits() ||
+       (totalBits > 64 - 7 - (LLFSELog+MLFSELog+OffFSELog)) ) BITv06_reloadDStream(&(seqState->DStream));
+
+    /* ANS state update */
+    FSEv06_updateState(&(seqState->stateLL), &(seqState->DStream));   /* <=  9 bits */
+    FSEv06_updateState(&(seqState->stateML), &(seqState->DStream));   /* <=  9 bits */
+    if (MEM_32bits()) BITv06_reloadDStream(&(seqState->DStream));     /* <= 18 bits */
+    FSEv06_updateState(&(seqState->stateOffb), &(seqState->DStream)); /* <=  8 bits */
+}
+
+
+size_t ZSTDv06_execSequence(BYTE* op,
+                                BYTE* const oend, seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_8 = oend-8;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+
+    /* copy Literals */
+    ZSTDv06_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = iLitEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base)) {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - vBase)) return ERROR(corruption_detected);
+        match = dictEnd - (base-match);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+            if (op > oend_8 || sequence.matchLength < MINMATCH) {
+              while (op < oMatchEnd) *op++ = *match++;
+              return sequenceLength;
+            }
+    }   }
+    /* Requirement: op <= oend_8 */
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+        static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* substracted */
+        int const sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTDv06_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTDv06_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH)) {
+        if (op < oend_8) {
+            ZSTDv06_wildcopy(op, match, oend_8 - op);
+            match += oend_8 - op;
+            op = oend_8;
+        }
+        while (op < oMatchEnd) *op++ = *match++;
+    } else {
+        ZSTDv06_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+
+static size_t ZSTDv06_decompressSequences(
+                               ZSTDv06_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + maxDstSize;
+    BYTE* op = ostart;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    FSEv06_DTable* DTableLL = dctx->LLTable;
+    FSEv06_DTable* DTableML = dctx->MLTable;
+    FSEv06_DTable* DTableOffb = dctx->OffTable;
+    const BYTE* const base = (const BYTE*) (dctx->base);
+    const BYTE* const vBase = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+    int nbSeq;
+
+    /* Build Decoding Tables */
+    {   size_t const seqHSize = ZSTDv06_decodeSeqHeaders(&nbSeq, DTableLL, DTableML, DTableOffb, dctx->flagRepeatTable, ip, seqSize);
+        if (ZSTDv06_isError(seqHSize)) return seqHSize;
+        ip += seqHSize;
+        dctx->flagRepeatTable = 0;
+    }
+
+    /* Regen sequences */
+    if (nbSeq) {
+        seq_t sequence;
+        seqState_t seqState;
+
+        memset(&sequence, 0, sizeof(sequence));
+        sequence.offset = REPCODE_STARTVALUE;
+        { U32 i; for (i=0; i<ZSTDv06_REP_INIT; i++) seqState.prevOffset[i] = REPCODE_STARTVALUE; }
+        { size_t const errorCode = BITv06_initDStream(&(seqState.DStream), ip, iend-ip);
+          if (ERR_isError(errorCode)) return ERROR(corruption_detected); }
+        FSEv06_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSEv06_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSEv06_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BITv06_reloadDStream(&(seqState.DStream)) <= BITv06_DStream_completed) && nbSeq ; ) {
+            nbSeq--;
+            ZSTDv06_decodeSequence(&sequence, &seqState);
+
+#if 0  /* debug */
+            static BYTE* start = NULL;
+            if (start==NULL) start = op;
+            size_t pos = (size_t)(op-start);
+            if ((pos >= 5810037) && (pos < 5810400))
+                printf("Dpos %6u :%5u literals & match %3u bytes at distance %6u \n",
+                       pos, (U32)sequence.litLength, (U32)sequence.matchLength, (U32)sequence.offset);
+#endif
+
+            {   size_t const oneSeqSize = ZSTDv06_execSequence(op, oend, sequence, &litPtr, litEnd, base, vBase, dictEnd);
+                if (ZSTDv06_isError(oneSeqSize)) return oneSeqSize;
+                op += oneSeqSize;
+        }   }
+
+        /* check if reached exact end */
+        if (nbSeq) return ERROR(corruption_detected);
+    }
+
+    /* last literal segment */
+    {   size_t const lastLLSize = litEnd - litPtr;
+        if (litPtr > litEnd) return ERROR(corruption_detected);   /* too many literals already used */
+        if (op+lastLLSize > oend) return ERROR(dstSize_tooSmall);
+        memcpy(op, litPtr, lastLLSize);
+        op += lastLLSize;
+    }
+
+    return op-ostart;
+}
+
+
+static void ZSTDv06_checkContinuity(ZSTDv06_DCtx* dctx, const void* dst)
+{
+    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+        dctx->dictEnd = dctx->previousDstEnd;
+        dctx->vBase = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+        dctx->base = dst;
+        dctx->previousDstEnd = dst;
+    }
+}
+
+
+static size_t ZSTDv06_decompressBlock_internal(ZSTDv06_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{   /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+
+    if (srcSize >= ZSTDv06_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);
+
+    /* Decode literals sub-block */
+    {   size_t const litCSize = ZSTDv06_decodeLiteralsBlock(dctx, src, srcSize);
+        if (ZSTDv06_isError(litCSize)) return litCSize;
+        ip += litCSize;
+        srcSize -= litCSize;
+    }
+    return ZSTDv06_decompressSequences(dctx, dst, dstCapacity, ip, srcSize);
+}
+
+
+size_t ZSTDv06_decompressBlock(ZSTDv06_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{
+    ZSTDv06_checkContinuity(dctx, dst);
+    return ZSTDv06_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+/*! ZSTDv06_decompressFrame() :
+*   `dctx` must be properly initialized */
+static size_t ZSTDv06_decompressFrame(ZSTDv06_DCtx* dctx,
+                                 void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* const iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* op = ostart;
+    BYTE* const oend = ostart + dstCapacity;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties = { bt_compressed, 0 };
+
+    /* check */
+    if (srcSize < ZSTDv06_frameHeaderSize_min+ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
+
+    /* Frame Header */
+    {   size_t const frameHeaderSize = ZSTDv06_frameHeaderSize(src, ZSTDv06_frameHeaderSize_min);
+        if (ZSTDv06_isError(frameHeaderSize)) return frameHeaderSize;
+        if (srcSize < frameHeaderSize+ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
+        if (ZSTDv06_decodeFrameHeader(dctx, src, frameHeaderSize)) return ERROR(corruption_detected);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1) {
+        size_t decodedSize=0;
+        size_t const cBlockSize = ZSTDv06_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTDv06_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv06_blockHeaderSize;
+        remainingSize -= ZSTDv06_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTDv06_decompressBlock_internal(dctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTDv06_copyRawBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            return ERROR(GENERIC);   /* not yet supported */
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        if (ZSTDv06_isError(decodedSize)) return decodedSize;
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+
+size_t ZSTDv06_decompress_usingPreparedDCtx(ZSTDv06_DCtx* dctx, const ZSTDv06_DCtx* refDCtx,
+                                         void* dst, size_t dstCapacity,
+                                   const void* src, size_t srcSize)
+{
+    ZSTDv06_copyDCtx(dctx, refDCtx);
+    ZSTDv06_checkContinuity(dctx, dst);
+    return ZSTDv06_decompressFrame(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+size_t ZSTDv06_decompress_usingDict(ZSTDv06_DCtx* dctx,
+                                 void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize,
+                                 const void* dict, size_t dictSize)
+{
+    ZSTDv06_decompressBegin_usingDict(dctx, dict, dictSize);
+    ZSTDv06_checkContinuity(dctx, dst);
+    return ZSTDv06_decompressFrame(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+size_t ZSTDv06_decompressDCtx(ZSTDv06_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    return ZSTDv06_decompress_usingDict(dctx, dst, dstCapacity, src, srcSize, NULL, 0);
+}
+
+
+size_t ZSTDv06_decompress(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+#if defined(ZSTDv06_HEAPMODE) && (ZSTDv06_HEAPMODE==1)
+    size_t regenSize;
+    ZSTDv06_DCtx* dctx = ZSTDv06_createDCtx();
+    if (dctx==NULL) return ERROR(memory_allocation);
+    regenSize = ZSTDv06_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
+    ZSTDv06_freeDCtx(dctx);
+    return regenSize;
+#else   /* stack mode */
+    ZSTDv06_DCtx dctx;
+    return ZSTDv06_decompressDCtx(&dctx, dst, dstCapacity, src, srcSize);
+#endif
+}
+
+size_t ZSTDv06_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+    blockProperties_t blockProperties = { bt_compressed, 0 };
+
+    /* Frame Header */
+    {   size_t const frameHeaderSize = ZSTDv06_frameHeaderSize(src, ZSTDv06_frameHeaderSize_min);
+        if (ZSTDv06_isError(frameHeaderSize)) return frameHeaderSize;
+        if (MEM_readLE32(src) != ZSTDv06_MAGICNUMBER) return ERROR(prefix_unknown);
+        if (srcSize < frameHeaderSize+ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1) {
+        size_t const cBlockSize = ZSTDv06_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTDv06_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv06_blockHeaderSize;
+        remainingSize -= ZSTDv06_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        if (cBlockSize == 0) break;   /* bt_end */
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/*_******************************
+*  Streaming Decompression API
+********************************/
+size_t ZSTDv06_nextSrcSizeToDecompress(ZSTDv06_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+size_t ZSTDv06_decompressContinue(ZSTDv06_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != dctx->expected) return ERROR(srcSize_wrong);
+    if (dstCapacity) ZSTDv06_checkContinuity(dctx, dst);
+
+    /* Decompress : frame header; part 1 */
+    switch (dctx->stage)
+    {
+    case ZSTDds_getFrameHeaderSize :
+        if (srcSize != ZSTDv06_frameHeaderSize_min) return ERROR(srcSize_wrong);   /* impossible */
+        dctx->headerSize = ZSTDv06_frameHeaderSize(src, ZSTDv06_frameHeaderSize_min);
+        if (ZSTDv06_isError(dctx->headerSize)) return dctx->headerSize;
+        memcpy(dctx->headerBuffer, src, ZSTDv06_frameHeaderSize_min);
+        if (dctx->headerSize > ZSTDv06_frameHeaderSize_min) {
+            dctx->expected = dctx->headerSize - ZSTDv06_frameHeaderSize_min;
+            dctx->stage = ZSTDds_decodeFrameHeader;
+            return 0;
+        }
+        dctx->expected = 0;   /* not necessary to copy more */
+	/* fall-through */
+    case ZSTDds_decodeFrameHeader:
+        {   size_t result;
+            memcpy(dctx->headerBuffer + ZSTDv06_frameHeaderSize_min, src, dctx->expected);
+            result = ZSTDv06_decodeFrameHeader(dctx, dctx->headerBuffer, dctx->headerSize);
+            if (ZSTDv06_isError(result)) return result;
+            dctx->expected = ZSTDv06_blockHeaderSize;
+            dctx->stage = ZSTDds_decodeBlockHeader;
+            return 0;
+        }
+    case ZSTDds_decodeBlockHeader:
+        {   blockProperties_t bp;
+            size_t const cBlockSize = ZSTDv06_getcBlockSize(src, ZSTDv06_blockHeaderSize, &bp);
+            if (ZSTDv06_isError(cBlockSize)) return cBlockSize;
+            if (bp.blockType == bt_end) {
+                dctx->expected = 0;
+                dctx->stage = ZSTDds_getFrameHeaderSize;
+            } else {
+                dctx->expected = cBlockSize;
+                dctx->bType = bp.blockType;
+                dctx->stage = ZSTDds_decompressBlock;
+            }
+            return 0;
+        }
+    case ZSTDds_decompressBlock:
+        {   size_t rSize;
+            switch(dctx->bType)
+            {
+            case bt_compressed:
+                rSize = ZSTDv06_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize);
+                break;
+            case bt_raw :
+                rSize = ZSTDv06_copyRawBlock(dst, dstCapacity, src, srcSize);
+                break;
+            case bt_rle :
+                return ERROR(GENERIC);   /* not yet handled */
+                break;
+            case bt_end :   /* should never happen (filtered at phase 1) */
+                rSize = 0;
+                break;
+            default:
+                return ERROR(GENERIC);   /* impossible */
+            }
+            dctx->stage = ZSTDds_decodeBlockHeader;
+            dctx->expected = ZSTDv06_blockHeaderSize;
+            dctx->previousDstEnd = (char*)dst + rSize;
+            return rSize;
+        }
+    default:
+        return ERROR(GENERIC);   /* impossible */
+    }
+}
+
+
+static void ZSTDv06_refDictContent(ZSTDv06_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    dctx->dictEnd = dctx->previousDstEnd;
+    dctx->vBase = (const char*)dict - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+    dctx->base = dict;
+    dctx->previousDstEnd = (const char*)dict + dictSize;
+}
+
+static size_t ZSTDv06_loadEntropy(ZSTDv06_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    size_t hSize, offcodeHeaderSize, matchlengthHeaderSize, litlengthHeaderSize;
+
+    hSize = HUFv06_readDTableX4(dctx->hufTableX4, dict, dictSize);
+    if (HUFv06_isError(hSize)) return ERROR(dictionary_corrupted);
+    dict = (const char*)dict + hSize;
+    dictSize -= hSize;
+
+    {   short offcodeNCount[MaxOff+1];
+        U32 offcodeMaxValue=MaxOff, offcodeLog;
+        offcodeHeaderSize = FSEv06_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dict, dictSize);
+        if (FSEv06_isError(offcodeHeaderSize)) return ERROR(dictionary_corrupted);
+        if (offcodeLog > OffFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv06_buildDTable(dctx->OffTable, offcodeNCount, offcodeMaxValue, offcodeLog);
+          if (FSEv06_isError(errorCode)) return ERROR(dictionary_corrupted); }
+        dict = (const char*)dict + offcodeHeaderSize;
+        dictSize -= offcodeHeaderSize;
+    }
+
+    {   short matchlengthNCount[MaxML+1];
+        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
+        matchlengthHeaderSize = FSEv06_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dict, dictSize);
+        if (FSEv06_isError(matchlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (matchlengthLog > MLFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv06_buildDTable(dctx->MLTable, matchlengthNCount, matchlengthMaxValue, matchlengthLog);
+          if (FSEv06_isError(errorCode)) return ERROR(dictionary_corrupted); }
+        dict = (const char*)dict + matchlengthHeaderSize;
+        dictSize -= matchlengthHeaderSize;
+    }
+
+    {   short litlengthNCount[MaxLL+1];
+        unsigned litlengthMaxValue = MaxLL, litlengthLog;
+        litlengthHeaderSize = FSEv06_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dict, dictSize);
+        if (FSEv06_isError(litlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (litlengthLog > LLFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv06_buildDTable(dctx->LLTable, litlengthNCount, litlengthMaxValue, litlengthLog);
+          if (FSEv06_isError(errorCode)) return ERROR(dictionary_corrupted); }
+    }
+
+    dctx->flagRepeatTable = 1;
+    return hSize + offcodeHeaderSize + matchlengthHeaderSize + litlengthHeaderSize;
+}
+
+static size_t ZSTDv06_decompress_insertDictionary(ZSTDv06_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    size_t eSize;
+    U32 const magic = MEM_readLE32(dict);
+    if (magic != ZSTDv06_DICT_MAGIC) {
+        /* pure content mode */
+        ZSTDv06_refDictContent(dctx, dict, dictSize);
+        return 0;
+    }
+    /* load entropy tables */
+    dict = (const char*)dict + 4;
+    dictSize -= 4;
+    eSize = ZSTDv06_loadEntropy(dctx, dict, dictSize);
+    if (ZSTDv06_isError(eSize)) return ERROR(dictionary_corrupted);
+
+    /* reference dictionary content */
+    dict = (const char*)dict + eSize;
+    dictSize -= eSize;
+    ZSTDv06_refDictContent(dctx, dict, dictSize);
+
+    return 0;
+}
+
+
+size_t ZSTDv06_decompressBegin_usingDict(ZSTDv06_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    { size_t const errorCode = ZSTDv06_decompressBegin(dctx);
+      if (ZSTDv06_isError(errorCode)) return errorCode; }
+
+    if (dict && dictSize) {
+        size_t const errorCode = ZSTDv06_decompress_insertDictionary(dctx, dict, dictSize);
+        if (ZSTDv06_isError(errorCode)) return ERROR(dictionary_corrupted);
+    }
+
+    return 0;
+}
+
+/*
+    Buffered version of Zstd compression library
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net/
+*/
+
+
+/*-***************************************************************************
+*  Streaming decompression howto
+*
+*  A ZBUFFv06_DCtx object is required to track streaming operations.
+*  Use ZBUFFv06_createDCtx() and ZBUFFv06_freeDCtx() to create/release resources.
+*  Use ZBUFFv06_decompressInit() to start a new decompression operation,
+*   or ZBUFFv06_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFFv06_DCtx objects can be re-init multiple times.
+*
+*  Use ZBUFFv06_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of @dst will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change @dst.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to help latency),
+*            or 0 when a frame is completely decoded,
+*            or an error code, which can be tested using ZBUFFv06_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFFv06_recommendedDInSize() and ZBUFFv06_recommendedDOutSize()
+*  output : ZBUFFv06_recommendedDOutSize==128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFFv06_recommendedDInSize == 128KB + 3;
+*           just follow indications from ZBUFFv06_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+typedef enum { ZBUFFds_init, ZBUFFds_loadHeader,
+               ZBUFFds_read, ZBUFFds_load, ZBUFFds_flush } ZBUFFv06_dStage;
+
+/* *** Resource management *** */
+struct ZBUFFv06_DCtx_s {
+    ZSTDv06_DCtx* zd;
+    ZSTDv06_frameParams fParams;
+    ZBUFFv06_dStage stage;
+    char*  inBuff;
+    size_t inBuffSize;
+    size_t inPos;
+    char*  outBuff;
+    size_t outBuffSize;
+    size_t outStart;
+    size_t outEnd;
+    size_t blockSize;
+    BYTE headerBuffer[ZSTDv06_FRAMEHEADERSIZE_MAX];
+    size_t lhSize;
+};   /* typedef'd to ZBUFFv06_DCtx within "zstd_buffered.h" */
+
+
+ZBUFFv06_DCtx* ZBUFFv06_createDCtx(void)
+{
+    ZBUFFv06_DCtx* zbd = (ZBUFFv06_DCtx*)malloc(sizeof(ZBUFFv06_DCtx));
+    if (zbd==NULL) return NULL;
+    memset(zbd, 0, sizeof(*zbd));
+    zbd->zd = ZSTDv06_createDCtx();
+    zbd->stage = ZBUFFds_init;
+    return zbd;
+}
+
+size_t ZBUFFv06_freeDCtx(ZBUFFv06_DCtx* zbd)
+{
+    if (zbd==NULL) return 0;   /* support free on null */
+    ZSTDv06_freeDCtx(zbd->zd);
+    free(zbd->inBuff);
+    free(zbd->outBuff);
+    free(zbd);
+    return 0;
+}
+
+
+/* *** Initialization *** */
+
+size_t ZBUFFv06_decompressInitDictionary(ZBUFFv06_DCtx* zbd, const void* dict, size_t dictSize)
+{
+    zbd->stage = ZBUFFds_loadHeader;
+    zbd->lhSize = zbd->inPos = zbd->outStart = zbd->outEnd = 0;
+    return ZSTDv06_decompressBegin_usingDict(zbd->zd, dict, dictSize);
+}
+
+size_t ZBUFFv06_decompressInit(ZBUFFv06_DCtx* zbd)
+{
+    return ZBUFFv06_decompressInitDictionary(zbd, NULL, 0);
+}
+
+
+
+MEM_STATIC size_t ZBUFFv06_limitCopy(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    size_t length = MIN(dstCapacity, srcSize);
+    memcpy(dst, src, length);
+    return length;
+}
+
+
+/* *** Decompression *** */
+
+size_t ZBUFFv06_decompressContinue(ZBUFFv06_DCtx* zbd,
+                                void* dst, size_t* dstCapacityPtr,
+                          const void* src, size_t* srcSizePtr)
+{
+    const char* const istart = (const char*)src;
+    const char* const iend = istart + *srcSizePtr;
+    const char* ip = istart;
+    char* const ostart = (char*)dst;
+    char* const oend = ostart + *dstCapacityPtr;
+    char* op = ostart;
+    U32 notDone = 1;
+
+    while (notDone) {
+        switch(zbd->stage)
+        {
+        case ZBUFFds_init :
+            return ERROR(init_missing);
+
+        case ZBUFFds_loadHeader :
+            {   size_t const hSize = ZSTDv06_getFrameParams(&(zbd->fParams), zbd->headerBuffer, zbd->lhSize);
+                if (hSize != 0) {
+                    size_t const toLoad = hSize - zbd->lhSize;   /* if hSize!=0, hSize > zbd->lhSize */
+                    if (ZSTDv06_isError(hSize)) return hSize;
+                    if (toLoad > (size_t)(iend-ip)) {   /* not enough input to load full header */
+                        memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
+                        zbd->lhSize += iend-ip; ip = iend; notDone = 0;
+                        *dstCapacityPtr = 0;
+                        return (hSize - zbd->lhSize) + ZSTDv06_blockHeaderSize;   /* remaining header bytes + next block header */
+                    }
+                    memcpy(zbd->headerBuffer + zbd->lhSize, ip, toLoad); zbd->lhSize = hSize; ip += toLoad;
+                    break;
+            }   }
+
+            /* Consume header */
+            {   size_t const h1Size = ZSTDv06_nextSrcSizeToDecompress(zbd->zd);  /* == ZSTDv06_frameHeaderSize_min */
+                size_t const h1Result = ZSTDv06_decompressContinue(zbd->zd, NULL, 0, zbd->headerBuffer, h1Size);
+                if (ZSTDv06_isError(h1Result)) return h1Result;
+                if (h1Size < zbd->lhSize) {   /* long header */
+                    size_t const h2Size = ZSTDv06_nextSrcSizeToDecompress(zbd->zd);
+                    size_t const h2Result = ZSTDv06_decompressContinue(zbd->zd, NULL, 0, zbd->headerBuffer+h1Size, h2Size);
+                    if (ZSTDv06_isError(h2Result)) return h2Result;
+            }   }
+
+            /* Frame header instruct buffer sizes */
+            {   size_t const blockSize = MIN(1 << zbd->fParams.windowLog, ZSTDv06_BLOCKSIZE_MAX);
+                zbd->blockSize = blockSize;
+                if (zbd->inBuffSize < blockSize) {
+                    free(zbd->inBuff);
+                    zbd->inBuffSize = blockSize;
+                    zbd->inBuff = (char*)malloc(blockSize);
+                    if (zbd->inBuff == NULL) return ERROR(memory_allocation);
+                }
+                {   size_t const neededOutSize = ((size_t)1 << zbd->fParams.windowLog) + blockSize + WILDCOPY_OVERLENGTH * 2;
+                    if (zbd->outBuffSize < neededOutSize) {
+                        free(zbd->outBuff);
+                        zbd->outBuffSize = neededOutSize;
+                        zbd->outBuff = (char*)malloc(neededOutSize);
+                        if (zbd->outBuff == NULL) return ERROR(memory_allocation);
+            }   }   }
+            zbd->stage = ZBUFFds_read;
+	    /* fall-through */
+        case ZBUFFds_read:
+            {   size_t const neededInSize = ZSTDv06_nextSrcSizeToDecompress(zbd->zd);
+                if (neededInSize==0) {  /* end of frame */
+                    zbd->stage = ZBUFFds_init;
+                    notDone = 0;
+                    break;
+                }
+                if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
+                    size_t const decodedSize = ZSTDv06_decompressContinue(zbd->zd,
+                        zbd->outBuff + zbd->outStart, zbd->outBuffSize - zbd->outStart,
+                        ip, neededInSize);
+                    if (ZSTDv06_isError(decodedSize)) return decodedSize;
+                    ip += neededInSize;
+                    if (!decodedSize) break;   /* this was just a header */
+                    zbd->outEnd = zbd->outStart +  decodedSize;
+                    zbd->stage = ZBUFFds_flush;
+                    break;
+                }
+                if (ip==iend) { notDone = 0; break; }   /* no more input */
+                zbd->stage = ZBUFFds_load;
+            }
+	    /* fall-through */
+        case ZBUFFds_load:
+            {   size_t const neededInSize = ZSTDv06_nextSrcSizeToDecompress(zbd->zd);
+                size_t const toLoad = neededInSize - zbd->inPos;   /* should always be <= remaining space within inBuff */
+                size_t loadedSize;
+                if (toLoad > zbd->inBuffSize - zbd->inPos) return ERROR(corruption_detected);   /* should never happen */
+                loadedSize = ZBUFFv06_limitCopy(zbd->inBuff + zbd->inPos, toLoad, ip, iend-ip);
+                ip += loadedSize;
+                zbd->inPos += loadedSize;
+                if (loadedSize < toLoad) { notDone = 0; break; }   /* not enough input, wait for more */
+
+                /* decode loaded input */
+                {   size_t const decodedSize = ZSTDv06_decompressContinue(zbd->zd,
+                        zbd->outBuff + zbd->outStart, zbd->outBuffSize - zbd->outStart,
+                        zbd->inBuff, neededInSize);
+                    if (ZSTDv06_isError(decodedSize)) return decodedSize;
+                    zbd->inPos = 0;   /* input is consumed */
+                    if (!decodedSize) { zbd->stage = ZBUFFds_read; break; }   /* this was just a header */
+                    zbd->outEnd = zbd->outStart +  decodedSize;
+                    zbd->stage = ZBUFFds_flush;
+                    // break; /* ZBUFFds_flush follows */
+                }
+	    }
+	    /* fall-through */
+        case ZBUFFds_flush:
+            {   size_t const toFlushSize = zbd->outEnd - zbd->outStart;
+                size_t const flushedSize = ZBUFFv06_limitCopy(op, oend-op, zbd->outBuff + zbd->outStart, toFlushSize);
+                op += flushedSize;
+                zbd->outStart += flushedSize;
+                if (flushedSize == toFlushSize) {
+                    zbd->stage = ZBUFFds_read;
+                    if (zbd->outStart + zbd->blockSize > zbd->outBuffSize)
+                        zbd->outStart = zbd->outEnd = 0;
+                    break;
+                }
+                /* cannot flush everything */
+                notDone = 0;
+                break;
+            }
+        default: return ERROR(GENERIC);   /* impossible */
+    }   }
+
+    /* result */
+    *srcSizePtr = ip-istart;
+    *dstCapacityPtr = op-ostart;
+    {   size_t nextSrcSizeHint = ZSTDv06_nextSrcSizeToDecompress(zbd->zd);
+        if (nextSrcSizeHint > ZSTDv06_blockHeaderSize) nextSrcSizeHint+= ZSTDv06_blockHeaderSize;   /* get following block header too */
+        nextSrcSizeHint -= zbd->inPos;   /* already loaded*/
+        return nextSrcSizeHint;
+    }
+}
+
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+size_t ZBUFFv06_recommendedDInSize(void)  { return ZSTDv06_BLOCKSIZE_MAX + ZSTDv06_blockHeaderSize /* block header size*/ ; }
+size_t ZBUFFv06_recommendedDOutSize(void) { return ZSTDv06_BLOCKSIZE_MAX; }
diff --git a/vendor/github.com/DataDog/zstd/zstd_v06.h b/vendor/github.com/DataDog/zstd/zstd_v06.h
new file mode 100644
index 0000000..fb4eb37
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v06.h
@@ -0,0 +1,167 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTDv06_H
+#define ZSTDv06_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*======  Dependency  ======*/
+#include <stddef.h>   /* size_t */
+
+
+/*======  Export for Windows  ======*/
+/*!
+*  ZSTDv06_DLL_EXPORT :
+*  Enable exporting of functions when building a Windows DLL
+*/
+#if defined(_WIN32) && defined(ZSTDv06_DLL_EXPORT) && (ZSTDv06_DLL_EXPORT==1)
+#  define ZSTDLIBv06_API __declspec(dllexport)
+#else
+#  define ZSTDLIBv06_API
+#endif
+
+
+/* *************************************
+*  Simple functions
+***************************************/
+/*! ZSTDv06_decompress() :
+    `compressedSize` : is the _exact_ size of the compressed blob, otherwise decompression will fail.
+    `dstCapacity` must be large enough, equal or larger than originalSize.
+    @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
+              or an errorCode if it fails (which can be tested using ZSTDv06_isError()) */
+ZSTDLIBv06_API size_t ZSTDv06_decompress( void* dst, size_t dstCapacity,
+                                    const void* src, size_t compressedSize);
+
+/**
+ZSTDv06_getFrameSrcSize() : get the source length of a ZSTD frame
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv06_isError())
+*/
+size_t ZSTDv06_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/* *************************************
+*  Helper functions
+***************************************/
+ZSTDLIBv06_API size_t      ZSTDv06_compressBound(size_t srcSize); /*!< maximum compressed size (worst case scenario) */
+
+/* Error Management */
+ZSTDLIBv06_API unsigned    ZSTDv06_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
+ZSTDLIBv06_API const char* ZSTDv06_getErrorName(size_t code);     /*!< provides readable string for an error code */
+
+
+/* *************************************
+*  Explicit memory management
+***************************************/
+/** Decompression context */
+typedef struct ZSTDv06_DCtx_s ZSTDv06_DCtx;
+ZSTDLIBv06_API ZSTDv06_DCtx* ZSTDv06_createDCtx(void);
+ZSTDLIBv06_API size_t     ZSTDv06_freeDCtx(ZSTDv06_DCtx* dctx);      /*!< @return : errorCode */
+
+/** ZSTDv06_decompressDCtx() :
+*   Same as ZSTDv06_decompress(), but requires an already allocated ZSTDv06_DCtx (see ZSTDv06_createDCtx()) */
+ZSTDLIBv06_API size_t ZSTDv06_decompressDCtx(ZSTDv06_DCtx* ctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+/*-***********************
+*  Dictionary API
+*************************/
+/*! ZSTDv06_decompress_usingDict() :
+*   Decompression using a pre-defined Dictionary content (see dictBuilder).
+*   Dictionary must be identical to the one used during compression, otherwise regenerated data will be corrupted.
+*   Note : dict can be NULL, in which case, it's equivalent to ZSTDv06_decompressDCtx() */
+ZSTDLIBv06_API size_t ZSTDv06_decompress_usingDict(ZSTDv06_DCtx* dctx,
+                                                   void* dst, size_t dstCapacity,
+                                             const void* src, size_t srcSize,
+                                             const void* dict,size_t dictSize);
+
+
+/*-************************
+*  Advanced Streaming API
+***************************/
+struct ZSTDv06_frameParams_s { unsigned long long frameContentSize; unsigned windowLog; };
+typedef struct ZSTDv06_frameParams_s ZSTDv06_frameParams;
+
+ZSTDLIBv06_API size_t ZSTDv06_getFrameParams(ZSTDv06_frameParams* fparamsPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
+ZSTDLIBv06_API size_t ZSTDv06_decompressBegin_usingDict(ZSTDv06_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIBv06_API void   ZSTDv06_copyDCtx(ZSTDv06_DCtx* dctx, const ZSTDv06_DCtx* preparedDCtx);
+
+ZSTDLIBv06_API size_t ZSTDv06_nextSrcSizeToDecompress(ZSTDv06_DCtx* dctx);
+ZSTDLIBv06_API size_t ZSTDv06_decompressContinue(ZSTDv06_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+
+/* *************************************
+*  ZBUFF API
+***************************************/
+
+typedef struct ZBUFFv06_DCtx_s ZBUFFv06_DCtx;
+ZSTDLIBv06_API ZBUFFv06_DCtx* ZBUFFv06_createDCtx(void);
+ZSTDLIBv06_API size_t         ZBUFFv06_freeDCtx(ZBUFFv06_DCtx* dctx);
+
+ZSTDLIBv06_API size_t ZBUFFv06_decompressInit(ZBUFFv06_DCtx* dctx);
+ZSTDLIBv06_API size_t ZBUFFv06_decompressInitDictionary(ZBUFFv06_DCtx* dctx, const void* dict, size_t dictSize);
+
+ZSTDLIBv06_API size_t ZBUFFv06_decompressContinue(ZBUFFv06_DCtx* dctx,
+                                                  void* dst, size_t* dstCapacityPtr,
+                                            const void* src, size_t* srcSizePtr);
+
+/*-***************************************************************************
+*  Streaming decompression howto
+*
+*  A ZBUFFv06_DCtx object is required to track streaming operations.
+*  Use ZBUFFv06_createDCtx() and ZBUFFv06_freeDCtx() to create/release resources.
+*  Use ZBUFFv06_decompressInit() to start a new decompression operation,
+*   or ZBUFFv06_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFFv06_DCtx objects can be re-init multiple times.
+*
+*  Use ZBUFFv06_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change `dst`.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to help latency),
+*            or 0 when a frame is completely decoded,
+*            or an error code, which can be tested using ZBUFFv06_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFFv06_recommendedDInSize() and ZBUFFv06_recommendedDOutSize()
+*  output : ZBUFFv06_recommendedDOutSize== 128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFFv06_recommendedDInSize == 128KB + 3;
+*           just follow indications from ZBUFFv06_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+ZSTDLIBv06_API unsigned ZBUFFv06_isError(size_t errorCode);
+ZSTDLIBv06_API const char* ZBUFFv06_getErrorName(size_t errorCode);
+
+/** Functions below provide recommended buffer sizes for Compression or Decompression operations.
+*   These sizes are just hints, they tend to offer better latency */
+ZSTDLIBv06_API size_t ZBUFFv06_recommendedDInSize(void);
+ZSTDLIBv06_API size_t ZBUFFv06_recommendedDOutSize(void);
+
+
+/*-*************************************
+*  Constants
+***************************************/
+#define ZSTDv06_MAGICNUMBER 0xFD2FB526   /* v0.6 */
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* ZSTDv06_BUFFERED_H */
diff --git a/vendor/github.com/DataDog/zstd/zstd_v07.c b/vendor/github.com/DataDog/zstd/zstd_v07.c
new file mode 100644
index 0000000..70b170f
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v07.c
@@ -0,0 +1,4502 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/*- Dependencies -*/
+#include <stddef.h>     /* size_t, ptrdiff_t */
+#include <string.h>     /* memcpy */
+#include <stdlib.h>     /* malloc, free, qsort */
+
+#ifndef XXH_STATIC_LINKING_ONLY
+#  define XXH_STATIC_LINKING_ONLY    /* XXH64_state_t */
+#endif
+#include "xxhash.h"                  /* XXH64_* */
+#include "zstd_v07.h"
+
+#define FSEv07_STATIC_LINKING_ONLY   /* FSEv07_MIN_TABLELOG */
+#define HUFv07_STATIC_LINKING_ONLY   /* HUFv07_TABLELOG_ABSOLUTEMAX */
+#define ZSTDv07_STATIC_LINKING_ONLY
+
+#include "error_private.h"
+
+
+#ifdef ZSTDv07_STATIC_LINKING_ONLY
+
+/* ====================================================================================
+ * The definitions in this section are considered experimental.
+ * They should never be used with a dynamic library, as they may change in the future.
+ * They are provided for advanced usages.
+ * Use them only in association with static linking.
+ * ==================================================================================== */
+
+/*--- Constants ---*/
+#define ZSTDv07_MAGIC_SKIPPABLE_START  0x184D2A50U
+
+#define ZSTDv07_WINDOWLOG_MAX_32  25
+#define ZSTDv07_WINDOWLOG_MAX_64  27
+#define ZSTDv07_WINDOWLOG_MAX    ((U32)(MEM_32bits() ? ZSTDv07_WINDOWLOG_MAX_32 : ZSTDv07_WINDOWLOG_MAX_64))
+#define ZSTDv07_WINDOWLOG_MIN     18
+#define ZSTDv07_CHAINLOG_MAX     (ZSTDv07_WINDOWLOG_MAX+1)
+#define ZSTDv07_CHAINLOG_MIN       4
+#define ZSTDv07_HASHLOG_MAX       ZSTDv07_WINDOWLOG_MAX
+#define ZSTDv07_HASHLOG_MIN       12
+#define ZSTDv07_HASHLOG3_MAX      17
+#define ZSTDv07_SEARCHLOG_MAX    (ZSTDv07_WINDOWLOG_MAX-1)
+#define ZSTDv07_SEARCHLOG_MIN      1
+#define ZSTDv07_SEARCHLENGTH_MAX   7
+#define ZSTDv07_SEARCHLENGTH_MIN   3
+#define ZSTDv07_TARGETLENGTH_MIN   4
+#define ZSTDv07_TARGETLENGTH_MAX 999
+
+#define ZSTDv07_FRAMEHEADERSIZE_MAX 18    /* for static allocation */
+static const size_t ZSTDv07_frameHeaderSize_min = 5;
+static const size_t ZSTDv07_frameHeaderSize_max = ZSTDv07_FRAMEHEADERSIZE_MAX;
+static const size_t ZSTDv07_skippableHeaderSize = 8;  /* magic number + skippable frame length */
+
+
+/* custom memory allocation functions */
+typedef void* (*ZSTDv07_allocFunction) (void* opaque, size_t size);
+typedef void  (*ZSTDv07_freeFunction) (void* opaque, void* address);
+typedef struct { ZSTDv07_allocFunction customAlloc; ZSTDv07_freeFunction customFree; void* opaque; } ZSTDv07_customMem;
+
+
+/*--- Advanced Decompression functions ---*/
+
+/*! ZSTDv07_estimateDCtxSize() :
+ *  Gives the potential amount of memory allocated to create a ZSTDv07_DCtx */
+ZSTDLIBv07_API size_t ZSTDv07_estimateDCtxSize(void);
+
+/*! ZSTDv07_createDCtx_advanced() :
+ *  Create a ZSTD decompression context using external alloc and free functions */
+ZSTDLIBv07_API ZSTDv07_DCtx* ZSTDv07_createDCtx_advanced(ZSTDv07_customMem customMem);
+
+/*! ZSTDv07_sizeofDCtx() :
+ *  Gives the amount of memory used by a given ZSTDv07_DCtx */
+ZSTDLIBv07_API size_t ZSTDv07_sizeofDCtx(const ZSTDv07_DCtx* dctx);
+
+
+/* ******************************************************************
+*  Buffer-less streaming functions (synchronous mode)
+********************************************************************/
+
+ZSTDLIBv07_API size_t ZSTDv07_decompressBegin(ZSTDv07_DCtx* dctx);
+ZSTDLIBv07_API size_t ZSTDv07_decompressBegin_usingDict(ZSTDv07_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIBv07_API void   ZSTDv07_copyDCtx(ZSTDv07_DCtx* dctx, const ZSTDv07_DCtx* preparedDCtx);
+
+ZSTDLIBv07_API size_t ZSTDv07_nextSrcSizeToDecompress(ZSTDv07_DCtx* dctx);
+ZSTDLIBv07_API size_t ZSTDv07_decompressContinue(ZSTDv07_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+/*
+  Buffer-less streaming decompression (synchronous mode)
+
+  A ZSTDv07_DCtx object is required to track streaming operations.
+  Use ZSTDv07_createDCtx() / ZSTDv07_freeDCtx() to manage it.
+  A ZSTDv07_DCtx object can be re-used multiple times.
+
+  First optional operation is to retrieve frame parameters, using ZSTDv07_getFrameParams(), which doesn't consume the input.
+  It can provide the minimum size of rolling buffer required to properly decompress data (`windowSize`),
+  and optionally the final size of uncompressed content.
+  (Note : content size is an optional info that may not be present. 0 means : content size unknown)
+  Frame parameters are extracted from the beginning of compressed frame.
+  The amount of data to read is variable, from ZSTDv07_frameHeaderSize_min to ZSTDv07_frameHeaderSize_max (so if `srcSize` >= ZSTDv07_frameHeaderSize_max, it will always work)
+  If `srcSize` is too small for operation to succeed, function will return the minimum size it requires to produce a result.
+  Result : 0 when successful, it means the ZSTDv07_frameParams structure has been filled.
+          >0 : means there is not enough data into `src`. Provides the expected size to successfully decode header.
+           errorCode, which can be tested using ZSTDv07_isError()
+
+  Start decompression, with ZSTDv07_decompressBegin() or ZSTDv07_decompressBegin_usingDict().
+  Alternatively, you can copy a prepared context, using ZSTDv07_copyDCtx().
+
+  Then use ZSTDv07_nextSrcSizeToDecompress() and ZSTDv07_decompressContinue() alternatively.
+  ZSTDv07_nextSrcSizeToDecompress() tells how much bytes to provide as 'srcSize' to ZSTDv07_decompressContinue().
+  ZSTDv07_decompressContinue() requires this exact amount of bytes, or it will fail.
+
+  @result of ZSTDv07_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
+  It can be zero, which is not an error; it just means ZSTDv07_decompressContinue() has decoded some header.
+
+  ZSTDv07_decompressContinue() needs previous data blocks during decompression, up to `windowSize`.
+  They should preferably be located contiguously, prior to current block.
+  Alternatively, a round buffer of sufficient size is also possible. Sufficient size is determined by frame parameters.
+  ZSTDv07_decompressContinue() is very sensitive to contiguity,
+  if 2 blocks don't follow each other, make sure that either the compressor breaks contiguity at the same place,
+    or that previous contiguous segment is large enough to properly handle maximum back-reference.
+
+  A frame is fully decoded when ZSTDv07_nextSrcSizeToDecompress() returns zero.
+  Context can then be reset to start a new decompression.
+
+
+  == Special case : skippable frames ==
+
+  Skippable frames allow the integration of user-defined data into a flow of concatenated frames.
+  Skippable frames will be ignored (skipped) by a decompressor. The format of skippable frame is following:
+  a) Skippable frame ID - 4 Bytes, Little endian format, any value from 0x184D2A50 to 0x184D2A5F
+  b) Frame Size - 4 Bytes, Little endian format, unsigned 32-bits
+  c) Frame Content - any content (User Data) of length equal to Frame Size
+  For skippable frames ZSTDv07_decompressContinue() always returns 0.
+  For skippable frames ZSTDv07_getFrameParams() returns fparamsPtr->windowLog==0 what means that a frame is skippable.
+  It also returns Frame Size as fparamsPtr->frameContentSize.
+*/
+
+
+/* **************************************
+*  Block functions
+****************************************/
+/*! Block functions produce and decode raw zstd blocks, without frame metadata.
+    Frame metadata cost is typically ~18 bytes, which can be non-negligible for very small blocks (< 100 bytes).
+    User will have to take in charge required information to regenerate data, such as compressed and content sizes.
+
+    A few rules to respect :
+    - Compressing and decompressing require a context structure
+      + Use ZSTDv07_createCCtx() and ZSTDv07_createDCtx()
+    - It is necessary to init context before starting
+      + compression : ZSTDv07_compressBegin()
+      + decompression : ZSTDv07_decompressBegin()
+      + variants _usingDict() are also allowed
+      + copyCCtx() and copyDCtx() work too
+    - Block size is limited, it must be <= ZSTDv07_getBlockSizeMax()
+      + If you need to compress more, cut data into multiple blocks
+      + Consider using the regular ZSTDv07_compress() instead, as frame metadata costs become negligible when source size is large.
+    - When a block is considered not compressible enough, ZSTDv07_compressBlock() result will be zero.
+      In which case, nothing is produced into `dst`.
+      + User must test for such outcome and deal directly with uncompressed data
+      + ZSTDv07_decompressBlock() doesn't accept uncompressed data as input !!!
+      + In case of multiple successive blocks, decoder must be informed of uncompressed block existence to follow proper history.
+        Use ZSTDv07_insertBlock() in such a case.
+*/
+
+#define ZSTDv07_BLOCKSIZE_ABSOLUTEMAX (128 * 1024)   /* define, for static allocation */
+ZSTDLIBv07_API size_t ZSTDv07_decompressBlock(ZSTDv07_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIBv07_API size_t ZSTDv07_insertBlock(ZSTDv07_DCtx* dctx, const void* blockStart, size_t blockSize);  /**< insert block into `dctx` history. Useful for uncompressed blocks */
+
+
+#endif   /* ZSTDv07_STATIC_LINKING_ONLY */
+
+
+/* ******************************************************************
+   mem.h
+   low-level memory access routines
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-****************************************
+*  Compiler specifics
+******************************************/
+#if defined(_MSC_VER)   /* Visual Studio */
+#   include <stdlib.h>  /* _byteswap_ulong */
+#   include <intrin.h>  /* _byteswap_* */
+#endif
+#if defined(__GNUC__)
+#  define MEM_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
+#else
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/*-**************************************************************
+*  Basic Types
+*****************************************************************/
+#if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef  int16_t S16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef  int64_t S64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef   signed short      S16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef   signed long long  S64;
+#endif
+
+
+/*-**************************************************************
+*  Memory I/O
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets depending on alignment.
+ *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define MEM_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
+#    define MEM_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(size_t)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard, by lying on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+
+#else
+
+/* default method, safe and standard.
+   can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+    memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif /* MEM_FORCE_MEMORY_ACCESS */
+
+MEM_STATIC U32 MEM_swap32(U32 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_ulong(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap32(in);
+#else
+    return  ((in << 24) & 0xff000000 ) |
+            ((in <<  8) & 0x00ff0000 ) |
+            ((in >>  8) & 0x0000ff00 ) |
+            ((in >> 24) & 0x000000ff );
+#endif
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_uint64(in);
+#elif defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)
+    return __builtin_bswap64(in);
+#else
+    return  ((in << 56) & 0xff00000000000000ULL) |
+            ((in << 40) & 0x00ff000000000000ULL) |
+            ((in << 24) & 0x0000ff0000000000ULL) |
+            ((in << 8)  & 0x000000ff00000000ULL) |
+            ((in >> 8)  & 0x00000000ff000000ULL) |
+            ((in >> 24) & 0x0000000000ff0000ULL) |
+            ((in >> 40) & 0x000000000000ff00ULL) |
+            ((in >> 56) & 0x00000000000000ffULL);
+#endif
+}
+
+
+/*=== Little endian r/w ===*/
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read16(memPtr);
+    else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)(p[0] + (p[1]<<8));
+    }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+    if (MEM_isLittleEndian()) {
+        MEM_write16(memPtr, val);
+    } else {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE)val;
+        p[1] = (BYTE)(val>>8);
+    }
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read32(memPtr);
+    else
+        return MEM_swap32(MEM_read32(memPtr));
+}
+
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+    if (MEM_isLittleEndian())
+        return MEM_read64(memPtr);
+    else
+        return MEM_swap64(MEM_read64(memPtr));
+}
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+    if (MEM_32bits())
+        return (size_t)MEM_readLE32(memPtr);
+    else
+        return (size_t)MEM_readLE64(memPtr);
+}
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
+/* ******************************************************************
+   bitstream
+   Part of FSE library
+   header file (to include)
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/*
+*  This API consists of small unitary functions, which must be inlined for best performance.
+*  Since link-time-optimization is not available for all compilers,
+*  these functions are defined into a .h to be included.
+*/
+
+
+/*=========================================
+*  Target specific
+=========================================*/
+#if defined(__BMI__) && defined(__GNUC__)
+#  include <immintrin.h>   /* support for bextr (experimental) */
+#endif
+
+/*-********************************************
+*  bitStream decoding API (read backward)
+**********************************************/
+typedef struct
+{
+    size_t   bitContainer;
+    unsigned bitsConsumed;
+    const char* ptr;
+    const char* start;
+} BITv07_DStream_t;
+
+typedef enum { BITv07_DStream_unfinished = 0,
+               BITv07_DStream_endOfBuffer = 1,
+               BITv07_DStream_completed = 2,
+               BITv07_DStream_overflow = 3 } BITv07_DStream_status;  /* result of BITv07_reloadDStream() */
+               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+
+MEM_STATIC size_t   BITv07_initDStream(BITv07_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t   BITv07_readBits(BITv07_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BITv07_DStream_status BITv07_reloadDStream(BITv07_DStream_t* bitD);
+MEM_STATIC unsigned BITv07_endOfDStream(const BITv07_DStream_t* bitD);
+
+
+
+/*-****************************************
+*  unsafe API
+******************************************/
+MEM_STATIC size_t BITv07_readBitsFast(BITv07_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+
+
+/*-**************************************************************
+*  Internal functions
+****************************************************************/
+MEM_STATIC unsigned BITv07_highbit32 (U32 val)
+{
+#   if defined(_MSC_VER)   /* Visual */
+    unsigned long r=0;
+    _BitScanReverse ( &r, val );
+    return (unsigned) r;
+#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
+    return 31 - __builtin_clz (val);
+#   else   /* Software version */
+    static const unsigned DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+    U32 v = val;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    return DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
+#   endif
+}
+
+
+
+/*-********************************************************
+* bitStream decoding
+**********************************************************/
+/*! BITv07_initDStream() :
+*   Initialize a BITv07_DStream_t.
+*   `bitD` : a pointer to an already allocated BITv07_DStream_t structure.
+*   `srcSize` must be the *exact* size of the bitStream, in bytes.
+*   @return : size of stream (== srcSize) or an errorCode if a problem is detected
+*/
+MEM_STATIC size_t BITv07_initDStream(BITv07_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+    if (srcSize >=  sizeof(bitD->bitContainer)) {  /* normal case */
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+          bitD->bitsConsumed = lastByte ? 8 - BITv07_highbit32(lastByte) : 0;
+          if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
+    } else {
+        bitD->start = (const char*)srcBuffer;
+        bitD->ptr   = bitD->start;
+        bitD->bitContainer = *(const BYTE*)(bitD->start);
+        switch(srcSize)
+        {
+            case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);/* fall-through */
+            case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);/* fall-through */
+            case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);/* fall-through */
+            case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24; /* fall-through */
+            case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16; /* fall-through */
+            case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8; /* fall-through */
+            default: break;
+        }
+        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+          bitD->bitsConsumed = lastByte ? 8 - BITv07_highbit32(lastByte) : 0;
+          if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
+        bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
+    }
+
+    return srcSize;
+}
+
+
+ MEM_STATIC size_t BITv07_lookBits(const BITv07_DStream_t* bitD, U32 nbBits)
+{
+    U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return ((bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> 1) >> ((bitMask-nbBits) & bitMask);
+}
+
+/*! BITv07_lookBitsFast() :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv07_lookBitsFast(const BITv07_DStream_t* bitD, U32 nbBits)
+{
+    U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
+    return (bitD->bitContainer << (bitD->bitsConsumed & bitMask)) >> (((bitMask+1)-nbBits) & bitMask);
+}
+
+MEM_STATIC void BITv07_skipBits(BITv07_DStream_t* bitD, U32 nbBits)
+{
+    bitD->bitsConsumed += nbBits;
+}
+
+MEM_STATIC size_t BITv07_readBits(BITv07_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BITv07_lookBits(bitD, nbBits);
+    BITv07_skipBits(bitD, nbBits);
+    return value;
+}
+
+/*! BITv07_readBitsFast() :
+*   unsafe version; only works only if nbBits >= 1 */
+MEM_STATIC size_t BITv07_readBitsFast(BITv07_DStream_t* bitD, U32 nbBits)
+{
+    size_t const value = BITv07_lookBitsFast(bitD, nbBits);
+    BITv07_skipBits(bitD, nbBits);
+    return value;
+}
+
+MEM_STATIC BITv07_DStream_status BITv07_reloadDStream(BITv07_DStream_t* bitD)
+{
+    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* should not happen => corruption detected */
+        return BITv07_DStream_overflow;
+
+    if (bitD->ptr >= bitD->start + sizeof(bitD->bitContainer)) {
+        bitD->ptr -= bitD->bitsConsumed >> 3;
+        bitD->bitsConsumed &= 7;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);
+        return BITv07_DStream_unfinished;
+    }
+    if (bitD->ptr == bitD->start) {
+        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BITv07_DStream_endOfBuffer;
+        return BITv07_DStream_completed;
+    }
+    {   U32 nbBytes = bitD->bitsConsumed >> 3;
+        BITv07_DStream_status result = BITv07_DStream_unfinished;
+        if (bitD->ptr - nbBytes < bitD->start) {
+            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
+            result = BITv07_DStream_endOfBuffer;
+        }
+        bitD->ptr -= nbBytes;
+        bitD->bitsConsumed -= nbBytes*8;
+        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD) */
+        return result;
+    }
+}
+
+/*! BITv07_endOfDStream() :
+*   @return Tells if DStream has exactly reached its end (all bits consumed).
+*/
+MEM_STATIC unsigned BITv07_endOfDStream(const BITv07_DStream_t* DStream)
+{
+    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
+/* ******************************************************************
+   FSE : Finite State Entropy codec
+   Public Prototypes declaration
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef FSEv07_H
+#define FSEv07_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/*-****************************************
+*  FSE simple functions
+******************************************/
+
+/*! FSEv07_decompress():
+    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated destination buffer 'dst', of size 'dstCapacity'.
+    @return : size of regenerated data (<= maxDstSize),
+              or an error code, which can be tested using FSEv07_isError() .
+
+    ** Important ** : FSEv07_decompress() does not decompress non-compressible nor RLE data !!!
+    Why ? : making this distinction requires a header.
+    Header management is intentionally delegated to the user layer, which can better manage special cases.
+*/
+size_t FSEv07_decompress(void* dst,  size_t dstCapacity,
+                const void* cSrc, size_t cSrcSize);
+
+
+/* Error Management */
+unsigned    FSEv07_isError(size_t code);        /* tells if a return value is an error code */
+const char* FSEv07_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
+
+
+/*-*****************************************
+*  FSE detailed API
+******************************************/
+/*!
+FSEv07_decompress() does the following:
+1. read normalized counters with readNCount()
+2. build decoding table 'DTable' from normalized counters
+3. decode the data stream using decoding table 'DTable'
+
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and provide normalized distribution using external method.
+*/
+
+
+/* *** DECOMPRESSION *** */
+
+/*! FSEv07_readNCount():
+    Read compactly saved 'normalizedCounter' from 'rBuffer'.
+    @return : size read from 'rBuffer',
+              or an errorCode, which can be tested using FSEv07_isError().
+              maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+size_t FSEv07_readNCount (short* normalizedCounter, unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, const void* rBuffer, size_t rBuffSize);
+
+/*! Constructor and Destructor of FSEv07_DTable.
+    Note that its size depends on 'tableLog' */
+typedef unsigned FSEv07_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
+FSEv07_DTable* FSEv07_createDTable(unsigned tableLog);
+void        FSEv07_freeDTable(FSEv07_DTable* dt);
+
+/*! FSEv07_buildDTable():
+    Builds 'dt', which must be already allocated, using FSEv07_createDTable().
+    return : 0, or an errorCode, which can be tested using FSEv07_isError() */
+size_t FSEv07_buildDTable (FSEv07_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSEv07_decompress_usingDTable():
+    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
+    into `dst` which must be already allocated.
+    @return : size of regenerated data (necessarily <= `dstCapacity`),
+              or an errorCode, which can be tested using FSEv07_isError() */
+size_t FSEv07_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSEv07_DTable* dt);
+
+/*!
+Tutorial :
+----------
+(Note : these functions only decompress FSE-compressed blocks.
+ If block is uncompressed, use memcpy() instead
+ If block is a single repeated byte, use memset() instead )
+
+The first step is to obtain the normalized frequencies of symbols.
+This can be performed by FSEv07_readNCount() if it was saved using FSEv07_writeNCount().
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
+In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
+or size the table to handle worst case situations (typically 256).
+FSEv07_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
+The result of FSEv07_readNCount() is the number of bytes read from 'rBuffer'.
+Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
+If there is an error, the function will return an error code, which can be tested using FSEv07_isError().
+
+The next step is to build the decompression tables 'FSEv07_DTable' from 'normalizedCounter'.
+This is performed by the function FSEv07_buildDTable().
+The space required by 'FSEv07_DTable' must be already allocated using FSEv07_createDTable().
+If there is an error, the function will return an error code, which can be tested using FSEv07_isError().
+
+`FSEv07_DTable` can then be used to decompress `cSrc`, with FSEv07_decompress_usingDTable().
+`cSrcSize` must be strictly correct, otherwise decompression will fail.
+FSEv07_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
+If there is an error, the function will return an error code, which can be tested using FSEv07_isError(). (ex: dst buffer too small)
+*/
+
+
+#ifdef FSEv07_STATIC_LINKING_ONLY
+
+
+/* *****************************************
+*  Static allocation
+*******************************************/
+/* FSE buffer bounds */
+#define FSEv07_NCOUNTBOUND 512
+#define FSEv07_BLOCKBOUND(size) (size + (size>>7))
+
+/* It is possible to statically allocate FSE CTable/DTable as a table of unsigned using below macros */
+#define FSEv07_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+
+
+/* *****************************************
+*  FSE advanced API
+*******************************************/
+size_t FSEv07_countFast(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize);
+/**< same as FSEv07_count(), but blindly trusts that all byte values within src are <= *maxSymbolValuePtr  */
+
+unsigned FSEv07_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
+/**< same as FSEv07_optimalTableLog(), which used `minus==2` */
+
+size_t FSEv07_buildDTable_raw (FSEv07_DTable* dt, unsigned nbBits);
+/**< build a fake FSEv07_DTable, designed to read an uncompressed bitstream where each symbol uses nbBits */
+
+size_t FSEv07_buildDTable_rle (FSEv07_DTable* dt, unsigned char symbolValue);
+/**< build a fake FSEv07_DTable, designed to always generate the same symbolValue */
+
+
+
+/* *****************************************
+*  FSE symbol decompression API
+*******************************************/
+typedef struct
+{
+    size_t      state;
+    const void* table;   /* precise table may vary, depending on U16 */
+} FSEv07_DState_t;
+
+
+static void     FSEv07_initDState(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD, const FSEv07_DTable* dt);
+
+static unsigned char FSEv07_decodeSymbol(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD);
+
+
+
+/* *****************************************
+*  FSE unsafe API
+*******************************************/
+static unsigned char FSEv07_decodeSymbolFast(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* ======    Decompression    ====== */
+
+typedef struct {
+    U16 tableLog;
+    U16 fastMode;
+} FSEv07_DTableHeader;   /* sizeof U32 */
+
+typedef struct
+{
+    unsigned short newState;
+    unsigned char  symbol;
+    unsigned char  nbBits;
+} FSEv07_decode_t;   /* size == U32 */
+
+MEM_STATIC void FSEv07_initDState(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD, const FSEv07_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv07_DTableHeader* const DTableH = (const FSEv07_DTableHeader*)ptr;
+    DStatePtr->state = BITv07_readBits(bitD, DTableH->tableLog);
+    BITv07_reloadDStream(bitD);
+    DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSEv07_peekSymbol(const FSEv07_DState_t* DStatePtr)
+{
+    FSEv07_decode_t const DInfo = ((const FSEv07_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    return DInfo.symbol;
+}
+
+MEM_STATIC void FSEv07_updateState(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD)
+{
+    FSEv07_decode_t const DInfo = ((const FSEv07_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    size_t const lowBits = BITv07_readBits(bitD, nbBits);
+    DStatePtr->state = DInfo.newState + lowBits;
+}
+
+MEM_STATIC BYTE FSEv07_decodeSymbol(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD)
+{
+    FSEv07_decode_t const DInfo = ((const FSEv07_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BITv07_readBits(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+/*! FSEv07_decodeSymbolFast() :
+    unsafe, only works if no symbol has a probability > 50% */
+MEM_STATIC BYTE FSEv07_decodeSymbolFast(FSEv07_DState_t* DStatePtr, BITv07_DStream_t* bitD)
+{
+    FSEv07_decode_t const DInfo = ((const FSEv07_decode_t*)(DStatePtr->table))[DStatePtr->state];
+    U32 const nbBits = DInfo.nbBits;
+    BYTE const symbol = DInfo.symbol;
+    size_t const lowBits = BITv07_readBitsFast(bitD, nbBits);
+
+    DStatePtr->state = DInfo.newState + lowBits;
+    return symbol;
+}
+
+
+
+#ifndef FSEv07_COMMONDEFS_ONLY
+
+/* **************************************************************
+*  Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+*  Increasing memory usage improves compression ratio
+*  Reduced memory usage can improve speed, due to cache effect
+*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#define FSEv07_MAX_MEMORY_USAGE 14
+#define FSEv07_DEFAULT_MEMORY_USAGE 13
+
+/*!FSEv07_MAX_SYMBOL_VALUE :
+*  Maximum symbol value authorized.
+*  Required for proper stack allocation */
+#define FSEv07_MAX_SYMBOL_VALUE 255
+
+
+/* **************************************************************
+*  template functions type & suffix
+****************************************************************/
+#define FSEv07_FUNCTION_TYPE BYTE
+#define FSEv07_FUNCTION_EXTENSION
+#define FSEv07_DECODE_TYPE FSEv07_decode_t
+
+
+#endif   /* !FSEv07_COMMONDEFS_ONLY */
+
+
+/* ***************************************************************
+*  Constants
+*****************************************************************/
+#define FSEv07_MAX_TABLELOG  (FSEv07_MAX_MEMORY_USAGE-2)
+#define FSEv07_MAX_TABLESIZE (1U<<FSEv07_MAX_TABLELOG)
+#define FSEv07_MAXTABLESIZE_MASK (FSEv07_MAX_TABLESIZE-1)
+#define FSEv07_DEFAULT_TABLELOG (FSEv07_DEFAULT_MEMORY_USAGE-2)
+#define FSEv07_MIN_TABLELOG 5
+
+#define FSEv07_TABLELOG_ABSOLUTE_MAX 15
+#if FSEv07_MAX_TABLELOG > FSEv07_TABLELOG_ABSOLUTE_MAX
+#  error "FSEv07_MAX_TABLELOG > FSEv07_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+#define FSEv07_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
+
+
+#endif /* FSEv07_STATIC_LINKING_ONLY */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* FSEv07_H */
+/* ******************************************************************
+   Huffman coder, part of New Generation Entropy library
+   header file
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+****************************************************************** */
+#ifndef HUFv07_H_298734234
+#define HUFv07_H_298734234
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+
+/* *** simple functions *** */
+/**
+HUFv07_decompress() :
+    Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
+    into already allocated buffer 'dst', of minimum size 'dstSize'.
+    `dstSize` : **must** be the ***exact*** size of original (uncompressed) data.
+    Note : in contrast with FSE, HUFv07_decompress can regenerate
+           RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
+           because it knows size to regenerate.
+    @return : size of regenerated data (== dstSize),
+              or an error code, which can be tested using HUFv07_isError()
+*/
+size_t HUFv07_decompress(void* dst,  size_t dstSize,
+                const void* cSrc, size_t cSrcSize);
+
+
+/* ****************************************
+*  Tool functions
+******************************************/
+#define HUFv07_BLOCKSIZE_MAX (128 * 1024)
+
+/* Error Management */
+unsigned    HUFv07_isError(size_t code);        /**< tells if a return value is an error code */
+const char* HUFv07_getErrorName(size_t code);   /**< provides error code string (useful for debugging) */
+
+
+/* *** Advanced function *** */
+
+
+#ifdef HUFv07_STATIC_LINKING_ONLY
+
+
+/* *** Constants *** */
+#define HUFv07_TABLELOG_ABSOLUTEMAX  16   /* absolute limit of HUFv07_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUFv07_TABLELOG_MAX  12           /* max configured tableLog (for static allocation); can be modified up to HUFv07_ABSOLUTEMAX_TABLELOG */
+#define HUFv07_TABLELOG_DEFAULT  11       /* tableLog by default, when not specified */
+#define HUFv07_SYMBOLVALUE_MAX 255
+#if (HUFv07_TABLELOG_MAX > HUFv07_TABLELOG_ABSOLUTEMAX)
+#  error "HUFv07_TABLELOG_MAX is too large !"
+#endif
+
+
+/* ****************************************
+*  Static allocation
+******************************************/
+/* HUF buffer bounds */
+#define HUFv07_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true if incompressible pre-filtered with fast heuristic */
+
+/* static allocation of HUF's DTable */
+typedef U32 HUFv07_DTable;
+#define HUFv07_DTABLE_SIZE(maxTableLog)   (1 + (1<<(maxTableLog)))
+#define HUFv07_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+        HUFv07_DTable DTable[HUFv07_DTABLE_SIZE((maxTableLog)-1)] = { ((U32)((maxTableLog)-1)*0x1000001) }
+#define HUFv07_CREATE_STATIC_DTABLEX4(DTable, maxTableLog) \
+        HUFv07_DTable DTable[HUFv07_DTABLE_SIZE(maxTableLog)] = { ((U32)(maxTableLog)*0x1000001) }
+
+
+/* ****************************************
+*  Advanced decompression functions
+******************************************/
+size_t HUFv07_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUFv07_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+
+size_t HUFv07_decompress4X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< decodes RLE and uncompressed */
+size_t HUFv07_decompress4X_hufOnly(HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize); /**< considers RLE and uncompressed as errors */
+size_t HUFv07_decompress4X2_DCtx(HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUFv07_decompress4X4_DCtx(HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+
+size_t HUFv07_decompress1X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+size_t HUFv07_decompress1X2_DCtx(HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
+size_t HUFv07_decompress1X4_DCtx(HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
+
+
+/* ****************************************
+*  HUF detailed API
+******************************************/
+/*!
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and regenerate 'CTable' using external methods.
+*/
+/* FSEv07_count() : find it within "fse.h" */
+
+/*! HUFv07_readStats() :
+    Read compact Huffman tree, saved by HUFv07_writeCTable().
+    `huffWeight` is destination buffer.
+    @return : size read from `src` , or an error Code .
+    Note : Needed by HUFv07_readCTable() and HUFv07_readDTableXn() . */
+size_t HUFv07_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize);
+
+
+/*
+HUFv07_decompress() does the following:
+1. select the decompression algorithm (X2, X4) based on pre-computed heuristics
+2. build Huffman table from save, using HUFv07_readDTableXn()
+3. decode 1 or 4 segments in parallel using HUFv07_decompressSXn_usingDTable
+*/
+
+/** HUFv07_selectDecoder() :
+*   Tells which decoder is likely to decode faster,
+*   based on a set of pre-determined metrics.
+*   @return : 0==HUFv07_decompress4X2, 1==HUFv07_decompress4X4 .
+*   Assumption : 0 < cSrcSize < dstSize <= 128 KB */
+U32 HUFv07_selectDecoder (size_t dstSize, size_t cSrcSize);
+
+size_t HUFv07_readDTableX2 (HUFv07_DTable* DTable, const void* src, size_t srcSize);
+size_t HUFv07_readDTableX4 (HUFv07_DTable* DTable, const void* src, size_t srcSize);
+
+size_t HUFv07_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+size_t HUFv07_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+size_t HUFv07_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+
+
+/* single stream variants */
+size_t HUFv07_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUFv07_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+
+size_t HUFv07_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+size_t HUFv07_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+size_t HUFv07_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUFv07_DTable* DTable);
+
+
+#endif /* HUFv07_STATIC_LINKING_ONLY */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* HUFv07_H_298734234 */
+/*
+   Common functions of New Generation Entropy library
+   Copyright (C) 2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+*************************************************************************** */
+
+
+
+/*-****************************************
+*  FSE Error Management
+******************************************/
+unsigned FSEv07_isError(size_t code) { return ERR_isError(code); }
+
+const char* FSEv07_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/* **************************************************************
+*  HUF Error Management
+****************************************************************/
+unsigned HUFv07_isError(size_t code) { return ERR_isError(code); }
+
+const char* HUFv07_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/*-**************************************************************
+*  FSE NCount encoding-decoding
+****************************************************************/
+static short FSEv07_abs(short a) { return (short)(a<0 ? -a : a); }
+
+size_t FSEv07_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                 const void* headerBuffer, size_t hbSize)
+{
+    const BYTE* const istart = (const BYTE*) headerBuffer;
+    const BYTE* const iend = istart + hbSize;
+    const BYTE* ip = istart;
+    int nbBits;
+    int remaining;
+    int threshold;
+    U32 bitStream;
+    int bitCount;
+    unsigned charnum = 0;
+    int previous0 = 0;
+
+    if (hbSize < 4) return ERROR(srcSize_wrong);
+    bitStream = MEM_readLE32(ip);
+    nbBits = (bitStream & 0xF) + FSEv07_MIN_TABLELOG;   /* extract tableLog */
+    if (nbBits > FSEv07_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+    bitStream >>= 4;
+    bitCount = 4;
+    *tableLogPtr = nbBits;
+    remaining = (1<<nbBits)+1;
+    threshold = 1<<nbBits;
+    nbBits++;
+
+    while ((remaining>1) && (charnum<=*maxSVPtr)) {
+        if (previous0) {
+            unsigned n0 = charnum;
+            while ((bitStream & 0xFFFF) == 0xFFFF) {
+                n0+=24;
+                if (ip < iend-5) {
+                    ip+=2;
+                    bitStream = MEM_readLE32(ip) >> bitCount;
+                } else {
+                    bitStream >>= 16;
+                    bitCount+=16;
+            }   }
+            while ((bitStream & 3) == 3) {
+                n0+=3;
+                bitStream>>=2;
+                bitCount+=2;
+            }
+            n0 += bitStream & 3;
+            bitCount += 2;
+            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
+            while (charnum < n0) normalizedCounter[charnum++] = 0;
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+                bitStream = MEM_readLE32(ip) >> bitCount;
+            }
+            else
+                bitStream >>= 2;
+        }
+        {   short const max = (short)((2*threshold-1)-remaining);
+            short count;
+
+            if ((bitStream & (threshold-1)) < (U32)max) {
+                count = (short)(bitStream & (threshold-1));
+                bitCount   += nbBits-1;
+            } else {
+                count = (short)(bitStream & (2*threshold-1));
+                if (count >= threshold) count -= max;
+                bitCount   += nbBits;
+            }
+
+            count--;   /* extra accuracy */
+            remaining -= FSEv07_abs(count);
+            normalizedCounter[charnum++] = count;
+            previous0 = !count;
+            while (remaining < threshold) {
+                nbBits--;
+                threshold >>= 1;
+            }
+
+            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+                ip += bitCount>>3;
+                bitCount &= 7;
+            } else {
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
+    }   }   /* while ((remaining>1) && (charnum<=*maxSVPtr)) */
+    if (remaining != 1) return ERROR(GENERIC);
+    *maxSVPtr = charnum-1;
+
+    ip += (bitCount+7)>>3;
+    if ((size_t)(ip-istart) > hbSize) return ERROR(srcSize_wrong);
+    return ip-istart;
+}
+
+
+/*! HUFv07_readStats() :
+    Read compact Huffman tree, saved by HUFv07_writeCTable().
+    `huffWeight` is destination buffer.
+    @return : size read from `src` , or an error Code .
+    Note : Needed by HUFv07_readCTable() and HUFv07_readDTableXn() .
+*/
+size_t HUFv07_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize)
+{
+    U32 weightTotal;
+    const BYTE* ip = (const BYTE*) src;
+    size_t iSize;
+    size_t oSize;
+
+    if (!srcSize) return ERROR(srcSize_wrong);
+    iSize = ip[0];
+    //memset(huffWeight, 0, hwSize);   /* is not necessary, even though some analyzer complain ... */
+
+    if (iSize >= 128)  { /* special header */
+        if (iSize >= (242)) {  /* RLE */
+            static U32 l[14] = { 1, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 };
+            oSize = l[iSize-242];
+            memset(huffWeight, 1, hwSize);
+            iSize = 0;
+        }
+        else {   /* Incompressible */
+            oSize = iSize - 127;
+            iSize = ((oSize+1)/2);
+            if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+            if (oSize >= hwSize) return ERROR(corruption_detected);
+            ip += 1;
+            {   U32 n;
+                for (n=0; n<oSize; n+=2) {
+                    huffWeight[n]   = ip[n/2] >> 4;
+                    huffWeight[n+1] = ip[n/2] & 15;
+    }   }   }   }
+    else  {   /* header compressed with FSE (normal case) */
+        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+        oSize = FSEv07_decompress(huffWeight, hwSize-1, ip+1, iSize);   /* max (hwSize-1) values decoded, as last one is implied */
+        if (FSEv07_isError(oSize)) return oSize;
+    }
+
+    /* collect weight stats */
+    memset(rankStats, 0, (HUFv07_TABLELOG_ABSOLUTEMAX + 1) * sizeof(U32));
+    weightTotal = 0;
+    {   U32 n; for (n=0; n<oSize; n++) {
+            if (huffWeight[n] >= HUFv07_TABLELOG_ABSOLUTEMAX) return ERROR(corruption_detected);
+            rankStats[huffWeight[n]]++;
+            weightTotal += (1 << huffWeight[n]) >> 1;
+    }   }
+    if (weightTotal == 0) return ERROR(corruption_detected);
+
+    /* get last non-null symbol weight (implied, total must be 2^n) */
+    {   U32 const tableLog = BITv07_highbit32(weightTotal) + 1;
+        if (tableLog > HUFv07_TABLELOG_ABSOLUTEMAX) return ERROR(corruption_detected);
+        *tableLogPtr = tableLog;
+        /* determine last weight */
+        {   U32 const total = 1 << tableLog;
+            U32 const rest = total - weightTotal;
+            U32 const verif = 1 << BITv07_highbit32(rest);
+            U32 const lastWeight = BITv07_highbit32(rest) + 1;
+            if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
+            huffWeight[oSize] = (BYTE)lastWeight;
+            rankStats[lastWeight]++;
+    }   }
+
+    /* check tree construction validity */
+    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
+
+    /* results */
+    *nbSymbolsPtr = (U32)(oSize+1);
+    return iSize+1;
+}
+/* ******************************************************************
+   FSE : Finite State Entropy decoder
+   Copyright (C) 2013-2015, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  define FORCE_INLINE static __forceinline
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define FORCE_INLINE static inline __attribute__((always_inline))
+#    else
+#      define FORCE_INLINE static inline
+#    endif
+#  else
+#    define FORCE_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSEv07_isError ERR_isError
+#define FSEv07_STATIC_ASSERT(c) { enum { FSEv07_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Complex types
+****************************************************************/
+typedef U32 DTable_max_t[FSEv07_DTABLE_SIZE_U32(FSEv07_MAX_TABLELOG)];
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSEv07_FUNCTION_EXTENSION
+#  error "FSEv07_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSEv07_FUNCTION_TYPE
+#  error "FSEv07_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSEv07_CAT(X,Y) X##Y
+#define FSEv07_FUNCTION_NAME(X,Y) FSEv07_CAT(X,Y)
+#define FSEv07_TYPE_NAME(X,Y) FSEv07_CAT(X,Y)
+
+
+/* Function templates */
+FSEv07_DTable* FSEv07_createDTable (unsigned tableLog)
+{
+    if (tableLog > FSEv07_TABLELOG_ABSOLUTE_MAX) tableLog = FSEv07_TABLELOG_ABSOLUTE_MAX;
+    return (FSEv07_DTable*)malloc( FSEv07_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
+}
+
+void FSEv07_freeDTable (FSEv07_DTable* dt)
+{
+    free(dt);
+}
+
+size_t FSEv07_buildDTable(FSEv07_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+{
+    void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
+    FSEv07_DECODE_TYPE* const tableDecode = (FSEv07_DECODE_TYPE*) (tdPtr);
+    U16 symbolNext[FSEv07_MAX_SYMBOL_VALUE+1];
+
+    U32 const maxSV1 = maxSymbolValue + 1;
+    U32 const tableSize = 1 << tableLog;
+    U32 highThreshold = tableSize-1;
+
+    /* Sanity Checks */
+    if (maxSymbolValue > FSEv07_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSEv07_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    {   FSEv07_DTableHeader DTableH;
+        DTableH.tableLog = (U16)tableLog;
+        DTableH.fastMode = 1;
+        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
+            U32 s;
+            for (s=0; s<maxSV1; s++) {
+                if (normalizedCounter[s]==-1) {
+                    tableDecode[highThreshold--].symbol = (FSEv07_FUNCTION_TYPE)s;
+                    symbolNext[s] = 1;
+                } else {
+                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+                    symbolNext[s] = normalizedCounter[s];
+        }   }   }
+        memcpy(dt, &DTableH, sizeof(DTableH));
+    }
+
+    /* Spread symbols */
+    {   U32 const tableMask = tableSize-1;
+        U32 const step = FSEv07_TABLESTEP(tableSize);
+        U32 s, position = 0;
+        for (s=0; s<maxSV1; s++) {
+            int i;
+            for (i=0; i<normalizedCounter[s]; i++) {
+                tableDecode[position].symbol = (FSEv07_FUNCTION_TYPE)s;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }   }
+
+        if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+    }
+
+    /* Build Decoding table */
+    {   U32 u;
+        for (u=0; u<tableSize; u++) {
+            FSEv07_FUNCTION_TYPE const symbol = (FSEv07_FUNCTION_TYPE)(tableDecode[u].symbol);
+            U16 nextState = symbolNext[symbol]++;
+            tableDecode[u].nbBits = (BYTE) (tableLog - BITv07_highbit32 ((U32)nextState) );
+            tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+    }   }
+
+    return 0;
+}
+
+
+
+#ifndef FSEv07_COMMONDEFS_ONLY
+
+/*-*******************************************************
+*  Decompression (Byte symbols)
+*********************************************************/
+size_t FSEv07_buildDTable_rle (FSEv07_DTable* dt, BYTE symbolValue)
+{
+    void* ptr = dt;
+    FSEv07_DTableHeader* const DTableH = (FSEv07_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv07_decode_t* const cell = (FSEv07_decode_t*)dPtr;
+
+    DTableH->tableLog = 0;
+    DTableH->fastMode = 0;
+
+    cell->newState = 0;
+    cell->symbol = symbolValue;
+    cell->nbBits = 0;
+
+    return 0;
+}
+
+
+size_t FSEv07_buildDTable_raw (FSEv07_DTable* dt, unsigned nbBits)
+{
+    void* ptr = dt;
+    FSEv07_DTableHeader* const DTableH = (FSEv07_DTableHeader*)ptr;
+    void* dPtr = dt + 1;
+    FSEv07_decode_t* const dinfo = (FSEv07_decode_t*)dPtr;
+    const unsigned tableSize = 1 << nbBits;
+    const unsigned tableMask = tableSize - 1;
+    const unsigned maxSV1 = tableMask+1;
+    unsigned s;
+
+    /* Sanity checks */
+    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
+
+    /* Build Decoding Table */
+    DTableH->tableLog = (U16)nbBits;
+    DTableH->fastMode = 1;
+    for (s=0; s<maxSV1; s++) {
+        dinfo[s].newState = 0;
+        dinfo[s].symbol = (BYTE)s;
+        dinfo[s].nbBits = (BYTE)nbBits;
+    }
+
+    return 0;
+}
+
+FORCE_INLINE size_t FSEv07_decompress_usingDTable_generic(
+          void* dst, size_t maxDstSize,
+    const void* cSrc, size_t cSrcSize,
+    const FSEv07_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BITv07_DStream_t bitD;
+    FSEv07_DState_t state1;
+    FSEv07_DState_t state2;
+
+    /* Init */
+    { size_t const errorCode = BITv07_initDStream(&bitD, cSrc, cSrcSize);   /* replaced last arg by maxCompressed Size */
+      if (FSEv07_isError(errorCode)) return errorCode; }
+
+    FSEv07_initDState(&state1, &bitD, dt);
+    FSEv07_initDState(&state2, &bitD, dt);
+
+#define FSEv07_GETSYMBOL(statePtr) fast ? FSEv07_decodeSymbolFast(statePtr, &bitD) : FSEv07_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BITv07_reloadDStream(&bitD)==BITv07_DStream_unfinished) && (op<olimit) ; op+=4) {
+        op[0] = FSEv07_GETSYMBOL(&state1);
+
+        if (FSEv07_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv07_reloadDStream(&bitD);
+
+        op[1] = FSEv07_GETSYMBOL(&state2);
+
+        if (FSEv07_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BITv07_reloadDStream(&bitD) > BITv07_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSEv07_GETSYMBOL(&state1);
+
+        if (FSEv07_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BITv07_reloadDStream(&bitD);
+
+        op[3] = FSEv07_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BITv07_reloadDStream(&bitD) >= FSEv07_DStream_partiallyFilled; Ends at exactly BITv07_DStream_completed */
+    while (1) {
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+
+        *op++ = FSEv07_GETSYMBOL(&state1);
+
+        if (BITv07_reloadDStream(&bitD)==BITv07_DStream_overflow) {
+            *op++ = FSEv07_GETSYMBOL(&state2);
+            break;
+        }
+
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+
+        *op++ = FSEv07_GETSYMBOL(&state2);
+
+        if (BITv07_reloadDStream(&bitD)==BITv07_DStream_overflow) {
+            *op++ = FSEv07_GETSYMBOL(&state1);
+            break;
+    }   }
+
+    return op-ostart;
+}
+
+
+size_t FSEv07_decompress_usingDTable(void* dst, size_t originalSize,
+                            const void* cSrc, size_t cSrcSize,
+                            const FSEv07_DTable* dt)
+{
+    const void* ptr = dt;
+    const FSEv07_DTableHeader* DTableH = (const FSEv07_DTableHeader*)ptr;
+    const U32 fastMode = DTableH->fastMode;
+
+    /* select fast mode (static) */
+    if (fastMode) return FSEv07_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
+    return FSEv07_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
+}
+
+
+size_t FSEv07_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    short counting[FSEv07_MAX_SYMBOL_VALUE+1];
+    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSEv07_MAX_SYMBOL_VALUE;
+
+    if (cSrcSize<2) return ERROR(srcSize_wrong);   /* too small input size */
+
+    /* normal FSE decoding mode */
+    {   size_t const NCountLength = FSEv07_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
+        if (FSEv07_isError(NCountLength)) return NCountLength;
+        if (NCountLength >= cSrcSize) return ERROR(srcSize_wrong);   /* too small input size */
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
+
+    { size_t const errorCode = FSEv07_buildDTable (dt, counting, maxSymbolValue, tableLog);
+      if (FSEv07_isError(errorCode)) return errorCode; }
+
+    return FSEv07_decompress_usingDTable (dst, maxDstSize, ip, cSrcSize, dt);   /* always return, even if it is an error code */
+}
+
+
+
+#endif   /* FSEv07_COMMONDEFS_ONLY */
+
+/* ******************************************************************
+   Huffman decoder, part of New Generation Entropy library
+   Copyright (C) 2013-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+    - Public forum : https://groups.google.com/forum/#!forum/lz4c
+****************************************************************** */
+
+/* **************************************************************
+*  Compiler specifics
+****************************************************************/
+#if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+/* inline is defined */
+#elif defined(_MSC_VER)
+#  define inline __inline
+#else
+#  define inline /* disable inline */
+#endif
+
+
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define HUFv07_STATIC_ASSERT(c) { enum { HUFv07_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+
+/*-***************************/
+/*  generic DTableDesc       */
+/*-***************************/
+
+typedef struct { BYTE maxTableLog; BYTE tableType; BYTE tableLog; BYTE reserved; } DTableDesc;
+
+static DTableDesc HUFv07_getDTableDesc(const HUFv07_DTable* table)
+{
+    DTableDesc dtd;
+    memcpy(&dtd, table, sizeof(dtd));
+    return dtd;
+}
+
+
+/*-***************************/
+/*  single-symbol decoding   */
+/*-***************************/
+
+typedef struct { BYTE byte; BYTE nbBits; } HUFv07_DEltX2;   /* single-symbol decoding */
+
+size_t HUFv07_readDTableX2 (HUFv07_DTable* DTable, const void* src, size_t srcSize)
+{
+    BYTE huffWeight[HUFv07_SYMBOLVALUE_MAX + 1];
+    U32 rankVal[HUFv07_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
+    U32 tableLog = 0;
+    U32 nbSymbols = 0;
+    size_t iSize;
+    void* const dtPtr = DTable + 1;
+    HUFv07_DEltX2* const dt = (HUFv07_DEltX2*)dtPtr;
+
+    HUFv07_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUFv07_DTable));
+    //memset(huffWeight, 0, sizeof(huffWeight));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv07_readStats(huffWeight, HUFv07_SYMBOLVALUE_MAX + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv07_isError(iSize)) return iSize;
+
+    /* Table header */
+    {   DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+        if (tableLog > (U32)(dtd.maxTableLog+1)) return ERROR(tableLog_tooLarge);   /* DTable too small, huffman tree cannot fit in */
+        dtd.tableType = 0;
+        dtd.tableLog = (BYTE)tableLog;
+        memcpy(DTable, &dtd, sizeof(dtd));
+    }
+
+    /* Prepare ranks */
+    {   U32 n, nextRankStart = 0;
+        for (n=1; n<tableLog+1; n++) {
+            U32 current = nextRankStart;
+            nextRankStart += (rankVal[n] << (n-1));
+            rankVal[n] = current;
+    }   }
+
+    /* fill DTable */
+    {   U32 n;
+        for (n=0; n<nbSymbols; n++) {
+            U32 const w = huffWeight[n];
+            U32 const length = (1 << w) >> 1;
+            U32 i;
+            HUFv07_DEltX2 D;
+            D.byte = (BYTE)n; D.nbBits = (BYTE)(tableLog + 1 - w);
+            for (i = rankVal[w]; i < rankVal[w] + length; i++)
+                dt[i] = D;
+            rankVal[w] += length;
+    }   }
+
+    return iSize;
+}
+
+
+static BYTE HUFv07_decodeSymbolX2(BITv07_DStream_t* Dstream, const HUFv07_DEltX2* dt, const U32 dtLog)
+{
+    size_t const val = BITv07_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
+    BYTE const c = dt[val].byte;
+    BITv07_skipBits(Dstream, dt[val].nbBits);
+    return c;
+}
+
+#define HUFv07_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
+    *ptr++ = HUFv07_decodeSymbolX2(DStreamPtr, dt, dtLog)
+
+#define HUFv07_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv07_TABLELOG_MAX<=12)) \
+        HUFv07_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+#define HUFv07_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        HUFv07_DECODE_SYMBOLX2_0(ptr, DStreamPtr)
+
+static inline size_t HUFv07_decodeStreamX2(BYTE* p, BITv07_DStream_t* const bitDPtr, BYTE* const pEnd, const HUFv07_DEltX2* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 4 symbols at a time */
+    while ((BITv07_reloadDStream(bitDPtr) == BITv07_DStream_unfinished) && (p <= pEnd-4)) {
+        HUFv07_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX2_1(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX2_2(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX2_0(p, bitDPtr);
+    }
+
+    /* closer to the end */
+    while ((BITv07_reloadDStream(bitDPtr) == BITv07_DStream_unfinished) && (p < pEnd))
+        HUFv07_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    /* no more data to retrieve from bitstream, hence no need to reload */
+    while (p < pEnd)
+        HUFv07_DECODE_SYMBOLX2_0(p, bitDPtr);
+
+    return pEnd-pStart;
+}
+
+static size_t HUFv07_decompress1X2_usingDTable_internal(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + dstSize;
+    const void* dtPtr = DTable + 1;
+    const HUFv07_DEltX2* const dt = (const HUFv07_DEltX2*)dtPtr;
+    BITv07_DStream_t bitD;
+    DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+    U32 const dtLog = dtd.tableLog;
+
+    { size_t const errorCode = BITv07_initDStream(&bitD, cSrc, cSrcSize);
+      if (HUFv07_isError(errorCode)) return errorCode; }
+
+    HUFv07_decodeStreamX2(op, &bitD, oend, dt, dtLog);
+
+    /* check */
+    if (!BITv07_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    return dstSize;
+}
+
+size_t HUFv07_decompress1X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+    if (dtd.tableType != 0) return ERROR(GENERIC);
+    return HUFv07_decompress1X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+
+size_t HUFv07_decompress1X2_DCtx (HUFv07_DTable* DCtx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUFv07_readDTableX2 (DCtx, cSrc, cSrcSize);
+    if (HUFv07_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUFv07_decompress1X2_usingDTable_internal (dst, dstSize, ip, cSrcSize, DCtx);
+}
+
+size_t HUFv07_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv07_CREATE_STATIC_DTABLEX2(DTable, HUFv07_TABLELOG_MAX);
+    return HUFv07_decompress1X2_DCtx (DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+
+static size_t HUFv07_decompress4X2_usingDTable_internal(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    /* Check */
+    if (cSrcSize < 10) return ERROR(corruption_detected);  /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable + 1;
+        const HUFv07_DEltX2* const dt = (const HUFv07_DEltX2*)dtPtr;
+
+        /* Init */
+        BITv07_DStream_t bitD1;
+        BITv07_DStream_t bitD2;
+        BITv07_DStream_t bitD3;
+        BITv07_DStream_t bitD4;
+        size_t const length1 = MEM_readLE16(istart);
+        size_t const length2 = MEM_readLE16(istart+2);
+        size_t const length3 = MEM_readLE16(istart+4);
+        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        const size_t segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+        DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+        U32 const dtLog = dtd.tableLog;
+
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        { size_t const errorCode = BITv07_initDStream(&bitD1, istart1, length1);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD2, istart2, length2);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD3, istart3, length3);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD4, istart4, length4);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BITv07_reloadDStream(&bitD1) | BITv07_reloadDStream(&bitD2) | BITv07_reloadDStream(&bitD3) | BITv07_reloadDStream(&bitD4);
+        for ( ; (endSignal==BITv07_DStream_unfinished) && (op4<(oend-7)) ; ) {
+            HUFv07_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX2_1(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX2_1(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX2_1(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX2_1(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX2_2(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX2_2(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX2_2(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX2_2(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX2_0(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX2_0(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX2_0(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX2_0(op4, &bitD4);
+            endSignal = BITv07_reloadDStream(&bitD1) | BITv07_reloadDStream(&bitD2) | BITv07_reloadDStream(&bitD3) | BITv07_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUFv07_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
+        HUFv07_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
+        HUFv07_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
+        HUFv07_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        endSignal = BITv07_endOfDStream(&bitD1) & BITv07_endOfDStream(&bitD2) & BITv07_endOfDStream(&bitD3) & BITv07_endOfDStream(&bitD4);
+        if (!endSignal) return ERROR(corruption_detected);
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+size_t HUFv07_decompress4X2_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+    if (dtd.tableType != 0) return ERROR(GENERIC);
+    return HUFv07_decompress4X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+
+
+size_t HUFv07_decompress4X2_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUFv07_readDTableX2 (dctx, cSrc, cSrcSize);
+    if (HUFv07_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUFv07_decompress4X2_usingDTable_internal (dst, dstSize, ip, cSrcSize, dctx);
+}
+
+size_t HUFv07_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv07_CREATE_STATIC_DTABLEX2(DTable, HUFv07_TABLELOG_MAX);
+    return HUFv07_decompress4X2_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+
+/* *************************/
+/* double-symbols decoding */
+/* *************************/
+typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUFv07_DEltX4;  /* double-symbols decoding */
+
+typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+
+static void HUFv07_fillDTableX4Level2(HUFv07_DEltX4* DTable, U32 sizeLog, const U32 consumed,
+                           const U32* rankValOrigin, const int minWeight,
+                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    HUFv07_DEltX4 DElt;
+    U32 rankVal[HUFv07_TABLELOG_ABSOLUTEMAX + 1];
+
+    /* get pre-calculated rankVal */
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill skipped values */
+    if (minWeight>1) {
+        U32 i, skipSize = rankVal[minWeight];
+        MEM_writeLE16(&(DElt.sequence), baseSeq);
+        DElt.nbBits   = (BYTE)(consumed);
+        DElt.length   = 1;
+        for (i = 0; i < skipSize; i++)
+            DTable[i] = DElt;
+    }
+
+    /* fill DTable */
+    { U32 s; for (s=0; s<sortedListSize; s++) {   /* note : sortedSymbols already skipped */
+        const U32 symbol = sortedSymbols[s].symbol;
+        const U32 weight = sortedSymbols[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 length = 1 << (sizeLog-nbBits);
+        const U32 start = rankVal[weight];
+        U32 i = start;
+        const U32 end = start + length;
+
+        MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
+        DElt.nbBits = (BYTE)(nbBits + consumed);
+        DElt.length = 2;
+        do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
+
+        rankVal[weight] += length;
+    }}
+}
+
+typedef U32 rankVal_t[HUFv07_TABLELOG_ABSOLUTEMAX][HUFv07_TABLELOG_ABSOLUTEMAX + 1];
+
+static void HUFv07_fillDTableX4(HUFv07_DEltX4* DTable, const U32 targetLog,
+                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
+                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const U32 nbBitsBaseline)
+{
+    U32 rankVal[HUFv07_TABLELOG_ABSOLUTEMAX + 1];
+    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
+    const U32 minBits  = nbBitsBaseline - maxWeight;
+    U32 s;
+
+    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+
+    /* fill DTable */
+    for (s=0; s<sortedListSize; s++) {
+        const U16 symbol = sortedList[s].symbol;
+        const U32 weight = sortedList[s].weight;
+        const U32 nbBits = nbBitsBaseline - weight;
+        const U32 start = rankVal[weight];
+        const U32 length = 1 << (targetLog-nbBits);
+
+        if (targetLog-nbBits >= minBits) {   /* enough room for a second symbol */
+            U32 sortedRank;
+            int minWeight = nbBits + scaleLog;
+            if (minWeight < 1) minWeight = 1;
+            sortedRank = rankStart[minWeight];
+            HUFv07_fillDTableX4Level2(DTable+start, targetLog-nbBits, nbBits,
+                           rankValOrigin[nbBits], minWeight,
+                           sortedList+sortedRank, sortedListSize-sortedRank,
+                           nbBitsBaseline, symbol);
+        } else {
+            HUFv07_DEltX4 DElt;
+            MEM_writeLE16(&(DElt.sequence), symbol);
+            DElt.nbBits = (BYTE)(nbBits);
+            DElt.length = 1;
+            {   U32 u;
+                const U32 end = start + length;
+                for (u = start; u < end; u++) DTable[u] = DElt;
+        }   }
+        rankVal[weight] += length;
+    }
+}
+
+size_t HUFv07_readDTableX4 (HUFv07_DTable* DTable, const void* src, size_t srcSize)
+{
+    BYTE weightList[HUFv07_SYMBOLVALUE_MAX + 1];
+    sortedSymbol_t sortedSymbol[HUFv07_SYMBOLVALUE_MAX + 1];
+    U32 rankStats[HUFv07_TABLELOG_ABSOLUTEMAX + 1] = { 0 };
+    U32 rankStart0[HUFv07_TABLELOG_ABSOLUTEMAX + 2] = { 0 };
+    U32* const rankStart = rankStart0+1;
+    rankVal_t rankVal;
+    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+    U32 const maxTableLog = dtd.maxTableLog;
+    size_t iSize;
+    void* dtPtr = DTable+1;   /* force compiler to avoid strict-aliasing */
+    HUFv07_DEltX4* const dt = (HUFv07_DEltX4*)dtPtr;
+
+    HUFv07_STATIC_ASSERT(sizeof(HUFv07_DEltX4) == sizeof(HUFv07_DTable));   /* if compilation fails here, assertion is false */
+    if (maxTableLog > HUFv07_TABLELOG_ABSOLUTEMAX) return ERROR(tableLog_tooLarge);
+    //memset(weightList, 0, sizeof(weightList));   /* is not necessary, even though some analyzer complain ... */
+
+    iSize = HUFv07_readStats(weightList, HUFv07_SYMBOLVALUE_MAX + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    if (HUFv07_isError(iSize)) return iSize;
+
+    /* check result */
+    if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+
+    /* find maxWeight */
+    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+
+    /* Get start index of each weight */
+    {   U32 w, nextRankStart = 0;
+        for (w=1; w<maxW+1; w++) {
+            U32 current = nextRankStart;
+            nextRankStart += rankStats[w];
+            rankStart[w] = current;
+        }
+        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
+        sizeOfSort = nextRankStart;
+    }
+
+    /* sort symbols by weight */
+    {   U32 s;
+        for (s=0; s<nbSymbols; s++) {
+            U32 const w = weightList[s];
+            U32 const r = rankStart[w]++;
+            sortedSymbol[r].symbol = (BYTE)s;
+            sortedSymbol[r].weight = (BYTE)w;
+        }
+        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
+    }
+
+    /* Build rankVal */
+    {   U32* const rankVal0 = rankVal[0];
+        {   int const rescale = (maxTableLog-tableLog) - 1;   /* tableLog <= maxTableLog */
+            U32 nextRankVal = 0;
+            U32 w;
+            for (w=1; w<maxW+1; w++) {
+                U32 current = nextRankVal;
+                nextRankVal += rankStats[w] << (w+rescale);
+                rankVal0[w] = current;
+        }   }
+        {   U32 const minBits = tableLog+1 - maxW;
+            U32 consumed;
+            for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
+                U32* const rankValPtr = rankVal[consumed];
+                U32 w;
+                for (w = 1; w < maxW+1; w++) {
+                    rankValPtr[w] = rankVal0[w] >> consumed;
+    }   }   }   }
+
+    HUFv07_fillDTableX4(dt, maxTableLog,
+                   sortedSymbol, sizeOfSort,
+                   rankStart0, rankVal, maxW,
+                   tableLog+1);
+
+    dtd.tableLog = (BYTE)maxTableLog;
+    dtd.tableType = 1;
+    memcpy(DTable, &dtd, sizeof(dtd));
+    return iSize;
+}
+
+
+static U32 HUFv07_decodeSymbolX4(void* op, BITv07_DStream_t* DStream, const HUFv07_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv07_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 2);
+    BITv07_skipBits(DStream, dt[val].nbBits);
+    return dt[val].length;
+}
+
+static U32 HUFv07_decodeLastSymbolX4(void* op, BITv07_DStream_t* DStream, const HUFv07_DEltX4* dt, const U32 dtLog)
+{
+    const size_t val = BITv07_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
+    memcpy(op, dt+val, 1);
+    if (dt[val].length==1) BITv07_skipBits(DStream, dt[val].nbBits);
+    else {
+        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
+            BITv07_skipBits(DStream, dt[val].nbBits);
+            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
+                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);   /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
+    }   }
+    return 1;
+}
+
+
+#define HUFv07_DECODE_SYMBOLX4_0(ptr, DStreamPtr) \
+    ptr += HUFv07_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv07_DECODE_SYMBOLX4_1(ptr, DStreamPtr) \
+    if (MEM_64bits() || (HUFv07_TABLELOG_MAX<=12)) \
+        ptr += HUFv07_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+#define HUFv07_DECODE_SYMBOLX4_2(ptr, DStreamPtr) \
+    if (MEM_64bits()) \
+        ptr += HUFv07_decodeSymbolX4(ptr, DStreamPtr, dt, dtLog)
+
+static inline size_t HUFv07_decodeStreamX4(BYTE* p, BITv07_DStream_t* bitDPtr, BYTE* const pEnd, const HUFv07_DEltX4* const dt, const U32 dtLog)
+{
+    BYTE* const pStart = p;
+
+    /* up to 8 symbols at a time */
+    while ((BITv07_reloadDStream(bitDPtr) == BITv07_DStream_unfinished) && (p < pEnd-7)) {
+        HUFv07_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX4_1(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX4_2(p, bitDPtr);
+        HUFv07_DECODE_SYMBOLX4_0(p, bitDPtr);
+    }
+
+    /* closer to end : up to 2 symbols at a time */
+    while ((BITv07_reloadDStream(bitDPtr) == BITv07_DStream_unfinished) && (p <= pEnd-2))
+        HUFv07_DECODE_SYMBOLX4_0(p, bitDPtr);
+
+    while (p <= pEnd-2)
+        HUFv07_DECODE_SYMBOLX4_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+
+    if (p < pEnd)
+        p += HUFv07_decodeLastSymbolX4(p, bitDPtr, dt, dtLog);
+
+    return p-pStart;
+}
+
+
+static size_t HUFv07_decompress1X4_usingDTable_internal(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    BITv07_DStream_t bitD;
+
+    /* Init */
+    {   size_t const errorCode = BITv07_initDStream(&bitD, cSrc, cSrcSize);
+        if (HUFv07_isError(errorCode)) return errorCode;
+    }
+
+    /* decode */
+    {   BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable+1;   /* force compiler to not use strict-aliasing */
+        const HUFv07_DEltX4* const dt = (const HUFv07_DEltX4*)dtPtr;
+        DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+        HUFv07_decodeStreamX4(ostart, &bitD, oend, dt, dtd.tableLog);
+    }
+
+    /* check */
+    if (!BITv07_endOfDStream(&bitD)) return ERROR(corruption_detected);
+
+    /* decoded size */
+    return dstSize;
+}
+
+size_t HUFv07_decompress1X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+    if (dtd.tableType != 1) return ERROR(GENERIC);
+    return HUFv07_decompress1X4_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+
+size_t HUFv07_decompress1X4_DCtx (HUFv07_DTable* DCtx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t const hSize = HUFv07_readDTableX4 (DCtx, cSrc, cSrcSize);
+    if (HUFv07_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUFv07_decompress1X4_usingDTable_internal (dst, dstSize, ip, cSrcSize, DCtx);
+}
+
+size_t HUFv07_decompress1X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv07_CREATE_STATIC_DTABLEX4(DTable, HUFv07_TABLELOG_MAX);
+    return HUFv07_decompress1X4_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+static size_t HUFv07_decompress4X4_usingDTable_internal(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+
+    {   const BYTE* const istart = (const BYTE*) cSrc;
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* const oend = ostart + dstSize;
+        const void* const dtPtr = DTable+1;
+        const HUFv07_DEltX4* const dt = (const HUFv07_DEltX4*)dtPtr;
+
+        /* Init */
+        BITv07_DStream_t bitD1;
+        BITv07_DStream_t bitD2;
+        BITv07_DStream_t bitD3;
+        BITv07_DStream_t bitD4;
+        size_t const length1 = MEM_readLE16(istart);
+        size_t const length2 = MEM_readLE16(istart+2);
+        size_t const length3 = MEM_readLE16(istart+4);
+        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
+        const BYTE* const istart1 = istart + 6;  /* jumpTable */
+        const BYTE* const istart2 = istart1 + length1;
+        const BYTE* const istart3 = istart2 + length2;
+        const BYTE* const istart4 = istart3 + length3;
+        size_t const segmentSize = (dstSize+3) / 4;
+        BYTE* const opStart2 = ostart + segmentSize;
+        BYTE* const opStart3 = opStart2 + segmentSize;
+        BYTE* const opStart4 = opStart3 + segmentSize;
+        BYTE* op1 = ostart;
+        BYTE* op2 = opStart2;
+        BYTE* op3 = opStart3;
+        BYTE* op4 = opStart4;
+        U32 endSignal;
+        DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+        U32 const dtLog = dtd.tableLog;
+
+        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        { size_t const errorCode = BITv07_initDStream(&bitD1, istart1, length1);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD2, istart2, length2);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD3, istart3, length3);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+        { size_t const errorCode = BITv07_initDStream(&bitD4, istart4, length4);
+          if (HUFv07_isError(errorCode)) return errorCode; }
+
+        /* 16-32 symbols per loop (4-8 symbols per stream) */
+        endSignal = BITv07_reloadDStream(&bitD1) | BITv07_reloadDStream(&bitD2) | BITv07_reloadDStream(&bitD3) | BITv07_reloadDStream(&bitD4);
+        for ( ; (endSignal==BITv07_DStream_unfinished) && (op4<(oend-7)) ; ) {
+            HUFv07_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX4_1(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX4_1(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX4_1(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX4_1(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX4_2(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX4_2(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX4_2(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX4_2(op4, &bitD4);
+            HUFv07_DECODE_SYMBOLX4_0(op1, &bitD1);
+            HUFv07_DECODE_SYMBOLX4_0(op2, &bitD2);
+            HUFv07_DECODE_SYMBOLX4_0(op3, &bitD3);
+            HUFv07_DECODE_SYMBOLX4_0(op4, &bitD4);
+
+            endSignal = BITv07_reloadDStream(&bitD1) | BITv07_reloadDStream(&bitD2) | BITv07_reloadDStream(&bitD3) | BITv07_reloadDStream(&bitD4);
+        }
+
+        /* check corruption */
+        if (op1 > opStart2) return ERROR(corruption_detected);
+        if (op2 > opStart3) return ERROR(corruption_detected);
+        if (op3 > opStart4) return ERROR(corruption_detected);
+        /* note : op4 supposed already verified within main loop */
+
+        /* finish bitStreams one by one */
+        HUFv07_decodeStreamX4(op1, &bitD1, opStart2, dt, dtLog);
+        HUFv07_decodeStreamX4(op2, &bitD2, opStart3, dt, dtLog);
+        HUFv07_decodeStreamX4(op3, &bitD3, opStart4, dt, dtLog);
+        HUFv07_decodeStreamX4(op4, &bitD4, oend,     dt, dtLog);
+
+        /* check */
+        { U32 const endCheck = BITv07_endOfDStream(&bitD1) & BITv07_endOfDStream(&bitD2) & BITv07_endOfDStream(&bitD3) & BITv07_endOfDStream(&bitD4);
+          if (!endCheck) return ERROR(corruption_detected); }
+
+        /* decoded size */
+        return dstSize;
+    }
+}
+
+
+size_t HUFv07_decompress4X4_usingDTable(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUFv07_DTable* DTable)
+{
+    DTableDesc dtd = HUFv07_getDTableDesc(DTable);
+    if (dtd.tableType != 1) return ERROR(GENERIC);
+    return HUFv07_decompress4X4_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+
+
+size_t HUFv07_decompress4X4_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    const BYTE* ip = (const BYTE*) cSrc;
+
+    size_t hSize = HUFv07_readDTableX4 (dctx, cSrc, cSrcSize);
+    if (HUFv07_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
+
+    return HUFv07_decompress4X4_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx);
+}
+
+size_t HUFv07_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    HUFv07_CREATE_STATIC_DTABLEX4(DTable, HUFv07_TABLELOG_MAX);
+    return HUFv07_decompress4X4_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+}
+
+
+/* ********************************/
+/* Generic decompression selector */
+/* ********************************/
+
+size_t HUFv07_decompress1X_usingDTable(void* dst, size_t maxDstSize,
+                                    const void* cSrc, size_t cSrcSize,
+                                    const HUFv07_DTable* DTable)
+{
+    DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+    return dtd.tableType ? HUFv07_decompress1X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable) :
+                           HUFv07_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable);
+}
+
+size_t HUFv07_decompress4X_usingDTable(void* dst, size_t maxDstSize,
+                                    const void* cSrc, size_t cSrcSize,
+                                    const HUFv07_DTable* DTable)
+{
+    DTableDesc const dtd = HUFv07_getDTableDesc(DTable);
+    return dtd.tableType ? HUFv07_decompress4X4_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable) :
+                           HUFv07_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable);
+}
+
+
+typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
+static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+{
+    /* single, double, quad */
+    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
+    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
+    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
+    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
+    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
+    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
+    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
+    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
+    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
+    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
+    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
+    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
+    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
+    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
+    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+};
+
+/** HUFv07_selectDecoder() :
+*   Tells which decoder is likely to decode faster,
+*   based on a set of pre-determined metrics.
+*   @return : 0==HUFv07_decompress4X2, 1==HUFv07_decompress4X4 .
+*   Assumption : 0 < cSrcSize < dstSize <= 128 KB */
+U32 HUFv07_selectDecoder (size_t dstSize, size_t cSrcSize)
+{
+    /* decoder timing evaluation */
+    U32 const Q = (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 since dstSize > cSrcSize */
+    U32 const D256 = (U32)(dstSize >> 8);
+    U32 const DTime0 = algoTime[Q][0].tableTime + (algoTime[Q][0].decode256Time * D256);
+    U32 DTime1 = algoTime[Q][1].tableTime + (algoTime[Q][1].decode256Time * D256);
+    DTime1 += DTime1 >> 3;  /* advantage to algorithm using less memory, for cache eviction */
+
+    return DTime1 < DTime0;
+}
+
+
+typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
+
+size_t HUFv07_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    static const decompressionAlgo decompress[2] = { HUFv07_decompress4X2, HUFv07_decompress4X4 };
+
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUFv07_selectDecoder(dstSize, cSrcSize);
+        return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
+    }
+
+    //return HUFv07_decompress4X2(dst, dstSize, cSrc, cSrcSize);   /* multi-streams single-symbol decoding */
+    //return HUFv07_decompress4X4(dst, dstSize, cSrc, cSrcSize);   /* multi-streams double-symbols decoding */
+}
+
+size_t HUFv07_decompress4X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUFv07_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUFv07_decompress4X4_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) :
+                        HUFv07_decompress4X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) ;
+    }
+}
+
+size_t HUFv07_decompress4X_hufOnly (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if ((cSrcSize >= dstSize) || (cSrcSize <= 1)) return ERROR(corruption_detected);   /* invalid */
+
+    {   U32 const algoNb = HUFv07_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUFv07_decompress4X4_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) :
+                        HUFv07_decompress4X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) ;
+    }
+}
+
+size_t HUFv07_decompress1X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
+{
+    /* validation checks */
+    if (dstSize == 0) return ERROR(dstSize_tooSmall);
+    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
+    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+
+    {   U32 const algoNb = HUFv07_selectDecoder(dstSize, cSrcSize);
+        return algoNb ? HUFv07_decompress1X4_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) :
+                        HUFv07_decompress1X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) ;
+    }
+}
+/*
+    Common functions of Zstd compression library
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net/
+*/
+
+
+
+/*-****************************************
+*  ZSTD Error Management
+******************************************/
+/*! ZSTDv07_isError() :
+*   tells if a return value is an error code */
+unsigned ZSTDv07_isError(size_t code) { return ERR_isError(code); }
+
+/*! ZSTDv07_getErrorName() :
+*   provides error code string from function result (useful for debugging) */
+const char* ZSTDv07_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+
+/* **************************************************************
+*  ZBUFF Error Management
+****************************************************************/
+unsigned ZBUFFv07_isError(size_t errorCode) { return ERR_isError(errorCode); }
+
+const char* ZBUFFv07_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
+
+
+
+void* ZSTDv07_defaultAllocFunction(void* opaque, size_t size)
+{
+    void* address = malloc(size);
+    (void)opaque;
+    /* printf("alloc %p, %d opaque=%p \n", address, (int)size, opaque); */
+    return address;
+}
+
+void ZSTDv07_defaultFreeFunction(void* opaque, void* address)
+{
+    (void)opaque;
+    /* if (address) printf("free %p opaque=%p \n", address, opaque); */
+    free(address);
+}
+/*
+    zstd_internal - common functions to include
+    Header File for include
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : https://www.zstd.net
+*/
+#ifndef ZSTDv07_CCOMMON_H_MODULE
+#define ZSTDv07_CCOMMON_H_MODULE
+
+
+/*-*************************************
+*  Common macros
+***************************************/
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+
+
+/*-*************************************
+*  Common constants
+***************************************/
+#define ZSTDv07_OPT_NUM    (1<<12)
+#define ZSTDv07_DICT_MAGIC  0xEC30A437   /* v0.7 */
+
+#define ZSTDv07_REP_NUM    3
+#define ZSTDv07_REP_INIT   ZSTDv07_REP_NUM
+#define ZSTDv07_REP_MOVE   (ZSTDv07_REP_NUM-1)
+static const U32 repStartValue[ZSTDv07_REP_NUM] = { 1, 4, 8 };
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BIT7 128
+#define BIT6  64
+#define BIT5  32
+#define BIT4  16
+#define BIT1   2
+#define BIT0   1
+
+#define ZSTDv07_WINDOWLOG_ABSOLUTEMIN 10
+static const size_t ZSTDv07_fcs_fieldSize[4] = { 0, 2, 4, 8 };
+static const size_t ZSTDv07_did_fieldSize[4] = { 0, 1, 2, 4 };
+
+#define ZSTDv07_BLOCKHEADERSIZE 3   /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
+static const size_t ZSTDv07_blockHeaderSize = ZSTDv07_BLOCKHEADERSIZE;
+typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
+
+#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+
+#define HufLog 12
+typedef enum { lbt_huffman, lbt_repeat, lbt_raw, lbt_rle } litBlockType_t;
+
+#define LONGNBSEQ 0x7F00
+
+#define MINMATCH 3
+#define EQUAL_READ32 4
+
+#define Litbits  8
+#define MaxLit ((1<<Litbits) - 1)
+#define MaxML  52
+#define MaxLL  35
+#define MaxOff 28
+#define MaxSeq MAX(MaxLL, MaxML)   /* Assumption : MaxOff < MaxLL,MaxML */
+#define MLFSELog    9
+#define LLFSELog    9
+#define OffFSELog   8
+
+#define FSEv07_ENCODING_RAW     0
+#define FSEv07_ENCODING_RLE     1
+#define FSEv07_ENCODING_STATIC  2
+#define FSEv07_ENCODING_DYNAMIC 3
+
+static const U32 LL_bits[MaxLL+1] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3, 4, 6, 7, 8, 9,10,11,12,
+                                     13,14,15,16 };
+static const S16 LL_defaultNorm[MaxLL+1] = { 4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1,
+                                             2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1,
+                                            -1,-1,-1,-1 };
+static const U32 LL_defaultNormLog = 6;
+
+static const U32 ML_bits[MaxML+1] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                      1, 1, 1, 1, 2, 2, 3, 3, 4, 4, 5, 7, 8, 9,10,11,
+                                     12,13,14,15,16 };
+static const S16 ML_defaultNorm[MaxML+1] = { 1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                                             1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,-1,-1,
+                                            -1,-1,-1,-1,-1 };
+static const U32 ML_defaultNormLog = 6;
+
+static const S16 OF_defaultNorm[MaxOff+1] = { 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+                                              1, 1, 1, 1, 1, 1, 1, 1,-1,-1,-1,-1,-1 };
+static const U32 OF_defaultNormLog = 5;
+
+
+/*-*******************************************
+*  Shared functions to include for inlining
+*********************************************/
+static void ZSTDv07_copy8(void* dst, const void* src) { memcpy(dst, src, 8); }
+#define COPY8(d,s) { ZSTDv07_copy8(d,s); d+=8; s+=8; }
+
+/*! ZSTDv07_wildcopy() :
+*   custom version of memcpy(), can copy up to 7 bytes too many (8 bytes if length==0) */
+#define WILDCOPY_OVERLENGTH 8
+MEM_STATIC void ZSTDv07_wildcopy(void* dst, const void* src, ptrdiff_t length)
+{
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    BYTE* const oend = op + length;
+    do
+        COPY8(op, ip)
+    while (op < oend);
+}
+
+
+/*-*******************************************
+*  Private interfaces
+*********************************************/
+typedef struct ZSTDv07_stats_s ZSTDv07_stats_t;
+
+typedef struct {
+    U32 off;
+    U32 len;
+} ZSTDv07_match_t;
+
+typedef struct {
+    U32 price;
+    U32 off;
+    U32 mlen;
+    U32 litlen;
+    U32 rep[ZSTDv07_REP_INIT];
+} ZSTDv07_optimal_t;
+
+struct ZSTDv07_stats_s { U32 unused; };
+
+typedef struct {
+    void* buffer;
+    U32*  offsetStart;
+    U32*  offset;
+    BYTE* offCodeStart;
+    BYTE* litStart;
+    BYTE* lit;
+    U16*  litLengthStart;
+    U16*  litLength;
+    BYTE* llCodeStart;
+    U16*  matchLengthStart;
+    U16*  matchLength;
+    BYTE* mlCodeStart;
+    U32   longLengthID;   /* 0 == no longLength; 1 == Lit.longLength; 2 == Match.longLength; */
+    U32   longLengthPos;
+    /* opt */
+    ZSTDv07_optimal_t* priceTable;
+    ZSTDv07_match_t* matchTable;
+    U32* matchLengthFreq;
+    U32* litLengthFreq;
+    U32* litFreq;
+    U32* offCodeFreq;
+    U32  matchLengthSum;
+    U32  matchSum;
+    U32  litLengthSum;
+    U32  litSum;
+    U32  offCodeSum;
+    U32  log2matchLengthSum;
+    U32  log2matchSum;
+    U32  log2litLengthSum;
+    U32  log2litSum;
+    U32  log2offCodeSum;
+    U32  factor;
+    U32  cachedPrice;
+    U32  cachedLitLength;
+    const BYTE* cachedLiterals;
+    ZSTDv07_stats_t stats;
+} seqStore_t;
+
+void ZSTDv07_seqToCodes(const seqStore_t* seqStorePtr, size_t const nbSeq);
+
+/* custom memory allocation functions */
+static const ZSTDv07_customMem defaultCustomMem = { ZSTDv07_defaultAllocFunction, ZSTDv07_defaultFreeFunction, NULL };
+
+#endif   /* ZSTDv07_CCOMMON_H_MODULE */
+/*
+    zstd - standard compression library
+    Copyright (C) 2014-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net
+*/
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * HEAPMODE :
+ * Select how default decompression function ZSTDv07_decompress() will allocate memory,
+ * in memory stack (0), or in memory heap (1, requires malloc())
+ */
+#ifndef ZSTDv07_HEAPMODE
+#  define ZSTDv07_HEAPMODE 1
+#endif
+
+
+/*-*******************************************************
+*  Compiler specifics
+*********************************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  include <intrin.h>                    /* For Visual 2005 */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
+#  pragma warning(disable : 4100)        /* disable: C4100: unreferenced formal parameter */
+#endif
+
+
+/*-*************************************
+*  Macros
+***************************************/
+#define ZSTDv07_isError ERR_isError   /* for inlining */
+#define FSEv07_isError  ERR_isError
+#define HUFv07_isError  ERR_isError
+
+
+/*_*******************************************************
+*  Memory operations
+**********************************************************/
+static void ZSTDv07_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+
+
+/*-*************************************************************
+*   Context management
+***************************************************************/
+typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
+               ZSTDds_decodeBlockHeader, ZSTDds_decompressBlock,
+               ZSTDds_decodeSkippableHeader, ZSTDds_skipFrame } ZSTDv07_dStage;
+
+struct ZSTDv07_DCtx_s
+{
+    FSEv07_DTable LLTable[FSEv07_DTABLE_SIZE_U32(LLFSELog)];
+    FSEv07_DTable OffTable[FSEv07_DTABLE_SIZE_U32(OffFSELog)];
+    FSEv07_DTable MLTable[FSEv07_DTABLE_SIZE_U32(MLFSELog)];
+    HUFv07_DTable hufTable[HUFv07_DTABLE_SIZE(HufLog)];  /* can accommodate HUFv07_decompress4X */
+    const void* previousDstEnd;
+    const void* base;
+    const void* vBase;
+    const void* dictEnd;
+    size_t expected;
+    U32 rep[3];
+    ZSTDv07_frameParams fParams;
+    blockType_t bType;   /* used in ZSTDv07_decompressContinue(), to transfer blockType between header decoding and block decoding stages */
+    ZSTDv07_dStage stage;
+    U32 litEntropy;
+    U32 fseEntropy;
+    XXH64_state_t xxhState;
+    size_t headerSize;
+    U32 dictID;
+    const BYTE* litPtr;
+    ZSTDv07_customMem customMem;
+    size_t litSize;
+    BYTE litBuffer[ZSTDv07_BLOCKSIZE_ABSOLUTEMAX + WILDCOPY_OVERLENGTH];
+    BYTE headerBuffer[ZSTDv07_FRAMEHEADERSIZE_MAX];
+};  /* typedef'd to ZSTDv07_DCtx within "zstd_static.h" */
+
+int ZSTDv07_isSkipFrame(ZSTDv07_DCtx* dctx);
+
+size_t ZSTDv07_sizeofDCtx (const ZSTDv07_DCtx* dctx) { return sizeof(*dctx); }
+
+size_t ZSTDv07_estimateDCtxSize(void) { return sizeof(ZSTDv07_DCtx); }
+
+size_t ZSTDv07_decompressBegin(ZSTDv07_DCtx* dctx)
+{
+    dctx->expected = ZSTDv07_frameHeaderSize_min;
+    dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->previousDstEnd = NULL;
+    dctx->base = NULL;
+    dctx->vBase = NULL;
+    dctx->dictEnd = NULL;
+    dctx->hufTable[0] = (HUFv07_DTable)((HufLog)*0x1000001);
+    dctx->litEntropy = dctx->fseEntropy = 0;
+    dctx->dictID = 0;
+    { int i; for (i=0; i<ZSTDv07_REP_NUM; i++) dctx->rep[i] = repStartValue[i]; }
+    return 0;
+}
+
+ZSTDv07_DCtx* ZSTDv07_createDCtx_advanced(ZSTDv07_customMem customMem)
+{
+    ZSTDv07_DCtx* dctx;
+
+    if (!customMem.customAlloc && !customMem.customFree)
+        customMem = defaultCustomMem;
+
+    if (!customMem.customAlloc || !customMem.customFree)
+        return NULL;
+
+    dctx = (ZSTDv07_DCtx*) customMem.customAlloc(customMem.opaque, sizeof(ZSTDv07_DCtx));
+    if (!dctx) return NULL;
+    memcpy(&dctx->customMem, &customMem, sizeof(ZSTDv07_customMem));
+    ZSTDv07_decompressBegin(dctx);
+    return dctx;
+}
+
+ZSTDv07_DCtx* ZSTDv07_createDCtx(void)
+{
+    return ZSTDv07_createDCtx_advanced(defaultCustomMem);
+}
+
+size_t ZSTDv07_freeDCtx(ZSTDv07_DCtx* dctx)
+{
+    if (dctx==NULL) return 0;   /* support free on NULL */
+    dctx->customMem.customFree(dctx->customMem.opaque, dctx);
+    return 0;   /* reserved as a potential error code in the future */
+}
+
+void ZSTDv07_copyDCtx(ZSTDv07_DCtx* dstDCtx, const ZSTDv07_DCtx* srcDCtx)
+{
+    memcpy(dstDCtx, srcDCtx,
+           sizeof(ZSTDv07_DCtx) - (ZSTDv07_BLOCKSIZE_ABSOLUTEMAX+WILDCOPY_OVERLENGTH + ZSTDv07_frameHeaderSize_max));  /* no need to copy workspace */
+}
+
+
+/*-*************************************************************
+*   Decompression section
+***************************************************************/
+
+/* Frame format description
+   Frame Header -  [ Block Header - Block ] - Frame End
+   1) Frame Header
+      - 4 bytes - Magic Number : ZSTDv07_MAGICNUMBER (defined within zstd.h)
+      - 1 byte  - Frame Descriptor
+   2) Block Header
+      - 3 bytes, starting with a 2-bits descriptor
+                 Uncompressed, Compressed, Frame End, unused
+   3) Block
+      See Block Format Description
+   4) Frame End
+      - 3 bytes, compatible with Block Header
+*/
+
+
+/* Frame Header :
+
+   1 byte - FrameHeaderDescription :
+   bit 0-1 : dictID (0, 1, 2 or 4 bytes)
+   bit 2   : checksumFlag
+   bit 3   : reserved (must be zero)
+   bit 4   : reserved (unused, can be any value)
+   bit 5   : Single Segment (if 1, WindowLog byte is not present)
+   bit 6-7 : FrameContentFieldSize (0, 2, 4, or 8)
+             if (SkippedWindowLog && !FrameContentFieldsize) FrameContentFieldsize=1;
+
+   Optional : WindowLog (0 or 1 byte)
+   bit 0-2 : octal Fractional (1/8th)
+   bit 3-7 : Power of 2, with 0 = 1 KB (up to 2 TB)
+
+   Optional : dictID (0, 1, 2 or 4 bytes)
+   Automatic adaptation
+   0 : no dictID
+   1 : 1 - 255
+   2 : 256 - 65535
+   4 : all other values
+
+   Optional : content size (0, 1, 2, 4 or 8 bytes)
+   0 : unknown          (fcfs==0 and swl==0)
+   1 : 0-255 bytes      (fcfs==0 and swl==1)
+   2 : 256 - 65535+256  (fcfs==1)
+   4 : 0 - 4GB-1        (fcfs==2)
+   8 : 0 - 16EB-1       (fcfs==3)
+*/
+
+
+/* Compressed Block, format description
+
+   Block = Literal Section - Sequences Section
+   Prerequisite : size of (compressed) block, maximum size of regenerated data
+
+   1) Literal Section
+
+   1.1) Header : 1-5 bytes
+        flags: 2 bits
+            00 compressed by Huff0
+            01 unused
+            10 is Raw (uncompressed)
+            11 is Rle
+            Note : using 01 => Huff0 with precomputed table ?
+            Note : delta map ? => compressed ?
+
+   1.1.1) Huff0-compressed literal block : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+   1.1.2) Raw (uncompressed) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RAW<<6) + (0<<4) + size
+               12 bits: (IS_RAW<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RAW<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.3) Rle (repeated single byte) literal block header : 1-3 bytes
+        size :  5 bits: (IS_RLE<<6) + (0<<4) + size
+               12 bits: (IS_RLE<<6) + (2<<4) + (size>>8)
+                        size&255
+               20 bits: (IS_RLE<<6) + (3<<4) + (size>>16)
+                        size>>8&255
+                        size&255
+
+   1.1.4) Huff0-compressed literal block, using precomputed CTables : 3-5 bytes
+            srcSize < 1 KB => 3 bytes (2-2-10-10) => single stream
+            srcSize < 1 KB => 3 bytes (2-2-10-10)
+            srcSize < 16KB => 4 bytes (2-2-14-14)
+            else           => 5 bytes (2-2-18-18)
+            big endian convention
+
+        1- CTable available (stored into workspace ?)
+        2- Small input (fast heuristic ? Full comparison ? depend on clevel ?)
+
+
+   1.2) Literal block content
+
+   1.2.1) Huff0 block, using sizes from header
+        See Huff0 format
+
+   1.2.2) Huff0 block, using prepared table
+
+   1.2.3) Raw content
+
+   1.2.4) single byte
+
+
+   2) Sequences section
+      TO DO
+*/
+
+/** ZSTDv07_frameHeaderSize() :
+*   srcSize must be >= ZSTDv07_frameHeaderSize_min.
+*   @return : size of the Frame Header */
+static size_t ZSTDv07_frameHeaderSize(const void* src, size_t srcSize)
+{
+    if (srcSize < ZSTDv07_frameHeaderSize_min) return ERROR(srcSize_wrong);
+    {   BYTE const fhd = ((const BYTE*)src)[4];
+        U32 const dictID= fhd & 3;
+        U32 const directMode = (fhd >> 5) & 1;
+        U32 const fcsId = fhd >> 6;
+        return ZSTDv07_frameHeaderSize_min + !directMode + ZSTDv07_did_fieldSize[dictID] + ZSTDv07_fcs_fieldSize[fcsId]
+                + (directMode && !ZSTDv07_fcs_fieldSize[fcsId]);
+    }
+}
+
+
+/** ZSTDv07_getFrameParams() :
+*   decode Frame Header, or require larger `srcSize`.
+*   @return : 0, `fparamsPtr` is correctly filled,
+*            >0, `srcSize` is too small, result is expected `srcSize`,
+*             or an error code, which can be tested using ZSTDv07_isError() */
+size_t ZSTDv07_getFrameParams(ZSTDv07_frameParams* fparamsPtr, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+
+    if (srcSize < ZSTDv07_frameHeaderSize_min) return ZSTDv07_frameHeaderSize_min;
+    if (MEM_readLE32(src) != ZSTDv07_MAGICNUMBER) {
+        if ((MEM_readLE32(src) & 0xFFFFFFF0U) == ZSTDv07_MAGIC_SKIPPABLE_START) {
+            if (srcSize < ZSTDv07_skippableHeaderSize) return ZSTDv07_skippableHeaderSize; /* magic number + skippable frame length */
+            memset(fparamsPtr, 0, sizeof(*fparamsPtr));
+            fparamsPtr->frameContentSize = MEM_readLE32((const char *)src + 4);
+            fparamsPtr->windowSize = 0; /* windowSize==0 means a frame is skippable */
+            return 0;
+        }
+        return ERROR(prefix_unknown);
+    }
+
+    /* ensure there is enough `srcSize` to fully read/decode frame header */
+    { size_t const fhsize = ZSTDv07_frameHeaderSize(src, srcSize);
+      if (srcSize < fhsize) return fhsize; }
+
+    {   BYTE const fhdByte = ip[4];
+        size_t pos = 5;
+        U32 const dictIDSizeCode = fhdByte&3;
+        U32 const checksumFlag = (fhdByte>>2)&1;
+        U32 const directMode = (fhdByte>>5)&1;
+        U32 const fcsID = fhdByte>>6;
+        U32 const windowSizeMax = 1U << ZSTDv07_WINDOWLOG_MAX;
+        U32 windowSize = 0;
+        U32 dictID = 0;
+        U64 frameContentSize = 0;
+        if ((fhdByte & 0x08) != 0) return ERROR(frameParameter_unsupported);   /* reserved bits, which must be zero */
+        if (!directMode) {
+            BYTE const wlByte = ip[pos++];
+            U32 const windowLog = (wlByte >> 3) + ZSTDv07_WINDOWLOG_ABSOLUTEMIN;
+            if (windowLog > ZSTDv07_WINDOWLOG_MAX) return ERROR(frameParameter_unsupported);
+            windowSize = (1U << windowLog);
+            windowSize += (windowSize >> 3) * (wlByte&7);
+        }
+
+        switch(dictIDSizeCode)
+        {
+            default:   /* impossible */
+            case 0 : break;
+            case 1 : dictID = ip[pos]; pos++; break;
+            case 2 : dictID = MEM_readLE16(ip+pos); pos+=2; break;
+            case 3 : dictID = MEM_readLE32(ip+pos); pos+=4; break;
+        }
+        switch(fcsID)
+        {
+            default:   /* impossible */
+            case 0 : if (directMode) frameContentSize = ip[pos]; break;
+            case 1 : frameContentSize = MEM_readLE16(ip+pos)+256; break;
+            case 2 : frameContentSize = MEM_readLE32(ip+pos); break;
+            case 3 : frameContentSize = MEM_readLE64(ip+pos); break;
+        }
+        if (!windowSize) windowSize = (U32)frameContentSize;
+        if (windowSize > windowSizeMax) return ERROR(frameParameter_unsupported);
+        fparamsPtr->frameContentSize = frameContentSize;
+        fparamsPtr->windowSize = windowSize;
+        fparamsPtr->dictID = dictID;
+        fparamsPtr->checksumFlag = checksumFlag;
+    }
+    return 0;
+}
+
+
+/** ZSTDv07_getDecompressedSize() :
+*   compatible with legacy mode
+*   @return : decompressed size if known, 0 otherwise
+              note : 0 can mean any of the following :
+                   - decompressed size is not provided within frame header
+                   - frame header unknown / not supported
+                   - frame header not completely provided (`srcSize` too small) */
+unsigned long long ZSTDv07_getDecompressedSize(const void* src, size_t srcSize)
+{
+    {   ZSTDv07_frameParams fparams;
+        size_t const frResult = ZSTDv07_getFrameParams(&fparams, src, srcSize);
+        if (frResult!=0) return 0;
+        return fparams.frameContentSize;
+    }
+}
+
+
+/** ZSTDv07_decodeFrameHeader() :
+*   `srcSize` must be the size provided by ZSTDv07_frameHeaderSize().
+*   @return : 0 if success, or an error code, which can be tested using ZSTDv07_isError() */
+static size_t ZSTDv07_decodeFrameHeader(ZSTDv07_DCtx* dctx, const void* src, size_t srcSize)
+{
+    size_t const result = ZSTDv07_getFrameParams(&(dctx->fParams), src, srcSize);
+    if (dctx->fParams.dictID && (dctx->dictID != dctx->fParams.dictID)) return ERROR(dictionary_wrong);
+    if (dctx->fParams.checksumFlag) XXH64_reset(&dctx->xxhState, 0);
+    return result;
+}
+
+
+typedef struct
+{
+    blockType_t blockType;
+    U32 origSize;
+} blockProperties_t;
+
+/*! ZSTDv07_getcBlockSize() :
+*   Provides the size of compressed block from block header `src` */
+size_t ZSTDv07_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
+{
+    const BYTE* const in = (const BYTE* const)src;
+    U32 cSize;
+
+    if (srcSize < ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
+
+    bpPtr->blockType = (blockType_t)((*in) >> 6);
+    cSize = in[2] + (in[1]<<8) + ((in[0] & 7)<<16);
+    bpPtr->origSize = (bpPtr->blockType == bt_rle) ? cSize : 0;
+
+    if (bpPtr->blockType == bt_end) return 0;
+    if (bpPtr->blockType == bt_rle) return 1;
+    return cSize;
+}
+
+
+static size_t ZSTDv07_copyRawBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    if (srcSize > dstCapacity) return ERROR(dstSize_tooSmall);
+    memcpy(dst, src, srcSize);
+    return srcSize;
+}
+
+
+/*! ZSTDv07_decodeLiteralsBlock() :
+    @return : nb of bytes read from src (< srcSize ) */
+size_t ZSTDv07_decodeLiteralsBlock(ZSTDv07_DCtx* dctx,
+                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+{
+    const BYTE* const istart = (const BYTE*) src;
+
+    if (srcSize < MIN_CBLOCK_SIZE) return ERROR(corruption_detected);
+
+    switch((litBlockType_t)(istart[0]>> 6))
+    {
+    case lbt_huffman:
+        {   size_t litSize, litCSize, singleStream=0;
+            U32 lhSize = (istart[0] >> 4) & 3;
+            if (srcSize < 5) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for lhSize, + cSize (+nbSeq) */
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                /* 2 - 2 - 10 - 10 */
+                lhSize=3;
+                singleStream = istart[0] & 16;
+                litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+                litCSize = ((istart[1] &  3) << 8) + istart[2];
+                break;
+            case 2:
+                /* 2 - 2 - 14 - 14 */
+                lhSize=4;
+                litSize  = ((istart[0] & 15) << 10) + (istart[1] << 2) + (istart[2] >> 6);
+                litCSize = ((istart[2] & 63) <<  8) + istart[3];
+                break;
+            case 3:
+                /* 2 - 2 - 18 - 18 */
+                lhSize=5;
+                litSize  = ((istart[0] & 15) << 14) + (istart[1] << 6) + (istart[2] >> 2);
+                litCSize = ((istart[2] &  3) << 16) + (istart[3] << 8) + istart[4];
+                break;
+            }
+            if (litSize > ZSTDv07_BLOCKSIZE_ABSOLUTEMAX) return ERROR(corruption_detected);
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            if (HUFv07_isError(singleStream ?
+                            HUFv07_decompress1X2_DCtx(dctx->hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize) :
+                            HUFv07_decompress4X_hufOnly (dctx->hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize) ))
+                return ERROR(corruption_detected);
+
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            dctx->litEntropy = 1;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case lbt_repeat:
+        {   size_t litSize, litCSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            if (lhSize != 1)  /* only case supported for now : small litSize, single stream */
+                return ERROR(corruption_detected);
+            if (dctx->litEntropy==0)
+                return ERROR(dictionary_corrupted);
+
+            /* 2 - 2 - 10 - 10 */
+            lhSize=3;
+            litSize  = ((istart[0] & 15) << 6) + (istart[1] >> 2);
+            litCSize = ((istart[1] &  3) << 8) + istart[2];
+            if (litCSize + lhSize > srcSize) return ERROR(corruption_detected);
+
+            {   size_t const errorCode = HUFv07_decompress1X4_usingDTable(dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->hufTable);
+                if (HUFv07_isError(errorCode)) return ERROR(corruption_detected);
+            }
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+            return litCSize + lhSize;
+        }
+    case lbt_raw:
+        {   size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize=1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                break;
+            }
+
+            if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
+                if (litSize+lhSize > srcSize) return ERROR(corruption_detected);
+                memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                dctx->litPtr = dctx->litBuffer;
+                dctx->litSize = litSize;
+                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                return lhSize+litSize;
+            }
+            /* direct reference into compressed stream */
+            dctx->litPtr = istart+lhSize;
+            dctx->litSize = litSize;
+            return lhSize+litSize;
+        }
+    case lbt_rle:
+        {   size_t litSize;
+            U32 lhSize = ((istart[0]) >> 4) & 3;
+            switch(lhSize)
+            {
+            case 0: case 1: default:   /* note : default is impossible, since lhSize into [0..3] */
+                lhSize = 1;
+                litSize = istart[0] & 31;
+                break;
+            case 2:
+                litSize = ((istart[0] & 15) << 8) + istart[1];
+                break;
+            case 3:
+                litSize = ((istart[0] & 15) << 16) + (istart[1] << 8) + istart[2];
+                if (srcSize<4) return ERROR(corruption_detected);   /* srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4 */
+                break;
+            }
+            if (litSize > ZSTDv07_BLOCKSIZE_ABSOLUTEMAX) return ERROR(corruption_detected);
+            memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+            dctx->litPtr = dctx->litBuffer;
+            dctx->litSize = litSize;
+            return lhSize+1;
+        }
+    default:
+        return ERROR(corruption_detected);   /* impossible */
+    }
+}
+
+
+/*! ZSTDv07_buildSeqTable() :
+    @return : nb bytes read from src,
+              or an error code if it fails, testable with ZSTDv07_isError()
+*/
+size_t ZSTDv07_buildSeqTable(FSEv07_DTable* DTable, U32 type, U32 max, U32 maxLog,
+                                 const void* src, size_t srcSize,
+                                 const S16* defaultNorm, U32 defaultLog, U32 flagRepeatTable)
+{
+    switch(type)
+    {
+    case FSEv07_ENCODING_RLE :
+        if (!srcSize) return ERROR(srcSize_wrong);
+        if ( (*(const BYTE*)src) > max) return ERROR(corruption_detected);
+        FSEv07_buildDTable_rle(DTable, *(const BYTE*)src);   /* if *src > max, data is corrupted */
+        return 1;
+    case FSEv07_ENCODING_RAW :
+        FSEv07_buildDTable(DTable, defaultNorm, max, defaultLog);
+        return 0;
+    case FSEv07_ENCODING_STATIC:
+        if (!flagRepeatTable) return ERROR(corruption_detected);
+        return 0;
+    default :   /* impossible */
+    case FSEv07_ENCODING_DYNAMIC :
+        {   U32 tableLog;
+            S16 norm[MaxSeq+1];
+            size_t const headerSize = FSEv07_readNCount(norm, &max, &tableLog, src, srcSize);
+            if (FSEv07_isError(headerSize)) return ERROR(corruption_detected);
+            if (tableLog > maxLog) return ERROR(corruption_detected);
+            FSEv07_buildDTable(DTable, norm, max, tableLog);
+            return headerSize;
+    }   }
+}
+
+
+size_t ZSTDv07_decodeSeqHeaders(int* nbSeqPtr,
+                             FSEv07_DTable* DTableLL, FSEv07_DTable* DTableML, FSEv07_DTable* DTableOffb, U32 flagRepeatTable,
+                             const void* src, size_t srcSize)
+{
+    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip = istart;
+
+    /* check */
+    if (srcSize < MIN_SEQUENCES_SIZE) return ERROR(srcSize_wrong);
+
+    /* SeqHead */
+    {   int nbSeq = *ip++;
+        if (!nbSeq) { *nbSeqPtr=0; return 1; }
+        if (nbSeq > 0x7F) {
+            if (nbSeq == 0xFF) {
+                if (ip+2 > iend) return ERROR(srcSize_wrong);
+                nbSeq = MEM_readLE16(ip) + LONGNBSEQ, ip+=2;
+            } else {
+                if (ip >= iend) return ERROR(srcSize_wrong);
+                nbSeq = ((nbSeq-0x80)<<8) + *ip++;
+            }
+        }
+        *nbSeqPtr = nbSeq;
+    }
+
+    /* FSE table descriptors */
+    {   U32 const LLtype  = *ip >> 6;
+        U32 const OFtype = (*ip >> 4) & 3;
+        U32 const MLtype  = (*ip >> 2) & 3;
+        ip++;
+
+        /* check */
+        if (ip > iend-3) return ERROR(srcSize_wrong); /* min : all 3 are "raw", hence no header, but at least xxLog bits per type */
+
+        /* Build DTables */
+        {   size_t const llhSize = ZSTDv07_buildSeqTable(DTableLL, LLtype, MaxLL, LLFSELog, ip, iend-ip, LL_defaultNorm, LL_defaultNormLog, flagRepeatTable);
+            if (ZSTDv07_isError(llhSize)) return ERROR(corruption_detected);
+            ip += llhSize;
+        }
+        {   size_t const ofhSize = ZSTDv07_buildSeqTable(DTableOffb, OFtype, MaxOff, OffFSELog, ip, iend-ip, OF_defaultNorm, OF_defaultNormLog, flagRepeatTable);
+            if (ZSTDv07_isError(ofhSize)) return ERROR(corruption_detected);
+            ip += ofhSize;
+        }
+        {   size_t const mlhSize = ZSTDv07_buildSeqTable(DTableML, MLtype, MaxML, MLFSELog, ip, iend-ip, ML_defaultNorm, ML_defaultNormLog, flagRepeatTable);
+            if (ZSTDv07_isError(mlhSize)) return ERROR(corruption_detected);
+            ip += mlhSize;
+    }   }
+
+    return ip-istart;
+}
+
+
+typedef struct {
+    size_t litLength;
+    size_t matchLength;
+    size_t offset;
+} seq_t;
+
+typedef struct {
+    BITv07_DStream_t DStream;
+    FSEv07_DState_t stateLL;
+    FSEv07_DState_t stateOffb;
+    FSEv07_DState_t stateML;
+    size_t prevOffset[ZSTDv07_REP_INIT];
+} seqState_t;
+
+
+static seq_t ZSTDv07_decodeSequence(seqState_t* seqState)
+{
+    seq_t seq;
+
+    U32 const llCode = FSEv07_peekSymbol(&(seqState->stateLL));
+    U32 const mlCode = FSEv07_peekSymbol(&(seqState->stateML));
+    U32 const ofCode = FSEv07_peekSymbol(&(seqState->stateOffb));   /* <= maxOff, by table construction */
+
+    U32 const llBits = LL_bits[llCode];
+    U32 const mlBits = ML_bits[mlCode];
+    U32 const ofBits = ofCode;
+    U32 const totalBits = llBits+mlBits+ofBits;
+
+    static const U32 LL_base[MaxLL+1] = {
+                             0,  1,  2,  3,  4,  5,  6,  7,  8,  9,   10,    11,    12,    13,    14,     15,
+                            16, 18, 20, 22, 24, 28, 32, 40, 48, 64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
+                            0x2000, 0x4000, 0x8000, 0x10000 };
+
+    static const U32 ML_base[MaxML+1] = {
+                             3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13,   14,    15,    16,    17,    18,
+                            19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,   30,    31,    32,    33,    34,
+                            35, 37, 39, 41, 43, 47, 51, 59, 67, 83, 99, 0x83, 0x103, 0x203, 0x403, 0x803,
+                            0x1003, 0x2003, 0x4003, 0x8003, 0x10003 };
+
+    static const U32 OF_base[MaxOff+1] = {
+                 0,        1,       1,       5,     0xD,     0x1D,     0x3D,     0x7D,
+                 0xFD,   0x1FD,   0x3FD,   0x7FD,   0xFFD,   0x1FFD,   0x3FFD,   0x7FFD,
+                 0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
+                 0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD };
+
+    /* sequence */
+    {   size_t offset;
+        if (!ofCode)
+            offset = 0;
+        else {
+            offset = OF_base[ofCode] + BITv07_readBits(&(seqState->DStream), ofBits);   /* <=  (ZSTDv07_WINDOWLOG_MAX-1) bits */
+            if (MEM_32bits()) BITv07_reloadDStream(&(seqState->DStream));
+        }
+
+        if (ofCode <= 1) {
+            if ((llCode == 0) & (offset <= 1)) offset = 1-offset;
+            if (offset) {
+                size_t const temp = seqState->prevOffset[offset];
+                if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset = temp;
+            } else {
+                offset = seqState->prevOffset[0];
+            }
+        } else {
+            seqState->prevOffset[2] = seqState->prevOffset[1];
+            seqState->prevOffset[1] = seqState->prevOffset[0];
+            seqState->prevOffset[0] = offset;
+        }
+        seq.offset = offset;
+    }
+
+    seq.matchLength = ML_base[mlCode] + ((mlCode>31) ? BITv07_readBits(&(seqState->DStream), mlBits) : 0);   /* <=  16 bits */
+    if (MEM_32bits() && (mlBits+llBits>24)) BITv07_reloadDStream(&(seqState->DStream));
+
+    seq.litLength = LL_base[llCode] + ((llCode>15) ? BITv07_readBits(&(seqState->DStream), llBits) : 0);   /* <=  16 bits */
+    if (MEM_32bits() ||
+       (totalBits > 64 - 7 - (LLFSELog+MLFSELog+OffFSELog)) ) BITv07_reloadDStream(&(seqState->DStream));
+
+    /* ANS state update */
+    FSEv07_updateState(&(seqState->stateLL), &(seqState->DStream));   /* <=  9 bits */
+    FSEv07_updateState(&(seqState->stateML), &(seqState->DStream));   /* <=  9 bits */
+    if (MEM_32bits()) BITv07_reloadDStream(&(seqState->DStream));     /* <= 18 bits */
+    FSEv07_updateState(&(seqState->stateOffb), &(seqState->DStream)); /* <=  8 bits */
+
+    return seq;
+}
+
+
+static
+size_t ZSTDv07_execSequence(BYTE* op,
+                                BYTE* const oend, seq_t sequence,
+                                const BYTE** litPtr, const BYTE* const litLimit,
+                                const BYTE* const base, const BYTE* const vBase, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    BYTE* const oend_w = oend-WILDCOPY_OVERLENGTH;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    /* check */
+    if ((oLitEnd>oend_w) | (oMatchEnd>oend)) return ERROR(dstSize_tooSmall); /* last match must start at a minimum distance of WILDCOPY_OVERLENGTH from oend */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+
+    /* copy Literals */
+    ZSTDv07_wildcopy(op, *litPtr, sequence.litLength);   /* note : since oLitEnd <= oend-WILDCOPY_OVERLENGTH, no risk of overwrite beyond oend */
+    op = oLitEnd;
+    *litPtr = iLitEnd;   /* update for next sequence */
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - base)) {
+        /* offset beyond prefix */
+        if (sequence.offset > (size_t)(oLitEnd - vBase)) return ERROR(corruption_detected);
+        match = dictEnd - (base-match);
+        if (match + sequence.matchLength <= dictEnd) {
+            memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+            memmove(oLitEnd, match, length1);
+            op = oLitEnd + length1;
+            sequence.matchLength -= length1;
+            match = base;
+            if (op > oend_w || sequence.matchLength < MINMATCH) {
+              while (op < oMatchEnd) *op++ = *match++;
+              return sequenceLength;
+            }
+    }   }
+    /* Requirement: op <= oend_w */
+
+    /* match within prefix */
+    if (sequence.offset < 8) {
+        /* close range match, overlap */
+        static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
+        static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* substracted */
+        int const sub2 = dec64table[sequence.offset];
+        op[0] = match[0];
+        op[1] = match[1];
+        op[2] = match[2];
+        op[3] = match[3];
+        match += dec32table[sequence.offset];
+        ZSTDv07_copy4(op+4, match);
+        match -= sub2;
+    } else {
+        ZSTDv07_copy8(op, match);
+    }
+    op += 8; match += 8;
+
+    if (oMatchEnd > oend-(16-MINMATCH)) {
+        if (op < oend_w) {
+            ZSTDv07_wildcopy(op, match, oend_w - op);
+            match += oend_w - op;
+            op = oend_w;
+        }
+        while (op < oMatchEnd) *op++ = *match++;
+    } else {
+        ZSTDv07_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8);   /* works even if matchLength < 8 */
+    }
+    return sequenceLength;
+}
+
+
+static size_t ZSTDv07_decompressSequences(
+                               ZSTDv07_DCtx* dctx,
+                               void* dst, size_t maxDstSize,
+                         const void* seqStart, size_t seqSize)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + maxDstSize;
+    BYTE* op = ostart;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    FSEv07_DTable* DTableLL = dctx->LLTable;
+    FSEv07_DTable* DTableML = dctx->MLTable;
+    FSEv07_DTable* DTableOffb = dctx->OffTable;
+    const BYTE* const base = (const BYTE*) (dctx->base);
+    const BYTE* const vBase = (const BYTE*) (dctx->vBase);
+    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
+    int nbSeq;
+
+    /* Build Decoding Tables */
+    {   size_t const seqHSize = ZSTDv07_decodeSeqHeaders(&nbSeq, DTableLL, DTableML, DTableOffb, dctx->fseEntropy, ip, seqSize);
+        if (ZSTDv07_isError(seqHSize)) return seqHSize;
+        ip += seqHSize;
+    }
+
+    /* Regen sequences */
+    if (nbSeq) {
+        seqState_t seqState;
+        dctx->fseEntropy = 1;
+        { U32 i; for (i=0; i<ZSTDv07_REP_INIT; i++) seqState.prevOffset[i] = dctx->rep[i]; }
+        { size_t const errorCode = BITv07_initDStream(&(seqState.DStream), ip, iend-ip);
+          if (ERR_isError(errorCode)) return ERROR(corruption_detected); }
+        FSEv07_initDState(&(seqState.stateLL), &(seqState.DStream), DTableLL);
+        FSEv07_initDState(&(seqState.stateOffb), &(seqState.DStream), DTableOffb);
+        FSEv07_initDState(&(seqState.stateML), &(seqState.DStream), DTableML);
+
+        for ( ; (BITv07_reloadDStream(&(seqState.DStream)) <= BITv07_DStream_completed) && nbSeq ; ) {
+            nbSeq--;
+            {   seq_t const sequence = ZSTDv07_decodeSequence(&seqState);
+                size_t const oneSeqSize = ZSTDv07_execSequence(op, oend, sequence, &litPtr, litEnd, base, vBase, dictEnd);
+                if (ZSTDv07_isError(oneSeqSize)) return oneSeqSize;
+                op += oneSeqSize;
+        }   }
+
+        /* check if reached exact end */
+        if (nbSeq) return ERROR(corruption_detected);
+        /* save reps for next block */
+        { U32 i; for (i=0; i<ZSTDv07_REP_INIT; i++) dctx->rep[i] = (U32)(seqState.prevOffset[i]); }
+    }
+
+    /* last literal segment */
+    {   size_t const lastLLSize = litEnd - litPtr;
+        //if (litPtr > litEnd) return ERROR(corruption_detected);   /* too many literals already used */
+        if (lastLLSize > (size_t)(oend-op)) return ERROR(dstSize_tooSmall);
+        memcpy(op, litPtr, lastLLSize);
+        op += lastLLSize;
+    }
+
+    return op-ostart;
+}
+
+
+static void ZSTDv07_checkContinuity(ZSTDv07_DCtx* dctx, const void* dst)
+{
+    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+        dctx->dictEnd = dctx->previousDstEnd;
+        dctx->vBase = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+        dctx->base = dst;
+        dctx->previousDstEnd = dst;
+    }
+}
+
+
+static size_t ZSTDv07_decompressBlock_internal(ZSTDv07_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{   /* blockType == blockCompressed */
+    const BYTE* ip = (const BYTE*)src;
+
+    if (srcSize >= ZSTDv07_BLOCKSIZE_ABSOLUTEMAX) return ERROR(srcSize_wrong);
+
+    /* Decode literals sub-block */
+    {   size_t const litCSize = ZSTDv07_decodeLiteralsBlock(dctx, src, srcSize);
+        if (ZSTDv07_isError(litCSize)) return litCSize;
+        ip += litCSize;
+        srcSize -= litCSize;
+    }
+    return ZSTDv07_decompressSequences(dctx, dst, dstCapacity, ip, srcSize);
+}
+
+
+size_t ZSTDv07_decompressBlock(ZSTDv07_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize)
+{
+    size_t dSize;
+    ZSTDv07_checkContinuity(dctx, dst);
+    dSize = ZSTDv07_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize);
+    dctx->previousDstEnd = (char*)dst + dSize;
+    return dSize;
+}
+
+
+/** ZSTDv07_insertBlock() :
+    insert `src` block into `dctx` history. Useful to track uncompressed blocks. */
+ZSTDLIBv07_API size_t ZSTDv07_insertBlock(ZSTDv07_DCtx* dctx, const void* blockStart, size_t blockSize)
+{
+    ZSTDv07_checkContinuity(dctx, blockStart);
+    dctx->previousDstEnd = (const char*)blockStart + blockSize;
+    return blockSize;
+}
+
+
+size_t ZSTDv07_generateNxBytes(void* dst, size_t dstCapacity, BYTE byte, size_t length)
+{
+    if (length > dstCapacity) return ERROR(dstSize_tooSmall);
+    memset(dst, byte, length);
+    return length;
+}
+
+
+/*! ZSTDv07_decompressFrame() :
+*   `dctx` must be properly initialized */
+static size_t ZSTDv07_decompressFrame(ZSTDv07_DCtx* dctx,
+                                 void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* const iend = ip + srcSize;
+    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const oend = ostart + dstCapacity;
+    BYTE* op = ostart;
+    size_t remainingSize = srcSize;
+
+    /* check */
+    if (srcSize < ZSTDv07_frameHeaderSize_min+ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
+
+    /* Frame Header */
+    {   size_t const frameHeaderSize = ZSTDv07_frameHeaderSize(src, ZSTDv07_frameHeaderSize_min);
+        if (ZSTDv07_isError(frameHeaderSize)) return frameHeaderSize;
+        if (srcSize < frameHeaderSize+ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
+        if (ZSTDv07_decodeFrameHeader(dctx, src, frameHeaderSize)) return ERROR(corruption_detected);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1) {
+        size_t decodedSize;
+        blockProperties_t blockProperties;
+        size_t const cBlockSize = ZSTDv07_getcBlockSize(ip, iend-ip, &blockProperties);
+        if (ZSTDv07_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv07_blockHeaderSize;
+        remainingSize -= ZSTDv07_blockHeaderSize;
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        switch(blockProperties.blockType)
+        {
+        case bt_compressed:
+            decodedSize = ZSTDv07_decompressBlock_internal(dctx, op, oend-op, ip, cBlockSize);
+            break;
+        case bt_raw :
+            decodedSize = ZSTDv07_copyRawBlock(op, oend-op, ip, cBlockSize);
+            break;
+        case bt_rle :
+            decodedSize = ZSTDv07_generateNxBytes(op, oend-op, *ip, blockProperties.origSize);
+            break;
+        case bt_end :
+            /* end of frame */
+            if (remainingSize) return ERROR(srcSize_wrong);
+            decodedSize = 0;
+            break;
+        default:
+            return ERROR(GENERIC);   /* impossible */
+        }
+        if (blockProperties.blockType == bt_end) break;   /* bt_end */
+
+        if (ZSTDv07_isError(decodedSize)) return decodedSize;
+        if (dctx->fParams.checksumFlag) XXH64_update(&dctx->xxhState, op, decodedSize);
+        op += decodedSize;
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return op-ostart;
+}
+
+
+/*! ZSTDv07_decompress_usingPreparedDCtx() :
+*   Same as ZSTDv07_decompress_usingDict, but using a reference context `preparedDCtx`, where dictionary has been loaded.
+*   It avoids reloading the dictionary each time.
+*   `preparedDCtx` must have been properly initialized using ZSTDv07_decompressBegin_usingDict().
+*   Requires 2 contexts : 1 for reference (preparedDCtx), which will not be modified, and 1 to run the decompression operation (dctx) */
+size_t ZSTDv07_decompress_usingPreparedDCtx(ZSTDv07_DCtx* dctx, const ZSTDv07_DCtx* refDCtx,
+                                         void* dst, size_t dstCapacity,
+                                   const void* src, size_t srcSize)
+{
+    ZSTDv07_copyDCtx(dctx, refDCtx);
+    ZSTDv07_checkContinuity(dctx, dst);
+    return ZSTDv07_decompressFrame(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+size_t ZSTDv07_decompress_usingDict(ZSTDv07_DCtx* dctx,
+                                 void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize,
+                                 const void* dict, size_t dictSize)
+{
+    ZSTDv07_decompressBegin_usingDict(dctx, dict, dictSize);
+    ZSTDv07_checkContinuity(dctx, dst);
+    return ZSTDv07_decompressFrame(dctx, dst, dstCapacity, src, srcSize);
+}
+
+
+size_t ZSTDv07_decompressDCtx(ZSTDv07_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    return ZSTDv07_decompress_usingDict(dctx, dst, dstCapacity, src, srcSize, NULL, 0);
+}
+
+
+size_t ZSTDv07_decompress(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+#if defined(ZSTDv07_HEAPMODE) && (ZSTDv07_HEAPMODE==1)
+    size_t regenSize;
+    ZSTDv07_DCtx* const dctx = ZSTDv07_createDCtx();
+    if (dctx==NULL) return ERROR(memory_allocation);
+    regenSize = ZSTDv07_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
+    ZSTDv07_freeDCtx(dctx);
+    return regenSize;
+#else   /* stack mode */
+    ZSTDv07_DCtx dctx;
+    return ZSTDv07_decompressDCtx(&dctx, dst, dstCapacity, src, srcSize);
+#endif
+}
+
+size_t ZSTDv07_findFrameCompressedSize(const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    size_t remainingSize = srcSize;
+
+    /* check */
+    if (srcSize < ZSTDv07_frameHeaderSize_min+ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
+
+    /* Frame Header */
+    {   size_t const frameHeaderSize = ZSTDv07_frameHeaderSize(src, ZSTDv07_frameHeaderSize_min);
+        if (ZSTDv07_isError(frameHeaderSize)) return frameHeaderSize;
+        if (MEM_readLE32(src) != ZSTDv07_MAGICNUMBER) return ERROR(prefix_unknown);
+        if (srcSize < frameHeaderSize+ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
+        ip += frameHeaderSize; remainingSize -= frameHeaderSize;
+    }
+
+    /* Loop on each block */
+    while (1) {
+        blockProperties_t blockProperties;
+        size_t const cBlockSize = ZSTDv07_getcBlockSize(ip, remainingSize, &blockProperties);
+        if (ZSTDv07_isError(cBlockSize)) return cBlockSize;
+
+        ip += ZSTDv07_blockHeaderSize;
+        remainingSize -= ZSTDv07_blockHeaderSize;
+
+        if (blockProperties.blockType == bt_end) break;
+
+        if (cBlockSize > remainingSize) return ERROR(srcSize_wrong);
+
+        ip += cBlockSize;
+        remainingSize -= cBlockSize;
+    }
+
+    return ip - (const BYTE*)src;
+}
+
+/*_******************************
+*  Streaming Decompression API
+********************************/
+size_t ZSTDv07_nextSrcSizeToDecompress(ZSTDv07_DCtx* dctx)
+{
+    return dctx->expected;
+}
+
+int ZSTDv07_isSkipFrame(ZSTDv07_DCtx* dctx)
+{
+    return dctx->stage == ZSTDds_skipFrame;
+}
+
+/** ZSTDv07_decompressContinue() :
+*   @return : nb of bytes generated into `dst` (necessarily <= `dstCapacity)
+*             or an error code, which can be tested using ZSTDv07_isError() */
+size_t ZSTDv07_decompressContinue(ZSTDv07_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    /* Sanity check */
+    if (srcSize != dctx->expected) return ERROR(srcSize_wrong);
+    if (dstCapacity) ZSTDv07_checkContinuity(dctx, dst);
+
+    switch (dctx->stage)
+    {
+    case ZSTDds_getFrameHeaderSize :
+        if (srcSize != ZSTDv07_frameHeaderSize_min) return ERROR(srcSize_wrong);   /* impossible */
+        if ((MEM_readLE32(src) & 0xFFFFFFF0U) == ZSTDv07_MAGIC_SKIPPABLE_START) {
+            memcpy(dctx->headerBuffer, src, ZSTDv07_frameHeaderSize_min);
+            dctx->expected = ZSTDv07_skippableHeaderSize - ZSTDv07_frameHeaderSize_min; /* magic number + skippable frame length */
+            dctx->stage = ZSTDds_decodeSkippableHeader;
+            return 0;
+        }
+        dctx->headerSize = ZSTDv07_frameHeaderSize(src, ZSTDv07_frameHeaderSize_min);
+        if (ZSTDv07_isError(dctx->headerSize)) return dctx->headerSize;
+        memcpy(dctx->headerBuffer, src, ZSTDv07_frameHeaderSize_min);
+        if (dctx->headerSize > ZSTDv07_frameHeaderSize_min) {
+            dctx->expected = dctx->headerSize - ZSTDv07_frameHeaderSize_min;
+            dctx->stage = ZSTDds_decodeFrameHeader;
+            return 0;
+        }
+        dctx->expected = 0;   /* not necessary to copy more */
+	/* fall-through */
+    case ZSTDds_decodeFrameHeader:
+        {   size_t result;
+            memcpy(dctx->headerBuffer + ZSTDv07_frameHeaderSize_min, src, dctx->expected);
+            result = ZSTDv07_decodeFrameHeader(dctx, dctx->headerBuffer, dctx->headerSize);
+            if (ZSTDv07_isError(result)) return result;
+            dctx->expected = ZSTDv07_blockHeaderSize;
+            dctx->stage = ZSTDds_decodeBlockHeader;
+            return 0;
+        }
+    case ZSTDds_decodeBlockHeader:
+        {   blockProperties_t bp;
+            size_t const cBlockSize = ZSTDv07_getcBlockSize(src, ZSTDv07_blockHeaderSize, &bp);
+            if (ZSTDv07_isError(cBlockSize)) return cBlockSize;
+            if (bp.blockType == bt_end) {
+                if (dctx->fParams.checksumFlag) {
+                    U64 const h64 = XXH64_digest(&dctx->xxhState);
+                    U32 const h32 = (U32)(h64>>11) & ((1<<22)-1);
+                    const BYTE* const ip = (const BYTE*)src;
+                    U32 const check32 = ip[2] + (ip[1] << 8) + ((ip[0] & 0x3F) << 16);
+                    if (check32 != h32) return ERROR(checksum_wrong);
+                }
+                dctx->expected = 0;
+                dctx->stage = ZSTDds_getFrameHeaderSize;
+            } else {
+                dctx->expected = cBlockSize;
+                dctx->bType = bp.blockType;
+                dctx->stage = ZSTDds_decompressBlock;
+            }
+            return 0;
+        }
+    case ZSTDds_decompressBlock:
+        {   size_t rSize;
+            switch(dctx->bType)
+            {
+            case bt_compressed:
+                rSize = ZSTDv07_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize);
+                break;
+            case bt_raw :
+                rSize = ZSTDv07_copyRawBlock(dst, dstCapacity, src, srcSize);
+                break;
+            case bt_rle :
+                return ERROR(GENERIC);   /* not yet handled */
+                break;
+            case bt_end :   /* should never happen (filtered at phase 1) */
+                rSize = 0;
+                break;
+            default:
+                return ERROR(GENERIC);   /* impossible */
+            }
+            dctx->stage = ZSTDds_decodeBlockHeader;
+            dctx->expected = ZSTDv07_blockHeaderSize;
+            dctx->previousDstEnd = (char*)dst + rSize;
+            if (ZSTDv07_isError(rSize)) return rSize;
+            if (dctx->fParams.checksumFlag) XXH64_update(&dctx->xxhState, dst, rSize);
+            return rSize;
+        }
+    case ZSTDds_decodeSkippableHeader:
+        {   memcpy(dctx->headerBuffer + ZSTDv07_frameHeaderSize_min, src, dctx->expected);
+            dctx->expected = MEM_readLE32(dctx->headerBuffer + 4);
+            dctx->stage = ZSTDds_skipFrame;
+            return 0;
+        }
+    case ZSTDds_skipFrame:
+        {   dctx->expected = 0;
+            dctx->stage = ZSTDds_getFrameHeaderSize;
+            return 0;
+        }
+    default:
+        return ERROR(GENERIC);   /* impossible */
+    }
+}
+
+
+static size_t ZSTDv07_refDictContent(ZSTDv07_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    dctx->dictEnd = dctx->previousDstEnd;
+    dctx->vBase = (const char*)dict - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->base));
+    dctx->base = dict;
+    dctx->previousDstEnd = (const char*)dict + dictSize;
+    return 0;
+}
+
+static size_t ZSTDv07_loadEntropy(ZSTDv07_DCtx* dctx, const void* const dict, size_t const dictSize)
+{
+    const BYTE* dictPtr = (const BYTE*)dict;
+    const BYTE* const dictEnd = dictPtr + dictSize;
+
+    {   size_t const hSize = HUFv07_readDTableX4(dctx->hufTable, dict, dictSize);
+        if (HUFv07_isError(hSize)) return ERROR(dictionary_corrupted);
+        dictPtr += hSize;
+    }
+
+    {   short offcodeNCount[MaxOff+1];
+        U32 offcodeMaxValue=MaxOff, offcodeLog;
+        size_t const offcodeHeaderSize = FSEv07_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        if (FSEv07_isError(offcodeHeaderSize)) return ERROR(dictionary_corrupted);
+        if (offcodeLog > OffFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv07_buildDTable(dctx->OffTable, offcodeNCount, offcodeMaxValue, offcodeLog);
+          if (FSEv07_isError(errorCode)) return ERROR(dictionary_corrupted); }
+        dictPtr += offcodeHeaderSize;
+    }
+
+    {   short matchlengthNCount[MaxML+1];
+        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
+        size_t const matchlengthHeaderSize = FSEv07_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSEv07_isError(matchlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (matchlengthLog > MLFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv07_buildDTable(dctx->MLTable, matchlengthNCount, matchlengthMaxValue, matchlengthLog);
+          if (FSEv07_isError(errorCode)) return ERROR(dictionary_corrupted); }
+        dictPtr += matchlengthHeaderSize;
+    }
+
+    {   short litlengthNCount[MaxLL+1];
+        unsigned litlengthMaxValue = MaxLL, litlengthLog;
+        size_t const litlengthHeaderSize = FSEv07_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
+        if (FSEv07_isError(litlengthHeaderSize)) return ERROR(dictionary_corrupted);
+        if (litlengthLog > LLFSELog) return ERROR(dictionary_corrupted);
+        { size_t const errorCode = FSEv07_buildDTable(dctx->LLTable, litlengthNCount, litlengthMaxValue, litlengthLog);
+          if (FSEv07_isError(errorCode)) return ERROR(dictionary_corrupted); }
+        dictPtr += litlengthHeaderSize;
+    }
+
+    if (dictPtr+12 > dictEnd) return ERROR(dictionary_corrupted);
+    dctx->rep[0] = MEM_readLE32(dictPtr+0); if (dctx->rep[0] == 0 || dctx->rep[0] >= dictSize) return ERROR(dictionary_corrupted);
+    dctx->rep[1] = MEM_readLE32(dictPtr+4); if (dctx->rep[1] == 0 || dctx->rep[1] >= dictSize) return ERROR(dictionary_corrupted);
+    dctx->rep[2] = MEM_readLE32(dictPtr+8); if (dctx->rep[2] == 0 || dctx->rep[2] >= dictSize) return ERROR(dictionary_corrupted);
+    dictPtr += 12;
+
+    dctx->litEntropy = dctx->fseEntropy = 1;
+    return dictPtr - (const BYTE*)dict;
+}
+
+static size_t ZSTDv07_decompress_insertDictionary(ZSTDv07_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    if (dictSize < 8) return ZSTDv07_refDictContent(dctx, dict, dictSize);
+    {   U32 const magic = MEM_readLE32(dict);
+        if (magic != ZSTDv07_DICT_MAGIC) {
+            return ZSTDv07_refDictContent(dctx, dict, dictSize);   /* pure content mode */
+    }   }
+    dctx->dictID = MEM_readLE32((const char*)dict + 4);
+
+    /* load entropy tables */
+    dict = (const char*)dict + 8;
+    dictSize -= 8;
+    {   size_t const eSize = ZSTDv07_loadEntropy(dctx, dict, dictSize);
+        if (ZSTDv07_isError(eSize)) return ERROR(dictionary_corrupted);
+        dict = (const char*)dict + eSize;
+        dictSize -= eSize;
+    }
+
+    /* reference dictionary content */
+    return ZSTDv07_refDictContent(dctx, dict, dictSize);
+}
+
+
+size_t ZSTDv07_decompressBegin_usingDict(ZSTDv07_DCtx* dctx, const void* dict, size_t dictSize)
+{
+    { size_t const errorCode = ZSTDv07_decompressBegin(dctx);
+      if (ZSTDv07_isError(errorCode)) return errorCode; }
+
+    if (dict && dictSize) {
+        size_t const errorCode = ZSTDv07_decompress_insertDictionary(dctx, dict, dictSize);
+        if (ZSTDv07_isError(errorCode)) return ERROR(dictionary_corrupted);
+    }
+
+    return 0;
+}
+
+
+struct ZSTDv07_DDict_s {
+    void* dict;
+    size_t dictSize;
+    ZSTDv07_DCtx* refContext;
+};  /* typedef'd tp ZSTDv07_CDict within zstd.h */
+
+ZSTDv07_DDict* ZSTDv07_createDDict_advanced(const void* dict, size_t dictSize, ZSTDv07_customMem customMem)
+{
+    if (!customMem.customAlloc && !customMem.customFree)
+        customMem = defaultCustomMem;
+
+    if (!customMem.customAlloc || !customMem.customFree)
+        return NULL;
+
+    {   ZSTDv07_DDict* const ddict = (ZSTDv07_DDict*) customMem.customAlloc(customMem.opaque, sizeof(*ddict));
+        void* const dictContent = customMem.customAlloc(customMem.opaque, dictSize);
+        ZSTDv07_DCtx* const dctx = ZSTDv07_createDCtx_advanced(customMem);
+
+        if (!dictContent || !ddict || !dctx) {
+            customMem.customFree(customMem.opaque, dictContent);
+            customMem.customFree(customMem.opaque, ddict);
+            customMem.customFree(customMem.opaque, dctx);
+            return NULL;
+        }
+
+        memcpy(dictContent, dict, dictSize);
+        {   size_t const errorCode = ZSTDv07_decompressBegin_usingDict(dctx, dictContent, dictSize);
+            if (ZSTDv07_isError(errorCode)) {
+                customMem.customFree(customMem.opaque, dictContent);
+                customMem.customFree(customMem.opaque, ddict);
+                customMem.customFree(customMem.opaque, dctx);
+                return NULL;
+        }   }
+
+        ddict->dict = dictContent;
+        ddict->dictSize = dictSize;
+        ddict->refContext = dctx;
+        return ddict;
+    }
+}
+
+/*! ZSTDv07_createDDict() :
+*   Create a digested dictionary, ready to start decompression without startup delay.
+*   `dict` can be released after `ZSTDv07_DDict` creation */
+ZSTDv07_DDict* ZSTDv07_createDDict(const void* dict, size_t dictSize)
+{
+    ZSTDv07_customMem const allocator = { NULL, NULL, NULL };
+    return ZSTDv07_createDDict_advanced(dict, dictSize, allocator);
+}
+
+size_t ZSTDv07_freeDDict(ZSTDv07_DDict* ddict)
+{
+    ZSTDv07_freeFunction const cFree = ddict->refContext->customMem.customFree;
+    void* const opaque = ddict->refContext->customMem.opaque;
+    ZSTDv07_freeDCtx(ddict->refContext);
+    cFree(opaque, ddict->dict);
+    cFree(opaque, ddict);
+    return 0;
+}
+
+/*! ZSTDv07_decompress_usingDDict() :
+*   Decompression using a pre-digested Dictionary
+*   Use dictionary without significant overhead. */
+ZSTDLIBv07_API size_t ZSTDv07_decompress_usingDDict(ZSTDv07_DCtx* dctx,
+                                           void* dst, size_t dstCapacity,
+                                     const void* src, size_t srcSize,
+                                     const ZSTDv07_DDict* ddict)
+{
+    return ZSTDv07_decompress_usingPreparedDCtx(dctx, ddict->refContext,
+                                           dst, dstCapacity,
+                                           src, srcSize);
+}
+/*
+    Buffered version of Zstd compression library
+    Copyright (C) 2015-2016, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+    - zstd homepage : http://www.zstd.net/
+*/
+
+
+
+/*-***************************************************************************
+*  Streaming decompression howto
+*
+*  A ZBUFFv07_DCtx object is required to track streaming operations.
+*  Use ZBUFFv07_createDCtx() and ZBUFFv07_freeDCtx() to create/release resources.
+*  Use ZBUFFv07_decompressInit() to start a new decompression operation,
+*   or ZBUFFv07_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFFv07_DCtx objects can be re-init multiple times.
+*
+*  Use ZBUFFv07_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of @dst will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change @dst.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to help latency),
+*            or 0 when a frame is completely decoded,
+*            or an error code, which can be tested using ZBUFFv07_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFFv07_recommendedDInSize() and ZBUFFv07_recommendedDOutSize()
+*  output : ZBUFFv07_recommendedDOutSize==128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFFv07_recommendedDInSize == 128KB + 3;
+*           just follow indications from ZBUFFv07_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+typedef enum { ZBUFFds_init, ZBUFFds_loadHeader,
+               ZBUFFds_read, ZBUFFds_load, ZBUFFds_flush } ZBUFFv07_dStage;
+
+/* *** Resource management *** */
+struct ZBUFFv07_DCtx_s {
+    ZSTDv07_DCtx* zd;
+    ZSTDv07_frameParams fParams;
+    ZBUFFv07_dStage stage;
+    char*  inBuff;
+    size_t inBuffSize;
+    size_t inPos;
+    char*  outBuff;
+    size_t outBuffSize;
+    size_t outStart;
+    size_t outEnd;
+    size_t blockSize;
+    BYTE headerBuffer[ZSTDv07_FRAMEHEADERSIZE_MAX];
+    size_t lhSize;
+    ZSTDv07_customMem customMem;
+};   /* typedef'd to ZBUFFv07_DCtx within "zstd_buffered.h" */
+
+ZSTDLIBv07_API ZBUFFv07_DCtx* ZBUFFv07_createDCtx_advanced(ZSTDv07_customMem customMem);
+
+ZBUFFv07_DCtx* ZBUFFv07_createDCtx(void)
+{
+    return ZBUFFv07_createDCtx_advanced(defaultCustomMem);
+}
+
+ZBUFFv07_DCtx* ZBUFFv07_createDCtx_advanced(ZSTDv07_customMem customMem)
+{
+    ZBUFFv07_DCtx* zbd;
+
+    if (!customMem.customAlloc && !customMem.customFree)
+        customMem = defaultCustomMem;
+
+    if (!customMem.customAlloc || !customMem.customFree)
+        return NULL;
+
+    zbd = (ZBUFFv07_DCtx*)customMem.customAlloc(customMem.opaque, sizeof(ZBUFFv07_DCtx));
+    if (zbd==NULL) return NULL;
+    memset(zbd, 0, sizeof(ZBUFFv07_DCtx));
+    memcpy(&zbd->customMem, &customMem, sizeof(ZSTDv07_customMem));
+    zbd->zd = ZSTDv07_createDCtx_advanced(customMem);
+    if (zbd->zd == NULL) { ZBUFFv07_freeDCtx(zbd); return NULL; }
+    zbd->stage = ZBUFFds_init;
+    return zbd;
+}
+
+size_t ZBUFFv07_freeDCtx(ZBUFFv07_DCtx* zbd)
+{
+    if (zbd==NULL) return 0;   /* support free on null */
+    ZSTDv07_freeDCtx(zbd->zd);
+    if (zbd->inBuff) zbd->customMem.customFree(zbd->customMem.opaque, zbd->inBuff);
+    if (zbd->outBuff) zbd->customMem.customFree(zbd->customMem.opaque, zbd->outBuff);
+    zbd->customMem.customFree(zbd->customMem.opaque, zbd);
+    return 0;
+}
+
+
+/* *** Initialization *** */
+
+size_t ZBUFFv07_decompressInitDictionary(ZBUFFv07_DCtx* zbd, const void* dict, size_t dictSize)
+{
+    zbd->stage = ZBUFFds_loadHeader;
+    zbd->lhSize = zbd->inPos = zbd->outStart = zbd->outEnd = 0;
+    return ZSTDv07_decompressBegin_usingDict(zbd->zd, dict, dictSize);
+}
+
+size_t ZBUFFv07_decompressInit(ZBUFFv07_DCtx* zbd)
+{
+    return ZBUFFv07_decompressInitDictionary(zbd, NULL, 0);
+}
+
+
+/* internal util function */
+MEM_STATIC size_t ZBUFFv07_limitCopy(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    size_t const length = MIN(dstCapacity, srcSize);
+    memcpy(dst, src, length);
+    return length;
+}
+
+
+/* *** Decompression *** */
+
+size_t ZBUFFv07_decompressContinue(ZBUFFv07_DCtx* zbd,
+                                void* dst, size_t* dstCapacityPtr,
+                          const void* src, size_t* srcSizePtr)
+{
+    const char* const istart = (const char*)src;
+    const char* const iend = istart + *srcSizePtr;
+    const char* ip = istart;
+    char* const ostart = (char*)dst;
+    char* const oend = ostart + *dstCapacityPtr;
+    char* op = ostart;
+    U32 notDone = 1;
+
+    while (notDone) {
+        switch(zbd->stage)
+        {
+        case ZBUFFds_init :
+            return ERROR(init_missing);
+
+        case ZBUFFds_loadHeader :
+            {   size_t const hSize = ZSTDv07_getFrameParams(&(zbd->fParams), zbd->headerBuffer, zbd->lhSize);
+                if (ZSTDv07_isError(hSize)) return hSize;
+                if (hSize != 0) {
+                    size_t const toLoad = hSize - zbd->lhSize;   /* if hSize!=0, hSize > zbd->lhSize */
+                    if (toLoad > (size_t)(iend-ip)) {   /* not enough input to load full header */
+                        memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
+                        zbd->lhSize += iend-ip;
+                        *dstCapacityPtr = 0;
+                        return (hSize - zbd->lhSize) + ZSTDv07_blockHeaderSize;   /* remaining header bytes + next block header */
+                    }
+                    memcpy(zbd->headerBuffer + zbd->lhSize, ip, toLoad); zbd->lhSize = hSize; ip += toLoad;
+                    break;
+            }   }
+
+            /* Consume header */
+            {   size_t const h1Size = ZSTDv07_nextSrcSizeToDecompress(zbd->zd);  /* == ZSTDv07_frameHeaderSize_min */
+                size_t const h1Result = ZSTDv07_decompressContinue(zbd->zd, NULL, 0, zbd->headerBuffer, h1Size);
+                if (ZSTDv07_isError(h1Result)) return h1Result;
+                if (h1Size < zbd->lhSize) {   /* long header */
+                    size_t const h2Size = ZSTDv07_nextSrcSizeToDecompress(zbd->zd);
+                    size_t const h2Result = ZSTDv07_decompressContinue(zbd->zd, NULL, 0, zbd->headerBuffer+h1Size, h2Size);
+                    if (ZSTDv07_isError(h2Result)) return h2Result;
+            }   }
+
+            zbd->fParams.windowSize = MAX(zbd->fParams.windowSize, 1U << ZSTDv07_WINDOWLOG_ABSOLUTEMIN);
+
+            /* Frame header instruct buffer sizes */
+            {   size_t const blockSize = MIN(zbd->fParams.windowSize, ZSTDv07_BLOCKSIZE_ABSOLUTEMAX);
+                zbd->blockSize = blockSize;
+                if (zbd->inBuffSize < blockSize) {
+                    zbd->customMem.customFree(zbd->customMem.opaque, zbd->inBuff);
+                    zbd->inBuffSize = blockSize;
+                    zbd->inBuff = (char*)zbd->customMem.customAlloc(zbd->customMem.opaque, blockSize);
+                    if (zbd->inBuff == NULL) return ERROR(memory_allocation);
+                }
+                {   size_t const neededOutSize = zbd->fParams.windowSize + blockSize + WILDCOPY_OVERLENGTH * 2;
+                    if (zbd->outBuffSize < neededOutSize) {
+                        zbd->customMem.customFree(zbd->customMem.opaque, zbd->outBuff);
+                        zbd->outBuffSize = neededOutSize;
+                        zbd->outBuff = (char*)zbd->customMem.customAlloc(zbd->customMem.opaque, neededOutSize);
+                        if (zbd->outBuff == NULL) return ERROR(memory_allocation);
+            }   }   }
+            zbd->stage = ZBUFFds_read;
+            /* pass-through */
+	    /* fall-through */
+        case ZBUFFds_read:
+            {   size_t const neededInSize = ZSTDv07_nextSrcSizeToDecompress(zbd->zd);
+                if (neededInSize==0) {  /* end of frame */
+                    zbd->stage = ZBUFFds_init;
+                    notDone = 0;
+                    break;
+                }
+                if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
+                    const int isSkipFrame = ZSTDv07_isSkipFrame(zbd->zd);
+                    size_t const decodedSize = ZSTDv07_decompressContinue(zbd->zd,
+                        zbd->outBuff + zbd->outStart, (isSkipFrame ? 0 : zbd->outBuffSize - zbd->outStart),
+                        ip, neededInSize);
+                    if (ZSTDv07_isError(decodedSize)) return decodedSize;
+                    ip += neededInSize;
+                    if (!decodedSize && !isSkipFrame) break;   /* this was just a header */
+                    zbd->outEnd = zbd->outStart +  decodedSize;
+                    zbd->stage = ZBUFFds_flush;
+                    break;
+                }
+                if (ip==iend) { notDone = 0; break; }   /* no more input */
+                zbd->stage = ZBUFFds_load;
+            }
+	    /* fall-through */
+        case ZBUFFds_load:
+            {   size_t const neededInSize = ZSTDv07_nextSrcSizeToDecompress(zbd->zd);
+                size_t const toLoad = neededInSize - zbd->inPos;   /* should always be <= remaining space within inBuff */
+                size_t loadedSize;
+                if (toLoad > zbd->inBuffSize - zbd->inPos) return ERROR(corruption_detected);   /* should never happen */
+                loadedSize = ZBUFFv07_limitCopy(zbd->inBuff + zbd->inPos, toLoad, ip, iend-ip);
+                ip += loadedSize;
+                zbd->inPos += loadedSize;
+                if (loadedSize < toLoad) { notDone = 0; break; }   /* not enough input, wait for more */
+
+                /* decode loaded input */
+                {  const int isSkipFrame = ZSTDv07_isSkipFrame(zbd->zd);
+                   size_t const decodedSize = ZSTDv07_decompressContinue(zbd->zd,
+                        zbd->outBuff + zbd->outStart, zbd->outBuffSize - zbd->outStart,
+                        zbd->inBuff, neededInSize);
+                    if (ZSTDv07_isError(decodedSize)) return decodedSize;
+                    zbd->inPos = 0;   /* input is consumed */
+                    if (!decodedSize && !isSkipFrame) { zbd->stage = ZBUFFds_read; break; }   /* this was just a header */
+                    zbd->outEnd = zbd->outStart +  decodedSize;
+                    zbd->stage = ZBUFFds_flush;
+                    /* break; */
+                    /* pass-through */
+                }
+	    }
+	    /* fall-through */
+        case ZBUFFds_flush:
+            {   size_t const toFlushSize = zbd->outEnd - zbd->outStart;
+                size_t const flushedSize = ZBUFFv07_limitCopy(op, oend-op, zbd->outBuff + zbd->outStart, toFlushSize);
+                op += flushedSize;
+                zbd->outStart += flushedSize;
+                if (flushedSize == toFlushSize) {
+                    zbd->stage = ZBUFFds_read;
+                    if (zbd->outStart + zbd->blockSize > zbd->outBuffSize)
+                        zbd->outStart = zbd->outEnd = 0;
+                    break;
+                }
+                /* cannot flush everything */
+                notDone = 0;
+                break;
+            }
+        default: return ERROR(GENERIC);   /* impossible */
+    }   }
+
+    /* result */
+    *srcSizePtr = ip-istart;
+    *dstCapacityPtr = op-ostart;
+    {   size_t nextSrcSizeHint = ZSTDv07_nextSrcSizeToDecompress(zbd->zd);
+        nextSrcSizeHint -= zbd->inPos;   /* already loaded*/
+        return nextSrcSizeHint;
+    }
+}
+
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+size_t ZBUFFv07_recommendedDInSize(void)  { return ZSTDv07_BLOCKSIZE_ABSOLUTEMAX + ZSTDv07_blockHeaderSize /* block header size*/ ; }
+size_t ZBUFFv07_recommendedDOutSize(void) { return ZSTDv07_BLOCKSIZE_ABSOLUTEMAX; }
diff --git a/vendor/github.com/DataDog/zstd/zstd_v07.h b/vendor/github.com/DataDog/zstd/zstd_v07.h
new file mode 100644
index 0000000..6591cd3
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstd_v07.h
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTDv07_H_235446
+#define ZSTDv07_H_235446
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*======  Dependency  ======*/
+#include <stddef.h>   /* size_t */
+
+
+/*======  Export for Windows  ======*/
+/*!
+*  ZSTDv07_DLL_EXPORT :
+*  Enable exporting of functions when building a Windows DLL
+*/
+#if defined(_WIN32) && defined(ZSTDv07_DLL_EXPORT) && (ZSTDv07_DLL_EXPORT==1)
+#  define ZSTDLIBv07_API __declspec(dllexport)
+#else
+#  define ZSTDLIBv07_API
+#endif
+
+
+/* *************************************
+*  Simple API
+***************************************/
+/*! ZSTDv07_getDecompressedSize() :
+*   @return : decompressed size if known, 0 otherwise.
+       note 1 : if `0`, follow up with ZSTDv07_getFrameParams() to know precise failure cause.
+       note 2 : decompressed size could be wrong or intentionally modified !
+                always ensure results fit within application's authorized limits */
+unsigned long long ZSTDv07_getDecompressedSize(const void* src, size_t srcSize);
+
+/*! ZSTDv07_decompress() :
+    `compressedSize` : must be _exact_ size of compressed input, otherwise decompression will fail.
+    `dstCapacity` must be equal or larger than originalSize.
+    @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
+              or an errorCode if it fails (which can be tested using ZSTDv07_isError()) */
+ZSTDLIBv07_API size_t ZSTDv07_decompress( void* dst, size_t dstCapacity,
+                                    const void* src, size_t compressedSize);
+
+/**
+ZSTDv07_getFrameSrcSize() : get the source length of a ZSTD frame
+    compressedSize : The size of the 'src' buffer, at least as large as the frame pointed to by 'src'
+    return : the number of bytes that would be read to decompress this frame
+             or an errorCode if it fails (which can be tested using ZSTDv07_isError())
+*/
+size_t ZSTDv07_findFrameCompressedSize(const void* src, size_t compressedSize);
+
+/*======  Helper functions  ======*/
+ZSTDLIBv07_API unsigned    ZSTDv07_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
+ZSTDLIBv07_API const char* ZSTDv07_getErrorName(size_t code);     /*!< provides readable string from an error code */
+
+
+/*-*************************************
+*  Explicit memory management
+***************************************/
+/** Decompression context */
+typedef struct ZSTDv07_DCtx_s ZSTDv07_DCtx;
+ZSTDLIBv07_API ZSTDv07_DCtx* ZSTDv07_createDCtx(void);
+ZSTDLIBv07_API size_t     ZSTDv07_freeDCtx(ZSTDv07_DCtx* dctx);      /*!< @return : errorCode */
+
+/** ZSTDv07_decompressDCtx() :
+*   Same as ZSTDv07_decompress(), requires an allocated ZSTDv07_DCtx (see ZSTDv07_createDCtx()) */
+ZSTDLIBv07_API size_t ZSTDv07_decompressDCtx(ZSTDv07_DCtx* ctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
+/*-************************
+*  Simple dictionary API
+***************************/
+/*! ZSTDv07_decompress_usingDict() :
+*   Decompression using a pre-defined Dictionary content (see dictBuilder).
+*   Dictionary must be identical to the one used during compression.
+*   Note : This function load the dictionary, resulting in a significant startup time */
+ZSTDLIBv07_API size_t ZSTDv07_decompress_usingDict(ZSTDv07_DCtx* dctx,
+                                                   void* dst, size_t dstCapacity,
+                                             const void* src, size_t srcSize,
+                                             const void* dict,size_t dictSize);
+
+
+/*-**************************
+*  Advanced Dictionary API
+****************************/
+/*! ZSTDv07_createDDict() :
+*   Create a digested dictionary, ready to start decompression operation without startup delay.
+*   `dict` can be released after creation */
+typedef struct ZSTDv07_DDict_s ZSTDv07_DDict;
+ZSTDLIBv07_API ZSTDv07_DDict* ZSTDv07_createDDict(const void* dict, size_t dictSize);
+ZSTDLIBv07_API size_t      ZSTDv07_freeDDict(ZSTDv07_DDict* ddict);
+
+/*! ZSTDv07_decompress_usingDDict() :
+*   Decompression using a pre-digested Dictionary
+*   Faster startup than ZSTDv07_decompress_usingDict(), recommended when same dictionary is used multiple times. */
+ZSTDLIBv07_API size_t ZSTDv07_decompress_usingDDict(ZSTDv07_DCtx* dctx,
+                                                    void* dst, size_t dstCapacity,
+                                              const void* src, size_t srcSize,
+                                              const ZSTDv07_DDict* ddict);
+
+typedef struct {
+    unsigned long long frameContentSize;
+    unsigned windowSize;
+    unsigned dictID;
+    unsigned checksumFlag;
+} ZSTDv07_frameParams;
+
+ZSTDLIBv07_API size_t ZSTDv07_getFrameParams(ZSTDv07_frameParams* fparamsPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
+
+
+
+
+/* *************************************
+*  Streaming functions
+***************************************/
+typedef struct ZBUFFv07_DCtx_s ZBUFFv07_DCtx;
+ZSTDLIBv07_API ZBUFFv07_DCtx* ZBUFFv07_createDCtx(void);
+ZSTDLIBv07_API size_t      ZBUFFv07_freeDCtx(ZBUFFv07_DCtx* dctx);
+
+ZSTDLIBv07_API size_t ZBUFFv07_decompressInit(ZBUFFv07_DCtx* dctx);
+ZSTDLIBv07_API size_t ZBUFFv07_decompressInitDictionary(ZBUFFv07_DCtx* dctx, const void* dict, size_t dictSize);
+
+ZSTDLIBv07_API size_t ZBUFFv07_decompressContinue(ZBUFFv07_DCtx* dctx,
+                                            void* dst, size_t* dstCapacityPtr,
+                                      const void* src, size_t* srcSizePtr);
+
+/*-***************************************************************************
+*  Streaming decompression howto
+*
+*  A ZBUFFv07_DCtx object is required to track streaming operations.
+*  Use ZBUFFv07_createDCtx() and ZBUFFv07_freeDCtx() to create/release resources.
+*  Use ZBUFFv07_decompressInit() to start a new decompression operation,
+*   or ZBUFFv07_decompressInitDictionary() if decompression requires a dictionary.
+*  Note that ZBUFFv07_DCtx objects can be re-init multiple times.
+*
+*  Use ZBUFFv07_decompressContinue() repetitively to consume your input.
+*  *srcSizePtr and *dstCapacityPtr can be any size.
+*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
+*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
+*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change `dst`.
+*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to help latency),
+*            or 0 when a frame is completely decoded,
+*            or an error code, which can be tested using ZBUFFv07_isError().
+*
+*  Hint : recommended buffer sizes (not compulsory) : ZBUFFv07_recommendedDInSize() and ZBUFFv07_recommendedDOutSize()
+*  output : ZBUFFv07_recommendedDOutSize== 128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
+*  input  : ZBUFFv07_recommendedDInSize == 128KB + 3;
+*           just follow indications from ZBUFFv07_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
+* *******************************************************************************/
+
+
+/* *************************************
+*  Tool functions
+***************************************/
+ZSTDLIBv07_API unsigned ZBUFFv07_isError(size_t errorCode);
+ZSTDLIBv07_API const char* ZBUFFv07_getErrorName(size_t errorCode);
+
+/** Functions below provide recommended buffer sizes for Compression or Decompression operations.
+*   These sizes are just hints, they tend to offer better latency */
+ZSTDLIBv07_API size_t ZBUFFv07_recommendedDInSize(void);
+ZSTDLIBv07_API size_t ZBUFFv07_recommendedDOutSize(void);
+
+
+/*-*************************************
+*  Constants
+***************************************/
+#define ZSTDv07_MAGICNUMBER            0xFD2FB527   /* v0.7 */
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* ZSTDv07_H_235446 */
diff --git a/vendor/github.com/DataDog/zstd/zstdmt_compress.c b/vendor/github.com/DataDog/zstd/zstdmt_compress.c
new file mode 100755
index 0000000..c7a205d
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstdmt_compress.c
@@ -0,0 +1,1831 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/* ======   Tuning parameters   ====== */
+#define ZSTDMT_NBWORKERS_MAX 200
+#define ZSTDMT_JOBSIZE_MAX  (MEM_32bits() ? (512 MB) : (2 GB))  /* note : limited by `jobSize` type, which is `unsigned` */
+#define ZSTDMT_OVERLAPLOG_DEFAULT 6
+
+
+/* ======   Compiler specifics   ====== */
+#if defined(_MSC_VER)
+#  pragma warning(disable : 4204)   /* disable: C4204: non-constant aggregate initializer */
+#endif
+
+
+/* ======   Dependencies   ====== */
+#include <string.h>      /* memcpy, memset */
+#include <limits.h>      /* INT_MAX */
+#include "pool.h"        /* threadpool */
+#include "threading.h"   /* mutex */
+#include "zstd_compress_internal.h"  /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
+#include "zstd_ldm.h"
+#include "zstdmt_compress.h"
+
+/* Guards code to support resizing the SeqPool.
+ * We will want to resize the SeqPool to save memory in the future.
+ * Until then, comment the code out since it is unused.
+ */
+#define ZSTD_RESIZE_SEQPOOL 0
+
+/* ======   Debug   ====== */
+#if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
+
+#  include <stdio.h>
+#  include <unistd.h>
+#  include <sys/times.h>
+#  define DEBUGLOGRAW(l, ...) if (l<=ZSTD_DEBUG) { fprintf(stderr, __VA_ARGS__); }
+
+#  define DEBUG_PRINTHEX(l,p,n) {            \
+    unsigned debug_u;                        \
+    for (debug_u=0; debug_u<(n); debug_u++)  \
+        DEBUGLOGRAW(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
+    DEBUGLOGRAW(l, " \n");                   \
+}
+
+static unsigned long long GetCurrentClockTimeMicroseconds(void)
+{
+   static clock_t _ticksPerSecond = 0;
+   if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);
+
+   { struct tms junk; clock_t newTicks = (clock_t) times(&junk);
+     return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond); }
+}
+
+#define MUTEX_WAIT_TIME_DLEVEL 6
+#define ZSTD_PTHREAD_MUTEX_LOCK(mutex) {          \
+    if (ZSTD_DEBUG >= MUTEX_WAIT_TIME_DLEVEL) {   \
+        unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
+        ZSTD_pthread_mutex_lock(mutex);           \
+        {   unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
+            unsigned long long const elapsedTime = (afterTime-beforeTime); \
+            if (elapsedTime > 1000) {  /* or whatever threshold you like; I'm using 1 millisecond here */ \
+                DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
+                   elapsedTime, #mutex);          \
+        }   }                                     \
+    } else {                                      \
+        ZSTD_pthread_mutex_lock(mutex);           \
+    }                                             \
+}
+
+#else
+
+#  define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
+#  define DEBUG_PRINTHEX(l,p,n) {}
+
+#endif
+
+
+/* =====   Buffer Pool   ===== */
+/* a single Buffer Pool can be invoked from multiple threads in parallel */
+
+typedef struct buffer_s {
+    void* start;
+    size_t capacity;
+} buffer_t;
+
+static const buffer_t g_nullBuffer = { NULL, 0 };
+
+typedef struct ZSTDMT_bufferPool_s {
+    ZSTD_pthread_mutex_t poolMutex;
+    size_t bufferSize;
+    unsigned totalBuffers;
+    unsigned nbBuffers;
+    ZSTD_customMem cMem;
+    buffer_t bTable[1];   /* variable size */
+} ZSTDMT_bufferPool;
+
+static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem)
+{
+    unsigned const maxNbBuffers = 2*nbWorkers + 3;
+    ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc(
+        sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
+    if (bufPool==NULL) return NULL;
+    if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) {
+        ZSTD_free(bufPool, cMem);
+        return NULL;
+    }
+    bufPool->bufferSize = 64 KB;
+    bufPool->totalBuffers = maxNbBuffers;
+    bufPool->nbBuffers = 0;
+    bufPool->cMem = cMem;
+    return bufPool;
+}
+
+static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
+{
+    unsigned u;
+    DEBUGLOG(3, "ZSTDMT_freeBufferPool (address:%08X)", (U32)(size_t)bufPool);
+    if (!bufPool) return;   /* compatibility with free on NULL */
+    for (u=0; u<bufPool->totalBuffers; u++) {
+        DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->bTable[u].start);
+        ZSTD_free(bufPool->bTable[u].start, bufPool->cMem);
+    }
+    ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
+    ZSTD_free(bufPool, bufPool->cMem);
+}
+
+/* only works at initialization, not during compression */
+static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
+{
+    size_t const poolSize = sizeof(*bufPool)
+                          + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
+    unsigned u;
+    size_t totalBufferSize = 0;
+    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
+    for (u=0; u<bufPool->totalBuffers; u++)
+        totalBufferSize += bufPool->bTable[u].capacity;
+    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+
+    return poolSize + totalBufferSize;
+}
+
+/* ZSTDMT_setBufferSize() :
+ * all future buffers provided by this buffer pool will have _at least_ this size
+ * note : it's better for all buffers to have same size,
+ * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
+static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
+{
+    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
+    DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
+    bufPool->bufferSize = bSize;
+    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+}
+
+/** ZSTDMT_getBuffer() :
+ *  assumption : bufPool must be valid
+ * @return : a buffer, with start pointer and size
+ *  note: allocation may fail, in this case, start==NULL and size==0 */
+static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
+{
+    size_t const bSize = bufPool->bufferSize;
+    DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
+    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
+    if (bufPool->nbBuffers) {   /* try to use an existing buffer */
+        buffer_t const buf = bufPool->bTable[--(bufPool->nbBuffers)];
+        size_t const availBufferSize = buf.capacity;
+        bufPool->bTable[bufPool->nbBuffers] = g_nullBuffer;
+        if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
+            /* large enough, but not too much */
+            DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
+                        bufPool->nbBuffers, (U32)buf.capacity);
+            ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+            return buf;
+        }
+        /* size conditions not respected : scratch this buffer, create new one */
+        DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
+        ZSTD_free(buf.start, bufPool->cMem);
+    }
+    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+    /* create new buffer */
+    DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
+    {   buffer_t buffer;
+        void* const start = ZSTD_malloc(bSize, bufPool->cMem);
+        buffer.start = start;   /* note : start can be NULL if malloc fails ! */
+        buffer.capacity = (start==NULL) ? 0 : bSize;
+        if (start==NULL) {
+            DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
+        } else {
+            DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
+        }
+        return buffer;
+    }
+}
+
+#if ZSTD_RESIZE_SEQPOOL
+/** ZSTDMT_resizeBuffer() :
+ * assumption : bufPool must be valid
+ * @return : a buffer that is at least the buffer pool buffer size.
+ *           If a reallocation happens, the data in the input buffer is copied.
+ */
+static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer)
+{
+    size_t const bSize = bufPool->bufferSize;
+    if (buffer.capacity < bSize) {
+        void* const start = ZSTD_malloc(bSize, bufPool->cMem);
+        buffer_t newBuffer;
+        newBuffer.start = start;
+        newBuffer.capacity = start == NULL ? 0 : bSize;
+        if (start != NULL) {
+            assert(newBuffer.capacity >= buffer.capacity);
+            memcpy(newBuffer.start, buffer.start, buffer.capacity);
+            DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
+            return newBuffer;
+        }
+        DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
+    }
+    return buffer;
+}
+#endif
+
+/* store buffer for later re-use, up to pool capacity */
+static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
+{
+    if (buf.start == NULL) return;   /* compatible with release on NULL */
+    DEBUGLOG(5, "ZSTDMT_releaseBuffer");
+    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
+    if (bufPool->nbBuffers < bufPool->totalBuffers) {
+        bufPool->bTable[bufPool->nbBuffers++] = buf;  /* stored for later use */
+        DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
+                    (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
+        ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+        return;
+    }
+    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
+    /* Reached bufferPool capacity (should not happen) */
+    DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
+    ZSTD_free(buf.start, bufPool->cMem);
+}
+
+
+/* =====   Seq Pool Wrapper   ====== */
+
+static rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0};
+
+typedef ZSTDMT_bufferPool ZSTDMT_seqPool;
+
+static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
+{
+    return ZSTDMT_sizeof_bufferPool(seqPool);
+}
+
+static rawSeqStore_t bufferToSeq(buffer_t buffer)
+{
+    rawSeqStore_t seq = {NULL, 0, 0, 0};
+    seq.seq = (rawSeq*)buffer.start;
+    seq.capacity = buffer.capacity / sizeof(rawSeq);
+    return seq;
+}
+
+static buffer_t seqToBuffer(rawSeqStore_t seq)
+{
+    buffer_t buffer;
+    buffer.start = seq.seq;
+    buffer.capacity = seq.capacity * sizeof(rawSeq);
+    return buffer;
+}
+
+static rawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
+{
+    if (seqPool->bufferSize == 0) {
+        return kNullRawSeqStore;
+    }
+    return bufferToSeq(ZSTDMT_getBuffer(seqPool));
+}
+
+#if ZSTD_RESIZE_SEQPOOL
+static rawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
+{
+  return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
+}
+#endif
+
+static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
+{
+  ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
+}
+
+static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
+{
+  ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
+}
+
+static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
+{
+    ZSTDMT_seqPool* seqPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
+    ZSTDMT_setNbSeq(seqPool, 0);
+    return seqPool;
+}
+
+static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
+{
+    ZSTDMT_freeBufferPool(seqPool);
+}
+
+
+
+/* =====   CCtx Pool   ===== */
+/* a single CCtx Pool can be invoked from multiple threads in parallel */
+
+typedef struct {
+    ZSTD_pthread_mutex_t poolMutex;
+    unsigned totalCCtx;
+    unsigned availCCtx;
+    ZSTD_customMem cMem;
+    ZSTD_CCtx* cctx[1];   /* variable size */
+} ZSTDMT_CCtxPool;
+
+/* note : all CCtx borrowed from the pool should be released back to the pool _before_ freeing the pool */
+static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
+{
+    unsigned u;
+    for (u=0; u<pool->totalCCtx; u++)
+        ZSTD_freeCCtx(pool->cctx[u]);  /* note : compatible with free on NULL */
+    ZSTD_pthread_mutex_destroy(&pool->poolMutex);
+    ZSTD_free(pool, pool->cMem);
+}
+
+/* ZSTDMT_createCCtxPool() :
+ * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
+static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbWorkers,
+                                              ZSTD_customMem cMem)
+{
+    ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc(
+        sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
+    assert(nbWorkers > 0);
+    if (!cctxPool) return NULL;
+    if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
+        ZSTD_free(cctxPool, cMem);
+        return NULL;
+    }
+    cctxPool->cMem = cMem;
+    cctxPool->totalCCtx = nbWorkers;
+    cctxPool->availCCtx = 1;   /* at least one cctx for single-thread mode */
+    cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
+    if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
+    DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
+    return cctxPool;
+}
+
+/* only works during initialization phase, not during compression */
+static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
+{
+    ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
+    {   unsigned const nbWorkers = cctxPool->totalCCtx;
+        size_t const poolSize = sizeof(*cctxPool)
+                                + (nbWorkers-1) * sizeof(ZSTD_CCtx*);
+        unsigned u;
+        size_t totalCCtxSize = 0;
+        for (u=0; u<nbWorkers; u++) {
+            totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
+        }
+        ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
+        assert(nbWorkers > 0);
+        return poolSize + totalCCtxSize;
+    }
+}
+
+static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
+{
+    DEBUGLOG(5, "ZSTDMT_getCCtx");
+    ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
+    if (cctxPool->availCCtx) {
+        cctxPool->availCCtx--;
+        {   ZSTD_CCtx* const cctx = cctxPool->cctx[cctxPool->availCCtx];
+            ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
+            return cctx;
+    }   }
+    ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
+    DEBUGLOG(5, "create one more CCtx");
+    return ZSTD_createCCtx_advanced(cctxPool->cMem);   /* note : can be NULL, when creation fails ! */
+}
+
+static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
+{
+    if (cctx==NULL) return;   /* compatibility with release on NULL */
+    ZSTD_pthread_mutex_lock(&pool->poolMutex);
+    if (pool->availCCtx < pool->totalCCtx)
+        pool->cctx[pool->availCCtx++] = cctx;
+    else {
+        /* pool overflow : should not happen, since totalCCtx==nbWorkers */
+        DEBUGLOG(4, "CCtx pool overflow : free cctx");
+        ZSTD_freeCCtx(cctx);
+    }
+    ZSTD_pthread_mutex_unlock(&pool->poolMutex);
+}
+
+/* ====   Serial State   ==== */
+
+typedef struct {
+    void const* start;
+    size_t size;
+} range_t;
+
+typedef struct {
+    /* All variables in the struct are protected by mutex. */
+    ZSTD_pthread_mutex_t mutex;
+    ZSTD_pthread_cond_t cond;
+    ZSTD_CCtx_params params;
+    ldmState_t ldmState;
+    XXH64_state_t xxhState;
+    unsigned nextJobID;
+    /* Protects ldmWindow.
+     * Must be acquired after the main mutex when acquiring both.
+     */
+    ZSTD_pthread_mutex_t ldmWindowMutex;
+    ZSTD_pthread_cond_t ldmWindowCond;  /* Signaled when ldmWindow is udpated */
+    ZSTD_window_t ldmWindow;  /* A thread-safe copy of ldmState.window */
+} serialState_t;
+
+static int ZSTDMT_serialState_reset(serialState_t* serialState, ZSTDMT_seqPool* seqPool, ZSTD_CCtx_params params)
+{
+    /* Adjust parameters */
+    if (params.ldmParams.enableLdm) {
+        DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
+        params.ldmParams.windowLog = params.cParams.windowLog;
+        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
+        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
+        assert(params.ldmParams.hashEveryLog < 32);
+        serialState->ldmState.hashPower =
+                ZSTD_ldm_getHashPower(params.ldmParams.minMatchLength);
+    } else {
+        memset(&params.ldmParams, 0, sizeof(params.ldmParams));
+    }
+    serialState->nextJobID = 0;
+    if (params.fParams.checksumFlag)
+        XXH64_reset(&serialState->xxhState, 0);
+    if (params.ldmParams.enableLdm) {
+        ZSTD_customMem cMem = params.customMem;
+        unsigned const hashLog = params.ldmParams.hashLog;
+        size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
+        unsigned const bucketLog =
+            params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
+        size_t const bucketSize = (size_t)1 << bucketLog;
+        unsigned const prevBucketLog =
+            serialState->params.ldmParams.hashLog -
+            serialState->params.ldmParams.bucketSizeLog;
+        /* Size the seq pool tables */
+        ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, params.jobSize));
+        /* Reset the window */
+        ZSTD_window_clear(&serialState->ldmState.window);
+        serialState->ldmWindow = serialState->ldmState.window;
+        /* Resize tables and output space if necessary. */
+        if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
+            ZSTD_free(serialState->ldmState.hashTable, cMem);
+            serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_malloc(hashSize, cMem);
+        }
+        if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
+            ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
+            serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_malloc(bucketSize, cMem);
+        }
+        if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
+            return 1;
+        /* Zero the tables */
+        memset(serialState->ldmState.hashTable, 0, hashSize);
+        memset(serialState->ldmState.bucketOffsets, 0, bucketSize);
+    }
+    serialState->params = params;
+    return 0;
+}
+
+static int ZSTDMT_serialState_init(serialState_t* serialState)
+{
+    int initError = 0;
+    memset(serialState, 0, sizeof(*serialState));
+    initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
+    initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
+    initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
+    initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
+    return initError;
+}
+
+static void ZSTDMT_serialState_free(serialState_t* serialState)
+{
+    ZSTD_customMem cMem = serialState->params.customMem;
+    ZSTD_pthread_mutex_destroy(&serialState->mutex);
+    ZSTD_pthread_cond_destroy(&serialState->cond);
+    ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
+    ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
+    ZSTD_free(serialState->ldmState.hashTable, cMem);
+    ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
+}
+
+static void ZSTDMT_serialState_update(serialState_t* serialState,
+                                      ZSTD_CCtx* jobCCtx, rawSeqStore_t seqStore,
+                                      range_t src, unsigned jobID)
+{
+    /* Wait for our turn */
+    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
+    while (serialState->nextJobID < jobID) {
+        ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
+    }
+    /* A future job may error and skip our job */
+    if (serialState->nextJobID == jobID) {
+        /* It is now our turn, do any processing necessary */
+        if (serialState->params.ldmParams.enableLdm) {
+            size_t error;
+            assert(seqStore.seq != NULL && seqStore.pos == 0 &&
+                   seqStore.size == 0 && seqStore.capacity > 0);
+            ZSTD_window_update(&serialState->ldmState.window, src.start, src.size);
+            error = ZSTD_ldm_generateSequences(
+                &serialState->ldmState, &seqStore,
+                &serialState->params.ldmParams, src.start, src.size);
+            /* We provide a large enough buffer to never fail. */
+            assert(!ZSTD_isError(error)); (void)error;
+            /* Update ldmWindow to match the ldmState.window and signal the main
+             * thread if it is waiting for a buffer.
+             */
+            ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
+            serialState->ldmWindow = serialState->ldmState.window;
+            ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
+            ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
+        }
+        if (serialState->params.fParams.checksumFlag && src.size > 0)
+            XXH64_update(&serialState->xxhState, src.start, src.size);
+    }
+    /* Now it is the next jobs turn */
+    serialState->nextJobID++;
+    ZSTD_pthread_cond_broadcast(&serialState->cond);
+    ZSTD_pthread_mutex_unlock(&serialState->mutex);
+
+    if (seqStore.size > 0) {
+        size_t const err = ZSTD_referenceExternalSequences(
+            jobCCtx, seqStore.seq, seqStore.size);
+        assert(serialState->params.ldmParams.enableLdm);
+        assert(!ZSTD_isError(err));
+        (void)err;
+    }
+}
+
+static void ZSTDMT_serialState_ensureFinished(serialState_t* serialState,
+                                              unsigned jobID, size_t cSize)
+{
+    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
+    if (serialState->nextJobID <= jobID) {
+        assert(ZSTD_isError(cSize)); (void)cSize;
+        DEBUGLOG(5, "Skipping past job %u because of error", jobID);
+        serialState->nextJobID = jobID + 1;
+        ZSTD_pthread_cond_broadcast(&serialState->cond);
+
+        ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
+        ZSTD_window_clear(&serialState->ldmWindow);
+        ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
+        ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
+    }
+    ZSTD_pthread_mutex_unlock(&serialState->mutex);
+
+}
+
+
+/* ------------------------------------------ */
+/* =====          Worker thread         ===== */
+/* ------------------------------------------ */
+
+static const range_t kNullRange = { NULL, 0 };
+
+typedef struct {
+    size_t   consumed;                   /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
+    size_t   cSize;                      /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
+    ZSTD_pthread_mutex_t job_mutex;      /* Thread-safe - used by mtctx and worker */
+    ZSTD_pthread_cond_t job_cond;        /* Thread-safe - used by mtctx and worker */
+    ZSTDMT_CCtxPool* cctxPool;           /* Thread-safe - used by mtctx and (all) workers */
+    ZSTDMT_bufferPool* bufPool;          /* Thread-safe - used by mtctx and (all) workers */
+    ZSTDMT_seqPool* seqPool;             /* Thread-safe - used by mtctx and (all) workers */
+    serialState_t* serial;               /* Thread-safe - used by mtctx and (all) workers */
+    buffer_t dstBuff;                    /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
+    range_t prefix;                      /* set by mtctx, then read by worker & mtctx => no barrier */
+    range_t src;                         /* set by mtctx, then read by worker & mtctx => no barrier */
+    unsigned jobID;                      /* set by mtctx, then read by worker => no barrier */
+    unsigned firstJob;                   /* set by mtctx, then read by worker => no barrier */
+    unsigned lastJob;                    /* set by mtctx, then read by worker => no barrier */
+    ZSTD_CCtx_params params;             /* set by mtctx, then read by worker => no barrier */
+    const ZSTD_CDict* cdict;             /* set by mtctx, then read by worker => no barrier */
+    unsigned long long fullFrameSize;    /* set by mtctx, then read by worker => no barrier */
+    size_t   dstFlushed;                 /* used only by mtctx */
+    unsigned frameChecksumNeeded;        /* used only by mtctx */
+} ZSTDMT_jobDescription;
+
+/* ZSTDMT_compressionJob() is a POOL_function type */
+void ZSTDMT_compressionJob(void* jobDescription)
+{
+    ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
+    ZSTD_CCtx_params jobParams = job->params;   /* do not modify job->params ! copy it, modify the copy */
+    ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
+    rawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
+    buffer_t dstBuff = job->dstBuff;
+
+    /* Don't compute the checksum for chunks, since we compute it externally,
+     * but write it in the header.
+     */
+    if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
+    /* Don't run LDM for the chunks, since we handle it externally */
+    jobParams.ldmParams.enableLdm = 0;
+
+    /* ressources */
+    if (cctx==NULL) {
+        job->cSize = ERROR(memory_allocation);
+        goto _endJob;
+    }
+    if (dstBuff.start == NULL) {   /* streaming job : doesn't provide a dstBuffer */
+        dstBuff = ZSTDMT_getBuffer(job->bufPool);
+        if (dstBuff.start==NULL) {
+            job->cSize = ERROR(memory_allocation);
+            goto _endJob;
+        }
+        job->dstBuff = dstBuff;   /* this value can be read in ZSTDMT_flush, when it copies the whole job */
+    }
+
+    /* init */
+    if (job->cdict) {
+        size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, job->cdict, jobParams, job->fullFrameSize);
+        assert(job->firstJob);  /* only allowed for first job */
+        if (ZSTD_isError(initError)) { job->cSize = initError; goto _endJob; }
+    } else {  /* srcStart points at reloaded section */
+        U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
+        {   size_t const forceWindowError = ZSTD_CCtxParam_setParameter(&jobParams, ZSTD_p_forceMaxWindow, !job->firstJob);
+            if (ZSTD_isError(forceWindowError)) {
+                job->cSize = forceWindowError;
+                goto _endJob;
+        }   }
+        {   size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
+                                        job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
+                                        NULL, /*cdict*/
+                                        jobParams, pledgedSrcSize);
+            if (ZSTD_isError(initError)) {
+                job->cSize = initError;
+                goto _endJob;
+    }   }   }
+
+    /* Perform serial step as early as possible, but after CCtx initialization */
+    ZSTDMT_serialState_update(job->serial, cctx, rawSeqStore, job->src, job->jobID);
+
+    if (!job->firstJob) {  /* flush and overwrite frame header when it's not first job */
+        size_t const hSize = ZSTD_compressContinue(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
+        if (ZSTD_isError(hSize)) { job->cSize = hSize; /* save error code */ goto _endJob; }
+        DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
+        ZSTD_invalidateRepCodes(cctx);
+    }
+
+    /* compress */
+    {   size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
+        int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
+        const BYTE* ip = (const BYTE*) job->src.start;
+        BYTE* const ostart = (BYTE*)dstBuff.start;
+        BYTE* op = ostart;
+        BYTE* oend = op + dstBuff.capacity;
+        int chunkNb;
+        if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize);   /* check overflow */
+        DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
+        assert(job->cSize == 0);
+        for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
+            size_t const cSize = ZSTD_compressContinue(cctx, op, oend-op, ip, chunkSize);
+            if (ZSTD_isError(cSize)) { job->cSize = cSize; goto _endJob; }
+            ip += chunkSize;
+            op += cSize; assert(op < oend);
+            /* stats */
+            ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
+            job->cSize += cSize;
+            job->consumed = chunkSize * chunkNb;
+            DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
+                        (U32)cSize, (U32)job->cSize);
+            ZSTD_pthread_cond_signal(&job->job_cond);   /* warns some more data is ready to be flushed */
+            ZSTD_pthread_mutex_unlock(&job->job_mutex);
+        }
+        /* last block */
+        assert(chunkSize > 0); assert((chunkSize & (chunkSize - 1)) == 0);  /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
+        if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
+            size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
+            size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
+            size_t const cSize = (job->lastJob) ?
+                 ZSTD_compressEnd     (cctx, op, oend-op, ip, lastBlockSize) :
+                 ZSTD_compressContinue(cctx, op, oend-op, ip, lastBlockSize);
+            if (ZSTD_isError(cSize)) { job->cSize = cSize; goto _endJob; }
+            /* stats */
+            ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
+            job->cSize += cSize;
+            ZSTD_pthread_mutex_unlock(&job->job_mutex);
+    }   }
+
+_endJob:
+    ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
+    if (job->prefix.size > 0)
+        DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
+    DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
+    /* release resources */
+    ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
+    ZSTDMT_releaseCCtx(job->cctxPool, cctx);
+    /* report */
+    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
+    job->consumed = job->src.size;
+    ZSTD_pthread_cond_signal(&job->job_cond);
+    ZSTD_pthread_mutex_unlock(&job->job_mutex);
+}
+
+
+/* ------------------------------------------ */
+/* =====   Multi-threaded compression   ===== */
+/* ------------------------------------------ */
+
+typedef struct {
+    range_t prefix;         /* read-only non-owned prefix buffer */
+    buffer_t buffer;
+    size_t filled;
+} inBuff_t;
+
+typedef struct {
+  BYTE* buffer;     /* The round input buffer. All jobs get references
+                     * to pieces of the buffer. ZSTDMT_tryGetInputRange()
+                     * handles handing out job input buffers, and makes
+                     * sure it doesn't overlap with any pieces still in use.
+                     */
+  size_t capacity;  /* The capacity of buffer. */
+  size_t pos;       /* The position of the current inBuff in the round
+                     * buffer. Updated past the end if the inBuff once
+                     * the inBuff is sent to the worker thread.
+                     * pos <= capacity.
+                     */
+} roundBuff_t;
+
+static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
+
+struct ZSTDMT_CCtx_s {
+    POOL_ctx* factory;
+    ZSTDMT_jobDescription* jobs;
+    ZSTDMT_bufferPool* bufPool;
+    ZSTDMT_CCtxPool* cctxPool;
+    ZSTDMT_seqPool* seqPool;
+    ZSTD_CCtx_params params;
+    size_t targetSectionSize;
+    size_t targetPrefixSize;
+    roundBuff_t roundBuff;
+    inBuff_t inBuff;
+    int jobReady;        /* 1 => one job is already prepared, but pool has shortage of workers. Don't create another one. */
+    serialState_t serial;
+    unsigned singleBlockingThread;
+    unsigned jobIDMask;
+    unsigned doneJobID;
+    unsigned nextJobID;
+    unsigned frameEnded;
+    unsigned allJobsCompleted;
+    unsigned long long frameContentSize;
+    unsigned long long consumed;
+    unsigned long long produced;
+    ZSTD_customMem cMem;
+    ZSTD_CDict* cdictLocal;
+    const ZSTD_CDict* cdict;
+};
+
+static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
+{
+    U32 jobNb;
+    if (jobTable == NULL) return;
+    for (jobNb=0; jobNb<nbJobs; jobNb++) {
+        ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
+        ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
+    }
+    ZSTD_free(jobTable, cMem);
+}
+
+/* ZSTDMT_allocJobsTable()
+ * allocate and init a job table.
+ * update *nbJobsPtr to next power of 2 value, as size of table */
+static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
+{
+    U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
+    U32 const nbJobs = 1 << nbJobsLog2;
+    U32 jobNb;
+    ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
+                ZSTD_calloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
+    int initError = 0;
+    if (jobTable==NULL) return NULL;
+    *nbJobsPtr = nbJobs;
+    for (jobNb=0; jobNb<nbJobs; jobNb++) {
+        initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
+        initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
+    }
+    if (initError != 0) {
+        ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
+        return NULL;
+    }
+    return jobTable;
+}
+
+/* ZSTDMT_CCtxParam_setNbWorkers():
+ * Internal use only */
+size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
+{
+    if (nbWorkers > ZSTDMT_NBWORKERS_MAX) nbWorkers = ZSTDMT_NBWORKERS_MAX;
+    params->nbWorkers = nbWorkers;
+    params->overlapSizeLog = ZSTDMT_OVERLAPLOG_DEFAULT;
+    params->jobSize = 0;
+    return nbWorkers;
+}
+
+ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem)
+{
+    ZSTDMT_CCtx* mtctx;
+    U32 nbJobs = nbWorkers + 2;
+    int initError;
+    DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
+
+    if (nbWorkers < 1) return NULL;
+    nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
+    if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
+        /* invalid custom allocator */
+        return NULL;
+
+    mtctx = (ZSTDMT_CCtx*) ZSTD_calloc(sizeof(ZSTDMT_CCtx), cMem);
+    if (!mtctx) return NULL;
+    ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
+    mtctx->cMem = cMem;
+    mtctx->allJobsCompleted = 1;
+    mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
+    mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
+    assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0);  /* ensure nbJobs is a power of 2 */
+    mtctx->jobIDMask = nbJobs - 1;
+    mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
+    mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
+    mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
+    initError = ZSTDMT_serialState_init(&mtctx->serial);
+    mtctx->roundBuff = kNullRoundBuff;
+    if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
+        ZSTDMT_freeCCtx(mtctx);
+        return NULL;
+    }
+    DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
+    return mtctx;
+}
+
+ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers)
+{
+    return ZSTDMT_createCCtx_advanced(nbWorkers, ZSTD_defaultCMem);
+}
+
+
+/* ZSTDMT_releaseAllJobResources() :
+ * note : ensure all workers are killed first ! */
+static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
+{
+    unsigned jobID;
+    DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
+    for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
+        DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
+        ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
+        mtctx->jobs[jobID].dstBuff = g_nullBuffer;
+        mtctx->jobs[jobID].cSize = 0;
+    }
+    memset(mtctx->jobs, 0, (mtctx->jobIDMask+1)*sizeof(ZSTDMT_jobDescription));
+    mtctx->inBuff.buffer = g_nullBuffer;
+    mtctx->inBuff.filled = 0;
+    mtctx->allJobsCompleted = 1;
+}
+
+static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
+{
+    DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
+    while (mtctx->doneJobID < mtctx->nextJobID) {
+        unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
+        ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
+        while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
+            DEBUGLOG(5, "waiting for jobCompleted signal from job %u", mtctx->doneJobID);   /* we want to block when waiting for data to flush */
+            ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
+        }
+        ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
+        mtctx->doneJobID++;
+    }
+}
+
+size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
+{
+    if (mtctx==NULL) return 0;   /* compatible with free on NULL */
+    POOL_free(mtctx->factory);   /* stop and free worker threads */
+    ZSTDMT_releaseAllJobResources(mtctx);  /* release job resources into pools first */
+    ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
+    ZSTDMT_freeBufferPool(mtctx->bufPool);
+    ZSTDMT_freeCCtxPool(mtctx->cctxPool);
+    ZSTDMT_freeSeqPool(mtctx->seqPool);
+    ZSTDMT_serialState_free(&mtctx->serial);
+    ZSTD_freeCDict(mtctx->cdictLocal);
+    if (mtctx->roundBuff.buffer)
+        ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
+    ZSTD_free(mtctx, mtctx->cMem);
+    return 0;
+}
+
+size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
+{
+    if (mtctx == NULL) return 0;   /* supports sizeof NULL */
+    return sizeof(*mtctx)
+            + POOL_sizeof(mtctx->factory)
+            + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
+            + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
+            + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
+            + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
+            + ZSTD_sizeof_CDict(mtctx->cdictLocal)
+            + mtctx->roundBuff.capacity;
+}
+
+/* Internal only */
+size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params,
+                                ZSTDMT_parameter parameter, unsigned value) {
+    DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter");
+    switch(parameter)
+    {
+    case ZSTDMT_p_jobSize :
+        DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter : set jobSize to %u", value);
+        if ( (value > 0)  /* value==0 => automatic job size */
+           & (value < ZSTDMT_JOBSIZE_MIN) )
+            value = ZSTDMT_JOBSIZE_MIN;
+        params->jobSize = value;
+        return value;
+    case ZSTDMT_p_overlapSectionLog :
+        if (value > 9) value = 9;
+        DEBUGLOG(4, "ZSTDMT_p_overlapSectionLog : %u", value);
+        params->overlapSizeLog = (value >= 9) ? 9 : value;
+        return value;
+    default :
+        return ERROR(parameter_unsupported);
+    }
+}
+
+size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned value)
+{
+    DEBUGLOG(4, "ZSTDMT_setMTCtxParameter");
+    switch(parameter)
+    {
+    case ZSTDMT_p_jobSize :
+        return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
+    case ZSTDMT_p_overlapSectionLog :
+        return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
+    default :
+        return ERROR(parameter_unsupported);
+    }
+}
+
+/* Sets parameters relevant to the compression job,
+ * initializing others to default values. */
+static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
+{
+    ZSTD_CCtx_params jobParams;
+    memset(&jobParams, 0, sizeof(jobParams));
+
+    jobParams.cParams = params.cParams;
+    jobParams.fParams = params.fParams;
+    jobParams.compressionLevel = params.compressionLevel;
+    jobParams.disableLiteralCompression = params.disableLiteralCompression;
+
+    return jobParams;
+}
+
+/*! ZSTDMT_updateCParams_whileCompressing() :
+ *  Updates only a selected set of compression parameters, to remain compatible with current frame.
+ *  New parameters will be applied to next compression job. */
+void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
+{
+    U32 const saved_wlog = mtctx->params.cParams.windowLog;   /* Do not modify windowLog while compressing */
+    int const compressionLevel = cctxParams->compressionLevel;
+    DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
+                compressionLevel);
+    mtctx->params.compressionLevel = compressionLevel;
+    {   ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, 0, 0);
+        cParams.windowLog = saved_wlog;
+        mtctx->params.cParams = cParams;
+    }
+}
+
+/* ZSTDMT_getNbWorkers():
+ * @return nb threads currently active in mtctx.
+ * mtctx must be valid */
+unsigned ZSTDMT_getNbWorkers(const ZSTDMT_CCtx* mtctx)
+{
+    assert(mtctx != NULL);
+    return mtctx->params.nbWorkers;
+}
+
+/* ZSTDMT_getFrameProgression():
+ * tells how much data has been consumed (input) and produced (output) for current frame.
+ * able to count progression inside worker threads.
+ * Note : mutex will be acquired during statistics collection. */
+ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
+{
+    ZSTD_frameProgression fps;
+    DEBUGLOG(6, "ZSTDMT_getFrameProgression");
+    fps.consumed = mtctx->consumed;
+    fps.produced = mtctx->produced;
+    fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
+    {   unsigned jobNb;
+        unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
+        DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
+                    mtctx->doneJobID, lastJobNb, mtctx->jobReady)
+        for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
+            unsigned const wJobID = jobNb & mtctx->jobIDMask;
+            ZSTD_pthread_mutex_lock(&mtctx->jobs[wJobID].job_mutex);
+            {   size_t const cResult = mtctx->jobs[wJobID].cSize;
+                size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
+                fps.consumed += mtctx->jobs[wJobID].consumed;
+                fps.ingested += mtctx->jobs[wJobID].src.size;
+                fps.produced += produced;
+            }
+            ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
+        }
+    }
+    return fps;
+}
+
+
+/* ------------------------------------------ */
+/* =====   Multi-threaded compression   ===== */
+/* ------------------------------------------ */
+
+static size_t ZSTDMT_computeTargetJobLog(ZSTD_CCtx_params const params)
+{
+    if (params.ldmParams.enableLdm)
+        return MAX(21, params.cParams.chainLog + 4);
+    return MAX(20, params.cParams.windowLog + 2);
+}
+
+static size_t ZSTDMT_computeOverlapLog(ZSTD_CCtx_params const params)
+{
+    unsigned const overlapRLog = (params.overlapSizeLog>9) ? 0 : 9-params.overlapSizeLog;
+    if (params.ldmParams.enableLdm)
+        return (MIN(params.cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2) - overlapRLog);
+    return overlapRLog >= 9 ? 0 : (params.cParams.windowLog - overlapRLog);
+}
+
+static unsigned ZSTDMT_computeNbJobs(ZSTD_CCtx_params params, size_t srcSize, unsigned nbWorkers) {
+    assert(nbWorkers>0);
+    {   size_t const jobSizeTarget = (size_t)1 << ZSTDMT_computeTargetJobLog(params);
+        size_t const jobMaxSize = jobSizeTarget << 2;
+        size_t const passSizeMax = jobMaxSize * nbWorkers;
+        unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1;
+        unsigned const nbJobsLarge = multiplier * nbWorkers;
+        unsigned const nbJobsMax = (unsigned)(srcSize / jobSizeTarget) + 1;
+        unsigned const nbJobsSmall = MIN(nbJobsMax, nbWorkers);
+        return (multiplier>1) ? nbJobsLarge : nbJobsSmall;
+}   }
+
+/* ZSTDMT_compress_advanced_internal() :
+ * This is a blocking function : it will only give back control to caller after finishing its compression job.
+ */
+static size_t ZSTDMT_compress_advanced_internal(
+                ZSTDMT_CCtx* mtctx,
+                void* dst, size_t dstCapacity,
+          const void* src, size_t srcSize,
+          const ZSTD_CDict* cdict,
+                ZSTD_CCtx_params params)
+{
+    ZSTD_CCtx_params const jobParams = ZSTDMT_initJobCCtxParams(params);
+    size_t const overlapSize = (size_t)1 << ZSTDMT_computeOverlapLog(params);
+    unsigned const nbJobs = ZSTDMT_computeNbJobs(params, srcSize, params.nbWorkers);
+    size_t const proposedJobSize = (srcSize + (nbJobs-1)) / nbJobs;
+    size_t const avgJobSize = (((proposedJobSize-1) & 0x1FFFF) < 0x7FFF) ? proposedJobSize + 0xFFFF : proposedJobSize;   /* avoid too small last block */
+    const char* const srcStart = (const char*)src;
+    size_t remainingSrcSize = srcSize;
+    unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbJobs : (unsigned)(dstCapacity / ZSTD_compressBound(avgJobSize));  /* presumes avgJobSize >= 256 KB, which should be the case */
+    size_t frameStartPos = 0, dstBufferPos = 0;
+    assert(jobParams.nbWorkers == 0);
+    assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
+
+    params.jobSize = (U32)avgJobSize;
+    DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: nbJobs=%2u (rawSize=%u bytes; fixedSize=%u) ",
+                nbJobs, (U32)proposedJobSize, (U32)avgJobSize);
+
+    if ((nbJobs==1) | (params.nbWorkers<=1)) {   /* fallback to single-thread mode : this is a blocking invocation anyway */
+        ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0];
+        DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: fallback to single-thread mode");
+        if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, jobParams.fParams);
+        return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, NULL, 0, jobParams);
+    }
+
+    assert(avgJobSize >= 256 KB);  /* condition for ZSTD_compressBound(A) + ZSTD_compressBound(B) <= ZSTD_compressBound(A+B), required to compress directly into Dst (no additional buffer) */
+    ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(avgJobSize) );
+    if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params))
+        return ERROR(memory_allocation);
+
+    if (nbJobs > mtctx->jobIDMask+1) {  /* enlarge job table */
+        U32 jobsTableSize = nbJobs;
+        ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
+        mtctx->jobIDMask = 0;
+        mtctx->jobs = ZSTDMT_createJobsTable(&jobsTableSize, mtctx->cMem);
+        if (mtctx->jobs==NULL) return ERROR(memory_allocation);
+        assert((jobsTableSize != 0) && ((jobsTableSize & (jobsTableSize - 1)) == 0));  /* ensure jobsTableSize is a power of 2 */
+        mtctx->jobIDMask = jobsTableSize - 1;
+    }
+
+    {   unsigned u;
+        for (u=0; u<nbJobs; u++) {
+            size_t const jobSize = MIN(remainingSrcSize, avgJobSize);
+            size_t const dstBufferCapacity = ZSTD_compressBound(jobSize);
+            buffer_t const dstAsBuffer = { (char*)dst + dstBufferPos, dstBufferCapacity };
+            buffer_t const dstBuffer = u < compressWithinDst ? dstAsBuffer : g_nullBuffer;
+            size_t dictSize = u ? overlapSize : 0;
+
+            mtctx->jobs[u].prefix.start = srcStart + frameStartPos - dictSize;
+            mtctx->jobs[u].prefix.size = dictSize;
+            mtctx->jobs[u].src.start = srcStart + frameStartPos;
+            mtctx->jobs[u].src.size = jobSize; assert(jobSize > 0);  /* avoid job.src.size == 0 */
+            mtctx->jobs[u].consumed = 0;
+            mtctx->jobs[u].cSize = 0;
+            mtctx->jobs[u].cdict = (u==0) ? cdict : NULL;
+            mtctx->jobs[u].fullFrameSize = srcSize;
+            mtctx->jobs[u].params = jobParams;
+            /* do not calculate checksum within sections, but write it in header for first section */
+            mtctx->jobs[u].dstBuff = dstBuffer;
+            mtctx->jobs[u].cctxPool = mtctx->cctxPool;
+            mtctx->jobs[u].bufPool = mtctx->bufPool;
+            mtctx->jobs[u].seqPool = mtctx->seqPool;
+            mtctx->jobs[u].serial = &mtctx->serial;
+            mtctx->jobs[u].jobID = u;
+            mtctx->jobs[u].firstJob = (u==0);
+            mtctx->jobs[u].lastJob = (u==nbJobs-1);
+
+            DEBUGLOG(5, "ZSTDMT_compress_advanced_internal: posting job %u  (%u bytes)", u, (U32)jobSize);
+            DEBUG_PRINTHEX(6, mtctx->jobs[u].prefix.start, 12);
+            POOL_add(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[u]);
+
+            frameStartPos += jobSize;
+            dstBufferPos += dstBufferCapacity;
+            remainingSrcSize -= jobSize;
+    }   }
+
+    /* collect result */
+    {   size_t error = 0, dstPos = 0;
+        unsigned jobID;
+        for (jobID=0; jobID<nbJobs; jobID++) {
+            DEBUGLOG(5, "waiting for job %u ", jobID);
+            ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
+            while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
+                DEBUGLOG(5, "waiting for jobCompleted signal from job %u", jobID);
+                ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
+            }
+            ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
+            DEBUGLOG(5, "ready to write job %u ", jobID);
+
+            {   size_t const cSize = mtctx->jobs[jobID].cSize;
+                if (ZSTD_isError(cSize)) error = cSize;
+                if ((!error) && (dstPos + cSize > dstCapacity)) error = ERROR(dstSize_tooSmall);
+                if (jobID) {   /* note : job 0 is written directly at dst, which is correct position */
+                    if (!error)
+                        memmove((char*)dst + dstPos, mtctx->jobs[jobID].dstBuff.start, cSize);  /* may overlap when job compressed within dst */
+                    if (jobID >= compressWithinDst) {  /* job compressed into its own buffer, which must be released */
+                        DEBUGLOG(5, "releasing buffer %u>=%u", jobID, compressWithinDst);
+                        ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
+                }   }
+                mtctx->jobs[jobID].dstBuff = g_nullBuffer;
+                mtctx->jobs[jobID].cSize = 0;
+                dstPos += cSize ;
+            }
+        }  /* for (jobID=0; jobID<nbJobs; jobID++) */
+
+        DEBUGLOG(4, "checksumFlag : %u ", params.fParams.checksumFlag);
+        if (params.fParams.checksumFlag) {
+            U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
+            if (dstPos + 4 > dstCapacity) {
+                error = ERROR(dstSize_tooSmall);
+            } else {
+                DEBUGLOG(4, "writing checksum : %08X \n", checksum);
+                MEM_writeLE32((char*)dst + dstPos, checksum);
+                dstPos += 4;
+        }   }
+
+        if (!error) DEBUGLOG(4, "compressed size : %u  ", (U32)dstPos);
+        return error ? error : dstPos;
+    }
+}
+
+size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize,
+                         const ZSTD_CDict* cdict,
+                               ZSTD_parameters params,
+                               unsigned overlapLog)
+{
+    ZSTD_CCtx_params cctxParams = mtctx->params;
+    cctxParams.cParams = params.cParams;
+    cctxParams.fParams = params.fParams;
+    cctxParams.overlapSizeLog = overlapLog;
+    return ZSTDMT_compress_advanced_internal(mtctx,
+                                             dst, dstCapacity,
+                                             src, srcSize,
+                                             cdict, cctxParams);
+}
+
+
+size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
+                           void* dst, size_t dstCapacity,
+                     const void* src, size_t srcSize,
+                           int compressionLevel)
+{
+    U32 const overlapLog = (compressionLevel >= ZSTD_maxCLevel()) ? 9 : ZSTDMT_OVERLAPLOG_DEFAULT;
+    ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0);
+    params.fParams.contentSizeFlag = 1;
+    return ZSTDMT_compress_advanced(mtctx, dst, dstCapacity, src, srcSize, NULL, params, overlapLog);
+}
+
+
+/* ====================================== */
+/* =======      Streaming API     ======= */
+/* ====================================== */
+
+size_t ZSTDMT_initCStream_internal(
+        ZSTDMT_CCtx* mtctx,
+        const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
+        const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
+        unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u, disableLiteralCompression=%i)",
+                (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx, params.disableLiteralCompression);
+    /* params are supposed to be fully validated at this point */
+    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
+    assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
+
+    /* init */
+    if (params.jobSize == 0) {
+        params.jobSize = 1U << ZSTDMT_computeTargetJobLog(params);
+    }
+    if (params.jobSize > ZSTDMT_JOBSIZE_MAX) params.jobSize = ZSTDMT_JOBSIZE_MAX;
+
+    mtctx->singleBlockingThread = (pledgedSrcSize <= ZSTDMT_JOBSIZE_MIN);  /* do not trigger multi-threading when srcSize is too small */
+    if (mtctx->singleBlockingThread) {
+        ZSTD_CCtx_params const singleThreadParams = ZSTDMT_initJobCCtxParams(params);
+        DEBUGLOG(5, "ZSTDMT_initCStream_internal: switch to single blocking thread mode");
+        assert(singleThreadParams.nbWorkers == 0);
+        return ZSTD_initCStream_internal(mtctx->cctxPool->cctx[0],
+                                         dict, dictSize, cdict,
+                                         singleThreadParams, pledgedSrcSize);
+    }
+
+    DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
+
+    if (mtctx->allJobsCompleted == 0) {   /* previous compression not correctly finished */
+        ZSTDMT_waitForAllJobsCompleted(mtctx);
+        ZSTDMT_releaseAllJobResources(mtctx);
+        mtctx->allJobsCompleted = 1;
+    }
+
+    mtctx->params = params;
+    mtctx->frameContentSize = pledgedSrcSize;
+    if (dict) {
+        ZSTD_freeCDict(mtctx->cdictLocal);
+        mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
+                                                    ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
+                                                    params.cParams, mtctx->cMem);
+        mtctx->cdict = mtctx->cdictLocal;
+        if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
+    } else {
+        ZSTD_freeCDict(mtctx->cdictLocal);
+        mtctx->cdictLocal = NULL;
+        mtctx->cdict = cdict;
+    }
+
+    mtctx->targetPrefixSize = (size_t)1 << ZSTDMT_computeOverlapLog(params);
+    DEBUGLOG(4, "overlapLog=%u => %u KB", params.overlapSizeLog, (U32)(mtctx->targetPrefixSize>>10));
+    mtctx->targetSectionSize = params.jobSize;
+    if (mtctx->targetSectionSize < ZSTDMT_JOBSIZE_MIN) mtctx->targetSectionSize = ZSTDMT_JOBSIZE_MIN;
+    if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize;  /* job size must be >= overlap size */
+    DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), params.jobSize);
+    DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
+    ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
+    {
+        /* If ldm is enabled we need windowSize space. */
+        size_t const windowSize = mtctx->params.ldmParams.enableLdm ? (1U << mtctx->params.cParams.windowLog) : 0;
+        /* Two buffers of slack, plus extra space for the overlap
+         * This is the minimum slack that LDM works with. One extra because
+         * flush might waste up to targetSectionSize-1 bytes. Another extra
+         * for the overlap (if > 0), then one to fill which doesn't overlap
+         * with the LDM window.
+         */
+        size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
+        size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
+        /* Compute the total size, and always have enough slack */
+        size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
+        size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
+        size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
+        if (mtctx->roundBuff.capacity < capacity) {
+            if (mtctx->roundBuff.buffer)
+                ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
+            mtctx->roundBuff.buffer = (BYTE*)ZSTD_malloc(capacity, mtctx->cMem);
+            if (mtctx->roundBuff.buffer == NULL) {
+                mtctx->roundBuff.capacity = 0;
+                return ERROR(memory_allocation);
+            }
+            mtctx->roundBuff.capacity = capacity;
+        }
+    }
+    DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
+    mtctx->roundBuff.pos = 0;
+    mtctx->inBuff.buffer = g_nullBuffer;
+    mtctx->inBuff.filled = 0;
+    mtctx->inBuff.prefix = kNullRange;
+    mtctx->doneJobID = 0;
+    mtctx->nextJobID = 0;
+    mtctx->frameEnded = 0;
+    mtctx->allJobsCompleted = 0;
+    mtctx->consumed = 0;
+    mtctx->produced = 0;
+    if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params))
+        return ERROR(memory_allocation);
+    return 0;
+}
+
+size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
+                             const void* dict, size_t dictSize,
+                                   ZSTD_parameters params,
+                                   unsigned long long pledgedSrcSize)
+{
+    ZSTD_CCtx_params cctxParams = mtctx->params;  /* retrieve sticky params */
+    DEBUGLOG(4, "ZSTDMT_initCStream_advanced (pledgedSrcSize=%u)", (U32)pledgedSrcSize);
+    cctxParams.cParams = params.cParams;
+    cctxParams.fParams = params.fParams;
+    return ZSTDMT_initCStream_internal(mtctx, dict, dictSize, ZSTD_dct_auto, NULL,
+                                       cctxParams, pledgedSrcSize);
+}
+
+size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
+                               const ZSTD_CDict* cdict,
+                                     ZSTD_frameParameters fParams,
+                                     unsigned long long pledgedSrcSize)
+{
+    ZSTD_CCtx_params cctxParams = mtctx->params;
+    if (cdict==NULL) return ERROR(dictionary_wrong);   /* method incompatible with NULL cdict */
+    cctxParams.cParams = ZSTD_getCParamsFromCDict(cdict);
+    cctxParams.fParams = fParams;
+    return ZSTDMT_initCStream_internal(mtctx, NULL, 0 /*dictSize*/, ZSTD_dct_auto, cdict,
+                                       cctxParams, pledgedSrcSize);
+}
+
+
+/* ZSTDMT_resetCStream() :
+ * pledgedSrcSize can be zero == unknown (for the time being)
+ * prefer using ZSTD_CONTENTSIZE_UNKNOWN,
+ * as `0` might mean "empty" in the future */
+size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize)
+{
+    if (!pledgedSrcSize) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
+    return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, 0, mtctx->params,
+                                       pledgedSrcSize);
+}
+
+size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel) {
+    ZSTD_parameters const params = ZSTD_getParams(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0);
+    ZSTD_CCtx_params cctxParams = mtctx->params;   /* retrieve sticky params */
+    DEBUGLOG(4, "ZSTDMT_initCStream (cLevel=%i)", compressionLevel);
+    cctxParams.cParams = params.cParams;
+    cctxParams.fParams = params.fParams;
+    return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, NULL, cctxParams, ZSTD_CONTENTSIZE_UNKNOWN);
+}
+
+
+/* ZSTDMT_writeLastEmptyBlock()
+ * Write a single empty block with an end-of-frame to finish a frame.
+ * Job must be created from streaming variant.
+ * This function is always successfull if expected conditions are fulfilled.
+ */
+static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
+{
+    assert(job->lastJob == 1);
+    assert(job->src.size == 0);   /* last job is empty -> will be simplified into a last empty block */
+    assert(job->firstJob == 0);   /* cannot be first job, as it also needs to create frame header */
+    assert(job->dstBuff.start == NULL);   /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
+    job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
+    if (job->dstBuff.start == NULL) {
+      job->cSize = ERROR(memory_allocation);
+      return;
+    }
+    assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize);   /* no buffer should ever be that small */
+    job->src = kNullRange;
+    job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
+    assert(!ZSTD_isError(job->cSize));
+    assert(job->consumed == 0);
+}
+
+static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
+{
+    unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
+    int const endFrame = (endOp == ZSTD_e_end);
+
+    if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
+        DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
+        assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
+        return 0;
+    }
+
+    if (!mtctx->jobReady) {
+        BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
+        DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
+                    mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
+        mtctx->jobs[jobID].src.start = src;
+        mtctx->jobs[jobID].src.size = srcSize;
+        assert(mtctx->inBuff.filled >= srcSize);
+        mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
+        mtctx->jobs[jobID].consumed = 0;
+        mtctx->jobs[jobID].cSize = 0;
+        mtctx->jobs[jobID].params = mtctx->params;
+        mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
+        mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
+        mtctx->jobs[jobID].dstBuff = g_nullBuffer;
+        mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
+        mtctx->jobs[jobID].bufPool = mtctx->bufPool;
+        mtctx->jobs[jobID].seqPool = mtctx->seqPool;
+        mtctx->jobs[jobID].serial = &mtctx->serial;
+        mtctx->jobs[jobID].jobID = mtctx->nextJobID;
+        mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
+        mtctx->jobs[jobID].lastJob = endFrame;
+        mtctx->jobs[jobID].frameChecksumNeeded = endFrame && (mtctx->nextJobID>0) && mtctx->params.fParams.checksumFlag;
+        mtctx->jobs[jobID].dstFlushed = 0;
+
+        /* Update the round buffer pos and clear the input buffer to be reset */
+        mtctx->roundBuff.pos += srcSize;
+        mtctx->inBuff.buffer = g_nullBuffer;
+        mtctx->inBuff.filled = 0;
+        /* Set the prefix */
+        if (!endFrame) {
+            size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
+            mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
+            mtctx->inBuff.prefix.size = newPrefixSize;
+        } else {   /* endFrame==1 => no need for another input buffer */
+            mtctx->inBuff.prefix = kNullRange;
+            mtctx->frameEnded = endFrame;
+            if (mtctx->nextJobID == 0) {
+                /* single job exception : checksum is already calculated directly within worker thread */
+                mtctx->params.fParams.checksumFlag = 0;
+        }   }
+
+        if ( (srcSize == 0)
+          && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
+            DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
+            assert(endOp == ZSTD_e_end);  /* only possible case : need to end the frame with an empty last block */
+            ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
+            mtctx->nextJobID++;
+            return 0;
+        }
+    }
+
+    DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes  (end:%u, jobNb == %u (mod:%u))",
+                mtctx->nextJobID,
+                (U32)mtctx->jobs[jobID].src.size,
+                mtctx->jobs[jobID].lastJob,
+                mtctx->nextJobID,
+                jobID);
+    if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
+        mtctx->nextJobID++;
+        mtctx->jobReady = 0;
+    } else {
+        DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
+        mtctx->jobReady = 1;
+    }
+    return 0;
+}
+
+
+/*! ZSTDMT_flushProduced() :
+ * `output` : `pos` will be updated with amount of data flushed .
+ * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
+ * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
+static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
+{
+    unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
+    DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
+                blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
+    assert(output->size >= output->pos);
+
+    ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
+    if (  blockToFlush
+      && (mtctx->doneJobID < mtctx->nextJobID) ) {
+        assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
+        while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) {  /* nothing to flush */
+            if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
+                DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
+                            mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
+                break;
+            }
+            DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
+                        mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
+            ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex);  /* block when nothing to flush but some to come */
+    }   }
+
+    /* try to flush something */
+    {   size_t cSize = mtctx->jobs[wJobID].cSize;                  /* shared */
+        size_t const srcConsumed = mtctx->jobs[wJobID].consumed;   /* shared */
+        size_t const srcSize = mtctx->jobs[wJobID].src.size;        /* read-only, could be done after mutex lock, but no-declaration-after-statement */
+        ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
+        if (ZSTD_isError(cSize)) {
+            DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
+                        mtctx->doneJobID, ZSTD_getErrorName(cSize));
+            ZSTDMT_waitForAllJobsCompleted(mtctx);
+            ZSTDMT_releaseAllJobResources(mtctx);
+            return cSize;
+        }
+        /* add frame checksum if necessary (can only happen once) */
+        assert(srcConsumed <= srcSize);
+        if ( (srcConsumed == srcSize)   /* job completed -> worker no longer active */
+          && mtctx->jobs[wJobID].frameChecksumNeeded ) {
+            U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
+            DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
+            MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
+            cSize += 4;
+            mtctx->jobs[wJobID].cSize += 4;  /* can write this shared value, as worker is no longer active */
+            mtctx->jobs[wJobID].frameChecksumNeeded = 0;
+        }
+        if (cSize > 0) {   /* compression is ongoing or completed */
+            size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
+            DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
+                        (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
+            assert(mtctx->doneJobID < mtctx->nextJobID);
+            assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
+            assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
+            memcpy((char*)output->dst + output->pos,
+                   (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
+                   toFlush);
+            output->pos += toFlush;
+            mtctx->jobs[wJobID].dstFlushed += toFlush;  /* can write : this value is only used by mtctx */
+
+            if ( (srcConsumed == srcSize)    /* job completed */
+              && (mtctx->jobs[wJobID].dstFlushed == cSize) ) {   /* output buffer fully flushed => free this job position */
+                DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
+                        mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
+                ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
+                mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
+                mtctx->jobs[wJobID].cSize = 0;   /* ensure this job slot is considered "not started" in future check */
+                mtctx->consumed += srcSize;
+                mtctx->produced += cSize;
+                mtctx->doneJobID++;
+        }   }
+
+        /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
+        if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
+        if (srcSize > srcConsumed) return 1;   /* current job not completely compressed */
+    }
+    if (mtctx->doneJobID < mtctx->nextJobID) return 1;   /* some more jobs ongoing */
+    if (mtctx->jobReady) return 1;      /* one job is ready to push, just not yet in the list */
+    if (mtctx->inBuff.filled > 0) return 1;   /* input is not empty, and still needs to be converted into a job */
+    mtctx->allJobsCompleted = mtctx->frameEnded;   /* all jobs are entirely flushed => if this one is last one, frame is completed */
+    if (end == ZSTD_e_end) return !mtctx->frameEnded;  /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
+    return 0;   /* internal buffers fully flushed */
+}
+
+/**
+ * Returns the range of data used by the earliest job that is not yet complete.
+ * If the data of the first job is broken up into two segments, we cover both
+ * sections.
+ */
+static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
+{
+    unsigned const firstJobID = mtctx->doneJobID;
+    unsigned const lastJobID = mtctx->nextJobID;
+    unsigned jobID;
+
+    for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
+        unsigned const wJobID = jobID & mtctx->jobIDMask;
+        size_t consumed;
+
+        ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
+        consumed = mtctx->jobs[wJobID].consumed;
+        ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
+
+        if (consumed < mtctx->jobs[wJobID].src.size) {
+            range_t range = mtctx->jobs[wJobID].prefix;
+            if (range.size == 0) {
+                /* Empty prefix */
+                range = mtctx->jobs[wJobID].src;
+            }
+            /* Job source in multiple segments not supported yet */
+            assert(range.start <= mtctx->jobs[wJobID].src.start);
+            return range;
+        }
+    }
+    return kNullRange;
+}
+
+/**
+ * Returns non-zero iff buffer and range overlap.
+ */
+static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
+{
+    BYTE const* const bufferStart = (BYTE const*)buffer.start;
+    BYTE const* const bufferEnd = bufferStart + buffer.capacity;
+    BYTE const* const rangeStart = (BYTE const*)range.start;
+    BYTE const* const rangeEnd = rangeStart + range.size;
+
+    if (rangeStart == NULL || bufferStart == NULL)
+        return 0;
+    /* Empty ranges cannot overlap */
+    if (bufferStart == bufferEnd || rangeStart == rangeEnd)
+        return 0;
+
+    return bufferStart < rangeEnd && rangeStart < bufferEnd;
+}
+
+static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
+{
+    range_t extDict;
+    range_t prefix;
+
+    extDict.start = window.dictBase + window.lowLimit;
+    extDict.size = window.dictLimit - window.lowLimit;
+
+    prefix.start = window.base + window.dictLimit;
+    prefix.size = window.nextSrc - (window.base + window.dictLimit);
+    DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
+                (size_t)extDict.start,
+                (size_t)extDict.start + extDict.size);
+    DEBUGLOG(5, "prefix  [0x%zx, 0x%zx)",
+                (size_t)prefix.start,
+                (size_t)prefix.start + prefix.size);
+
+    return ZSTDMT_isOverlapped(buffer, extDict)
+        || ZSTDMT_isOverlapped(buffer, prefix);
+}
+
+static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
+{
+    if (mtctx->params.ldmParams.enableLdm) {
+        ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
+        DEBUGLOG(5, "source  [0x%zx, 0x%zx)",
+                    (size_t)buffer.start,
+                    (size_t)buffer.start + buffer.capacity);
+        ZSTD_PTHREAD_MUTEX_LOCK(mutex);
+        while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
+            DEBUGLOG(6, "Waiting for LDM to finish...");
+            ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
+        }
+        DEBUGLOG(6, "Done waiting for LDM to finish");
+        ZSTD_pthread_mutex_unlock(mutex);
+    }
+}
+
+/**
+ * Attempts to set the inBuff to the next section to fill.
+ * If any part of the new section is still in use we give up.
+ * Returns non-zero if the buffer is filled.
+ */
+static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
+{
+    range_t const inUse = ZSTDMT_getInputDataInUse(mtctx);
+    size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
+    size_t const target = mtctx->targetSectionSize;
+    buffer_t buffer;
+
+    assert(mtctx->inBuff.buffer.start == NULL);
+    assert(mtctx->roundBuff.capacity >= target);
+
+    if (spaceLeft < target) {
+        /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
+         * Simply copy the prefix to the beginning in that case.
+         */
+        BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
+        size_t const prefixSize = mtctx->inBuff.prefix.size;
+
+        buffer.start = start;
+        buffer.capacity = prefixSize;
+        if (ZSTDMT_isOverlapped(buffer, inUse)) {
+            DEBUGLOG(6, "Waiting for buffer...");
+            return 0;
+        }
+        ZSTDMT_waitForLdmComplete(mtctx, buffer);
+        memmove(start, mtctx->inBuff.prefix.start, prefixSize);
+        mtctx->inBuff.prefix.start = start;
+        mtctx->roundBuff.pos = prefixSize;
+    }
+    buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
+    buffer.capacity = target;
+
+    if (ZSTDMT_isOverlapped(buffer, inUse)) {
+        DEBUGLOG(6, "Waiting for buffer...");
+        return 0;
+    }
+    assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));
+
+    ZSTDMT_waitForLdmComplete(mtctx, buffer);
+
+    DEBUGLOG(5, "Using prefix range [%zx, %zx)",
+                (size_t)mtctx->inBuff.prefix.start,
+                (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
+    DEBUGLOG(5, "Using source range [%zx, %zx)",
+                (size_t)buffer.start,
+                (size_t)buffer.start + buffer.capacity);
+
+
+    mtctx->inBuff.buffer = buffer;
+    mtctx->inBuff.filled = 0;
+    assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
+    return 1;
+}
+
+
+/** ZSTDMT_compressStream_generic() :
+ *  internal use only - exposed to be invoked from zstd_compress.c
+ *  assumption : output and input are valid (pos <= size)
+ * @return : minimum amount of data remaining to flush, 0 if none */
+size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
+                                     ZSTD_outBuffer* output,
+                                     ZSTD_inBuffer* input,
+                                     ZSTD_EndDirective endOp)
+{
+    unsigned forwardInputProgress = 0;
+    DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
+                (U32)endOp, (U32)(input->size - input->pos));
+    assert(output->pos <= output->size);
+    assert(input->pos  <= input->size);
+
+    if (mtctx->singleBlockingThread) {  /* delegate to single-thread (synchronous) */
+        return ZSTD_compressStream_generic(mtctx->cctxPool->cctx[0], output, input, endOp);
+    }
+
+    if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
+        /* current frame being ended. Only flush/end are allowed */
+        return ERROR(stage_wrong);
+    }
+
+    /* single-pass shortcut (note : synchronous-mode) */
+    if ( (mtctx->nextJobID == 0)      /* just started */
+      && (mtctx->inBuff.filled == 0)  /* nothing buffered */
+      && (!mtctx->jobReady)           /* no job already created */
+      && (endOp == ZSTD_e_end)        /* end order */
+      && (output->size - output->pos >= ZSTD_compressBound(input->size - input->pos)) ) { /* enough space in dst */
+        size_t const cSize = ZSTDMT_compress_advanced_internal(mtctx,
+                (char*)output->dst + output->pos, output->size - output->pos,
+                (const char*)input->src + input->pos, input->size - input->pos,
+                mtctx->cdict, mtctx->params);
+        if (ZSTD_isError(cSize)) return cSize;
+        input->pos = input->size;
+        output->pos += cSize;
+        mtctx->allJobsCompleted = 1;
+        mtctx->frameEnded = 1;
+        return 0;
+    }
+
+    /* fill input buffer */
+    if ( (!mtctx->jobReady)
+      && (input->size > input->pos) ) {   /* support NULL input */
+        if (mtctx->inBuff.buffer.start == NULL) {
+            assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
+            if (!ZSTDMT_tryGetInputRange(mtctx)) {
+                /* It is only possible for this operation to fail if there are
+                 * still compression jobs ongoing.
+                 */
+                assert(mtctx->doneJobID != mtctx->nextJobID);
+            }
+        }
+        if (mtctx->inBuff.buffer.start != NULL) {
+            size_t const toLoad = MIN(input->size - input->pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
+            assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
+            DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
+                        (U32)toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
+            memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, toLoad);
+            input->pos += toLoad;
+            mtctx->inBuff.filled += toLoad;
+            forwardInputProgress = toLoad>0;
+        }
+        if ((input->pos < input->size) && (endOp == ZSTD_e_end))
+            endOp = ZSTD_e_flush;   /* can't end now : not all input consumed */
+    }
+
+    if ( (mtctx->jobReady)
+      || (mtctx->inBuff.filled >= mtctx->targetSectionSize)  /* filled enough : let's compress */
+      || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0))  /* something to flush : let's go */
+      || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) {   /* must finish the frame with a zero-size block */
+        size_t const jobSize = mtctx->inBuff.filled;
+        assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
+        CHECK_F( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) );
+    }
+
+    /* check for potential compressed data ready to be flushed */
+    {   size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
+        if (input->pos < input->size) return MAX(remainingToFlush, 1);  /* input not consumed : do not end flush yet */
+        return remainingToFlush;
+    }
+}
+
+
+size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
+{
+    CHECK_F( ZSTDMT_compressStream_generic(mtctx, output, input, ZSTD_e_continue) );
+
+    /* recommended next input size : fill current input buffer */
+    return mtctx->targetSectionSize - mtctx->inBuff.filled;   /* note : could be zero when input buffer is fully filled and no more availability to create new job */
+}
+
+
+static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_EndDirective endFrame)
+{
+    size_t const srcSize = mtctx->inBuff.filled;
+    DEBUGLOG(5, "ZSTDMT_flushStream_internal");
+
+    if ( mtctx->jobReady     /* one job ready for a worker to pick up */
+      || (srcSize > 0)       /* still some data within input buffer */
+      || ((endFrame==ZSTD_e_end) && !mtctx->frameEnded)) {  /* need a last 0-size block to end frame */
+           DEBUGLOG(5, "ZSTDMT_flushStream_internal : create a new job (%u bytes, end:%u)",
+                        (U32)srcSize, (U32)endFrame);
+        CHECK_F( ZSTDMT_createCompressionJob(mtctx, srcSize, endFrame) );
+    }
+
+    /* check if there is any data available to flush */
+    return ZSTDMT_flushProduced(mtctx, output, 1 /* blockToFlush */, endFrame);
+}
+
+
+size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
+{
+    DEBUGLOG(5, "ZSTDMT_flushStream");
+    if (mtctx->singleBlockingThread)
+        return ZSTD_flushStream(mtctx->cctxPool->cctx[0], output);
+    return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_flush);
+}
+
+size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
+{
+    DEBUGLOG(4, "ZSTDMT_endStream");
+    if (mtctx->singleBlockingThread)
+        return ZSTD_endStream(mtctx->cctxPool->cctx[0], output);
+    return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_end);
+}
diff --git a/vendor/github.com/DataDog/zstd/zstdmt_compress.h b/vendor/github.com/DataDog/zstd/zstdmt_compress.h
new file mode 100755
index 0000000..f79e3b4
--- /dev/null
+++ b/vendor/github.com/DataDog/zstd/zstdmt_compress.h
@@ -0,0 +1,156 @@
+/*
+ * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+ #ifndef ZSTDMT_COMPRESS_H
+ #define ZSTDMT_COMPRESS_H
+
+ #if defined (__cplusplus)
+ extern "C" {
+ #endif
+
+
+/* Note : This is an internal API.
+ *        Some methods are still exposed (ZSTDLIB_API),
+ *        because it used to be the only way to invoke MT compression.
+ *        Now, it's recommended to use ZSTD_compress_generic() instead.
+ *        These methods will stop being exposed in a future version */
+
+/* ===   Dependencies   === */
+#include <stddef.h>                /* size_t */
+#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_parameters */
+#include "zstd.h"            /* ZSTD_inBuffer, ZSTD_outBuffer, ZSTDLIB_API */
+
+
+/* ===   Memory management   === */
+typedef struct ZSTDMT_CCtx_s ZSTDMT_CCtx;
+ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers);
+ZSTDLIB_API ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers,
+                                                    ZSTD_customMem cMem);
+ZSTDLIB_API size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx);
+
+ZSTDLIB_API size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx);
+
+
+/* ===   Simple one-pass compression function   === */
+
+ZSTDLIB_API size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
+                                       void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize,
+                                       int compressionLevel);
+
+
+
+/* ===   Streaming functions   === */
+
+ZSTDLIB_API size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel);
+ZSTDLIB_API size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize);  /**< if srcSize is not known at reset time, use ZSTD_CONTENTSIZE_UNKNOWN. Note: for compatibility with older programs, 0 means the same as ZSTD_CONTENTSIZE_UNKNOWN, but it will change in the future to mean "empty" */
+
+ZSTDLIB_API size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
+
+ZSTDLIB_API size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output);   /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */
+ZSTDLIB_API size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output);     /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */
+
+
+/* ===   Advanced functions and parameters  === */
+
+#ifndef ZSTDMT_JOBSIZE_MIN
+#  define ZSTDMT_JOBSIZE_MIN (1U << 20)   /* 1 MB - Minimum size of each compression job */
+#endif
+
+ZSTDLIB_API size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
+                                           void* dst, size_t dstCapacity,
+                                     const void* src, size_t srcSize,
+                                     const ZSTD_CDict* cdict,
+                                           ZSTD_parameters params,
+                                           unsigned overlapLog);
+
+ZSTDLIB_API size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
+                                        const void* dict, size_t dictSize,   /* dict can be released after init, a local copy is preserved within zcs */
+                                        ZSTD_parameters params,
+                                        unsigned long long pledgedSrcSize);  /* pledgedSrcSize is optional and can be zero == unknown */
+
+ZSTDLIB_API size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
+                                        const ZSTD_CDict* cdict,
+                                        ZSTD_frameParameters fparams,
+                                        unsigned long long pledgedSrcSize);  /* note : zero means empty */
+
+/* ZSTDMT_parameter :
+ * List of parameters that can be set using ZSTDMT_setMTCtxParameter() */
+typedef enum {
+    ZSTDMT_p_jobSize,           /* Each job is compressed in parallel. By default, this value is dynamically determined depending on compression parameters. Can be set explicitly here. */
+    ZSTDMT_p_overlapSectionLog  /* Each job may reload a part of previous job to enhance compressionr ratio; 0 == no overlap, 6(default) == use 1/8th of window, >=9 == use full window. This is a "sticky" parameter : its value will be re-used on next compression job */
+} ZSTDMT_parameter;
+
+/* ZSTDMT_setMTCtxParameter() :
+ * allow setting individual parameters, one at a time, among a list of enums defined in ZSTDMT_parameter.
+ * The function must be called typically after ZSTD_createCCtx() but __before ZSTDMT_init*() !__
+ * Parameters not explicitly reset by ZSTDMT_init*() remain the same in consecutive compression sessions.
+ * @return : 0, or an error code (which can be tested using ZSTD_isError()) */
+ZSTDLIB_API size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned value);
+
+
+/*! ZSTDMT_compressStream_generic() :
+ *  Combines ZSTDMT_compressStream() with optional ZSTDMT_flushStream() or ZSTDMT_endStream()
+ *  depending on flush directive.
+ * @return : minimum amount of data still to be flushed
+ *           0 if fully flushed
+ *           or an error code
+ *  note : needs to be init using any ZSTD_initCStream*() variant */
+ZSTDLIB_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
+                                                ZSTD_outBuffer* output,
+                                                ZSTD_inBuffer* input,
+                                                ZSTD_EndDirective endOp);
+
+
+/* ========================================================
+ * ===  Private interface, for use by ZSTD_compress.c   ===
+ * ===  Not exposed in libzstd. Never invoke directly   ===
+ * ======================================================== */
+
+size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params, ZSTDMT_parameter parameter, unsigned value);
+
+/* ZSTDMT_CCtxParam_setNbWorkers()
+ * Set nbWorkers, and clamp it.
+ * Also reset jobSize and overlapLog */
+size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers);
+
+/*! ZSTDMT_updateCParams_whileCompressing() :
+ *  Updates only a selected set of compression parameters, to remain compatible with current frame.
+ *  New parameters will be applied to next compression job. */
+void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams);
+
+/* ZSTDMT_getNbWorkers():
+ * @return nb threads currently active in mtctx.
+ * mtctx must be valid */
+unsigned ZSTDMT_getNbWorkers(const ZSTDMT_CCtx* mtctx);
+
+/* ZSTDMT_getFrameProgression():
+ * tells how much data has been consumed (input) and produced (output) for current frame.
+ * able to count progression inside worker threads.
+ */
+ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx);
+
+
+/*! ZSTDMT_initCStream_internal() :
+ *  Private use only. Init streaming operation.
+ *  expects params to be valid.
+ *  must receive dict, or cdict, or none, but not both.
+ *  @return : 0, or an error code */
+size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
+                    const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
+                    const ZSTD_CDict* cdict,
+                    ZSTD_CCtx_params params, unsigned long long pledgedSrcSize);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* ZSTDMT_COMPRESS_H */
diff --git a/vendor/github.com/Shopify/sarama/.gitignore b/vendor/github.com/Shopify/sarama/.gitignore
new file mode 100644
index 0000000..6e362e4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/.gitignore
@@ -0,0 +1,27 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+*.test
+
+# Folders
+_obj
+_test
+.vagrant
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+
+coverage.txt
+profile.out
diff --git a/vendor/github.com/Shopify/sarama/.travis.yml b/vendor/github.com/Shopify/sarama/.travis.yml
new file mode 100644
index 0000000..b7f7874
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/.travis.yml
@@ -0,0 +1,35 @@
+language: go
+go:
+- 1.10.x
+- 1.11.x
+
+env:
+  global:
+  - KAFKA_PEERS=localhost:9091,localhost:9092,localhost:9093,localhost:9094,localhost:9095
+  - TOXIPROXY_ADDR=http://localhost:8474
+  - KAFKA_INSTALL_ROOT=/home/travis/kafka
+  - KAFKA_HOSTNAME=localhost
+  - DEBUG=true
+  matrix:
+  - KAFKA_VERSION=1.1.1 KAFKA_SCALA_VERSION=2.11
+  - KAFKA_VERSION=2.0.1 KAFKA_SCALA_VERSION=2.12
+  - KAFKA_VERSION=2.1.0 KAFKA_SCALA_VERSION=2.12
+
+before_install:
+- export REPOSITORY_ROOT=${TRAVIS_BUILD_DIR}
+- vagrant/install_cluster.sh
+- vagrant/boot_cluster.sh
+- vagrant/create_topics.sh
+
+install: make install_dependencies
+
+script:
+- make test
+- make vet
+- make errcheck
+- if [[ "$TRAVIS_GO_VERSION" == 1.11* ]]; then make fmt; fi
+
+after_success:
+- bash <(curl -s https://codecov.io/bash)
+
+after_script: vagrant/halt_cluster.sh
diff --git a/vendor/github.com/Shopify/sarama/CHANGELOG.md b/vendor/github.com/Shopify/sarama/CHANGELOG.md
new file mode 100644
index 0000000..0e69c2d
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/CHANGELOG.md
@@ -0,0 +1,683 @@
+# Changelog
+
+#### Version 1.21.0 (2019-02-24)
+
+New Features:
+- Add CreateAclRequest, DescribeAclRequest, DeleteAclRequest
+  ([1236](https://github.com/Shopify/sarama/pull/1236)).
+- Add DescribeTopic, DescribeConsumerGroup, ListConsumerGroups, ListConsumerGroupOffsets admin requests
+  ([1178](https://github.com/Shopify/sarama/pull/1178)).
+- Implement SASL/OAUTHBEARER
+  ([1240](https://github.com/Shopify/sarama/pull/1240)).
+
+Improvements:
+- Add Go mod support
+  ([1282](https://github.com/Shopify/sarama/pull/1282)).
+- Add error codes 73—76
+  ([1239](https://github.com/Shopify/sarama/pull/1239)).
+- Add retry backoff function
+  ([1160](https://github.com/Shopify/sarama/pull/1160)).
+- Maintain metadata in the producer even when retries are disabled
+  ([1189](https://github.com/Shopify/sarama/pull/1189)).
+- Include ReplicaAssignment in ListTopics
+  ([1274](https://github.com/Shopify/sarama/pull/1274)).
+- Add producer performance tool
+  ([1222](https://github.com/Shopify/sarama/pull/1222)).
+- Add support LogAppend timestamps
+  ([1258](https://github.com/Shopify/sarama/pull/1258)).
+
+Bug Fixes:
+- Fix potential deadlock when a heartbeat request fails
+  ([1286](https://github.com/Shopify/sarama/pull/1286)).
+- Fix consuming compacted topic
+  ([1227](https://github.com/Shopify/sarama/pull/1227)).
+- Set correct Kafka version for DescribeConfigsRequest v1
+  ([1277](https://github.com/Shopify/sarama/pull/1277)).
+- Update kafka test version
+  ([1273](https://github.com/Shopify/sarama/pull/1273)).
+
+#### Version 1.20.1 (2019-01-10)
+
+New Features:
+- Add optional replica id in offset request
+  ([1100](https://github.com/Shopify/sarama/pull/1100)).
+
+Improvements:
+- Implement DescribeConfigs Request + Response v1 & v2
+  ([1230](https://github.com/Shopify/sarama/pull/1230)).
+- Reuse compression objects
+  ([1185](https://github.com/Shopify/sarama/pull/1185)).
+- Switch from png to svg for GoDoc link in README
+  ([1243](https://github.com/Shopify/sarama/pull/1243)).
+- Fix typo in deprecation notice for FetchResponseBlock.Records
+  ([1242](https://github.com/Shopify/sarama/pull/1242)).
+- Fix typos in consumer metadata response file
+  ([1244](https://github.com/Shopify/sarama/pull/1244)).
+
+Bug Fixes:
+- Revert to individual msg retries for non-idempotent
+  ([1203](https://github.com/Shopify/sarama/pull/1203)).
+- Respect MaxMessageBytes limit for uncompressed messages
+  ([1141](https://github.com/Shopify/sarama/pull/1141)).
+
+#### Version 1.20.0 (2018-12-10)
+
+New Features:
+ - Add support for zstd compression
+   ([#1170](https://github.com/Shopify/sarama/pull/1170)).
+ - Add support for Idempotent Producer
+   ([#1152](https://github.com/Shopify/sarama/pull/1152)).
+ - Add support support for Kafka 2.1.0
+   ([#1229](https://github.com/Shopify/sarama/pull/1229)).
+ - Add support support for OffsetCommit request/response pairs versions v1 to v5
+   ([#1201](https://github.com/Shopify/sarama/pull/1201)).
+ - Add support support for OffsetFetch request/response pair up to version v5
+   ([#1198](https://github.com/Shopify/sarama/pull/1198)).
+
+Improvements:
+ - Export broker's Rack setting
+   ([#1173](https://github.com/Shopify/sarama/pull/1173)).
+ - Always use latest patch version of Go on CI
+   ([#1202](https://github.com/Shopify/sarama/pull/1202)).
+ - Add error codes 61 to 72
+   ([#1195](https://github.com/Shopify/sarama/pull/1195)).
+
+Bug Fixes:
+ - Fix build without cgo
+   ([#1182](https://github.com/Shopify/sarama/pull/1182)).
+ - Fix go vet suggestion in consumer group file
+   ([#1209](https://github.com/Shopify/sarama/pull/1209)).
+ - Fix typos in code and comments
+   ([#1228](https://github.com/Shopify/sarama/pull/1228)).
+
+#### Version 1.19.0 (2018-09-27)
+
+New Features:
+ - Implement a higher-level consumer group
+   ([#1099](https://github.com/Shopify/sarama/pull/1099)).
+
+Improvements:
+ - Add support for Go 1.11
+   ([#1176](https://github.com/Shopify/sarama/pull/1176)).
+
+Bug Fixes:
+ - Fix encoding of `MetadataResponse` with version 2 and higher
+   ([#1174](https://github.com/Shopify/sarama/pull/1174)).
+ - Fix race condition in mock async producer
+   ([#1174](https://github.com/Shopify/sarama/pull/1174)).
+
+#### Version 1.18.0 (2018-09-07)
+
+New Features:
+ - Make `Partitioner.RequiresConsistency` vary per-message
+   ([#1112](https://github.com/Shopify/sarama/pull/1112)).
+ - Add customizable partitioner
+   ([#1118](https://github.com/Shopify/sarama/pull/1118)).
+ - Add `ClusterAdmin` support for `CreateTopic`, `DeleteTopic`, `CreatePartitions`,
+   `DeleteRecords`, `DescribeConfig`, `AlterConfig`, `CreateACL`, `ListAcls`, `DeleteACL`
+   ([#1055](https://github.com/Shopify/sarama/pull/1055)).
+
+Improvements:
+ - Add support for Kafka 2.0.0
+   ([#1149](https://github.com/Shopify/sarama/pull/1149)).
+ - Allow setting `LocalAddr` when dialing an address to support multi-homed hosts
+   ([#1123](https://github.com/Shopify/sarama/pull/1123)).
+ - Simpler offset management
+   ([#1127](https://github.com/Shopify/sarama/pull/1127)).
+
+Bug Fixes:
+ - Fix mutation of `ProducerMessage.MetaData` when producing to Kafka
+   ([#1110](https://github.com/Shopify/sarama/pull/1110)).
+ - Fix consumer block when response did not contain all the
+   expected topic/partition blocks
+   ([#1086](https://github.com/Shopify/sarama/pull/1086)).
+ - Fix consumer block when response contains only constrol messages
+   ([#1115](https://github.com/Shopify/sarama/pull/1115)).
+ - Add timeout config for ClusterAdmin requests
+   ([#1142](https://github.com/Shopify/sarama/pull/1142)).
+ - Add version check when producing message with headers
+   ([#1117](https://github.com/Shopify/sarama/pull/1117)).
+ - Fix `MetadataRequest` for empty list of topics
+   ([#1132](https://github.com/Shopify/sarama/pull/1132)).
+ - Fix producer topic metadata on-demand fetch when topic error happens in metadata response
+   ([#1125](https://github.com/Shopify/sarama/pull/1125)).
+
+#### Version 1.17.0 (2018-05-30)
+
+New Features:
+ - Add support for gzip compression levels
+   ([#1044](https://github.com/Shopify/sarama/pull/1044)).
+ - Add support for Metadata request/response pairs versions v1 to v5
+   ([#1047](https://github.com/Shopify/sarama/pull/1047),
+    [#1069](https://github.com/Shopify/sarama/pull/1069)).
+ - Add versioning to JoinGroup request/response pairs
+   ([#1098](https://github.com/Shopify/sarama/pull/1098))
+ - Add support for CreatePartitions, DeleteGroups, DeleteRecords request/response pairs
+   ([#1065](https://github.com/Shopify/sarama/pull/1065),
+    [#1096](https://github.com/Shopify/sarama/pull/1096),
+    [#1027](https://github.com/Shopify/sarama/pull/1027)).
+ - Add `Controller()` method to Client interface
+   ([#1063](https://github.com/Shopify/sarama/pull/1063)).
+
+Improvements:
+ - ConsumerMetadataReq/Resp has been migrated to FindCoordinatorReq/Resp
+   ([#1010](https://github.com/Shopify/sarama/pull/1010)).
+ - Expose missing protocol parts: `msgSet` and `recordBatch`
+   ([#1049](https://github.com/Shopify/sarama/pull/1049)).
+ - Add support for v1 DeleteTopics Request
+   ([#1052](https://github.com/Shopify/sarama/pull/1052)).
+ - Add support for Go 1.10
+   ([#1064](https://github.com/Shopify/sarama/pull/1064)).
+ - Claim support for Kafka 1.1.0
+   ([#1073](https://github.com/Shopify/sarama/pull/1073)).
+
+Bug Fixes:
+ - Fix FindCoordinatorResponse.encode to allow nil Coordinator
+   ([#1050](https://github.com/Shopify/sarama/pull/1050),
+    [#1051](https://github.com/Shopify/sarama/pull/1051)).
+ - Clear all metadata when we have the latest topic info
+   ([#1033](https://github.com/Shopify/sarama/pull/1033)).
+ - Make `PartitionConsumer.Close` idempotent
+   ([#1092](https://github.com/Shopify/sarama/pull/1092)).
+
+#### Version 1.16.0 (2018-02-12)
+
+New Features:
+ - Add support for the Create/Delete Topics request/response pairs
+   ([#1007](https://github.com/Shopify/sarama/pull/1007),
+    [#1008](https://github.com/Shopify/sarama/pull/1008)).
+ - Add support for the Describe/Create/Delete ACL request/response pairs
+   ([#1009](https://github.com/Shopify/sarama/pull/1009)).
+ - Add support for the five transaction-related request/response pairs
+   ([#1016](https://github.com/Shopify/sarama/pull/1016)).
+
+Improvements:
+ - Permit setting version on mock producer responses
+   ([#999](https://github.com/Shopify/sarama/pull/999)).
+ - Add `NewMockBrokerListener` helper for testing TLS connections
+   ([#1019](https://github.com/Shopify/sarama/pull/1019)).
+ - Changed the default value for `Consumer.Fetch.Default` from 32KiB to 1MiB
+   which results in much higher throughput in most cases
+   ([#1024](https://github.com/Shopify/sarama/pull/1024)).
+ - Reuse the `time.Ticker` across fetch requests in the PartitionConsumer to
+   reduce CPU and memory usage when processing many partitions
+   ([#1028](https://github.com/Shopify/sarama/pull/1028)).
+ - Assign relative offsets to messages in the producer to save the brokers a
+   recompression pass
+   ([#1002](https://github.com/Shopify/sarama/pull/1002),
+    [#1015](https://github.com/Shopify/sarama/pull/1015)).
+
+Bug Fixes:
+ - Fix producing uncompressed batches with the new protocol format
+   ([#1032](https://github.com/Shopify/sarama/issues/1032)).
+ - Fix consuming compacted topics with the new protocol format
+   ([#1005](https://github.com/Shopify/sarama/issues/1005)).
+ - Fix consuming topics with a mix of protocol formats
+   ([#1021](https://github.com/Shopify/sarama/issues/1021)).
+ - Fix consuming when the broker includes multiple batches in a single response
+   ([#1022](https://github.com/Shopify/sarama/issues/1022)).
+ - Fix detection of `PartialTrailingMessage` when the partial message was
+   truncated before the magic value indicating its version
+   ([#1030](https://github.com/Shopify/sarama/pull/1030)).
+ - Fix expectation-checking in the mock of `SyncProducer.SendMessages`
+   ([#1035](https://github.com/Shopify/sarama/pull/1035)).
+
+#### Version 1.15.0 (2017-12-08)
+
+New Features:
+ - Claim official support for Kafka 1.0, though it did already work
+   ([#984](https://github.com/Shopify/sarama/pull/984)).
+ - Helper methods for Kafka version numbers to/from strings
+   ([#989](https://github.com/Shopify/sarama/pull/989)).
+ - Implement CreatePartitions request/response
+   ([#985](https://github.com/Shopify/sarama/pull/985)).
+
+Improvements:
+ - Add error codes 45-60
+   ([#986](https://github.com/Shopify/sarama/issues/986)).
+
+Bug Fixes:
+ - Fix slow consuming for certain Kafka 0.11/1.0 configurations
+   ([#982](https://github.com/Shopify/sarama/pull/982)).
+ - Correctly determine when a FetchResponse contains the new message format
+   ([#990](https://github.com/Shopify/sarama/pull/990)).
+ - Fix producing with multiple headers
+   ([#996](https://github.com/Shopify/sarama/pull/996)).
+ - Fix handling of truncated record batches
+   ([#998](https://github.com/Shopify/sarama/pull/998)).
+ - Fix leaking metrics when closing brokers
+   ([#991](https://github.com/Shopify/sarama/pull/991)).
+
+#### Version 1.14.0 (2017-11-13)
+
+New Features:
+ - Add support for the new Kafka 0.11 record-batch format, including the wire
+   protocol and the necessary behavioural changes in the producer and consumer.
+   Transactions and idempotency are not yet supported, but producing and
+   consuming should work with all the existing bells and whistles (batching,
+   compression, etc) as well as the new custom headers. Thanks to Vlad Hanciuta
+   of Arista Networks for this work. Part of
+   ([#901](https://github.com/Shopify/sarama/issues/901)).
+
+Bug Fixes:
+ - Fix encoding of ProduceResponse versions in test
+   ([#970](https://github.com/Shopify/sarama/pull/970)).
+ - Return partial replicas list when we have it
+   ([#975](https://github.com/Shopify/sarama/pull/975)).
+
+#### Version 1.13.0 (2017-10-04)
+
+New Features:
+ - Support for FetchRequest version 3
+   ([#905](https://github.com/Shopify/sarama/pull/905)).
+ - Permit setting version on mock FetchResponses
+   ([#939](https://github.com/Shopify/sarama/pull/939)).
+ - Add a configuration option to support storing only minimal metadata for
+   extremely large clusters
+   ([#937](https://github.com/Shopify/sarama/pull/937)).
+ - Add `PartitionOffsetManager.ResetOffset` for backtracking tracked offsets
+   ([#932](https://github.com/Shopify/sarama/pull/932)).
+
+Improvements:
+ - Provide the block-level timestamp when consuming compressed messages
+   ([#885](https://github.com/Shopify/sarama/issues/885)).
+ - `Client.Replicas` and `Client.InSyncReplicas` now respect the order returned
+   by the broker, which can be meaningful
+   ([#930](https://github.com/Shopify/sarama/pull/930)).
+ - Use a `Ticker` to reduce consumer timer overhead at the cost of higher
+   variance in the actual timeout
+   ([#933](https://github.com/Shopify/sarama/pull/933)).
+
+Bug Fixes:
+ - Gracefully handle messages with negative timestamps
+   ([#907](https://github.com/Shopify/sarama/pull/907)).
+ - Raise a proper error when encountering an unknown message version
+   ([#940](https://github.com/Shopify/sarama/pull/940)).
+
+#### Version 1.12.0 (2017-05-08)
+
+New Features:
+ - Added support for the `ApiVersions` request and response pair, and Kafka
+   version 0.10.2 ([#867](https://github.com/Shopify/sarama/pull/867)). Note
+   that you still need to specify the Kafka version in the Sarama configuration
+   for the time being.
+ - Added a `Brokers` method to the Client which returns the complete set of
+   active brokers ([#813](https://github.com/Shopify/sarama/pull/813)).
+ - Added an `InSyncReplicas` method to the Client which returns the set of all
+   in-sync broker IDs for the given partition, now that the Kafka versions for
+   which this was misleading are no longer in our supported set
+   ([#872](https://github.com/Shopify/sarama/pull/872)).
+ - Added a `NewCustomHashPartitioner` method which allows constructing a hash
+   partitioner with a custom hash method in case the default (FNV-1a) is not
+   suitable
+   ([#837](https://github.com/Shopify/sarama/pull/837),
+    [#841](https://github.com/Shopify/sarama/pull/841)).
+
+Improvements:
+ - Recognize more Kafka error codes
+   ([#859](https://github.com/Shopify/sarama/pull/859)).
+
+Bug Fixes:
+ - Fix an issue where decoding a malformed FetchRequest would not return the
+   correct error ([#818](https://github.com/Shopify/sarama/pull/818)).
+ - Respect ordering of group protocols in JoinGroupRequests. This fix is
+   transparent if you're using the `AddGroupProtocol` or
+   `AddGroupProtocolMetadata` helpers; otherwise you will need to switch from
+   the `GroupProtocols` field (now deprecated) to use `OrderedGroupProtocols`
+   ([#812](https://github.com/Shopify/sarama/issues/812)).
+ - Fix an alignment-related issue with atomics on 32-bit architectures
+   ([#859](https://github.com/Shopify/sarama/pull/859)).
+
+#### Version 1.11.0 (2016-12-20)
+
+_Important:_ As of Sarama 1.11 it is necessary to set the config value of
+`Producer.Return.Successes` to true in order to use the SyncProducer. Previous
+versions would silently override this value when instantiating a SyncProducer
+which led to unexpected values and data races.
+
+New Features:
+ - Metrics! Thanks to Sébastien Launay for all his work on this feature
+   ([#701](https://github.com/Shopify/sarama/pull/701),
+    [#746](https://github.com/Shopify/sarama/pull/746),
+    [#766](https://github.com/Shopify/sarama/pull/766)).
+ - Add support for LZ4 compression
+   ([#786](https://github.com/Shopify/sarama/pull/786)).
+ - Add support for ListOffsetRequest v1 and Kafka 0.10.1
+   ([#775](https://github.com/Shopify/sarama/pull/775)).
+ - Added a `HighWaterMarks` method to the Consumer which aggregates the
+   `HighWaterMarkOffset` values of its child topic/partitions
+   ([#769](https://github.com/Shopify/sarama/pull/769)).
+
+Bug Fixes:
+ - Fixed producing when using timestamps, compression and Kafka 0.10
+   ([#759](https://github.com/Shopify/sarama/pull/759)).
+ - Added missing decoder methods to DescribeGroups response
+   ([#756](https://github.com/Shopify/sarama/pull/756)).
+ - Fix producer shutdown when `Return.Errors` is disabled
+   ([#787](https://github.com/Shopify/sarama/pull/787)).
+ - Don't mutate configuration in SyncProducer
+   ([#790](https://github.com/Shopify/sarama/pull/790)).
+ - Fix crash on SASL initialization failure
+   ([#795](https://github.com/Shopify/sarama/pull/795)).
+
+#### Version 1.10.1 (2016-08-30)
+
+Bug Fixes:
+ - Fix the documentation for `HashPartitioner` which was incorrect
+   ([#717](https://github.com/Shopify/sarama/pull/717)).
+ - Permit client creation even when it is limited by ACLs
+   ([#722](https://github.com/Shopify/sarama/pull/722)).
+ - Several fixes to the consumer timer optimization code, regressions introduced
+   in v1.10.0. Go's timers are finicky
+   ([#730](https://github.com/Shopify/sarama/pull/730),
+    [#733](https://github.com/Shopify/sarama/pull/733),
+    [#734](https://github.com/Shopify/sarama/pull/734)).
+ - Handle consuming compressed relative offsets with Kafka 0.10
+   ([#735](https://github.com/Shopify/sarama/pull/735)).
+
+#### Version 1.10.0 (2016-08-02)
+
+_Important:_ As of Sarama 1.10 it is necessary to tell Sarama the version of
+Kafka you are running against (via the `config.Version` value) in order to use
+features that may not be compatible with old Kafka versions. If you don't
+specify this value it will default to 0.8.2 (the minimum supported), and trying
+to use more recent features (like the offset manager) will fail with an error.
+
+_Also:_ The offset-manager's behaviour has been changed to match the upstream
+java consumer (see [#705](https://github.com/Shopify/sarama/pull/705) and
+[#713](https://github.com/Shopify/sarama/pull/713)). If you use the
+offset-manager, please ensure that you are committing one *greater* than the
+last consumed message offset or else you may end up consuming duplicate
+messages.
+
+New Features:
+ - Support for Kafka 0.10
+   ([#672](https://github.com/Shopify/sarama/pull/672),
+    [#678](https://github.com/Shopify/sarama/pull/678),
+    [#681](https://github.com/Shopify/sarama/pull/681), and others).
+ - Support for configuring the target Kafka version
+   ([#676](https://github.com/Shopify/sarama/pull/676)).
+ - Batch producing support in the SyncProducer
+   ([#677](https://github.com/Shopify/sarama/pull/677)).
+ - Extend producer mock to allow setting expectations on message contents
+   ([#667](https://github.com/Shopify/sarama/pull/667)).
+
+Improvements:
+ - Support `nil` compressed messages for deleting in compacted topics
+   ([#634](https://github.com/Shopify/sarama/pull/634)).
+ - Pre-allocate decoding errors, greatly reducing heap usage and GC time against
+   misbehaving brokers ([#690](https://github.com/Shopify/sarama/pull/690)).
+ - Re-use consumer expiry timers, removing one allocation per consumed message
+   ([#707](https://github.com/Shopify/sarama/pull/707)).
+
+Bug Fixes:
+ - Actually default the client ID to "sarama" like we say we do
+   ([#664](https://github.com/Shopify/sarama/pull/664)).
+ - Fix a rare issue where `Client.Leader` could return the wrong error
+   ([#685](https://github.com/Shopify/sarama/pull/685)).
+ - Fix a possible tight loop in the consumer
+   ([#693](https://github.com/Shopify/sarama/pull/693)).
+ - Match upstream's offset-tracking behaviour
+   ([#705](https://github.com/Shopify/sarama/pull/705)).
+ - Report UnknownTopicOrPartition errors from the offset manager
+   ([#706](https://github.com/Shopify/sarama/pull/706)).
+ - Fix possible negative partition value from the HashPartitioner
+   ([#709](https://github.com/Shopify/sarama/pull/709)).
+
+#### Version 1.9.0 (2016-05-16)
+
+New Features:
+ - Add support for custom offset manager retention durations
+   ([#602](https://github.com/Shopify/sarama/pull/602)).
+ - Publish low-level mocks to enable testing of third-party producer/consumer
+   implementations ([#570](https://github.com/Shopify/sarama/pull/570)).
+ - Declare support for Golang 1.6
+   ([#611](https://github.com/Shopify/sarama/pull/611)).
+ - Support for SASL plain-text auth
+   ([#648](https://github.com/Shopify/sarama/pull/648)).
+
+Improvements:
+ - Simplified broker locking scheme slightly
+   ([#604](https://github.com/Shopify/sarama/pull/604)).
+ - Documentation cleanup
+   ([#605](https://github.com/Shopify/sarama/pull/605),
+    [#621](https://github.com/Shopify/sarama/pull/621),
+    [#654](https://github.com/Shopify/sarama/pull/654)).
+
+Bug Fixes:
+ - Fix race condition shutting down the OffsetManager
+   ([#658](https://github.com/Shopify/sarama/pull/658)).
+
+#### Version 1.8.0 (2016-02-01)
+
+New Features:
+ - Full support for Kafka 0.9:
+   - All protocol messages and fields
+   ([#586](https://github.com/Shopify/sarama/pull/586),
+   [#588](https://github.com/Shopify/sarama/pull/588),
+   [#590](https://github.com/Shopify/sarama/pull/590)).
+   - Verified that TLS support works
+   ([#581](https://github.com/Shopify/sarama/pull/581)).
+   - Fixed the OffsetManager compatibility
+   ([#585](https://github.com/Shopify/sarama/pull/585)).
+
+Improvements:
+ - Optimize for fewer system calls when reading from the network
+   ([#584](https://github.com/Shopify/sarama/pull/584)).
+ - Automatically retry `InvalidMessage` errors to match upstream behaviour
+   ([#589](https://github.com/Shopify/sarama/pull/589)).
+
+#### Version 1.7.0 (2015-12-11)
+
+New Features:
+ - Preliminary support for Kafka 0.9
+   ([#572](https://github.com/Shopify/sarama/pull/572)). This comes with several
+   caveats:
+   - Protocol-layer support is mostly in place
+     ([#577](https://github.com/Shopify/sarama/pull/577)), however Kafka 0.9
+     renamed some messages and fields, which we did not in order to preserve API
+     compatibility.
+   - The producer and consumer work against 0.9, but the offset manager does
+     not ([#573](https://github.com/Shopify/sarama/pull/573)).
+   - TLS support may or may not work
+     ([#581](https://github.com/Shopify/sarama/pull/581)).
+
+Improvements:
+ - Don't wait for request timeouts on dead brokers, greatly speeding recovery
+   when the TCP connection is left hanging
+   ([#548](https://github.com/Shopify/sarama/pull/548)).
+ - Refactored part of the producer. The new version provides a much more elegant
+   solution to [#449](https://github.com/Shopify/sarama/pull/449). It is also
+   slightly more efficient, and much more precise in calculating batch sizes
+   when compression is used
+   ([#549](https://github.com/Shopify/sarama/pull/549),
+   [#550](https://github.com/Shopify/sarama/pull/550),
+   [#551](https://github.com/Shopify/sarama/pull/551)).
+
+Bug Fixes:
+ - Fix race condition in consumer test mock
+   ([#553](https://github.com/Shopify/sarama/pull/553)).
+
+#### Version 1.6.1 (2015-09-25)
+
+Bug Fixes:
+ - Fix panic that could occur if a user-supplied message value failed to encode
+   ([#449](https://github.com/Shopify/sarama/pull/449)).
+
+#### Version 1.6.0 (2015-09-04)
+
+New Features:
+ - Implementation of a consumer offset manager using the APIs introduced in
+   Kafka 0.8.2. The API is designed mainly for integration into a future
+   high-level consumer, not for direct use, although it is *possible* to use it
+   directly.
+   ([#461](https://github.com/Shopify/sarama/pull/461)).
+
+Improvements:
+ - CRC32 calculation is much faster on machines with SSE4.2 instructions,
+   removing a major hotspot from most profiles
+   ([#255](https://github.com/Shopify/sarama/pull/255)).
+
+Bug Fixes:
+ - Make protocol decoding more robust against some malformed packets generated
+   by go-fuzz ([#523](https://github.com/Shopify/sarama/pull/523),
+   [#525](https://github.com/Shopify/sarama/pull/525)) or found in other ways
+   ([#528](https://github.com/Shopify/sarama/pull/528)).
+ - Fix a potential race condition panic in the consumer on shutdown
+   ([#529](https://github.com/Shopify/sarama/pull/529)).
+
+#### Version 1.5.0 (2015-08-17)
+
+New Features:
+ - TLS-encrypted network connections are now supported. This feature is subject
+   to change when Kafka releases built-in TLS support, but for now this is
+   enough to work with TLS-terminating proxies
+   ([#154](https://github.com/Shopify/sarama/pull/154)).
+
+Improvements:
+ - The consumer will not block if a single partition is not drained by the user;
+   all other partitions will continue to consume normally
+   ([#485](https://github.com/Shopify/sarama/pull/485)).
+ - Formatting of error strings has been much improved
+   ([#495](https://github.com/Shopify/sarama/pull/495)).
+ - Internal refactoring of the producer for code cleanliness and to enable
+   future work ([#300](https://github.com/Shopify/sarama/pull/300)).
+
+Bug Fixes:
+ - Fix a potential deadlock in the consumer on shutdown
+   ([#475](https://github.com/Shopify/sarama/pull/475)).
+
+#### Version 1.4.3 (2015-07-21)
+
+Bug Fixes:
+ - Don't include the partitioner in the producer's "fetch partitions"
+   circuit-breaker ([#466](https://github.com/Shopify/sarama/pull/466)).
+ - Don't retry messages until the broker is closed when abandoning a broker in
+   the producer ([#468](https://github.com/Shopify/sarama/pull/468)).
+ - Update the import path for snappy-go, it has moved again and the API has
+   changed slightly ([#486](https://github.com/Shopify/sarama/pull/486)).
+
+#### Version 1.4.2 (2015-05-27)
+
+Bug Fixes:
+ - Update the import path for snappy-go, it has moved from google code to github
+   ([#456](https://github.com/Shopify/sarama/pull/456)).
+
+#### Version 1.4.1 (2015-05-25)
+
+Improvements:
+ - Optimizations when decoding snappy messages, thanks to John Potocny
+   ([#446](https://github.com/Shopify/sarama/pull/446)).
+
+Bug Fixes:
+ - Fix hypothetical race conditions on producer shutdown
+   ([#450](https://github.com/Shopify/sarama/pull/450),
+   [#451](https://github.com/Shopify/sarama/pull/451)).
+
+#### Version 1.4.0 (2015-05-01)
+
+New Features:
+ - The consumer now implements `Topics()` and `Partitions()` methods to enable
+   users to dynamically choose what topics/partitions to consume without
+   instantiating a full client
+   ([#431](https://github.com/Shopify/sarama/pull/431)).
+ - The partition-consumer now exposes the high water mark offset value returned
+   by the broker via the `HighWaterMarkOffset()` method ([#339](https://github.com/Shopify/sarama/pull/339)).
+ - Added a `kafka-console-consumer` tool capable of handling multiple
+   partitions, and deprecated the now-obsolete `kafka-console-partitionConsumer`
+   ([#439](https://github.com/Shopify/sarama/pull/439),
+   [#442](https://github.com/Shopify/sarama/pull/442)).
+
+Improvements:
+ - The producer's logging during retry scenarios is more consistent, more
+   useful, and slightly less verbose
+   ([#429](https://github.com/Shopify/sarama/pull/429)).
+ - The client now shuffles its initial list of seed brokers in order to prevent
+   thundering herd on the first broker in the list
+   ([#441](https://github.com/Shopify/sarama/pull/441)).
+
+Bug Fixes:
+ - The producer now correctly manages its state if retries occur when it is
+   shutting down, fixing several instances of confusing behaviour and at least
+   one potential deadlock ([#419](https://github.com/Shopify/sarama/pull/419)).
+ - The consumer now handles messages for different partitions asynchronously,
+   making it much more resilient to specific user code ordering
+   ([#325](https://github.com/Shopify/sarama/pull/325)).
+
+#### Version 1.3.0 (2015-04-16)
+
+New Features:
+ - The client now tracks consumer group coordinators using
+   ConsumerMetadataRequests similar to how it tracks partition leadership using
+   regular MetadataRequests ([#411](https://github.com/Shopify/sarama/pull/411)).
+   This adds two methods to the client API:
+   - `Coordinator(consumerGroup string) (*Broker, error)`
+   - `RefreshCoordinator(consumerGroup string) error`
+
+Improvements:
+ - ConsumerMetadataResponses now automatically create a Broker object out of the
+   ID/address/port combination for the Coordinator; accessing the fields
+   individually has been deprecated
+   ([#413](https://github.com/Shopify/sarama/pull/413)).
+ - Much improved handling of `OffsetOutOfRange` errors in the consumer.
+   Consumers will fail to start if the provided offset is out of range
+   ([#418](https://github.com/Shopify/sarama/pull/418))
+   and they will automatically shut down if the offset falls out of range
+   ([#424](https://github.com/Shopify/sarama/pull/424)).
+ - Small performance improvement in encoding and decoding protocol messages
+   ([#427](https://github.com/Shopify/sarama/pull/427)).
+
+Bug Fixes:
+ - Fix a rare race condition in the client's background metadata refresher if
+   it happens to be activated while the client is being closed
+   ([#422](https://github.com/Shopify/sarama/pull/422)).
+
+#### Version 1.2.0 (2015-04-07)
+
+Improvements:
+ - The producer's behaviour when `Flush.Frequency` is set is now more intuitive
+   ([#389](https://github.com/Shopify/sarama/pull/389)).
+ - The producer is now somewhat more memory-efficient during and after retrying
+   messages due to an improved queue implementation
+   ([#396](https://github.com/Shopify/sarama/pull/396)).
+ - The consumer produces much more useful logging output when leadership
+   changes ([#385](https://github.com/Shopify/sarama/pull/385)).
+ - The client's `GetOffset` method will now automatically refresh metadata and
+   retry once in the event of stale information or similar
+   ([#394](https://github.com/Shopify/sarama/pull/394)).
+ - Broker connections now have support for using TCP keepalives
+   ([#407](https://github.com/Shopify/sarama/issues/407)).
+
+Bug Fixes:
+ - The OffsetCommitRequest message now correctly implements all three possible
+   API versions ([#390](https://github.com/Shopify/sarama/pull/390),
+   [#400](https://github.com/Shopify/sarama/pull/400)).
+
+#### Version 1.1.0 (2015-03-20)
+
+Improvements:
+ - Wrap the producer's partitioner call in a circuit-breaker so that repeatedly
+   broken topics don't choke throughput
+   ([#373](https://github.com/Shopify/sarama/pull/373)).
+
+Bug Fixes:
+ - Fix the producer's internal reference counting in certain unusual scenarios
+   ([#367](https://github.com/Shopify/sarama/pull/367)).
+ - Fix the consumer's internal reference counting in certain unusual scenarios
+   ([#369](https://github.com/Shopify/sarama/pull/369)).
+ - Fix a condition where the producer's internal control messages could have
+   gotten stuck ([#368](https://github.com/Shopify/sarama/pull/368)).
+ - Fix an issue where invalid partition lists would be cached when asking for
+   metadata for a non-existant topic ([#372](https://github.com/Shopify/sarama/pull/372)).
+
+
+#### Version 1.0.0 (2015-03-17)
+
+Version 1.0.0 is the first tagged version, and is almost a complete rewrite. The primary differences with previous untagged versions are:
+
+- The producer has been rewritten; there is now a `SyncProducer` with a blocking API, and an `AsyncProducer` that is non-blocking.
+- The consumer has been rewritten to only open one connection per broker instead of one connection per partition.
+- The main types of Sarama are now interfaces to make depedency injection easy; mock implementations for `Consumer`, `SyncProducer` and `AsyncProducer` are provided in the `github.com/Shopify/sarama/mocks` package.
+- For most uses cases, it is no longer necessary to open a `Client`; this will be done for you.
+- All the configuration values have been unified in the `Config` struct.
+- Much improved test suite.
diff --git a/vendor/github.com/Shopify/sarama/LICENSE b/vendor/github.com/Shopify/sarama/LICENSE
new file mode 100644
index 0000000..d2bf435
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/LICENSE
@@ -0,0 +1,20 @@
+Copyright (c) 2013 Shopify
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/vendor/github.com/Shopify/sarama/Makefile b/vendor/github.com/Shopify/sarama/Makefile
new file mode 100644
index 0000000..09f743c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/Makefile
@@ -0,0 +1,32 @@
+export GO111MODULE=on
+
+default: fmt vet errcheck test
+
+# Taken from https://github.com/codecov/example-go#caveat-multiple-files
+test:
+	echo "" > coverage.txt
+	for d in `go list ./... | grep -v vendor`; do \
+		go test -p 1 -v -timeout 240s -race -coverprofile=profile.out -covermode=atomic $$d || exit 1; \
+		if [ -f profile.out ]; then \
+			cat profile.out >> coverage.txt; \
+			rm profile.out; \
+		fi \
+	done
+
+vet:
+	go vet ./...
+
+# See https://github.com/kisielk/errcheck/pull/141 for details on ignorepkg
+errcheck:
+	errcheck -ignorepkg fmt github.com/Shopify/sarama/...
+
+fmt:
+	@if [ -n "$$(go fmt ./...)" ]; then echo 'Please run go fmt on your code.' && exit 1; fi
+
+install_dependencies: install_errcheck get
+
+install_errcheck:
+	go get github.com/kisielk/errcheck
+
+get:
+	go get -t
diff --git a/vendor/github.com/Shopify/sarama/README.md b/vendor/github.com/Shopify/sarama/README.md
new file mode 100644
index 0000000..f241b89
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/README.md
@@ -0,0 +1,39 @@
+sarama
+======
+
+[![GoDoc](https://godoc.org/github.com/Shopify/sarama?status.svg)](https://godoc.org/github.com/Shopify/sarama)
+[![Build Status](https://travis-ci.org/Shopify/sarama.svg?branch=master)](https://travis-ci.org/Shopify/sarama)
+[![Coverage](https://codecov.io/gh/Shopify/sarama/branch/master/graph/badge.svg)](https://codecov.io/gh/Shopify/sarama)
+
+Sarama is an MIT-licensed Go client library for [Apache Kafka](https://kafka.apache.org/) version 0.8 (and later).
+
+### Getting started
+
+- API documentation and examples are available via [godoc](https://godoc.org/github.com/Shopify/sarama).
+- Mocks for testing are available in the [mocks](./mocks) subpackage.
+- The [examples](./examples) directory contains more elaborate example applications.
+- The [tools](./tools) directory contains command line tools that can be useful for testing, diagnostics, and instrumentation.
+
+You might also want to look at the [Frequently Asked Questions](https://github.com/Shopify/sarama/wiki/Frequently-Asked-Questions).
+
+### Compatibility and API stability
+
+Sarama provides a "2 releases + 2 months" compatibility guarantee: we support
+the two latest stable releases of Kafka and Go, and we provide a two month
+grace period for older releases. This means we currently officially support
+Go 1.8 through 1.11, and Kafka 1.0 through 2.0, although older releases are
+still likely to work.
+
+Sarama follows semantic versioning and provides API stability via the gopkg.in service.
+You can import a version with a guaranteed stable API via http://gopkg.in/Shopify/sarama.v1.
+A changelog is available [here](CHANGELOG.md).
+
+### Contributing
+
+* Get started by checking our [contribution guidelines](https://github.com/Shopify/sarama/blob/master/.github/CONTRIBUTING.md).
+* Read the [Sarama wiki](https://github.com/Shopify/sarama/wiki) for more
+  technical and design details.
+* The [Kafka Protocol Specification](https://cwiki.apache.org/confluence/display/KAFKA/A+Guide+To+The+Kafka+Protocol)
+  contains a wealth of useful information.
+* For more general issues, there is [a google group](https://groups.google.com/forum/#!forum/kafka-clients) for Kafka client developers.
+* If you have any questions, just ask!
diff --git a/vendor/github.com/Shopify/sarama/Vagrantfile b/vendor/github.com/Shopify/sarama/Vagrantfile
new file mode 100644
index 0000000..f4b848a
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/Vagrantfile
@@ -0,0 +1,20 @@
+# -*- mode: ruby -*-
+# vi: set ft=ruby :
+
+# Vagrantfile API/syntax version. Don't touch unless you know what you're doing!
+VAGRANTFILE_API_VERSION = "2"
+
+# We have 5 * 192MB ZK processes and 5 * 320MB Kafka processes => 2560MB
+MEMORY = 3072
+
+Vagrant.configure(VAGRANTFILE_API_VERSION) do |config|
+  config.vm.box = "ubuntu/trusty64"
+
+  config.vm.provision :shell, path: "vagrant/provision.sh"
+
+  config.vm.network "private_network", ip: "192.168.100.67"
+
+  config.vm.provider "virtualbox" do |v|
+    v.memory = MEMORY
+  end
+end
diff --git a/vendor/github.com/Shopify/sarama/acl_bindings.go b/vendor/github.com/Shopify/sarama/acl_bindings.go
new file mode 100644
index 0000000..b91c282
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_bindings.go
@@ -0,0 +1,135 @@
+package sarama
+
+type Resource struct {
+	ResourceType       AclResourceType
+	ResourceName       string
+	ResoucePatternType AclResourcePatternType
+}
+
+func (r *Resource) encode(pe packetEncoder, version int16) error {
+	pe.putInt8(int8(r.ResourceType))
+
+	if err := pe.putString(r.ResourceName); err != nil {
+		return err
+	}
+
+	if version == 1 {
+		if r.ResoucePatternType == AclPatternUnknown {
+			Logger.Print("Cannot encode an unknown resource pattern type, using Literal instead")
+			r.ResoucePatternType = AclPatternLiteral
+		}
+		pe.putInt8(int8(r.ResoucePatternType))
+	}
+
+	return nil
+}
+
+func (r *Resource) decode(pd packetDecoder, version int16) (err error) {
+	resourceType, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	r.ResourceType = AclResourceType(resourceType)
+
+	if r.ResourceName, err = pd.getString(); err != nil {
+		return err
+	}
+	if version == 1 {
+		pattern, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+		r.ResoucePatternType = AclResourcePatternType(pattern)
+	}
+
+	return nil
+}
+
+type Acl struct {
+	Principal      string
+	Host           string
+	Operation      AclOperation
+	PermissionType AclPermissionType
+}
+
+func (a *Acl) encode(pe packetEncoder) error {
+	if err := pe.putString(a.Principal); err != nil {
+		return err
+	}
+
+	if err := pe.putString(a.Host); err != nil {
+		return err
+	}
+
+	pe.putInt8(int8(a.Operation))
+	pe.putInt8(int8(a.PermissionType))
+
+	return nil
+}
+
+func (a *Acl) decode(pd packetDecoder, version int16) (err error) {
+	if a.Principal, err = pd.getString(); err != nil {
+		return err
+	}
+
+	if a.Host, err = pd.getString(); err != nil {
+		return err
+	}
+
+	operation, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	a.Operation = AclOperation(operation)
+
+	permissionType, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	a.PermissionType = AclPermissionType(permissionType)
+
+	return nil
+}
+
+type ResourceAcls struct {
+	Resource
+	Acls []*Acl
+}
+
+func (r *ResourceAcls) encode(pe packetEncoder, version int16) error {
+	if err := r.Resource.encode(pe, version); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(r.Acls)); err != nil {
+		return err
+	}
+	for _, acl := range r.Acls {
+		if err := acl.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *ResourceAcls) decode(pd packetDecoder, version int16) error {
+	if err := r.Resource.decode(pd, version); err != nil {
+		return err
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Acls = make([]*Acl, n)
+	for i := 0; i < n; i++ {
+		r.Acls[i] = new(Acl)
+		if err := r.Acls[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_create_request.go b/vendor/github.com/Shopify/sarama/acl_create_request.go
new file mode 100644
index 0000000..d3d5ad8
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_create_request.go
@@ -0,0 +1,83 @@
+package sarama
+
+type CreateAclsRequest struct {
+	Version      int16
+	AclCreations []*AclCreation
+}
+
+func (c *CreateAclsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(c.AclCreations)); err != nil {
+		return err
+	}
+
+	for _, aclCreation := range c.AclCreations {
+		if err := aclCreation.encode(pe, c.Version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *CreateAclsRequest) decode(pd packetDecoder, version int16) (err error) {
+	c.Version = version
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	c.AclCreations = make([]*AclCreation, n)
+
+	for i := 0; i < n; i++ {
+		c.AclCreations[i] = new(AclCreation)
+		if err := c.AclCreations[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *CreateAclsRequest) key() int16 {
+	return 30
+}
+
+func (d *CreateAclsRequest) version() int16 {
+	return d.Version
+}
+
+func (d *CreateAclsRequest) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
+
+type AclCreation struct {
+	Resource
+	Acl
+}
+
+func (a *AclCreation) encode(pe packetEncoder, version int16) error {
+	if err := a.Resource.encode(pe, version); err != nil {
+		return err
+	}
+	if err := a.Acl.encode(pe); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (a *AclCreation) decode(pd packetDecoder, version int16) (err error) {
+	if err := a.Resource.decode(pd, version); err != nil {
+		return err
+	}
+	if err := a.Acl.decode(pd, version); err != nil {
+		return err
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_create_response.go b/vendor/github.com/Shopify/sarama/acl_create_response.go
new file mode 100644
index 0000000..8a56f35
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_create_response.go
@@ -0,0 +1,88 @@
+package sarama
+
+import "time"
+
+type CreateAclsResponse struct {
+	ThrottleTime         time.Duration
+	AclCreationResponses []*AclCreationResponse
+}
+
+func (c *CreateAclsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(c.ThrottleTime / time.Millisecond))
+
+	if err := pe.putArrayLength(len(c.AclCreationResponses)); err != nil {
+		return err
+	}
+
+	for _, aclCreationResponse := range c.AclCreationResponses {
+		if err := aclCreationResponse.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *CreateAclsResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	c.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	c.AclCreationResponses = make([]*AclCreationResponse, n)
+	for i := 0; i < n; i++ {
+		c.AclCreationResponses[i] = new(AclCreationResponse)
+		if err := c.AclCreationResponses[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *CreateAclsResponse) key() int16 {
+	return 30
+}
+
+func (d *CreateAclsResponse) version() int16 {
+	return 0
+}
+
+func (d *CreateAclsResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type AclCreationResponse struct {
+	Err    KError
+	ErrMsg *string
+}
+
+func (a *AclCreationResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(a.Err))
+
+	if err := pe.putNullableString(a.ErrMsg); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (a *AclCreationResponse) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	a.Err = KError(kerr)
+
+	if a.ErrMsg, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_delete_request.go b/vendor/github.com/Shopify/sarama/acl_delete_request.go
new file mode 100644
index 0000000..5e94ad7
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_delete_request.go
@@ -0,0 +1,57 @@
+package sarama
+
+type DeleteAclsRequest struct {
+	Version int
+	Filters []*AclFilter
+}
+
+func (d *DeleteAclsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(d.Filters)); err != nil {
+		return err
+	}
+
+	for _, filter := range d.Filters {
+		filter.Version = d.Version
+		if err := filter.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *DeleteAclsRequest) decode(pd packetDecoder, version int16) (err error) {
+	d.Version = int(version)
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	d.Filters = make([]*AclFilter, n)
+	for i := 0; i < n; i++ {
+		d.Filters[i] = new(AclFilter)
+		d.Filters[i].Version = int(version)
+		if err := d.Filters[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *DeleteAclsRequest) key() int16 {
+	return 31
+}
+
+func (d *DeleteAclsRequest) version() int16 {
+	return int16(d.Version)
+}
+
+func (d *DeleteAclsRequest) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_delete_response.go b/vendor/github.com/Shopify/sarama/acl_delete_response.go
new file mode 100644
index 0000000..a885fe5
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_delete_response.go
@@ -0,0 +1,156 @@
+package sarama
+
+import "time"
+
+type DeleteAclsResponse struct {
+	Version         int16
+	ThrottleTime    time.Duration
+	FilterResponses []*FilterResponse
+}
+
+func (a *DeleteAclsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(a.ThrottleTime / time.Millisecond))
+
+	if err := pe.putArrayLength(len(a.FilterResponses)); err != nil {
+		return err
+	}
+
+	for _, filterResponse := range a.FilterResponses {
+		if err := filterResponse.encode(pe, a.Version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (a *DeleteAclsResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	a.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	a.FilterResponses = make([]*FilterResponse, n)
+
+	for i := 0; i < n; i++ {
+		a.FilterResponses[i] = new(FilterResponse)
+		if err := a.FilterResponses[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *DeleteAclsResponse) key() int16 {
+	return 31
+}
+
+func (d *DeleteAclsResponse) version() int16 {
+	return int16(d.Version)
+}
+
+func (d *DeleteAclsResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type FilterResponse struct {
+	Err          KError
+	ErrMsg       *string
+	MatchingAcls []*MatchingAcl
+}
+
+func (f *FilterResponse) encode(pe packetEncoder, version int16) error {
+	pe.putInt16(int16(f.Err))
+	if err := pe.putNullableString(f.ErrMsg); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(f.MatchingAcls)); err != nil {
+		return err
+	}
+	for _, matchingAcl := range f.MatchingAcls {
+		if err := matchingAcl.encode(pe, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (f *FilterResponse) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	f.Err = KError(kerr)
+
+	if f.ErrMsg, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	f.MatchingAcls = make([]*MatchingAcl, n)
+	for i := 0; i < n; i++ {
+		f.MatchingAcls[i] = new(MatchingAcl)
+		if err := f.MatchingAcls[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+type MatchingAcl struct {
+	Err    KError
+	ErrMsg *string
+	Resource
+	Acl
+}
+
+func (m *MatchingAcl) encode(pe packetEncoder, version int16) error {
+	pe.putInt16(int16(m.Err))
+	if err := pe.putNullableString(m.ErrMsg); err != nil {
+		return err
+	}
+
+	if err := m.Resource.encode(pe, version); err != nil {
+		return err
+	}
+
+	if err := m.Acl.encode(pe); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m *MatchingAcl) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	m.Err = KError(kerr)
+
+	if m.ErrMsg, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	if err := m.Resource.decode(pd, version); err != nil {
+		return err
+	}
+
+	if err := m.Acl.decode(pd, version); err != nil {
+		return err
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_describe_request.go b/vendor/github.com/Shopify/sarama/acl_describe_request.go
new file mode 100644
index 0000000..3c95320
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_describe_request.go
@@ -0,0 +1,34 @@
+package sarama
+
+type DescribeAclsRequest struct {
+	Version int
+	AclFilter
+}
+
+func (d *DescribeAclsRequest) encode(pe packetEncoder) error {
+	d.AclFilter.Version = d.Version
+	return d.AclFilter.encode(pe)
+}
+
+func (d *DescribeAclsRequest) decode(pd packetDecoder, version int16) (err error) {
+	d.Version = int(version)
+	d.AclFilter.Version = int(version)
+	return d.AclFilter.decode(pd, version)
+}
+
+func (d *DescribeAclsRequest) key() int16 {
+	return 29
+}
+
+func (d *DescribeAclsRequest) version() int16 {
+	return int16(d.Version)
+}
+
+func (d *DescribeAclsRequest) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_describe_response.go b/vendor/github.com/Shopify/sarama/acl_describe_response.go
new file mode 100644
index 0000000..db3a88c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_describe_response.go
@@ -0,0 +1,86 @@
+package sarama
+
+import "time"
+
+type DescribeAclsResponse struct {
+	Version      int16
+	ThrottleTime time.Duration
+	Err          KError
+	ErrMsg       *string
+	ResourceAcls []*ResourceAcls
+}
+
+func (d *DescribeAclsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(d.ThrottleTime / time.Millisecond))
+	pe.putInt16(int16(d.Err))
+
+	if err := pe.putNullableString(d.ErrMsg); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(d.ResourceAcls)); err != nil {
+		return err
+	}
+
+	for _, resourceAcl := range d.ResourceAcls {
+		if err := resourceAcl.encode(pe, d.Version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *DescribeAclsResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	d.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	d.Err = KError(kerr)
+
+	errmsg, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	if errmsg != "" {
+		d.ErrMsg = &errmsg
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	d.ResourceAcls = make([]*ResourceAcls, n)
+
+	for i := 0; i < n; i++ {
+		d.ResourceAcls[i] = new(ResourceAcls)
+		if err := d.ResourceAcls[i].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (d *DescribeAclsResponse) key() int16 {
+	return 29
+}
+
+func (d *DescribeAclsResponse) version() int16 {
+	return int16(d.Version)
+}
+
+func (d *DescribeAclsResponse) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_filter.go b/vendor/github.com/Shopify/sarama/acl_filter.go
new file mode 100644
index 0000000..fad5558
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_filter.go
@@ -0,0 +1,78 @@
+package sarama
+
+type AclFilter struct {
+	Version                   int
+	ResourceType              AclResourceType
+	ResourceName              *string
+	ResourcePatternTypeFilter AclResourcePatternType
+	Principal                 *string
+	Host                      *string
+	Operation                 AclOperation
+	PermissionType            AclPermissionType
+}
+
+func (a *AclFilter) encode(pe packetEncoder) error {
+	pe.putInt8(int8(a.ResourceType))
+	if err := pe.putNullableString(a.ResourceName); err != nil {
+		return err
+	}
+
+	if a.Version == 1 {
+		pe.putInt8(int8(a.ResourcePatternTypeFilter))
+	}
+
+	if err := pe.putNullableString(a.Principal); err != nil {
+		return err
+	}
+	if err := pe.putNullableString(a.Host); err != nil {
+		return err
+	}
+	pe.putInt8(int8(a.Operation))
+	pe.putInt8(int8(a.PermissionType))
+
+	return nil
+}
+
+func (a *AclFilter) decode(pd packetDecoder, version int16) (err error) {
+	resourceType, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	a.ResourceType = AclResourceType(resourceType)
+
+	if a.ResourceName, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	if a.Version == 1 {
+		pattern, err := pd.getInt8()
+
+		if err != nil {
+			return err
+		}
+
+		a.ResourcePatternTypeFilter = AclResourcePatternType(pattern)
+	}
+
+	if a.Principal, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	if a.Host, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	operation, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	a.Operation = AclOperation(operation)
+
+	permissionType, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	a.PermissionType = AclPermissionType(permissionType)
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/acl_types.go b/vendor/github.com/Shopify/sarama/acl_types.go
new file mode 100644
index 0000000..72b7985
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/acl_types.go
@@ -0,0 +1,54 @@
+package sarama
+
+type AclOperation int
+
+// ref: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/acl/AclOperation.java
+const (
+	AclOperationUnknown         AclOperation = 0
+	AclOperationAny             AclOperation = 1
+	AclOperationAll             AclOperation = 2
+	AclOperationRead            AclOperation = 3
+	AclOperationWrite           AclOperation = 4
+	AclOperationCreate          AclOperation = 5
+	AclOperationDelete          AclOperation = 6
+	AclOperationAlter           AclOperation = 7
+	AclOperationDescribe        AclOperation = 8
+	AclOperationClusterAction   AclOperation = 9
+	AclOperationDescribeConfigs AclOperation = 10
+	AclOperationAlterConfigs    AclOperation = 11
+	AclOperationIdempotentWrite AclOperation = 12
+)
+
+type AclPermissionType int
+
+// ref: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/acl/AclPermissionType.java
+const (
+	AclPermissionUnknown AclPermissionType = 0
+	AclPermissionAny     AclPermissionType = 1
+	AclPermissionDeny    AclPermissionType = 2
+	AclPermissionAllow   AclPermissionType = 3
+)
+
+type AclResourceType int
+
+// ref: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/resource/ResourceType.java
+const (
+	AclResourceUnknown         AclResourceType = 0
+	AclResourceAny             AclResourceType = 1
+	AclResourceTopic           AclResourceType = 2
+	AclResourceGroup           AclResourceType = 3
+	AclResourceCluster         AclResourceType = 4
+	AclResourceTransactionalID AclResourceType = 5
+)
+
+type AclResourcePatternType int
+
+// ref: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/resource/PatternType.java
+
+const (
+	AclPatternUnknown AclResourcePatternType = iota
+	AclPatternAny
+	AclPatternMatch
+	AclPatternLiteral
+	AclPatternPrefixed
+)
diff --git a/vendor/github.com/Shopify/sarama/add_offsets_to_txn_request.go b/vendor/github.com/Shopify/sarama/add_offsets_to_txn_request.go
new file mode 100644
index 0000000..6da166c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/add_offsets_to_txn_request.go
@@ -0,0 +1,52 @@
+package sarama
+
+type AddOffsetsToTxnRequest struct {
+	TransactionalID string
+	ProducerID      int64
+	ProducerEpoch   int16
+	GroupID         string
+}
+
+func (a *AddOffsetsToTxnRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(a.TransactionalID); err != nil {
+		return err
+	}
+
+	pe.putInt64(a.ProducerID)
+
+	pe.putInt16(a.ProducerEpoch)
+
+	if err := pe.putString(a.GroupID); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (a *AddOffsetsToTxnRequest) decode(pd packetDecoder, version int16) (err error) {
+	if a.TransactionalID, err = pd.getString(); err != nil {
+		return err
+	}
+	if a.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if a.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+	if a.GroupID, err = pd.getString(); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *AddOffsetsToTxnRequest) key() int16 {
+	return 25
+}
+
+func (a *AddOffsetsToTxnRequest) version() int16 {
+	return 0
+}
+
+func (a *AddOffsetsToTxnRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/add_offsets_to_txn_response.go b/vendor/github.com/Shopify/sarama/add_offsets_to_txn_response.go
new file mode 100644
index 0000000..3a46151
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/add_offsets_to_txn_response.go
@@ -0,0 +1,44 @@
+package sarama
+
+import (
+	"time"
+)
+
+type AddOffsetsToTxnResponse struct {
+	ThrottleTime time.Duration
+	Err          KError
+}
+
+func (a *AddOffsetsToTxnResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(a.ThrottleTime / time.Millisecond))
+	pe.putInt16(int16(a.Err))
+	return nil
+}
+
+func (a *AddOffsetsToTxnResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	a.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	a.Err = KError(kerr)
+
+	return nil
+}
+
+func (a *AddOffsetsToTxnResponse) key() int16 {
+	return 25
+}
+
+func (a *AddOffsetsToTxnResponse) version() int16 {
+	return 0
+}
+
+func (a *AddOffsetsToTxnResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/add_partitions_to_txn_request.go b/vendor/github.com/Shopify/sarama/add_partitions_to_txn_request.go
new file mode 100644
index 0000000..a8a5922
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/add_partitions_to_txn_request.go
@@ -0,0 +1,76 @@
+package sarama
+
+type AddPartitionsToTxnRequest struct {
+	TransactionalID string
+	ProducerID      int64
+	ProducerEpoch   int16
+	TopicPartitions map[string][]int32
+}
+
+func (a *AddPartitionsToTxnRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(a.TransactionalID); err != nil {
+		return err
+	}
+	pe.putInt64(a.ProducerID)
+	pe.putInt16(a.ProducerEpoch)
+
+	if err := pe.putArrayLength(len(a.TopicPartitions)); err != nil {
+		return err
+	}
+	for topic, partitions := range a.TopicPartitions {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putInt32Array(partitions); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (a *AddPartitionsToTxnRequest) decode(pd packetDecoder, version int16) (err error) {
+	if a.TransactionalID, err = pd.getString(); err != nil {
+		return err
+	}
+	if a.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if a.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	a.TopicPartitions = make(map[string][]int32)
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		partitions, err := pd.getInt32Array()
+		if err != nil {
+			return err
+		}
+
+		a.TopicPartitions[topic] = partitions
+	}
+
+	return nil
+}
+
+func (a *AddPartitionsToTxnRequest) key() int16 {
+	return 24
+}
+
+func (a *AddPartitionsToTxnRequest) version() int16 {
+	return 0
+}
+
+func (a *AddPartitionsToTxnRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/add_partitions_to_txn_response.go b/vendor/github.com/Shopify/sarama/add_partitions_to_txn_response.go
new file mode 100644
index 0000000..581c556
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/add_partitions_to_txn_response.go
@@ -0,0 +1,108 @@
+package sarama
+
+import (
+	"time"
+)
+
+type AddPartitionsToTxnResponse struct {
+	ThrottleTime time.Duration
+	Errors       map[string][]*PartitionError
+}
+
+func (a *AddPartitionsToTxnResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(a.ThrottleTime / time.Millisecond))
+	if err := pe.putArrayLength(len(a.Errors)); err != nil {
+		return err
+	}
+
+	for topic, e := range a.Errors {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(e)); err != nil {
+			return err
+		}
+		for _, partitionError := range e {
+			if err := partitionError.encode(pe); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (a *AddPartitionsToTxnResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	a.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	a.Errors = make(map[string][]*PartitionError)
+
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		m, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		a.Errors[topic] = make([]*PartitionError, m)
+
+		for j := 0; j < m; j++ {
+			a.Errors[topic][j] = new(PartitionError)
+			if err := a.Errors[topic][j].decode(pd, version); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (a *AddPartitionsToTxnResponse) key() int16 {
+	return 24
+}
+
+func (a *AddPartitionsToTxnResponse) version() int16 {
+	return 0
+}
+
+func (a *AddPartitionsToTxnResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type PartitionError struct {
+	Partition int32
+	Err       KError
+}
+
+func (p *PartitionError) encode(pe packetEncoder) error {
+	pe.putInt32(p.Partition)
+	pe.putInt16(int16(p.Err))
+	return nil
+}
+
+func (p *PartitionError) decode(pd packetDecoder, version int16) (err error) {
+	if p.Partition, err = pd.getInt32(); err != nil {
+		return err
+	}
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	p.Err = KError(kerr)
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/admin.go b/vendor/github.com/Shopify/sarama/admin.go
new file mode 100644
index 0000000..18b055a
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/admin.go
@@ -0,0 +1,614 @@
+package sarama
+
+import (
+	"errors"
+	"math/rand"
+	"sync"
+)
+
+// ClusterAdmin is the administrative client for Kafka, which supports managing and inspecting topics,
+// brokers, configurations and ACLs. The minimum broker version required is 0.10.0.0.
+// Methods with stricter requirements will specify the minimum broker version required.
+// You MUST call Close() on a client to avoid leaks
+type ClusterAdmin interface {
+	// Creates a new topic. This operation is supported by brokers with version 0.10.1.0 or higher.
+	// It may take several seconds after CreateTopic returns success for all the brokers
+	// to become aware that the topic has been created. During this time, listTopics
+	// may not return information about the new topic.The validateOnly option is supported from version 0.10.2.0.
+	CreateTopic(topic string, detail *TopicDetail, validateOnly bool) error
+
+	// List the topics available in the cluster with the default options.
+	ListTopics() (map[string]TopicDetail, error)
+
+	// Describe some topics in the cluster
+	DescribeTopics(topics []string) (metadata []*TopicMetadata, err error)
+
+	// Delete a topic. It may take several seconds after the DeleteTopic to returns success
+	// and for all the brokers to become aware that the topics are gone.
+	// During this time, listTopics  may continue to return information about the deleted topic.
+	// If delete.topic.enable is false on the brokers, deleteTopic will mark
+	// the topic for deletion, but not actually delete them.
+	// This operation is supported by brokers with version 0.10.1.0 or higher.
+	DeleteTopic(topic string) error
+
+	// Increase the number of partitions of the topics  according to the corresponding values.
+	// If partitions are increased for a topic that has a key, the partition logic or ordering of
+	// the messages will be affected. It may take several seconds after this method returns
+	// success for all the brokers to become aware that the partitions have been created.
+	// During this time, ClusterAdmin#describeTopics may not return information about the
+	// new partitions. This operation is supported by brokers with version 1.0.0 or higher.
+	CreatePartitions(topic string, count int32, assignment [][]int32, validateOnly bool) error
+
+	// Delete records whose offset is smaller than the given offset of the corresponding partition.
+	// This operation is supported by brokers with version 0.11.0.0 or higher.
+	DeleteRecords(topic string, partitionOffsets map[int32]int64) error
+
+	// Get the configuration for the specified resources.
+	// The returned configuration includes default values and the Default is true
+	// can be used to distinguish them from user supplied values.
+	// Config entries where ReadOnly is true cannot be updated.
+	// The value of config entries where Sensitive is true is always nil so
+	// sensitive information is not disclosed.
+	// This operation is supported by brokers with version 0.11.0.0 or higher.
+	DescribeConfig(resource ConfigResource) ([]ConfigEntry, error)
+
+	// Update the configuration for the specified resources with the default options.
+	// This operation is supported by brokers with version 0.11.0.0 or higher.
+	// The resources with their configs (topic is the only resource type with configs
+	// that can be updated currently Updates are not transactional so they may succeed
+	// for some resources while fail for others. The configs for a particular resource are updated automatically.
+	AlterConfig(resourceType ConfigResourceType, name string, entries map[string]*string, validateOnly bool) error
+
+	// Creates access control lists (ACLs) which are bound to specific resources.
+	// This operation is not transactional so it may succeed for some ACLs while fail for others.
+	// If you attempt to add an ACL that duplicates an existing ACL, no error will be raised, but
+	// no changes will be made. This operation is supported by brokers with version 0.11.0.0 or higher.
+	CreateACL(resource Resource, acl Acl) error
+
+	// Lists access control lists (ACLs) according to the supplied filter.
+	// it may take some time for changes made by createAcls or deleteAcls to be reflected in the output of ListAcls
+	// This operation is supported by brokers with version 0.11.0.0 or higher.
+	ListAcls(filter AclFilter) ([]ResourceAcls, error)
+
+	// Deletes access control lists (ACLs) according to the supplied filters.
+	// This operation is not transactional so it may succeed for some ACLs while fail for others.
+	// This operation is supported by brokers with version 0.11.0.0 or higher.
+	DeleteACL(filter AclFilter, validateOnly bool) ([]MatchingAcl, error)
+
+	// List the consumer groups available in the cluster.
+	ListConsumerGroups() (map[string]string, error)
+
+	// Describe the given consumer group
+	DescribeConsumerGroups(groups []string) ([]*GroupDescription, error)
+
+	// List the consumer group offsets available in the cluster.
+	ListConsumerGroupOffsets(group string, topicPartitions map[string][]int32) (*OffsetFetchResponse, error)
+
+	// Get information about the nodes in the cluster
+	DescribeCluster() (brokers []*Broker, controllerID int32, err error)
+
+	// Close shuts down the admin and closes underlying client.
+	Close() error
+}
+
+type clusterAdmin struct {
+	client Client
+	conf   *Config
+}
+
+// NewClusterAdmin creates a new ClusterAdmin using the given broker addresses and configuration.
+func NewClusterAdmin(addrs []string, conf *Config) (ClusterAdmin, error) {
+	client, err := NewClient(addrs, conf)
+	if err != nil {
+		return nil, err
+	}
+
+	//make sure we can retrieve the controller
+	_, err = client.Controller()
+	if err != nil {
+		return nil, err
+	}
+
+	ca := &clusterAdmin{
+		client: client,
+		conf:   client.Config(),
+	}
+	return ca, nil
+}
+
+func (ca *clusterAdmin) Close() error {
+	return ca.client.Close()
+}
+
+func (ca *clusterAdmin) Controller() (*Broker, error) {
+	return ca.client.Controller()
+}
+
+func (ca *clusterAdmin) CreateTopic(topic string, detail *TopicDetail, validateOnly bool) error {
+
+	if topic == "" {
+		return ErrInvalidTopic
+	}
+
+	if detail == nil {
+		return errors.New("You must specify topic details")
+	}
+
+	topicDetails := make(map[string]*TopicDetail)
+	topicDetails[topic] = detail
+
+	request := &CreateTopicsRequest{
+		TopicDetails: topicDetails,
+		ValidateOnly: validateOnly,
+		Timeout:      ca.conf.Admin.Timeout,
+	}
+
+	if ca.conf.Version.IsAtLeast(V0_11_0_0) {
+		request.Version = 1
+	}
+	if ca.conf.Version.IsAtLeast(V1_0_0_0) {
+		request.Version = 2
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	rsp, err := b.CreateTopics(request)
+	if err != nil {
+		return err
+	}
+
+	topicErr, ok := rsp.TopicErrors[topic]
+	if !ok {
+		return ErrIncompleteResponse
+	}
+
+	if topicErr.Err != ErrNoError {
+		return topicErr.Err
+	}
+
+	return nil
+}
+
+func (ca *clusterAdmin) DescribeTopics(topics []string) (metadata []*TopicMetadata, err error) {
+	controller, err := ca.Controller()
+	if err != nil {
+		return nil, err
+	}
+
+	request := &MetadataRequest{
+		Topics:                 topics,
+		AllowAutoTopicCreation: false,
+	}
+
+	if ca.conf.Version.IsAtLeast(V0_11_0_0) {
+		request.Version = 4
+	}
+
+	response, err := controller.GetMetadata(request)
+	if err != nil {
+		return nil, err
+	}
+	return response.Topics, nil
+}
+
+func (ca *clusterAdmin) DescribeCluster() (brokers []*Broker, controllerID int32, err error) {
+	controller, err := ca.Controller()
+	if err != nil {
+		return nil, int32(0), err
+	}
+
+	request := &MetadataRequest{
+		Topics: []string{},
+	}
+
+	response, err := controller.GetMetadata(request)
+	if err != nil {
+		return nil, int32(0), err
+	}
+
+	return response.Brokers, response.ControllerID, nil
+}
+
+func (ca *clusterAdmin) findAnyBroker() (*Broker, error) {
+	brokers := ca.client.Brokers()
+	if len(brokers) > 0 {
+		index := rand.Intn(len(brokers))
+		return brokers[index], nil
+	}
+	return nil, errors.New("no available broker")
+}
+
+func (ca *clusterAdmin) ListTopics() (map[string]TopicDetail, error) {
+	// In order to build TopicDetails we need to first get the list of all
+	// topics using a MetadataRequest and then get their configs using a
+	// DescribeConfigsRequest request. To avoid sending many requests to the
+	// broker, we use a single DescribeConfigsRequest.
+
+	// Send the all-topic MetadataRequest
+	b, err := ca.findAnyBroker()
+	if err != nil {
+		return nil, err
+	}
+	_ = b.Open(ca.client.Config())
+
+	metadataReq := &MetadataRequest{}
+	metadataResp, err := b.GetMetadata(metadataReq)
+	if err != nil {
+		return nil, err
+	}
+
+	topicsDetailsMap := make(map[string]TopicDetail)
+
+	var describeConfigsResources []*ConfigResource
+
+	for _, topic := range metadataResp.Topics {
+		topicDetails := TopicDetail{
+			NumPartitions: int32(len(topic.Partitions)),
+		}
+		if len(topic.Partitions) > 0 {
+			topicDetails.ReplicaAssignment = map[int32][]int32{}
+			for _, partition := range topic.Partitions {
+				topicDetails.ReplicaAssignment[partition.ID] = partition.Replicas
+			}
+			topicDetails.ReplicationFactor = int16(len(topic.Partitions[0].Replicas))
+		}
+		topicsDetailsMap[topic.Name] = topicDetails
+
+		// we populate the resources we want to describe from the MetadataResponse
+		topicResource := ConfigResource{
+			Type: TopicResource,
+			Name: topic.Name,
+		}
+		describeConfigsResources = append(describeConfigsResources, &topicResource)
+	}
+
+	// Send the DescribeConfigsRequest
+	describeConfigsReq := &DescribeConfigsRequest{
+		Resources: describeConfigsResources,
+	}
+	describeConfigsResp, err := b.DescribeConfigs(describeConfigsReq)
+	if err != nil {
+		return nil, err
+	}
+
+	for _, resource := range describeConfigsResp.Resources {
+		topicDetails := topicsDetailsMap[resource.Name]
+		topicDetails.ConfigEntries = make(map[string]*string)
+
+		for _, entry := range resource.Configs {
+			// only include non-default non-sensitive config
+			// (don't actually think topic config will ever be sensitive)
+			if entry.Default || entry.Sensitive {
+				continue
+			}
+			topicDetails.ConfigEntries[entry.Name] = &entry.Value
+		}
+
+		topicsDetailsMap[resource.Name] = topicDetails
+	}
+
+	return topicsDetailsMap, nil
+}
+
+func (ca *clusterAdmin) DeleteTopic(topic string) error {
+
+	if topic == "" {
+		return ErrInvalidTopic
+	}
+
+	request := &DeleteTopicsRequest{
+		Topics:  []string{topic},
+		Timeout: ca.conf.Admin.Timeout,
+	}
+
+	if ca.conf.Version.IsAtLeast(V0_11_0_0) {
+		request.Version = 1
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	rsp, err := b.DeleteTopics(request)
+	if err != nil {
+		return err
+	}
+
+	topicErr, ok := rsp.TopicErrorCodes[topic]
+	if !ok {
+		return ErrIncompleteResponse
+	}
+
+	if topicErr != ErrNoError {
+		return topicErr
+	}
+	return nil
+}
+
+func (ca *clusterAdmin) CreatePartitions(topic string, count int32, assignment [][]int32, validateOnly bool) error {
+	if topic == "" {
+		return ErrInvalidTopic
+	}
+
+	topicPartitions := make(map[string]*TopicPartition)
+	topicPartitions[topic] = &TopicPartition{Count: count, Assignment: assignment}
+
+	request := &CreatePartitionsRequest{
+		TopicPartitions: topicPartitions,
+		Timeout:         ca.conf.Admin.Timeout,
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	rsp, err := b.CreatePartitions(request)
+	if err != nil {
+		return err
+	}
+
+	topicErr, ok := rsp.TopicPartitionErrors[topic]
+	if !ok {
+		return ErrIncompleteResponse
+	}
+
+	if topicErr.Err != ErrNoError {
+		return topicErr.Err
+	}
+
+	return nil
+}
+
+func (ca *clusterAdmin) DeleteRecords(topic string, partitionOffsets map[int32]int64) error {
+
+	if topic == "" {
+		return ErrInvalidTopic
+	}
+
+	topics := make(map[string]*DeleteRecordsRequestTopic)
+	topics[topic] = &DeleteRecordsRequestTopic{PartitionOffsets: partitionOffsets}
+	request := &DeleteRecordsRequest{
+		Topics:  topics,
+		Timeout: ca.conf.Admin.Timeout,
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	rsp, err := b.DeleteRecords(request)
+	if err != nil {
+		return err
+	}
+
+	_, ok := rsp.Topics[topic]
+	if !ok {
+		return ErrIncompleteResponse
+	}
+
+	//todo since we are dealing with couple of partitions it would be good if we return slice of errors
+	//for each partition instead of one error
+	return nil
+}
+
+func (ca *clusterAdmin) DescribeConfig(resource ConfigResource) ([]ConfigEntry, error) {
+
+	var entries []ConfigEntry
+	var resources []*ConfigResource
+	resources = append(resources, &resource)
+
+	request := &DescribeConfigsRequest{
+		Resources: resources,
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return nil, err
+	}
+
+	rsp, err := b.DescribeConfigs(request)
+	if err != nil {
+		return nil, err
+	}
+
+	for _, rspResource := range rsp.Resources {
+		if rspResource.Name == resource.Name {
+			if rspResource.ErrorMsg != "" {
+				return nil, errors.New(rspResource.ErrorMsg)
+			}
+			for _, cfgEntry := range rspResource.Configs {
+				entries = append(entries, *cfgEntry)
+			}
+		}
+	}
+	return entries, nil
+}
+
+func (ca *clusterAdmin) AlterConfig(resourceType ConfigResourceType, name string, entries map[string]*string, validateOnly bool) error {
+
+	var resources []*AlterConfigsResource
+	resources = append(resources, &AlterConfigsResource{
+		Type:          resourceType,
+		Name:          name,
+		ConfigEntries: entries,
+	})
+
+	request := &AlterConfigsRequest{
+		Resources:    resources,
+		ValidateOnly: validateOnly,
+	}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	rsp, err := b.AlterConfigs(request)
+	if err != nil {
+		return err
+	}
+
+	for _, rspResource := range rsp.Resources {
+		if rspResource.Name == name {
+			if rspResource.ErrorMsg != "" {
+				return errors.New(rspResource.ErrorMsg)
+			}
+		}
+	}
+	return nil
+}
+
+func (ca *clusterAdmin) CreateACL(resource Resource, acl Acl) error {
+	var acls []*AclCreation
+	acls = append(acls, &AclCreation{resource, acl})
+	request := &CreateAclsRequest{AclCreations: acls}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return err
+	}
+
+	_, err = b.CreateAcls(request)
+	return err
+}
+
+func (ca *clusterAdmin) ListAcls(filter AclFilter) ([]ResourceAcls, error) {
+
+	request := &DescribeAclsRequest{AclFilter: filter}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return nil, err
+	}
+
+	rsp, err := b.DescribeAcls(request)
+	if err != nil {
+		return nil, err
+	}
+
+	var lAcls []ResourceAcls
+	for _, rAcl := range rsp.ResourceAcls {
+		lAcls = append(lAcls, *rAcl)
+	}
+	return lAcls, nil
+}
+
+func (ca *clusterAdmin) DeleteACL(filter AclFilter, validateOnly bool) ([]MatchingAcl, error) {
+	var filters []*AclFilter
+	filters = append(filters, &filter)
+	request := &DeleteAclsRequest{Filters: filters}
+
+	b, err := ca.Controller()
+	if err != nil {
+		return nil, err
+	}
+
+	rsp, err := b.DeleteAcls(request)
+	if err != nil {
+		return nil, err
+	}
+
+	var mAcls []MatchingAcl
+	for _, fr := range rsp.FilterResponses {
+		for _, mACL := range fr.MatchingAcls {
+			mAcls = append(mAcls, *mACL)
+		}
+
+	}
+	return mAcls, nil
+}
+
+func (ca *clusterAdmin) DescribeConsumerGroups(groups []string) (result []*GroupDescription, err error) {
+	groupsPerBroker := make(map[*Broker][]string)
+
+	for _, group := range groups {
+		controller, err := ca.client.Coordinator(group)
+		if err != nil {
+			return nil, err
+		}
+		groupsPerBroker[controller] = append(groupsPerBroker[controller], group)
+
+	}
+
+	for broker, brokerGroups := range groupsPerBroker {
+		response, err := broker.DescribeGroups(&DescribeGroupsRequest{
+			Groups: brokerGroups,
+		})
+		if err != nil {
+			return nil, err
+		}
+
+		result = append(result, response.Groups...)
+	}
+	return result, nil
+}
+
+func (ca *clusterAdmin) ListConsumerGroups() (allGroups map[string]string, err error) {
+	allGroups = make(map[string]string)
+
+	// Query brokers in parallel, since we have to query *all* brokers
+	brokers := ca.client.Brokers()
+	groupMaps := make(chan map[string]string, len(brokers))
+	errors := make(chan error, len(brokers))
+	wg := sync.WaitGroup{}
+
+	for _, b := range brokers {
+		wg.Add(1)
+		go func(b *Broker, conf *Config) {
+			defer wg.Done()
+			_ = b.Open(conf) // Ensure that broker is opened
+
+			response, err := b.ListGroups(&ListGroupsRequest{})
+			if err != nil {
+				errors <- err
+				return
+			}
+
+			groups := make(map[string]string)
+			for group, typ := range response.Groups {
+				groups[group] = typ
+			}
+
+			groupMaps <- groups
+
+		}(b, ca.conf)
+	}
+
+	wg.Wait()
+	close(groupMaps)
+	close(errors)
+
+	for groupMap := range groupMaps {
+		for group, protocolType := range groupMap {
+			allGroups[group] = protocolType
+		}
+	}
+
+	// Intentionally return only the first error for simplicity
+	err = <-errors
+	return
+}
+
+func (ca *clusterAdmin) ListConsumerGroupOffsets(group string, topicPartitions map[string][]int32) (*OffsetFetchResponse, error) {
+	coordinator, err := ca.client.Coordinator(group)
+	if err != nil {
+		return nil, err
+	}
+
+	request := &OffsetFetchRequest{
+		ConsumerGroup: group,
+		partitions:    topicPartitions,
+	}
+
+	if ca.conf.Version.IsAtLeast(V0_8_2_2) {
+		request.Version = 1
+	}
+
+	return coordinator.FetchOffset(request)
+}
diff --git a/vendor/github.com/Shopify/sarama/alter_configs_request.go b/vendor/github.com/Shopify/sarama/alter_configs_request.go
new file mode 100644
index 0000000..48c44ea
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/alter_configs_request.go
@@ -0,0 +1,120 @@
+package sarama
+
+type AlterConfigsRequest struct {
+	Resources    []*AlterConfigsResource
+	ValidateOnly bool
+}
+
+type AlterConfigsResource struct {
+	Type          ConfigResourceType
+	Name          string
+	ConfigEntries map[string]*string
+}
+
+func (acr *AlterConfigsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(acr.Resources)); err != nil {
+		return err
+	}
+
+	for _, r := range acr.Resources {
+		if err := r.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	pe.putBool(acr.ValidateOnly)
+	return nil
+}
+
+func (acr *AlterConfigsRequest) decode(pd packetDecoder, version int16) error {
+	resourceCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	acr.Resources = make([]*AlterConfigsResource, resourceCount)
+	for i := range acr.Resources {
+		r := &AlterConfigsResource{}
+		err = r.decode(pd, version)
+		if err != nil {
+			return err
+		}
+		acr.Resources[i] = r
+	}
+
+	validateOnly, err := pd.getBool()
+	if err != nil {
+		return err
+	}
+
+	acr.ValidateOnly = validateOnly
+
+	return nil
+}
+
+func (ac *AlterConfigsResource) encode(pe packetEncoder) error {
+	pe.putInt8(int8(ac.Type))
+
+	if err := pe.putString(ac.Name); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(ac.ConfigEntries)); err != nil {
+		return err
+	}
+	for configKey, configValue := range ac.ConfigEntries {
+		if err := pe.putString(configKey); err != nil {
+			return err
+		}
+		if err := pe.putNullableString(configValue); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (ac *AlterConfigsResource) decode(pd packetDecoder, version int16) error {
+	t, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	ac.Type = ConfigResourceType(t)
+
+	name, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	ac.Name = name
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		ac.ConfigEntries = make(map[string]*string, n)
+		for i := 0; i < n; i++ {
+			configKey, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			if ac.ConfigEntries[configKey], err = pd.getNullableString(); err != nil {
+				return err
+			}
+		}
+	}
+	return err
+}
+
+func (acr *AlterConfigsRequest) key() int16 {
+	return 33
+}
+
+func (acr *AlterConfigsRequest) version() int16 {
+	return 0
+}
+
+func (acr *AlterConfigsRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/alter_configs_response.go b/vendor/github.com/Shopify/sarama/alter_configs_response.go
new file mode 100644
index 0000000..29b09e1
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/alter_configs_response.go
@@ -0,0 +1,95 @@
+package sarama
+
+import "time"
+
+type AlterConfigsResponse struct {
+	ThrottleTime time.Duration
+	Resources    []*AlterConfigsResourceResponse
+}
+
+type AlterConfigsResourceResponse struct {
+	ErrorCode int16
+	ErrorMsg  string
+	Type      ConfigResourceType
+	Name      string
+}
+
+func (ct *AlterConfigsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(ct.ThrottleTime / time.Millisecond))
+
+	if err := pe.putArrayLength(len(ct.Resources)); err != nil {
+		return err
+	}
+
+	for i := range ct.Resources {
+		pe.putInt16(ct.Resources[i].ErrorCode)
+		err := pe.putString(ct.Resources[i].ErrorMsg)
+		if err != nil {
+			return nil
+		}
+		pe.putInt8(int8(ct.Resources[i].Type))
+		err = pe.putString(ct.Resources[i].Name)
+		if err != nil {
+			return nil
+		}
+	}
+
+	return nil
+}
+
+func (acr *AlterConfigsResponse) decode(pd packetDecoder, version int16) error {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	acr.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	responseCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	acr.Resources = make([]*AlterConfigsResourceResponse, responseCount)
+
+	for i := range acr.Resources {
+		acr.Resources[i] = new(AlterConfigsResourceResponse)
+
+		errCode, err := pd.getInt16()
+		if err != nil {
+			return err
+		}
+		acr.Resources[i].ErrorCode = errCode
+
+		e, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		acr.Resources[i].ErrorMsg = e
+
+		t, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+		acr.Resources[i].Type = ConfigResourceType(t)
+
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		acr.Resources[i].Name = name
+	}
+
+	return nil
+}
+
+func (r *AlterConfigsResponse) key() int16 {
+	return 32
+}
+
+func (r *AlterConfigsResponse) version() int16 {
+	return 0
+}
+
+func (r *AlterConfigsResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/api_versions_request.go b/vendor/github.com/Shopify/sarama/api_versions_request.go
new file mode 100644
index 0000000..ab65f01
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/api_versions_request.go
@@ -0,0 +1,24 @@
+package sarama
+
+type ApiVersionsRequest struct {
+}
+
+func (r *ApiVersionsRequest) encode(pe packetEncoder) error {
+	return nil
+}
+
+func (r *ApiVersionsRequest) decode(pd packetDecoder, version int16) (err error) {
+	return nil
+}
+
+func (r *ApiVersionsRequest) key() int16 {
+	return 18
+}
+
+func (r *ApiVersionsRequest) version() int16 {
+	return 0
+}
+
+func (r *ApiVersionsRequest) requiredVersion() KafkaVersion {
+	return V0_10_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/api_versions_response.go b/vendor/github.com/Shopify/sarama/api_versions_response.go
new file mode 100644
index 0000000..23bc326
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/api_versions_response.go
@@ -0,0 +1,87 @@
+package sarama
+
+type ApiVersionsResponseBlock struct {
+	ApiKey     int16
+	MinVersion int16
+	MaxVersion int16
+}
+
+func (b *ApiVersionsResponseBlock) encode(pe packetEncoder) error {
+	pe.putInt16(b.ApiKey)
+	pe.putInt16(b.MinVersion)
+	pe.putInt16(b.MaxVersion)
+	return nil
+}
+
+func (b *ApiVersionsResponseBlock) decode(pd packetDecoder) error {
+	var err error
+
+	if b.ApiKey, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	if b.MinVersion, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	if b.MaxVersion, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+type ApiVersionsResponse struct {
+	Err         KError
+	ApiVersions []*ApiVersionsResponseBlock
+}
+
+func (r *ApiVersionsResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	if err := pe.putArrayLength(len(r.ApiVersions)); err != nil {
+		return err
+	}
+	for _, apiVersion := range r.ApiVersions {
+		if err := apiVersion.encode(pe); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (r *ApiVersionsResponse) decode(pd packetDecoder, version int16) error {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	numBlocks, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.ApiVersions = make([]*ApiVersionsResponseBlock, numBlocks)
+	for i := 0; i < numBlocks; i++ {
+		block := new(ApiVersionsResponseBlock)
+		if err := block.decode(pd); err != nil {
+			return err
+		}
+		r.ApiVersions[i] = block
+	}
+
+	return nil
+}
+
+func (r *ApiVersionsResponse) key() int16 {
+	return 18
+}
+
+func (r *ApiVersionsResponse) version() int16 {
+	return 0
+}
+
+func (r *ApiVersionsResponse) requiredVersion() KafkaVersion {
+	return V0_10_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/async_producer.go b/vendor/github.com/Shopify/sarama/async_producer.go
new file mode 100644
index 0000000..5174a35
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/async_producer.go
@@ -0,0 +1,1099 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"fmt"
+	"sync"
+	"time"
+
+	"github.com/eapache/go-resiliency/breaker"
+	"github.com/eapache/queue"
+)
+
+// AsyncProducer publishes Kafka messages using a non-blocking API. It routes messages
+// to the correct broker for the provided topic-partition, refreshing metadata as appropriate,
+// and parses responses for errors. You must read from the Errors() channel or the
+// producer will deadlock. You must call Close() or AsyncClose() on a producer to avoid
+// leaks: it will not be garbage-collected automatically when it passes out of
+// scope.
+type AsyncProducer interface {
+
+	// AsyncClose triggers a shutdown of the producer. The shutdown has completed
+	// when both the Errors and Successes channels have been closed. When calling
+	// AsyncClose, you *must* continue to read from those channels in order to
+	// drain the results of any messages in flight.
+	AsyncClose()
+
+	// Close shuts down the producer and waits for any buffered messages to be
+	// flushed. You must call this function before a producer object passes out of
+	// scope, as it may otherwise leak memory. You must call this before calling
+	// Close on the underlying client.
+	Close() error
+
+	// Input is the input channel for the user to write messages to that they
+	// wish to send.
+	Input() chan<- *ProducerMessage
+
+	// Successes is the success output channel back to the user when Return.Successes is
+	// enabled. If Return.Successes is true, you MUST read from this channel or the
+	// Producer will deadlock. It is suggested that you send and read messages
+	// together in a single select statement.
+	Successes() <-chan *ProducerMessage
+
+	// Errors is the error output channel back to the user. You MUST read from this
+	// channel or the Producer will deadlock when the channel is full. Alternatively,
+	// you can set Producer.Return.Errors in your config to false, which prevents
+	// errors to be returned.
+	Errors() <-chan *ProducerError
+}
+
+// transactionManager keeps the state necessary to ensure idempotent production
+type transactionManager struct {
+	producerID      int64
+	producerEpoch   int16
+	sequenceNumbers map[string]int32
+	mutex           sync.Mutex
+}
+
+const (
+	noProducerID    = -1
+	noProducerEpoch = -1
+)
+
+func (t *transactionManager) getAndIncrementSequenceNumber(topic string, partition int32) int32 {
+	key := fmt.Sprintf("%s-%d", topic, partition)
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+	sequence := t.sequenceNumbers[key]
+	t.sequenceNumbers[key] = sequence + 1
+	return sequence
+}
+
+func newTransactionManager(conf *Config, client Client) (*transactionManager, error) {
+	txnmgr := &transactionManager{
+		producerID:    noProducerID,
+		producerEpoch: noProducerEpoch,
+	}
+
+	if conf.Producer.Idempotent {
+		initProducerIDResponse, err := client.InitProducerID()
+		if err != nil {
+			return nil, err
+		}
+		txnmgr.producerID = initProducerIDResponse.ProducerID
+		txnmgr.producerEpoch = initProducerIDResponse.ProducerEpoch
+		txnmgr.sequenceNumbers = make(map[string]int32)
+		txnmgr.mutex = sync.Mutex{}
+
+		Logger.Printf("Obtained a ProducerId: %d and ProducerEpoch: %d\n", txnmgr.producerID, txnmgr.producerEpoch)
+	}
+
+	return txnmgr, nil
+}
+
+type asyncProducer struct {
+	client    Client
+	conf      *Config
+	ownClient bool
+
+	errors                    chan *ProducerError
+	input, successes, retries chan *ProducerMessage
+	inFlight                  sync.WaitGroup
+
+	brokers    map[*Broker]*brokerProducer
+	brokerRefs map[*brokerProducer]int
+	brokerLock sync.Mutex
+
+	txnmgr *transactionManager
+}
+
+// NewAsyncProducer creates a new AsyncProducer using the given broker addresses and configuration.
+func NewAsyncProducer(addrs []string, conf *Config) (AsyncProducer, error) {
+	client, err := NewClient(addrs, conf)
+	if err != nil {
+		return nil, err
+	}
+
+	p, err := NewAsyncProducerFromClient(client)
+	if err != nil {
+		return nil, err
+	}
+	p.(*asyncProducer).ownClient = true
+	return p, nil
+}
+
+// NewAsyncProducerFromClient creates a new Producer using the given client. It is still
+// necessary to call Close() on the underlying client when shutting down this producer.
+func NewAsyncProducerFromClient(client Client) (AsyncProducer, error) {
+	// Check that we are not dealing with a closed Client before processing any other arguments
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	txnmgr, err := newTransactionManager(client.Config(), client)
+	if err != nil {
+		return nil, err
+	}
+
+	p := &asyncProducer{
+		client:     client,
+		conf:       client.Config(),
+		errors:     make(chan *ProducerError),
+		input:      make(chan *ProducerMessage),
+		successes:  make(chan *ProducerMessage),
+		retries:    make(chan *ProducerMessage),
+		brokers:    make(map[*Broker]*brokerProducer),
+		brokerRefs: make(map[*brokerProducer]int),
+		txnmgr:     txnmgr,
+	}
+
+	// launch our singleton dispatchers
+	go withRecover(p.dispatcher)
+	go withRecover(p.retryHandler)
+
+	return p, nil
+}
+
+type flagSet int8
+
+const (
+	syn      flagSet = 1 << iota // first message from partitionProducer to brokerProducer
+	fin                          // final message from partitionProducer to brokerProducer and back
+	shutdown                     // start the shutdown process
+)
+
+// ProducerMessage is the collection of elements passed to the Producer in order to send a message.
+type ProducerMessage struct {
+	Topic string // The Kafka topic for this message.
+	// The partitioning key for this message. Pre-existing Encoders include
+	// StringEncoder and ByteEncoder.
+	Key Encoder
+	// The actual message to store in Kafka. Pre-existing Encoders include
+	// StringEncoder and ByteEncoder.
+	Value Encoder
+
+	// The headers are key-value pairs that are transparently passed
+	// by Kafka between producers and consumers.
+	Headers []RecordHeader
+
+	// This field is used to hold arbitrary data you wish to include so it
+	// will be available when receiving on the Successes and Errors channels.
+	// Sarama completely ignores this field and is only to be used for
+	// pass-through data.
+	Metadata interface{}
+
+	// Below this point are filled in by the producer as the message is processed
+
+	// Offset is the offset of the message stored on the broker. This is only
+	// guaranteed to be defined if the message was successfully delivered and
+	// RequiredAcks is not NoResponse.
+	Offset int64
+	// Partition is the partition that the message was sent to. This is only
+	// guaranteed to be defined if the message was successfully delivered.
+	Partition int32
+	// Timestamp is the timestamp assigned to the message by the broker. This
+	// is only guaranteed to be defined if the message was successfully
+	// delivered, RequiredAcks is not NoResponse, and the Kafka broker is at
+	// least version 0.10.0.
+	Timestamp time.Time
+
+	retries        int
+	flags          flagSet
+	expectation    chan *ProducerError
+	sequenceNumber int32
+}
+
+const producerMessageOverhead = 26 // the metadata overhead of CRC, flags, etc.
+
+func (m *ProducerMessage) byteSize(version int) int {
+	var size int
+	if version >= 2 {
+		size = maximumRecordOverhead
+		for _, h := range m.Headers {
+			size += len(h.Key) + len(h.Value) + 2*binary.MaxVarintLen32
+		}
+	} else {
+		size = producerMessageOverhead
+	}
+	if m.Key != nil {
+		size += m.Key.Length()
+	}
+	if m.Value != nil {
+		size += m.Value.Length()
+	}
+	return size
+}
+
+func (m *ProducerMessage) clear() {
+	m.flags = 0
+	m.retries = 0
+}
+
+// ProducerError is the type of error generated when the producer fails to deliver a message.
+// It contains the original ProducerMessage as well as the actual error value.
+type ProducerError struct {
+	Msg *ProducerMessage
+	Err error
+}
+
+func (pe ProducerError) Error() string {
+	return fmt.Sprintf("kafka: Failed to produce message to topic %s: %s", pe.Msg.Topic, pe.Err)
+}
+
+// ProducerErrors is a type that wraps a batch of "ProducerError"s and implements the Error interface.
+// It can be returned from the Producer's Close method to avoid the need to manually drain the Errors channel
+// when closing a producer.
+type ProducerErrors []*ProducerError
+
+func (pe ProducerErrors) Error() string {
+	return fmt.Sprintf("kafka: Failed to deliver %d messages.", len(pe))
+}
+
+func (p *asyncProducer) Errors() <-chan *ProducerError {
+	return p.errors
+}
+
+func (p *asyncProducer) Successes() <-chan *ProducerMessage {
+	return p.successes
+}
+
+func (p *asyncProducer) Input() chan<- *ProducerMessage {
+	return p.input
+}
+
+func (p *asyncProducer) Close() error {
+	p.AsyncClose()
+
+	if p.conf.Producer.Return.Successes {
+		go withRecover(func() {
+			for range p.successes {
+			}
+		})
+	}
+
+	var errors ProducerErrors
+	if p.conf.Producer.Return.Errors {
+		for event := range p.errors {
+			errors = append(errors, event)
+		}
+	} else {
+		<-p.errors
+	}
+
+	if len(errors) > 0 {
+		return errors
+	}
+	return nil
+}
+
+func (p *asyncProducer) AsyncClose() {
+	go withRecover(p.shutdown)
+}
+
+// singleton
+// dispatches messages by topic
+func (p *asyncProducer) dispatcher() {
+	handlers := make(map[string]chan<- *ProducerMessage)
+	shuttingDown := false
+
+	for msg := range p.input {
+		if msg == nil {
+			Logger.Println("Something tried to send a nil message, it was ignored.")
+			continue
+		}
+
+		if msg.flags&shutdown != 0 {
+			shuttingDown = true
+			p.inFlight.Done()
+			continue
+		} else if msg.retries == 0 {
+			if shuttingDown {
+				// we can't just call returnError here because that decrements the wait group,
+				// which hasn't been incremented yet for this message, and shouldn't be
+				pErr := &ProducerError{Msg: msg, Err: ErrShuttingDown}
+				if p.conf.Producer.Return.Errors {
+					p.errors <- pErr
+				} else {
+					Logger.Println(pErr)
+				}
+				continue
+			}
+			p.inFlight.Add(1)
+		}
+
+		version := 1
+		if p.conf.Version.IsAtLeast(V0_11_0_0) {
+			version = 2
+		} else if msg.Headers != nil {
+			p.returnError(msg, ConfigurationError("Producing headers requires Kafka at least v0.11"))
+			continue
+		}
+		if msg.byteSize(version) > p.conf.Producer.MaxMessageBytes {
+			p.returnError(msg, ErrMessageSizeTooLarge)
+			continue
+		}
+
+		handler := handlers[msg.Topic]
+		if handler == nil {
+			handler = p.newTopicProducer(msg.Topic)
+			handlers[msg.Topic] = handler
+		}
+
+		handler <- msg
+	}
+
+	for _, handler := range handlers {
+		close(handler)
+	}
+}
+
+// one per topic
+// partitions messages, then dispatches them by partition
+type topicProducer struct {
+	parent *asyncProducer
+	topic  string
+	input  <-chan *ProducerMessage
+
+	breaker     *breaker.Breaker
+	handlers    map[int32]chan<- *ProducerMessage
+	partitioner Partitioner
+}
+
+func (p *asyncProducer) newTopicProducer(topic string) chan<- *ProducerMessage {
+	input := make(chan *ProducerMessage, p.conf.ChannelBufferSize)
+	tp := &topicProducer{
+		parent:      p,
+		topic:       topic,
+		input:       input,
+		breaker:     breaker.New(3, 1, 10*time.Second),
+		handlers:    make(map[int32]chan<- *ProducerMessage),
+		partitioner: p.conf.Producer.Partitioner(topic),
+	}
+	go withRecover(tp.dispatch)
+	return input
+}
+
+func (tp *topicProducer) dispatch() {
+	for msg := range tp.input {
+		if msg.retries == 0 {
+			if err := tp.partitionMessage(msg); err != nil {
+				tp.parent.returnError(msg, err)
+				continue
+			}
+		}
+		// All messages being retried (sent or not) have already had their retry count updated
+		if tp.parent.conf.Producer.Idempotent && msg.retries == 0 {
+			msg.sequenceNumber = tp.parent.txnmgr.getAndIncrementSequenceNumber(msg.Topic, msg.Partition)
+		}
+
+		handler := tp.handlers[msg.Partition]
+		if handler == nil {
+			handler = tp.parent.newPartitionProducer(msg.Topic, msg.Partition)
+			tp.handlers[msg.Partition] = handler
+		}
+
+		handler <- msg
+	}
+
+	for _, handler := range tp.handlers {
+		close(handler)
+	}
+}
+
+func (tp *topicProducer) partitionMessage(msg *ProducerMessage) error {
+	var partitions []int32
+
+	err := tp.breaker.Run(func() (err error) {
+		var requiresConsistency = false
+		if ep, ok := tp.partitioner.(DynamicConsistencyPartitioner); ok {
+			requiresConsistency = ep.MessageRequiresConsistency(msg)
+		} else {
+			requiresConsistency = tp.partitioner.RequiresConsistency()
+		}
+
+		if requiresConsistency {
+			partitions, err = tp.parent.client.Partitions(msg.Topic)
+		} else {
+			partitions, err = tp.parent.client.WritablePartitions(msg.Topic)
+		}
+		return
+	})
+
+	if err != nil {
+		return err
+	}
+
+	numPartitions := int32(len(partitions))
+
+	if numPartitions == 0 {
+		return ErrLeaderNotAvailable
+	}
+
+	choice, err := tp.partitioner.Partition(msg, numPartitions)
+
+	if err != nil {
+		return err
+	} else if choice < 0 || choice >= numPartitions {
+		return ErrInvalidPartition
+	}
+
+	msg.Partition = partitions[choice]
+
+	return nil
+}
+
+// one per partition per topic
+// dispatches messages to the appropriate broker
+// also responsible for maintaining message order during retries
+type partitionProducer struct {
+	parent    *asyncProducer
+	topic     string
+	partition int32
+	input     <-chan *ProducerMessage
+
+	leader         *Broker
+	breaker        *breaker.Breaker
+	brokerProducer *brokerProducer
+
+	// highWatermark tracks the "current" retry level, which is the only one where we actually let messages through,
+	// all other messages get buffered in retryState[msg.retries].buf to preserve ordering
+	// retryState[msg.retries].expectChaser simply tracks whether we've seen a fin message for a given level (and
+	// therefore whether our buffer is complete and safe to flush)
+	highWatermark int
+	retryState    []partitionRetryState
+}
+
+type partitionRetryState struct {
+	buf          []*ProducerMessage
+	expectChaser bool
+}
+
+func (p *asyncProducer) newPartitionProducer(topic string, partition int32) chan<- *ProducerMessage {
+	input := make(chan *ProducerMessage, p.conf.ChannelBufferSize)
+	pp := &partitionProducer{
+		parent:    p,
+		topic:     topic,
+		partition: partition,
+		input:     input,
+
+		breaker:    breaker.New(3, 1, 10*time.Second),
+		retryState: make([]partitionRetryState, p.conf.Producer.Retry.Max+1),
+	}
+	go withRecover(pp.dispatch)
+	return input
+}
+
+func (pp *partitionProducer) backoff(retries int) {
+	var backoff time.Duration
+	if pp.parent.conf.Producer.Retry.BackoffFunc != nil {
+		maxRetries := pp.parent.conf.Producer.Retry.Max
+		backoff = pp.parent.conf.Producer.Retry.BackoffFunc(retries, maxRetries)
+	} else {
+		backoff = pp.parent.conf.Producer.Retry.Backoff
+	}
+	if backoff > 0 {
+		time.Sleep(backoff)
+	}
+}
+
+func (pp *partitionProducer) dispatch() {
+	// try to prefetch the leader; if this doesn't work, we'll do a proper call to `updateLeader`
+	// on the first message
+	pp.leader, _ = pp.parent.client.Leader(pp.topic, pp.partition)
+	if pp.leader != nil {
+		pp.brokerProducer = pp.parent.getBrokerProducer(pp.leader)
+		pp.parent.inFlight.Add(1) // we're generating a syn message; track it so we don't shut down while it's still inflight
+		pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: syn}
+	}
+
+	defer func() {
+		if pp.brokerProducer != nil {
+			pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer)
+		}
+	}()
+
+	for msg := range pp.input {
+
+		if pp.brokerProducer != nil && pp.brokerProducer.abandoned != nil {
+			select {
+			case <-pp.brokerProducer.abandoned:
+				// a message on the abandoned channel means that our current broker selection is out of date
+				Logger.Printf("producer/leader/%s/%d abandoning broker %d\n", pp.topic, pp.partition, pp.leader.ID())
+				pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer)
+				pp.brokerProducer = nil
+				time.Sleep(pp.parent.conf.Producer.Retry.Backoff)
+			default:
+				// producer connection is still open.
+			}
+		}
+
+		if msg.retries > pp.highWatermark {
+			// a new, higher, retry level; handle it and then back off
+			pp.newHighWatermark(msg.retries)
+			pp.backoff(msg.retries)
+		} else if pp.highWatermark > 0 {
+			// we are retrying something (else highWatermark would be 0) but this message is not a *new* retry level
+			if msg.retries < pp.highWatermark {
+				// in fact this message is not even the current retry level, so buffer it for now (unless it's a just a fin)
+				if msg.flags&fin == fin {
+					pp.retryState[msg.retries].expectChaser = false
+					pp.parent.inFlight.Done() // this fin is now handled and will be garbage collected
+				} else {
+					pp.retryState[msg.retries].buf = append(pp.retryState[msg.retries].buf, msg)
+				}
+				continue
+			} else if msg.flags&fin == fin {
+				// this message is of the current retry level (msg.retries == highWatermark) and the fin flag is set,
+				// meaning this retry level is done and we can go down (at least) one level and flush that
+				pp.retryState[pp.highWatermark].expectChaser = false
+				pp.flushRetryBuffers()
+				pp.parent.inFlight.Done() // this fin is now handled and will be garbage collected
+				continue
+			}
+		}
+
+		// if we made it this far then the current msg contains real data, and can be sent to the next goroutine
+		// without breaking any of our ordering guarantees
+
+		if pp.brokerProducer == nil {
+			if err := pp.updateLeader(); err != nil {
+				pp.parent.returnError(msg, err)
+				pp.backoff(msg.retries)
+				continue
+			}
+			Logger.Printf("producer/leader/%s/%d selected broker %d\n", pp.topic, pp.partition, pp.leader.ID())
+		}
+
+		pp.brokerProducer.input <- msg
+	}
+}
+
+func (pp *partitionProducer) newHighWatermark(hwm int) {
+	Logger.Printf("producer/leader/%s/%d state change to [retrying-%d]\n", pp.topic, pp.partition, hwm)
+	pp.highWatermark = hwm
+
+	// send off a fin so that we know when everything "in between" has made it
+	// back to us and we can safely flush the backlog (otherwise we risk re-ordering messages)
+	pp.retryState[pp.highWatermark].expectChaser = true
+	pp.parent.inFlight.Add(1) // we're generating a fin message; track it so we don't shut down while it's still inflight
+	pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: fin, retries: pp.highWatermark - 1}
+
+	// a new HWM means that our current broker selection is out of date
+	Logger.Printf("producer/leader/%s/%d abandoning broker %d\n", pp.topic, pp.partition, pp.leader.ID())
+	pp.parent.unrefBrokerProducer(pp.leader, pp.brokerProducer)
+	pp.brokerProducer = nil
+}
+
+func (pp *partitionProducer) flushRetryBuffers() {
+	Logger.Printf("producer/leader/%s/%d state change to [flushing-%d]\n", pp.topic, pp.partition, pp.highWatermark)
+	for {
+		pp.highWatermark--
+
+		if pp.brokerProducer == nil {
+			if err := pp.updateLeader(); err != nil {
+				pp.parent.returnErrors(pp.retryState[pp.highWatermark].buf, err)
+				goto flushDone
+			}
+			Logger.Printf("producer/leader/%s/%d selected broker %d\n", pp.topic, pp.partition, pp.leader.ID())
+		}
+
+		for _, msg := range pp.retryState[pp.highWatermark].buf {
+			pp.brokerProducer.input <- msg
+		}
+
+	flushDone:
+		pp.retryState[pp.highWatermark].buf = nil
+		if pp.retryState[pp.highWatermark].expectChaser {
+			Logger.Printf("producer/leader/%s/%d state change to [retrying-%d]\n", pp.topic, pp.partition, pp.highWatermark)
+			break
+		} else if pp.highWatermark == 0 {
+			Logger.Printf("producer/leader/%s/%d state change to [normal]\n", pp.topic, pp.partition)
+			break
+		}
+	}
+}
+
+func (pp *partitionProducer) updateLeader() error {
+	return pp.breaker.Run(func() (err error) {
+		if err = pp.parent.client.RefreshMetadata(pp.topic); err != nil {
+			return err
+		}
+
+		if pp.leader, err = pp.parent.client.Leader(pp.topic, pp.partition); err != nil {
+			return err
+		}
+
+		pp.brokerProducer = pp.parent.getBrokerProducer(pp.leader)
+		pp.parent.inFlight.Add(1) // we're generating a syn message; track it so we don't shut down while it's still inflight
+		pp.brokerProducer.input <- &ProducerMessage{Topic: pp.topic, Partition: pp.partition, flags: syn}
+
+		return nil
+	})
+}
+
+// one per broker; also constructs an associated flusher
+func (p *asyncProducer) newBrokerProducer(broker *Broker) *brokerProducer {
+	var (
+		input     = make(chan *ProducerMessage)
+		bridge    = make(chan *produceSet)
+		responses = make(chan *brokerProducerResponse)
+	)
+
+	bp := &brokerProducer{
+		parent:         p,
+		broker:         broker,
+		input:          input,
+		output:         bridge,
+		responses:      responses,
+		buffer:         newProduceSet(p),
+		currentRetries: make(map[string]map[int32]error),
+	}
+	go withRecover(bp.run)
+
+	// minimal bridge to make the network response `select`able
+	go withRecover(func() {
+		for set := range bridge {
+			request := set.buildRequest()
+
+			response, err := broker.Produce(request)
+
+			responses <- &brokerProducerResponse{
+				set: set,
+				err: err,
+				res: response,
+			}
+		}
+		close(responses)
+	})
+
+	if p.conf.Producer.Retry.Max <= 0 {
+		bp.abandoned = make(chan struct{})
+	}
+
+	return bp
+}
+
+type brokerProducerResponse struct {
+	set *produceSet
+	err error
+	res *ProduceResponse
+}
+
+// groups messages together into appropriately-sized batches for sending to the broker
+// handles state related to retries etc
+type brokerProducer struct {
+	parent *asyncProducer
+	broker *Broker
+
+	input     chan *ProducerMessage
+	output    chan<- *produceSet
+	responses <-chan *brokerProducerResponse
+	abandoned chan struct{}
+
+	buffer     *produceSet
+	timer      <-chan time.Time
+	timerFired bool
+
+	closing        error
+	currentRetries map[string]map[int32]error
+}
+
+func (bp *brokerProducer) run() {
+	var output chan<- *produceSet
+	Logger.Printf("producer/broker/%d starting up\n", bp.broker.ID())
+
+	for {
+		select {
+		case msg := <-bp.input:
+			if msg == nil {
+				bp.shutdown()
+				return
+			}
+
+			if msg.flags&syn == syn {
+				Logger.Printf("producer/broker/%d state change to [open] on %s/%d\n",
+					bp.broker.ID(), msg.Topic, msg.Partition)
+				if bp.currentRetries[msg.Topic] == nil {
+					bp.currentRetries[msg.Topic] = make(map[int32]error)
+				}
+				bp.currentRetries[msg.Topic][msg.Partition] = nil
+				bp.parent.inFlight.Done()
+				continue
+			}
+
+			if reason := bp.needsRetry(msg); reason != nil {
+				bp.parent.retryMessage(msg, reason)
+
+				if bp.closing == nil && msg.flags&fin == fin {
+					// we were retrying this partition but we can start processing again
+					delete(bp.currentRetries[msg.Topic], msg.Partition)
+					Logger.Printf("producer/broker/%d state change to [closed] on %s/%d\n",
+						bp.broker.ID(), msg.Topic, msg.Partition)
+				}
+
+				continue
+			}
+
+			if bp.buffer.wouldOverflow(msg) {
+				if err := bp.waitForSpace(msg); err != nil {
+					bp.parent.retryMessage(msg, err)
+					continue
+				}
+			}
+
+			if err := bp.buffer.add(msg); err != nil {
+				bp.parent.returnError(msg, err)
+				continue
+			}
+
+			if bp.parent.conf.Producer.Flush.Frequency > 0 && bp.timer == nil {
+				bp.timer = time.After(bp.parent.conf.Producer.Flush.Frequency)
+			}
+		case <-bp.timer:
+			bp.timerFired = true
+		case output <- bp.buffer:
+			bp.rollOver()
+		case response := <-bp.responses:
+			bp.handleResponse(response)
+		}
+
+		if bp.timerFired || bp.buffer.readyToFlush() {
+			output = bp.output
+		} else {
+			output = nil
+		}
+	}
+}
+
+func (bp *brokerProducer) shutdown() {
+	for !bp.buffer.empty() {
+		select {
+		case response := <-bp.responses:
+			bp.handleResponse(response)
+		case bp.output <- bp.buffer:
+			bp.rollOver()
+		}
+	}
+	close(bp.output)
+	for response := range bp.responses {
+		bp.handleResponse(response)
+	}
+
+	Logger.Printf("producer/broker/%d shut down\n", bp.broker.ID())
+}
+
+func (bp *brokerProducer) needsRetry(msg *ProducerMessage) error {
+	if bp.closing != nil {
+		return bp.closing
+	}
+
+	return bp.currentRetries[msg.Topic][msg.Partition]
+}
+
+func (bp *brokerProducer) waitForSpace(msg *ProducerMessage) error {
+	Logger.Printf("producer/broker/%d maximum request accumulated, waiting for space\n", bp.broker.ID())
+
+	for {
+		select {
+		case response := <-bp.responses:
+			bp.handleResponse(response)
+			// handling a response can change our state, so re-check some things
+			if reason := bp.needsRetry(msg); reason != nil {
+				return reason
+			} else if !bp.buffer.wouldOverflow(msg) {
+				return nil
+			}
+		case bp.output <- bp.buffer:
+			bp.rollOver()
+			return nil
+		}
+	}
+}
+
+func (bp *brokerProducer) rollOver() {
+	bp.timer = nil
+	bp.timerFired = false
+	bp.buffer = newProduceSet(bp.parent)
+}
+
+func (bp *brokerProducer) handleResponse(response *brokerProducerResponse) {
+	if response.err != nil {
+		bp.handleError(response.set, response.err)
+	} else {
+		bp.handleSuccess(response.set, response.res)
+	}
+
+	if bp.buffer.empty() {
+		bp.rollOver() // this can happen if the response invalidated our buffer
+	}
+}
+
+func (bp *brokerProducer) handleSuccess(sent *produceSet, response *ProduceResponse) {
+	// we iterate through the blocks in the request set, not the response, so that we notice
+	// if the response is missing a block completely
+	var retryTopics []string
+	sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) {
+		if response == nil {
+			// this only happens when RequiredAcks is NoResponse, so we have to assume success
+			bp.parent.returnSuccesses(pSet.msgs)
+			return
+		}
+
+		block := response.GetBlock(topic, partition)
+		if block == nil {
+			bp.parent.returnErrors(pSet.msgs, ErrIncompleteResponse)
+			return
+		}
+
+		switch block.Err {
+		// Success
+		case ErrNoError:
+			if bp.parent.conf.Version.IsAtLeast(V0_10_0_0) && !block.Timestamp.IsZero() {
+				for _, msg := range pSet.msgs {
+					msg.Timestamp = block.Timestamp
+				}
+			}
+			for i, msg := range pSet.msgs {
+				msg.Offset = block.Offset + int64(i)
+			}
+			bp.parent.returnSuccesses(pSet.msgs)
+		// Duplicate
+		case ErrDuplicateSequenceNumber:
+			bp.parent.returnSuccesses(pSet.msgs)
+		// Retriable errors
+		case ErrInvalidMessage, ErrUnknownTopicOrPartition, ErrLeaderNotAvailable, ErrNotLeaderForPartition,
+			ErrRequestTimedOut, ErrNotEnoughReplicas, ErrNotEnoughReplicasAfterAppend:
+			if bp.parent.conf.Producer.Retry.Max <= 0 {
+				bp.parent.abandonBrokerConnection(bp.broker)
+				bp.parent.returnErrors(pSet.msgs, block.Err)
+			} else {
+				retryTopics = append(retryTopics, topic)
+			}
+		// Other non-retriable errors
+		default:
+			if bp.parent.conf.Producer.Retry.Max <= 0 {
+				bp.parent.abandonBrokerConnection(bp.broker)
+			}
+			bp.parent.returnErrors(pSet.msgs, block.Err)
+		}
+	})
+
+	if len(retryTopics) > 0 {
+		if bp.parent.conf.Producer.Idempotent {
+			err := bp.parent.client.RefreshMetadata(retryTopics...)
+			if err != nil {
+				Logger.Printf("Failed refreshing metadata because of %v\n", err)
+			}
+		}
+
+		sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) {
+			block := response.GetBlock(topic, partition)
+			if block == nil {
+				// handled in the previous "eachPartition" loop
+				return
+			}
+
+			switch block.Err {
+			case ErrInvalidMessage, ErrUnknownTopicOrPartition, ErrLeaderNotAvailable, ErrNotLeaderForPartition,
+				ErrRequestTimedOut, ErrNotEnoughReplicas, ErrNotEnoughReplicasAfterAppend:
+				Logger.Printf("producer/broker/%d state change to [retrying] on %s/%d because %v\n",
+					bp.broker.ID(), topic, partition, block.Err)
+				if bp.currentRetries[topic] == nil {
+					bp.currentRetries[topic] = make(map[int32]error)
+				}
+				bp.currentRetries[topic][partition] = block.Err
+				if bp.parent.conf.Producer.Idempotent {
+					go bp.parent.retryBatch(topic, partition, pSet, block.Err)
+				} else {
+					bp.parent.retryMessages(pSet.msgs, block.Err)
+				}
+				// dropping the following messages has the side effect of incrementing their retry count
+				bp.parent.retryMessages(bp.buffer.dropPartition(topic, partition), block.Err)
+			}
+		})
+	}
+}
+
+func (p *asyncProducer) retryBatch(topic string, partition int32, pSet *partitionSet, kerr KError) {
+	Logger.Printf("Retrying batch for %v-%d because of %s\n", topic, partition, kerr)
+	produceSet := newProduceSet(p)
+	produceSet.msgs[topic] = make(map[int32]*partitionSet)
+	produceSet.msgs[topic][partition] = pSet
+	produceSet.bufferBytes += pSet.bufferBytes
+	produceSet.bufferCount += len(pSet.msgs)
+	for _, msg := range pSet.msgs {
+		if msg.retries >= p.conf.Producer.Retry.Max {
+			p.returnError(msg, kerr)
+			return
+		}
+		msg.retries++
+	}
+
+	// it's expected that a metadata refresh has been requested prior to calling retryBatch
+	leader, err := p.client.Leader(topic, partition)
+	if err != nil {
+		Logger.Printf("Failed retrying batch for %v-%d because of %v while looking up for new leader\n", topic, partition, err)
+		for _, msg := range pSet.msgs {
+			p.returnError(msg, kerr)
+		}
+		return
+	}
+	bp := p.getBrokerProducer(leader)
+	bp.output <- produceSet
+}
+
+func (bp *brokerProducer) handleError(sent *produceSet, err error) {
+	switch err.(type) {
+	case PacketEncodingError:
+		sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) {
+			bp.parent.returnErrors(pSet.msgs, err)
+		})
+	default:
+		Logger.Printf("producer/broker/%d state change to [closing] because %s\n", bp.broker.ID(), err)
+		bp.parent.abandonBrokerConnection(bp.broker)
+		_ = bp.broker.Close()
+		bp.closing = err
+		sent.eachPartition(func(topic string, partition int32, pSet *partitionSet) {
+			bp.parent.retryMessages(pSet.msgs, err)
+		})
+		bp.buffer.eachPartition(func(topic string, partition int32, pSet *partitionSet) {
+			bp.parent.retryMessages(pSet.msgs, err)
+		})
+		bp.rollOver()
+	}
+}
+
+// singleton
+// effectively a "bridge" between the flushers and the dispatcher in order to avoid deadlock
+// based on https://godoc.org/github.com/eapache/channels#InfiniteChannel
+func (p *asyncProducer) retryHandler() {
+	var msg *ProducerMessage
+	buf := queue.New()
+
+	for {
+		if buf.Length() == 0 {
+			msg = <-p.retries
+		} else {
+			select {
+			case msg = <-p.retries:
+			case p.input <- buf.Peek().(*ProducerMessage):
+				buf.Remove()
+				continue
+			}
+		}
+
+		if msg == nil {
+			return
+		}
+
+		buf.Add(msg)
+	}
+}
+
+// utility functions
+
+func (p *asyncProducer) shutdown() {
+	Logger.Println("Producer shutting down.")
+	p.inFlight.Add(1)
+	p.input <- &ProducerMessage{flags: shutdown}
+
+	p.inFlight.Wait()
+
+	if p.ownClient {
+		err := p.client.Close()
+		if err != nil {
+			Logger.Println("producer/shutdown failed to close the embedded client:", err)
+		}
+	}
+
+	close(p.input)
+	close(p.retries)
+	close(p.errors)
+	close(p.successes)
+}
+
+func (p *asyncProducer) returnError(msg *ProducerMessage, err error) {
+	msg.clear()
+	pErr := &ProducerError{Msg: msg, Err: err}
+	if p.conf.Producer.Return.Errors {
+		p.errors <- pErr
+	} else {
+		Logger.Println(pErr)
+	}
+	p.inFlight.Done()
+}
+
+func (p *asyncProducer) returnErrors(batch []*ProducerMessage, err error) {
+	for _, msg := range batch {
+		p.returnError(msg, err)
+	}
+}
+
+func (p *asyncProducer) returnSuccesses(batch []*ProducerMessage) {
+	for _, msg := range batch {
+		if p.conf.Producer.Return.Successes {
+			msg.clear()
+			p.successes <- msg
+		}
+		p.inFlight.Done()
+	}
+}
+
+func (p *asyncProducer) retryMessage(msg *ProducerMessage, err error) {
+	if msg.retries >= p.conf.Producer.Retry.Max {
+		p.returnError(msg, err)
+	} else {
+		msg.retries++
+		p.retries <- msg
+	}
+}
+
+func (p *asyncProducer) retryMessages(batch []*ProducerMessage, err error) {
+	for _, msg := range batch {
+		p.retryMessage(msg, err)
+	}
+}
+
+func (p *asyncProducer) getBrokerProducer(broker *Broker) *brokerProducer {
+	p.brokerLock.Lock()
+	defer p.brokerLock.Unlock()
+
+	bp := p.brokers[broker]
+
+	if bp == nil {
+		bp = p.newBrokerProducer(broker)
+		p.brokers[broker] = bp
+		p.brokerRefs[bp] = 0
+	}
+
+	p.brokerRefs[bp]++
+
+	return bp
+}
+
+func (p *asyncProducer) unrefBrokerProducer(broker *Broker, bp *brokerProducer) {
+	p.brokerLock.Lock()
+	defer p.brokerLock.Unlock()
+
+	p.brokerRefs[bp]--
+	if p.brokerRefs[bp] == 0 {
+		close(bp.input)
+		delete(p.brokerRefs, bp)
+
+		if p.brokers[broker] == bp {
+			delete(p.brokers, broker)
+		}
+	}
+}
+
+func (p *asyncProducer) abandonBrokerConnection(broker *Broker) {
+	p.brokerLock.Lock()
+	defer p.brokerLock.Unlock()
+
+	bc, ok := p.brokers[broker]
+	if ok && bc.abandoned != nil {
+		close(bc.abandoned)
+	}
+
+	delete(p.brokers, broker)
+}
diff --git a/vendor/github.com/Shopify/sarama/balance_strategy.go b/vendor/github.com/Shopify/sarama/balance_strategy.go
new file mode 100644
index 0000000..e78988d
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/balance_strategy.go
@@ -0,0 +1,129 @@
+package sarama
+
+import (
+	"math"
+	"sort"
+)
+
+// BalanceStrategyPlan is the results of any BalanceStrategy.Plan attempt.
+// It contains an allocation of topic/partitions by memberID in the form of
+// a `memberID -> topic -> partitions` map.
+type BalanceStrategyPlan map[string]map[string][]int32
+
+// Add assigns a topic with a number partitions to a member.
+func (p BalanceStrategyPlan) Add(memberID, topic string, partitions ...int32) {
+	if len(partitions) == 0 {
+		return
+	}
+	if _, ok := p[memberID]; !ok {
+		p[memberID] = make(map[string][]int32, 1)
+	}
+	p[memberID][topic] = append(p[memberID][topic], partitions...)
+}
+
+// --------------------------------------------------------------------
+
+// BalanceStrategy is used to balance topics and partitions
+// across memebers of a consumer group
+type BalanceStrategy interface {
+	// Name uniquely identifies the strategy.
+	Name() string
+
+	// Plan accepts a map of `memberID -> metadata` and a map of `topic -> partitions`
+	// and returns a distribution plan.
+	Plan(members map[string]ConsumerGroupMemberMetadata, topics map[string][]int32) (BalanceStrategyPlan, error)
+}
+
+// --------------------------------------------------------------------
+
+// BalanceStrategyRange is the default and assigns partitions as ranges to consumer group members.
+// Example with one topic T with six partitions (0..5) and two members (M1, M2):
+//   M1: {T: [0, 1, 2]}
+//   M2: {T: [3, 4, 5]}
+var BalanceStrategyRange = &balanceStrategy{
+	name: "range",
+	coreFn: func(plan BalanceStrategyPlan, memberIDs []string, topic string, partitions []int32) {
+		step := float64(len(partitions)) / float64(len(memberIDs))
+
+		for i, memberID := range memberIDs {
+			pos := float64(i)
+			min := int(math.Floor(pos*step + 0.5))
+			max := int(math.Floor((pos+1)*step + 0.5))
+			plan.Add(memberID, topic, partitions[min:max]...)
+		}
+	},
+}
+
+// BalanceStrategyRoundRobin assigns partitions to members in alternating order.
+// Example with topic T with six partitions (0..5) and two members (M1, M2):
+//   M1: {T: [0, 2, 4]}
+//   M2: {T: [1, 3, 5]}
+var BalanceStrategyRoundRobin = &balanceStrategy{
+	name: "roundrobin",
+	coreFn: func(plan BalanceStrategyPlan, memberIDs []string, topic string, partitions []int32) {
+		for i, part := range partitions {
+			memberID := memberIDs[i%len(memberIDs)]
+			plan.Add(memberID, topic, part)
+		}
+	},
+}
+
+// --------------------------------------------------------------------
+
+type balanceStrategy struct {
+	name   string
+	coreFn func(plan BalanceStrategyPlan, memberIDs []string, topic string, partitions []int32)
+}
+
+// Name implements BalanceStrategy.
+func (s *balanceStrategy) Name() string { return s.name }
+
+// Balance implements BalanceStrategy.
+func (s *balanceStrategy) Plan(members map[string]ConsumerGroupMemberMetadata, topics map[string][]int32) (BalanceStrategyPlan, error) {
+	// Build members by topic map
+	mbt := make(map[string][]string)
+	for memberID, meta := range members {
+		for _, topic := range meta.Topics {
+			mbt[topic] = append(mbt[topic], memberID)
+		}
+	}
+
+	// Sort members for each topic
+	for topic, memberIDs := range mbt {
+		sort.Sort(&balanceStrategySortable{
+			topic:     topic,
+			memberIDs: memberIDs,
+		})
+	}
+
+	// Assemble plan
+	plan := make(BalanceStrategyPlan, len(members))
+	for topic, memberIDs := range mbt {
+		s.coreFn(plan, memberIDs, topic, topics[topic])
+	}
+	return plan, nil
+}
+
+type balanceStrategySortable struct {
+	topic     string
+	memberIDs []string
+}
+
+func (p balanceStrategySortable) Len() int { return len(p.memberIDs) }
+func (p balanceStrategySortable) Swap(i, j int) {
+	p.memberIDs[i], p.memberIDs[j] = p.memberIDs[j], p.memberIDs[i]
+}
+func (p balanceStrategySortable) Less(i, j int) bool {
+	return balanceStrategyHashValue(p.topic, p.memberIDs[i]) < balanceStrategyHashValue(p.topic, p.memberIDs[j])
+}
+
+func balanceStrategyHashValue(vv ...string) uint32 {
+	h := uint32(2166136261)
+	for _, s := range vv {
+		for _, c := range s {
+			h ^= uint32(c)
+			h *= 16777619
+		}
+	}
+	return h
+}
diff --git a/vendor/github.com/Shopify/sarama/broker.go b/vendor/github.com/Shopify/sarama/broker.go
new file mode 100644
index 0000000..9129089
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/broker.go
@@ -0,0 +1,1102 @@
+package sarama
+
+import (
+	"crypto/tls"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"net"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+// Broker represents a single Kafka broker connection. All operations on this object are entirely concurrency-safe.
+type Broker struct {
+	id   int32
+	addr string
+	rack *string
+
+	conf          *Config
+	correlationID int32
+	conn          net.Conn
+	connErr       error
+	lock          sync.Mutex
+	opened        int32
+
+	responses chan responsePromise
+	done      chan bool
+
+	incomingByteRate       metrics.Meter
+	requestRate            metrics.Meter
+	requestSize            metrics.Histogram
+	requestLatency         metrics.Histogram
+	outgoingByteRate       metrics.Meter
+	responseRate           metrics.Meter
+	responseSize           metrics.Histogram
+	brokerIncomingByteRate metrics.Meter
+	brokerRequestRate      metrics.Meter
+	brokerRequestSize      metrics.Histogram
+	brokerRequestLatency   metrics.Histogram
+	brokerOutgoingByteRate metrics.Meter
+	brokerResponseRate     metrics.Meter
+	brokerResponseSize     metrics.Histogram
+}
+
+// SASLMechanism specifies the SASL mechanism the client uses to authenticate with the broker
+type SASLMechanism string
+
+const (
+	// SASLTypeOAuth represents the SASL/OAUTHBEARER mechanism (Kafka 2.0.0+)
+	SASLTypeOAuth = "OAUTHBEARER"
+	// SASLTypePlaintext represents the SASL/PLAIN mechanism
+	SASLTypePlaintext = "PLAIN"
+	// SASLHandshakeV0 is v0 of the Kafka SASL handshake protocol. Client and
+	// server negotiate SASL auth using opaque packets.
+	SASLHandshakeV0 = int16(0)
+	// SASLHandshakeV1 is v1 of the Kafka SASL handshake protocol. Client and
+	// server negotiate SASL by wrapping tokens with Kafka protocol headers.
+	SASLHandshakeV1 = int16(1)
+	// SASLExtKeyAuth is the reserved extension key name sent as part of the
+	// SASL/OAUTHBEARER intial client response
+	SASLExtKeyAuth = "auth"
+)
+
+// AccessToken contains an access token used to authenticate a
+// SASL/OAUTHBEARER client along with associated metadata.
+type AccessToken struct {
+	// Token is the access token payload.
+	Token string
+	// Extensions is a optional map of arbitrary key-value pairs that can be
+	// sent with the SASL/OAUTHBEARER initial client response. These values are
+	// ignored by the SASL server if they are unexpected. This feature is only
+	// supported by Kafka >= 2.1.0.
+	Extensions map[string]string
+}
+
+// AccessTokenProvider is the interface that encapsulates how implementors
+// can generate access tokens for Kafka broker authentication.
+type AccessTokenProvider interface {
+	// Token returns an access token. The implementation should ensure token
+	// reuse so that multiple calls at connect time do not create multiple
+	// tokens. The implementation should also periodically refresh the token in
+	// order to guarantee that each call returns an unexpired token.  This
+	// method should not block indefinitely--a timeout error should be returned
+	// after a short period of inactivity so that the broker connection logic
+	// can log debugging information and retry.
+	Token() (*AccessToken, error)
+}
+
+type responsePromise struct {
+	requestTime   time.Time
+	correlationID int32
+	packets       chan []byte
+	errors        chan error
+}
+
+// NewBroker creates and returns a Broker targeting the given host:port address.
+// This does not attempt to actually connect, you have to call Open() for that.
+func NewBroker(addr string) *Broker {
+	return &Broker{id: -1, addr: addr}
+}
+
+// Open tries to connect to the Broker if it is not already connected or connecting, but does not block
+// waiting for the connection to complete. This means that any subsequent operations on the broker will
+// block waiting for the connection to succeed or fail. To get the effect of a fully synchronous Open call,
+// follow it by a call to Connected(). The only errors Open will return directly are ConfigurationError or
+// AlreadyConnected. If conf is nil, the result of NewConfig() is used.
+func (b *Broker) Open(conf *Config) error {
+	if !atomic.CompareAndSwapInt32(&b.opened, 0, 1) {
+		return ErrAlreadyConnected
+	}
+
+	if conf == nil {
+		conf = NewConfig()
+	}
+
+	err := conf.Validate()
+	if err != nil {
+		return err
+	}
+
+	b.lock.Lock()
+
+	go withRecover(func() {
+		defer b.lock.Unlock()
+
+		dialer := net.Dialer{
+			Timeout:   conf.Net.DialTimeout,
+			KeepAlive: conf.Net.KeepAlive,
+			LocalAddr: conf.Net.LocalAddr,
+		}
+
+		if conf.Net.TLS.Enable {
+			b.conn, b.connErr = tls.DialWithDialer(&dialer, "tcp", b.addr, conf.Net.TLS.Config)
+		} else {
+			b.conn, b.connErr = dialer.Dial("tcp", b.addr)
+		}
+		if b.connErr != nil {
+			Logger.Printf("Failed to connect to broker %s: %s\n", b.addr, b.connErr)
+			b.conn = nil
+			atomic.StoreInt32(&b.opened, 0)
+			return
+		}
+		b.conn = newBufConn(b.conn)
+
+		b.conf = conf
+
+		// Create or reuse the global metrics shared between brokers
+		b.incomingByteRate = metrics.GetOrRegisterMeter("incoming-byte-rate", conf.MetricRegistry)
+		b.requestRate = metrics.GetOrRegisterMeter("request-rate", conf.MetricRegistry)
+		b.requestSize = getOrRegisterHistogram("request-size", conf.MetricRegistry)
+		b.requestLatency = getOrRegisterHistogram("request-latency-in-ms", conf.MetricRegistry)
+		b.outgoingByteRate = metrics.GetOrRegisterMeter("outgoing-byte-rate", conf.MetricRegistry)
+		b.responseRate = metrics.GetOrRegisterMeter("response-rate", conf.MetricRegistry)
+		b.responseSize = getOrRegisterHistogram("response-size", conf.MetricRegistry)
+		// Do not gather metrics for seeded broker (only used during bootstrap) because they share
+		// the same id (-1) and are already exposed through the global metrics above
+		if b.id >= 0 {
+			b.brokerIncomingByteRate = getOrRegisterBrokerMeter("incoming-byte-rate", b, conf.MetricRegistry)
+			b.brokerRequestRate = getOrRegisterBrokerMeter("request-rate", b, conf.MetricRegistry)
+			b.brokerRequestSize = getOrRegisterBrokerHistogram("request-size", b, conf.MetricRegistry)
+			b.brokerRequestLatency = getOrRegisterBrokerHistogram("request-latency-in-ms", b, conf.MetricRegistry)
+			b.brokerOutgoingByteRate = getOrRegisterBrokerMeter("outgoing-byte-rate", b, conf.MetricRegistry)
+			b.brokerResponseRate = getOrRegisterBrokerMeter("response-rate", b, conf.MetricRegistry)
+			b.brokerResponseSize = getOrRegisterBrokerHistogram("response-size", b, conf.MetricRegistry)
+		}
+
+		if conf.Net.SASL.Enable {
+
+			b.connErr = b.authenticateViaSASL()
+
+			if b.connErr != nil {
+				err = b.conn.Close()
+				if err == nil {
+					Logger.Printf("Closed connection to broker %s\n", b.addr)
+				} else {
+					Logger.Printf("Error while closing connection to broker %s: %s\n", b.addr, err)
+				}
+				b.conn = nil
+				atomic.StoreInt32(&b.opened, 0)
+				return
+			}
+		}
+
+		b.done = make(chan bool)
+		b.responses = make(chan responsePromise, b.conf.Net.MaxOpenRequests-1)
+
+		if b.id >= 0 {
+			Logger.Printf("Connected to broker at %s (registered as #%d)\n", b.addr, b.id)
+		} else {
+			Logger.Printf("Connected to broker at %s (unregistered)\n", b.addr)
+		}
+		go withRecover(b.responseReceiver)
+	})
+
+	return nil
+}
+
+// Connected returns true if the broker is connected and false otherwise. If the broker is not
+// connected but it had tried to connect, the error from that connection attempt is also returned.
+func (b *Broker) Connected() (bool, error) {
+	b.lock.Lock()
+	defer b.lock.Unlock()
+
+	return b.conn != nil, b.connErr
+}
+
+func (b *Broker) Close() error {
+	b.lock.Lock()
+	defer b.lock.Unlock()
+
+	if b.conn == nil {
+		return ErrNotConnected
+	}
+
+	close(b.responses)
+	<-b.done
+
+	err := b.conn.Close()
+
+	b.conn = nil
+	b.connErr = nil
+	b.done = nil
+	b.responses = nil
+
+	if b.id >= 0 {
+		b.conf.MetricRegistry.Unregister(getMetricNameForBroker("incoming-byte-rate", b))
+		b.conf.MetricRegistry.Unregister(getMetricNameForBroker("request-rate", b))
+		b.conf.MetricRegistry.Unregister(getMetricNameForBroker("outgoing-byte-rate", b))
+		b.conf.MetricRegistry.Unregister(getMetricNameForBroker("response-rate", b))
+	}
+
+	if err == nil {
+		Logger.Printf("Closed connection to broker %s\n", b.addr)
+	} else {
+		Logger.Printf("Error while closing connection to broker %s: %s\n", b.addr, err)
+	}
+
+	atomic.StoreInt32(&b.opened, 0)
+
+	return err
+}
+
+// ID returns the broker ID retrieved from Kafka's metadata, or -1 if that is not known.
+func (b *Broker) ID() int32 {
+	return b.id
+}
+
+// Addr returns the broker address as either retrieved from Kafka's metadata or passed to NewBroker.
+func (b *Broker) Addr() string {
+	return b.addr
+}
+
+// Rack returns the broker's rack as retrieved from Kafka's metadata or the
+// empty string if it is not known.  The returned value corresponds to the
+// broker's broker.rack configuration setting.  Requires protocol version to be
+// at least v0.10.0.0.
+func (b *Broker) Rack() string {
+	if b.rack == nil {
+		return ""
+	}
+	return *b.rack
+}
+
+func (b *Broker) GetMetadata(request *MetadataRequest) (*MetadataResponse, error) {
+	response := new(MetadataResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) GetConsumerMetadata(request *ConsumerMetadataRequest) (*ConsumerMetadataResponse, error) {
+	response := new(ConsumerMetadataResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) FindCoordinator(request *FindCoordinatorRequest) (*FindCoordinatorResponse, error) {
+	response := new(FindCoordinatorResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) GetAvailableOffsets(request *OffsetRequest) (*OffsetResponse, error) {
+	response := new(OffsetResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) Produce(request *ProduceRequest) (*ProduceResponse, error) {
+	var response *ProduceResponse
+	var err error
+
+	if request.RequiredAcks == NoResponse {
+		err = b.sendAndReceive(request, nil)
+	} else {
+		response = new(ProduceResponse)
+		err = b.sendAndReceive(request, response)
+	}
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) Fetch(request *FetchRequest) (*FetchResponse, error) {
+	response := new(FetchResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) CommitOffset(request *OffsetCommitRequest) (*OffsetCommitResponse, error) {
+	response := new(OffsetCommitResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) FetchOffset(request *OffsetFetchRequest) (*OffsetFetchResponse, error) {
+	response := new(OffsetFetchResponse)
+
+	err := b.sendAndReceive(request, response)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) JoinGroup(request *JoinGroupRequest) (*JoinGroupResponse, error) {
+	response := new(JoinGroupResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) SyncGroup(request *SyncGroupRequest) (*SyncGroupResponse, error) {
+	response := new(SyncGroupResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) LeaveGroup(request *LeaveGroupRequest) (*LeaveGroupResponse, error) {
+	response := new(LeaveGroupResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) Heartbeat(request *HeartbeatRequest) (*HeartbeatResponse, error) {
+	response := new(HeartbeatResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) ListGroups(request *ListGroupsRequest) (*ListGroupsResponse, error) {
+	response := new(ListGroupsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DescribeGroups(request *DescribeGroupsRequest) (*DescribeGroupsResponse, error) {
+	response := new(DescribeGroupsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) ApiVersions(request *ApiVersionsRequest) (*ApiVersionsResponse, error) {
+	response := new(ApiVersionsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) CreateTopics(request *CreateTopicsRequest) (*CreateTopicsResponse, error) {
+	response := new(CreateTopicsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DeleteTopics(request *DeleteTopicsRequest) (*DeleteTopicsResponse, error) {
+	response := new(DeleteTopicsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) CreatePartitions(request *CreatePartitionsRequest) (*CreatePartitionsResponse, error) {
+	response := new(CreatePartitionsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DeleteRecords(request *DeleteRecordsRequest) (*DeleteRecordsResponse, error) {
+	response := new(DeleteRecordsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DescribeAcls(request *DescribeAclsRequest) (*DescribeAclsResponse, error) {
+	response := new(DescribeAclsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) CreateAcls(request *CreateAclsRequest) (*CreateAclsResponse, error) {
+	response := new(CreateAclsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DeleteAcls(request *DeleteAclsRequest) (*DeleteAclsResponse, error) {
+	response := new(DeleteAclsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) InitProducerID(request *InitProducerIDRequest) (*InitProducerIDResponse, error) {
+	response := new(InitProducerIDResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) AddPartitionsToTxn(request *AddPartitionsToTxnRequest) (*AddPartitionsToTxnResponse, error) {
+	response := new(AddPartitionsToTxnResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) AddOffsetsToTxn(request *AddOffsetsToTxnRequest) (*AddOffsetsToTxnResponse, error) {
+	response := new(AddOffsetsToTxnResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) EndTxn(request *EndTxnRequest) (*EndTxnResponse, error) {
+	response := new(EndTxnResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) TxnOffsetCommit(request *TxnOffsetCommitRequest) (*TxnOffsetCommitResponse, error) {
+	response := new(TxnOffsetCommitResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DescribeConfigs(request *DescribeConfigsRequest) (*DescribeConfigsResponse, error) {
+	response := new(DescribeConfigsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) AlterConfigs(request *AlterConfigsRequest) (*AlterConfigsResponse, error) {
+	response := new(AlterConfigsResponse)
+
+	err := b.sendAndReceive(request, response)
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) DeleteGroups(request *DeleteGroupsRequest) (*DeleteGroupsResponse, error) {
+	response := new(DeleteGroupsResponse)
+
+	if err := b.sendAndReceive(request, response); err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func (b *Broker) send(rb protocolBody, promiseResponse bool) (*responsePromise, error) {
+	b.lock.Lock()
+	defer b.lock.Unlock()
+
+	if b.conn == nil {
+		if b.connErr != nil {
+			return nil, b.connErr
+		}
+		return nil, ErrNotConnected
+	}
+
+	if !b.conf.Version.IsAtLeast(rb.requiredVersion()) {
+		return nil, ErrUnsupportedVersion
+	}
+
+	req := &request{correlationID: b.correlationID, clientID: b.conf.ClientID, body: rb}
+	buf, err := encode(req, b.conf.MetricRegistry)
+	if err != nil {
+		return nil, err
+	}
+
+	err = b.conn.SetWriteDeadline(time.Now().Add(b.conf.Net.WriteTimeout))
+	if err != nil {
+		return nil, err
+	}
+
+	requestTime := time.Now()
+	bytes, err := b.conn.Write(buf)
+	b.updateOutgoingCommunicationMetrics(bytes)
+	if err != nil {
+		return nil, err
+	}
+	b.correlationID++
+
+	if !promiseResponse {
+		// Record request latency without the response
+		b.updateRequestLatencyMetrics(time.Since(requestTime))
+		return nil, nil
+	}
+
+	promise := responsePromise{requestTime, req.correlationID, make(chan []byte), make(chan error)}
+	b.responses <- promise
+
+	return &promise, nil
+}
+
+func (b *Broker) sendAndReceive(req protocolBody, res versionedDecoder) error {
+	promise, err := b.send(req, res != nil)
+
+	if err != nil {
+		return err
+	}
+
+	if promise == nil {
+		return nil
+	}
+
+	select {
+	case buf := <-promise.packets:
+		return versionedDecode(buf, res, req.version())
+	case err = <-promise.errors:
+		return err
+	}
+}
+
+func (b *Broker) decode(pd packetDecoder, version int16) (err error) {
+	b.id, err = pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	host, err := pd.getString()
+	if err != nil {
+		return err
+	}
+
+	port, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	if version >= 1 {
+		b.rack, err = pd.getNullableString()
+		if err != nil {
+			return err
+		}
+	}
+
+	b.addr = net.JoinHostPort(host, fmt.Sprint(port))
+	if _, _, err := net.SplitHostPort(b.addr); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (b *Broker) encode(pe packetEncoder, version int16) (err error) {
+
+	host, portstr, err := net.SplitHostPort(b.addr)
+	if err != nil {
+		return err
+	}
+	port, err := strconv.Atoi(portstr)
+	if err != nil {
+		return err
+	}
+
+	pe.putInt32(b.id)
+
+	err = pe.putString(host)
+	if err != nil {
+		return err
+	}
+
+	pe.putInt32(int32(port))
+
+	if version >= 1 {
+		err = pe.putNullableString(b.rack)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (b *Broker) responseReceiver() {
+	var dead error
+	header := make([]byte, 8)
+	for response := range b.responses {
+		if dead != nil {
+			response.errors <- dead
+			continue
+		}
+
+		err := b.conn.SetReadDeadline(time.Now().Add(b.conf.Net.ReadTimeout))
+		if err != nil {
+			dead = err
+			response.errors <- err
+			continue
+		}
+
+		bytesReadHeader, err := io.ReadFull(b.conn, header)
+		requestLatency := time.Since(response.requestTime)
+		if err != nil {
+			b.updateIncomingCommunicationMetrics(bytesReadHeader, requestLatency)
+			dead = err
+			response.errors <- err
+			continue
+		}
+
+		decodedHeader := responseHeader{}
+		err = decode(header, &decodedHeader)
+		if err != nil {
+			b.updateIncomingCommunicationMetrics(bytesReadHeader, requestLatency)
+			dead = err
+			response.errors <- err
+			continue
+		}
+		if decodedHeader.correlationID != response.correlationID {
+			b.updateIncomingCommunicationMetrics(bytesReadHeader, requestLatency)
+			// TODO if decoded ID < cur ID, discard until we catch up
+			// TODO if decoded ID > cur ID, save it so when cur ID catches up we have a response
+			dead = PacketDecodingError{fmt.Sprintf("correlation ID didn't match, wanted %d, got %d", response.correlationID, decodedHeader.correlationID)}
+			response.errors <- dead
+			continue
+		}
+
+		buf := make([]byte, decodedHeader.length-4)
+		bytesReadBody, err := io.ReadFull(b.conn, buf)
+		b.updateIncomingCommunicationMetrics(bytesReadHeader+bytesReadBody, requestLatency)
+		if err != nil {
+			dead = err
+			response.errors <- err
+			continue
+		}
+
+		response.packets <- buf
+	}
+	close(b.done)
+}
+
+func (b *Broker) authenticateViaSASL() error {
+	if b.conf.Net.SASL.Mechanism == SASLTypeOAuth {
+		return b.sendAndReceiveSASLOAuth(b.conf.Net.SASL.TokenProvider)
+	}
+	return b.sendAndReceiveSASLPlainAuth()
+}
+
+func (b *Broker) sendAndReceiveSASLHandshake(saslType string, version int16) error {
+	rb := &SaslHandshakeRequest{Mechanism: saslType, Version: version}
+
+	req := &request{correlationID: b.correlationID, clientID: b.conf.ClientID, body: rb}
+	buf, err := encode(req, b.conf.MetricRegistry)
+	if err != nil {
+		return err
+	}
+
+	err = b.conn.SetWriteDeadline(time.Now().Add(b.conf.Net.WriteTimeout))
+	if err != nil {
+		return err
+	}
+
+	requestTime := time.Now()
+	bytes, err := b.conn.Write(buf)
+	b.updateOutgoingCommunicationMetrics(bytes)
+	if err != nil {
+		Logger.Printf("Failed to send SASL handshake %s: %s\n", b.addr, err.Error())
+		return err
+	}
+	b.correlationID++
+	//wait for the response
+	header := make([]byte, 8) // response header
+	_, err = io.ReadFull(b.conn, header)
+	if err != nil {
+		Logger.Printf("Failed to read SASL handshake header : %s\n", err.Error())
+		return err
+	}
+	length := binary.BigEndian.Uint32(header[:4])
+	payload := make([]byte, length-4)
+	n, err := io.ReadFull(b.conn, payload)
+	if err != nil {
+		Logger.Printf("Failed to read SASL handshake payload : %s\n", err.Error())
+		return err
+	}
+	b.updateIncomingCommunicationMetrics(n+8, time.Since(requestTime))
+	res := &SaslHandshakeResponse{}
+	err = versionedDecode(payload, res, 0)
+	if err != nil {
+		Logger.Printf("Failed to parse SASL handshake : %s\n", err.Error())
+		return err
+	}
+	if res.Err != ErrNoError {
+		Logger.Printf("Invalid SASL Mechanism : %s\n", res.Err.Error())
+		return res.Err
+	}
+	Logger.Print("Successful SASL handshake")
+	return nil
+}
+
+// Kafka 0.10.0 plans to support SASL Plain and Kerberos as per PR #812 (KIP-43)/(JIRA KAFKA-3149)
+// Some hosted kafka services such as IBM Message Hub already offer SASL/PLAIN auth with Kafka 0.9
+//
+// In SASL Plain, Kafka expects the auth header to be in the following format
+// Message format (from https://tools.ietf.org/html/rfc4616):
+//
+//   message   = [authzid] UTF8NUL authcid UTF8NUL passwd
+//   authcid   = 1*SAFE ; MUST accept up to 255 octets
+//   authzid   = 1*SAFE ; MUST accept up to 255 octets
+//   passwd    = 1*SAFE ; MUST accept up to 255 octets
+//   UTF8NUL   = %x00 ; UTF-8 encoded NUL character
+//
+//   SAFE      = UTF1 / UTF2 / UTF3 / UTF4
+//                  ;; any UTF-8 encoded Unicode character except NUL
+//
+// When credentials are valid, Kafka returns a 4 byte array of null characters.
+// When credentials are invalid, Kafka closes the connection. This does not seem to be the ideal way
+// of responding to bad credentials but thats how its being done today.
+func (b *Broker) sendAndReceiveSASLPlainAuth() error {
+	if b.conf.Net.SASL.Handshake {
+		handshakeErr := b.sendAndReceiveSASLHandshake(SASLTypePlaintext, SASLHandshakeV0)
+		if handshakeErr != nil {
+			Logger.Printf("Error while performing SASL handshake %s\n", b.addr)
+			return handshakeErr
+		}
+	}
+	length := 1 + len(b.conf.Net.SASL.User) + 1 + len(b.conf.Net.SASL.Password)
+	authBytes := make([]byte, length+4) //4 byte length header + auth data
+	binary.BigEndian.PutUint32(authBytes, uint32(length))
+	copy(authBytes[4:], []byte("\x00"+b.conf.Net.SASL.User+"\x00"+b.conf.Net.SASL.Password))
+
+	err := b.conn.SetWriteDeadline(time.Now().Add(b.conf.Net.WriteTimeout))
+	if err != nil {
+		Logger.Printf("Failed to set write deadline when doing SASL auth with broker %s: %s\n", b.addr, err.Error())
+		return err
+	}
+
+	requestTime := time.Now()
+	bytesWritten, err := b.conn.Write(authBytes)
+	b.updateOutgoingCommunicationMetrics(bytesWritten)
+	if err != nil {
+		Logger.Printf("Failed to write SASL auth header to broker %s: %s\n", b.addr, err.Error())
+		return err
+	}
+
+	header := make([]byte, 4)
+	n, err := io.ReadFull(b.conn, header)
+	b.updateIncomingCommunicationMetrics(n, time.Since(requestTime))
+	// If the credentials are valid, we would get a 4 byte response filled with null characters.
+	// Otherwise, the broker closes the connection and we get an EOF
+	if err != nil {
+		Logger.Printf("Failed to read response while authenticating with SASL to broker %s: %s\n", b.addr, err.Error())
+		return err
+	}
+
+	Logger.Printf("SASL authentication successful with broker %s:%v - %v\n", b.addr, n, header)
+	return nil
+}
+
+// sendAndReceiveSASLOAuth performs the authentication flow as described by KIP-255
+// https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=75968876
+func (b *Broker) sendAndReceiveSASLOAuth(provider AccessTokenProvider) error {
+
+	if err := b.sendAndReceiveSASLHandshake(SASLTypeOAuth, SASLHandshakeV1); err != nil {
+		return err
+	}
+
+	token, err := provider.Token()
+
+	if err != nil {
+		return err
+	}
+
+	requestTime := time.Now()
+
+	correlationID := b.correlationID
+
+	bytesWritten, err := b.sendSASLOAuthBearerClientResponse(token, correlationID)
+
+	if err != nil {
+		return err
+	}
+
+	b.updateOutgoingCommunicationMetrics(bytesWritten)
+
+	b.correlationID++
+
+	bytesRead, err := b.receiveSASLOAuthBearerServerResponse(correlationID)
+
+	if err != nil {
+		return err
+	}
+
+	requestLatency := time.Since(requestTime)
+	b.updateIncomingCommunicationMetrics(bytesRead, requestLatency)
+
+	return nil
+}
+
+// Build SASL/OAUTHBEARER initial client response as described by RFC-7628
+// https://tools.ietf.org/html/rfc7628
+func buildClientInitialResponse(token *AccessToken) ([]byte, error) {
+
+	var ext string
+
+	if token.Extensions != nil && len(token.Extensions) > 0 {
+		if _, ok := token.Extensions[SASLExtKeyAuth]; ok {
+			return []byte{}, fmt.Errorf("The extension `%s` is invalid", SASLExtKeyAuth)
+		}
+		ext = "\x01" + mapToString(token.Extensions, "=", "\x01")
+	}
+
+	resp := []byte(fmt.Sprintf("n,,\x01auth=Bearer %s%s\x01\x01", token.Token, ext))
+
+	return resp, nil
+}
+
+// mapToString returns a list of key-value pairs ordered by key.
+// keyValSep separates the key from the value. elemSep separates each pair.
+func mapToString(extensions map[string]string, keyValSep string, elemSep string) string {
+
+	buf := make([]string, 0, len(extensions))
+
+	for k, v := range extensions {
+		buf = append(buf, k+keyValSep+v)
+	}
+
+	sort.Strings(buf)
+
+	return strings.Join(buf, elemSep)
+}
+
+func (b *Broker) sendSASLOAuthBearerClientResponse(token *AccessToken, correlationID int32) (int, error) {
+
+	initialResp, err := buildClientInitialResponse(token)
+
+	if err != nil {
+		return 0, err
+	}
+
+	rb := &SaslAuthenticateRequest{initialResp}
+
+	req := &request{correlationID: correlationID, clientID: b.conf.ClientID, body: rb}
+
+	buf, err := encode(req, b.conf.MetricRegistry)
+
+	if err != nil {
+		return 0, err
+	}
+
+	if err := b.conn.SetWriteDeadline(time.Now().Add(b.conf.Net.WriteTimeout)); err != nil {
+		return 0, err
+	}
+
+	return b.conn.Write(buf)
+}
+
+func (b *Broker) receiveSASLOAuthBearerServerResponse(correlationID int32) (int, error) {
+
+	buf := make([]byte, 8)
+
+	bytesRead, err := io.ReadFull(b.conn, buf)
+
+	if err != nil {
+		return bytesRead, err
+	}
+
+	header := responseHeader{}
+
+	err = decode(buf, &header)
+
+	if err != nil {
+		return bytesRead, err
+	}
+
+	if header.correlationID != correlationID {
+		return bytesRead, fmt.Errorf("correlation ID didn't match, wanted %d, got %d", b.correlationID, header.correlationID)
+	}
+
+	buf = make([]byte, header.length-4)
+
+	c, err := io.ReadFull(b.conn, buf)
+
+	bytesRead += c
+
+	if err != nil {
+		return bytesRead, err
+	}
+
+	res := &SaslAuthenticateResponse{}
+
+	if err := versionedDecode(buf, res, 0); err != nil {
+		return bytesRead, err
+	}
+
+	if err != nil {
+		return bytesRead, err
+	}
+
+	if res.Err != ErrNoError {
+		return bytesRead, res.Err
+	}
+
+	if len(res.SaslAuthBytes) > 0 {
+		Logger.Printf("Received SASL auth response: %s", res.SaslAuthBytes)
+	}
+
+	return bytesRead, nil
+}
+
+func (b *Broker) updateIncomingCommunicationMetrics(bytes int, requestLatency time.Duration) {
+	b.updateRequestLatencyMetrics(requestLatency)
+	b.responseRate.Mark(1)
+	if b.brokerResponseRate != nil {
+		b.brokerResponseRate.Mark(1)
+	}
+	responseSize := int64(bytes)
+	b.incomingByteRate.Mark(responseSize)
+	if b.brokerIncomingByteRate != nil {
+		b.brokerIncomingByteRate.Mark(responseSize)
+	}
+	b.responseSize.Update(responseSize)
+	if b.brokerResponseSize != nil {
+		b.brokerResponseSize.Update(responseSize)
+	}
+}
+
+func (b *Broker) updateRequestLatencyMetrics(requestLatency time.Duration) {
+	requestLatencyInMs := int64(requestLatency / time.Millisecond)
+	b.requestLatency.Update(requestLatencyInMs)
+	if b.brokerRequestLatency != nil {
+		b.brokerRequestLatency.Update(requestLatencyInMs)
+	}
+}
+
+func (b *Broker) updateOutgoingCommunicationMetrics(bytes int) {
+	b.requestRate.Mark(1)
+	if b.brokerRequestRate != nil {
+		b.brokerRequestRate.Mark(1)
+	}
+	requestSize := int64(bytes)
+	b.outgoingByteRate.Mark(requestSize)
+	if b.brokerOutgoingByteRate != nil {
+		b.brokerOutgoingByteRate.Mark(requestSize)
+	}
+	b.requestSize.Update(requestSize)
+	if b.brokerRequestSize != nil {
+		b.brokerRequestSize.Update(requestSize)
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/client.go b/vendor/github.com/Shopify/sarama/client.go
new file mode 100644
index 0000000..0016f8f
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/client.go
@@ -0,0 +1,913 @@
+package sarama
+
+import (
+	"math/rand"
+	"sort"
+	"sync"
+	"time"
+)
+
+// Client is a generic Kafka client. It manages connections to one or more Kafka brokers.
+// You MUST call Close() on a client to avoid leaks, it will not be garbage-collected
+// automatically when it passes out of scope. It is safe to share a client amongst many
+// users, however Kafka will process requests from a single client strictly in serial,
+// so it is generally more efficient to use the default one client per producer/consumer.
+type Client interface {
+	// Config returns the Config struct of the client. This struct should not be
+	// altered after it has been created.
+	Config() *Config
+
+	// Controller returns the cluster controller broker. Requires Kafka 0.10 or higher.
+	Controller() (*Broker, error)
+
+	// Brokers returns the current set of active brokers as retrieved from cluster metadata.
+	Brokers() []*Broker
+
+	// Topics returns the set of available topics as retrieved from cluster metadata.
+	Topics() ([]string, error)
+
+	// Partitions returns the sorted list of all partition IDs for the given topic.
+	Partitions(topic string) ([]int32, error)
+
+	// WritablePartitions returns the sorted list of all writable partition IDs for
+	// the given topic, where "writable" means "having a valid leader accepting
+	// writes".
+	WritablePartitions(topic string) ([]int32, error)
+
+	// Leader returns the broker object that is the leader of the current
+	// topic/partition, as determined by querying the cluster metadata.
+	Leader(topic string, partitionID int32) (*Broker, error)
+
+	// Replicas returns the set of all replica IDs for the given partition.
+	Replicas(topic string, partitionID int32) ([]int32, error)
+
+	// InSyncReplicas returns the set of all in-sync replica IDs for the given
+	// partition. In-sync replicas are replicas which are fully caught up with
+	// the partition leader.
+	InSyncReplicas(topic string, partitionID int32) ([]int32, error)
+
+	// RefreshMetadata takes a list of topics and queries the cluster to refresh the
+	// available metadata for those topics. If no topics are provided, it will refresh
+	// metadata for all topics.
+	RefreshMetadata(topics ...string) error
+
+	// GetOffset queries the cluster to get the most recent available offset at the
+	// given time (in milliseconds) on the topic/partition combination.
+	// Time should be OffsetOldest for the earliest available offset,
+	// OffsetNewest for the offset of the message that will be produced next, or a time.
+	GetOffset(topic string, partitionID int32, time int64) (int64, error)
+
+	// Coordinator returns the coordinating broker for a consumer group. It will
+	// return a locally cached value if it's available. You can call
+	// RefreshCoordinator to update the cached value. This function only works on
+	// Kafka 0.8.2 and higher.
+	Coordinator(consumerGroup string) (*Broker, error)
+
+	// RefreshCoordinator retrieves the coordinator for a consumer group and stores it
+	// in local cache. This function only works on Kafka 0.8.2 and higher.
+	RefreshCoordinator(consumerGroup string) error
+
+	// InitProducerID retrieves information required for Idempotent Producer
+	InitProducerID() (*InitProducerIDResponse, error)
+
+	// Close shuts down all broker connections managed by this client. It is required
+	// to call this function before a client object passes out of scope, as it will
+	// otherwise leak memory. You must close any Producers or Consumers using a client
+	// before you close the client.
+	Close() error
+
+	// Closed returns true if the client has already had Close called on it
+	Closed() bool
+}
+
+const (
+	// OffsetNewest stands for the log head offset, i.e. the offset that will be
+	// assigned to the next message that will be produced to the partition. You
+	// can send this to a client's GetOffset method to get this offset, or when
+	// calling ConsumePartition to start consuming new messages.
+	OffsetNewest int64 = -1
+	// OffsetOldest stands for the oldest offset available on the broker for a
+	// partition. You can send this to a client's GetOffset method to get this
+	// offset, or when calling ConsumePartition to start consuming from the
+	// oldest offset that is still available on the broker.
+	OffsetOldest int64 = -2
+)
+
+type client struct {
+	conf           *Config
+	closer, closed chan none // for shutting down background metadata updater
+
+	// the broker addresses given to us through the constructor are not guaranteed to be returned in
+	// the cluster metadata (I *think* it only returns brokers who are currently leading partitions?)
+	// so we store them separately
+	seedBrokers []*Broker
+	deadSeeds   []*Broker
+
+	controllerID   int32                                   // cluster controller broker id
+	brokers        map[int32]*Broker                       // maps broker ids to brokers
+	metadata       map[string]map[int32]*PartitionMetadata // maps topics to partition ids to metadata
+	metadataTopics map[string]none                         // topics that need to collect metadata
+	coordinators   map[string]int32                        // Maps consumer group names to coordinating broker IDs
+
+	// If the number of partitions is large, we can get some churn calling cachedPartitions,
+	// so the result is cached.  It is important to update this value whenever metadata is changed
+	cachedPartitionsResults map[string][maxPartitionIndex][]int32
+
+	lock sync.RWMutex // protects access to the maps that hold cluster state.
+}
+
+// NewClient creates a new Client. It connects to one of the given broker addresses
+// and uses that broker to automatically fetch metadata on the rest of the kafka cluster. If metadata cannot
+// be retrieved from any of the given broker addresses, the client is not created.
+func NewClient(addrs []string, conf *Config) (Client, error) {
+	Logger.Println("Initializing new client")
+
+	if conf == nil {
+		conf = NewConfig()
+	}
+
+	if err := conf.Validate(); err != nil {
+		return nil, err
+	}
+
+	if len(addrs) < 1 {
+		return nil, ConfigurationError("You must provide at least one broker address")
+	}
+
+	client := &client{
+		conf:                    conf,
+		closer:                  make(chan none),
+		closed:                  make(chan none),
+		brokers:                 make(map[int32]*Broker),
+		metadata:                make(map[string]map[int32]*PartitionMetadata),
+		metadataTopics:          make(map[string]none),
+		cachedPartitionsResults: make(map[string][maxPartitionIndex][]int32),
+		coordinators:            make(map[string]int32),
+	}
+
+	random := rand.New(rand.NewSource(time.Now().UnixNano()))
+	for _, index := range random.Perm(len(addrs)) {
+		client.seedBrokers = append(client.seedBrokers, NewBroker(addrs[index]))
+	}
+
+	if conf.Metadata.Full {
+		// do an initial fetch of all cluster metadata by specifying an empty list of topics
+		err := client.RefreshMetadata()
+		switch err {
+		case nil:
+			break
+		case ErrLeaderNotAvailable, ErrReplicaNotAvailable, ErrTopicAuthorizationFailed, ErrClusterAuthorizationFailed:
+			// indicates that maybe part of the cluster is down, but is not fatal to creating the client
+			Logger.Println(err)
+		default:
+			close(client.closed) // we haven't started the background updater yet, so we have to do this manually
+			_ = client.Close()
+			return nil, err
+		}
+	}
+	go withRecover(client.backgroundMetadataUpdater)
+
+	Logger.Println("Successfully initialized new client")
+
+	return client, nil
+}
+
+func (client *client) Config() *Config {
+	return client.conf
+}
+
+func (client *client) Brokers() []*Broker {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+	brokers := make([]*Broker, 0, len(client.brokers))
+	for _, broker := range client.brokers {
+		brokers = append(brokers, broker)
+	}
+	return brokers
+}
+
+func (client *client) InitProducerID() (*InitProducerIDResponse, error) {
+	var err error
+	for broker := client.any(); broker != nil; broker = client.any() {
+
+		req := &InitProducerIDRequest{}
+
+		response, err := broker.InitProducerID(req)
+		switch err.(type) {
+		case nil:
+			return response, nil
+		default:
+			// some error, remove that broker and try again
+			Logger.Printf("Client got error from broker %d when issuing InitProducerID : %v\n", broker.ID(), err)
+			_ = broker.Close()
+			client.deregisterBroker(broker)
+		}
+	}
+	return nil, err
+}
+
+func (client *client) Close() error {
+	if client.Closed() {
+		// Chances are this is being called from a defer() and the error will go unobserved
+		// so we go ahead and log the event in this case.
+		Logger.Printf("Close() called on already closed client")
+		return ErrClosedClient
+	}
+
+	// shutdown and wait for the background thread before we take the lock, to avoid races
+	close(client.closer)
+	<-client.closed
+
+	client.lock.Lock()
+	defer client.lock.Unlock()
+	Logger.Println("Closing Client")
+
+	for _, broker := range client.brokers {
+		safeAsyncClose(broker)
+	}
+
+	for _, broker := range client.seedBrokers {
+		safeAsyncClose(broker)
+	}
+
+	client.brokers = nil
+	client.metadata = nil
+	client.metadataTopics = nil
+
+	return nil
+}
+
+func (client *client) Closed() bool {
+	return client.brokers == nil
+}
+
+func (client *client) Topics() ([]string, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	ret := make([]string, 0, len(client.metadata))
+	for topic := range client.metadata {
+		ret = append(ret, topic)
+	}
+
+	return ret, nil
+}
+
+func (client *client) MetadataTopics() ([]string, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	ret := make([]string, 0, len(client.metadataTopics))
+	for topic := range client.metadataTopics {
+		ret = append(ret, topic)
+	}
+
+	return ret, nil
+}
+
+func (client *client) Partitions(topic string) ([]int32, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	partitions := client.cachedPartitions(topic, allPartitions)
+
+	if len(partitions) == 0 {
+		err := client.RefreshMetadata(topic)
+		if err != nil {
+			return nil, err
+		}
+		partitions = client.cachedPartitions(topic, allPartitions)
+	}
+
+	if partitions == nil {
+		return nil, ErrUnknownTopicOrPartition
+	}
+
+	return partitions, nil
+}
+
+func (client *client) WritablePartitions(topic string) ([]int32, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	partitions := client.cachedPartitions(topic, writablePartitions)
+
+	// len==0 catches when it's nil (no such topic) and the odd case when every single
+	// partition is undergoing leader election simultaneously. Callers have to be able to handle
+	// this function returning an empty slice (which is a valid return value) but catching it
+	// here the first time (note we *don't* catch it below where we return ErrUnknownTopicOrPartition) triggers
+	// a metadata refresh as a nicety so callers can just try again and don't have to manually
+	// trigger a refresh (otherwise they'd just keep getting a stale cached copy).
+	if len(partitions) == 0 {
+		err := client.RefreshMetadata(topic)
+		if err != nil {
+			return nil, err
+		}
+		partitions = client.cachedPartitions(topic, writablePartitions)
+	}
+
+	if partitions == nil {
+		return nil, ErrUnknownTopicOrPartition
+	}
+
+	return partitions, nil
+}
+
+func (client *client) Replicas(topic string, partitionID int32) ([]int32, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	metadata := client.cachedMetadata(topic, partitionID)
+
+	if metadata == nil {
+		err := client.RefreshMetadata(topic)
+		if err != nil {
+			return nil, err
+		}
+		metadata = client.cachedMetadata(topic, partitionID)
+	}
+
+	if metadata == nil {
+		return nil, ErrUnknownTopicOrPartition
+	}
+
+	if metadata.Err == ErrReplicaNotAvailable {
+		return dupInt32Slice(metadata.Replicas), metadata.Err
+	}
+	return dupInt32Slice(metadata.Replicas), nil
+}
+
+func (client *client) InSyncReplicas(topic string, partitionID int32) ([]int32, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	metadata := client.cachedMetadata(topic, partitionID)
+
+	if metadata == nil {
+		err := client.RefreshMetadata(topic)
+		if err != nil {
+			return nil, err
+		}
+		metadata = client.cachedMetadata(topic, partitionID)
+	}
+
+	if metadata == nil {
+		return nil, ErrUnknownTopicOrPartition
+	}
+
+	if metadata.Err == ErrReplicaNotAvailable {
+		return dupInt32Slice(metadata.Isr), metadata.Err
+	}
+	return dupInt32Slice(metadata.Isr), nil
+}
+
+func (client *client) Leader(topic string, partitionID int32) (*Broker, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	leader, err := client.cachedLeader(topic, partitionID)
+
+	if leader == nil {
+		err = client.RefreshMetadata(topic)
+		if err != nil {
+			return nil, err
+		}
+		leader, err = client.cachedLeader(topic, partitionID)
+	}
+
+	return leader, err
+}
+
+func (client *client) RefreshMetadata(topics ...string) error {
+	if client.Closed() {
+		return ErrClosedClient
+	}
+
+	// Prior to 0.8.2, Kafka will throw exceptions on an empty topic and not return a proper
+	// error. This handles the case by returning an error instead of sending it
+	// off to Kafka. See: https://github.com/Shopify/sarama/pull/38#issuecomment-26362310
+	for _, topic := range topics {
+		if len(topic) == 0 {
+			return ErrInvalidTopic // this is the error that 0.8.2 and later correctly return
+		}
+	}
+
+	return client.tryRefreshMetadata(topics, client.conf.Metadata.Retry.Max)
+}
+
+func (client *client) GetOffset(topic string, partitionID int32, time int64) (int64, error) {
+	if client.Closed() {
+		return -1, ErrClosedClient
+	}
+
+	offset, err := client.getOffset(topic, partitionID, time)
+
+	if err != nil {
+		if err := client.RefreshMetadata(topic); err != nil {
+			return -1, err
+		}
+		return client.getOffset(topic, partitionID, time)
+	}
+
+	return offset, err
+}
+
+func (client *client) Controller() (*Broker, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	if !client.conf.Version.IsAtLeast(V0_10_0_0) {
+		return nil, ErrUnsupportedVersion
+	}
+
+	controller := client.cachedController()
+	if controller == nil {
+		if err := client.refreshMetadata(); err != nil {
+			return nil, err
+		}
+		controller = client.cachedController()
+	}
+
+	if controller == nil {
+		return nil, ErrControllerNotAvailable
+	}
+
+	_ = controller.Open(client.conf)
+	return controller, nil
+}
+
+func (client *client) Coordinator(consumerGroup string) (*Broker, error) {
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	coordinator := client.cachedCoordinator(consumerGroup)
+
+	if coordinator == nil {
+		if err := client.RefreshCoordinator(consumerGroup); err != nil {
+			return nil, err
+		}
+		coordinator = client.cachedCoordinator(consumerGroup)
+	}
+
+	if coordinator == nil {
+		return nil, ErrConsumerCoordinatorNotAvailable
+	}
+
+	_ = coordinator.Open(client.conf)
+	return coordinator, nil
+}
+
+func (client *client) RefreshCoordinator(consumerGroup string) error {
+	if client.Closed() {
+		return ErrClosedClient
+	}
+
+	response, err := client.getConsumerMetadata(consumerGroup, client.conf.Metadata.Retry.Max)
+	if err != nil {
+		return err
+	}
+
+	client.lock.Lock()
+	defer client.lock.Unlock()
+	client.registerBroker(response.Coordinator)
+	client.coordinators[consumerGroup] = response.Coordinator.ID()
+	return nil
+}
+
+// private broker management helpers
+
+// registerBroker makes sure a broker received by a Metadata or Coordinator request is registered
+// in the brokers map. It returns the broker that is registered, which may be the provided broker,
+// or a previously registered Broker instance. You must hold the write lock before calling this function.
+func (client *client) registerBroker(broker *Broker) {
+	if client.brokers[broker.ID()] == nil {
+		client.brokers[broker.ID()] = broker
+		Logger.Printf("client/brokers registered new broker #%d at %s", broker.ID(), broker.Addr())
+	} else if broker.Addr() != client.brokers[broker.ID()].Addr() {
+		safeAsyncClose(client.brokers[broker.ID()])
+		client.brokers[broker.ID()] = broker
+		Logger.Printf("client/brokers replaced registered broker #%d with %s", broker.ID(), broker.Addr())
+	}
+}
+
+// deregisterBroker removes a broker from the seedsBroker list, and if it's
+// not the seedbroker, removes it from brokers map completely.
+func (client *client) deregisterBroker(broker *Broker) {
+	client.lock.Lock()
+	defer client.lock.Unlock()
+
+	if len(client.seedBrokers) > 0 && broker == client.seedBrokers[0] {
+		client.deadSeeds = append(client.deadSeeds, broker)
+		client.seedBrokers = client.seedBrokers[1:]
+	} else {
+		// we do this so that our loop in `tryRefreshMetadata` doesn't go on forever,
+		// but we really shouldn't have to; once that loop is made better this case can be
+		// removed, and the function generally can be renamed from `deregisterBroker` to
+		// `nextSeedBroker` or something
+		Logger.Printf("client/brokers deregistered broker #%d at %s", broker.ID(), broker.Addr())
+		delete(client.brokers, broker.ID())
+	}
+}
+
+func (client *client) resurrectDeadBrokers() {
+	client.lock.Lock()
+	defer client.lock.Unlock()
+
+	Logger.Printf("client/brokers resurrecting %d dead seed brokers", len(client.deadSeeds))
+	client.seedBrokers = append(client.seedBrokers, client.deadSeeds...)
+	client.deadSeeds = nil
+}
+
+func (client *client) any() *Broker {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	if len(client.seedBrokers) > 0 {
+		_ = client.seedBrokers[0].Open(client.conf)
+		return client.seedBrokers[0]
+	}
+
+	// not guaranteed to be random *or* deterministic
+	for _, broker := range client.brokers {
+		_ = broker.Open(client.conf)
+		return broker
+	}
+
+	return nil
+}
+
+// private caching/lazy metadata helpers
+
+type partitionType int
+
+const (
+	allPartitions partitionType = iota
+	writablePartitions
+	// If you add any more types, update the partition cache in update()
+
+	// Ensure this is the last partition type value
+	maxPartitionIndex
+)
+
+func (client *client) cachedMetadata(topic string, partitionID int32) *PartitionMetadata {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	partitions := client.metadata[topic]
+	if partitions != nil {
+		return partitions[partitionID]
+	}
+
+	return nil
+}
+
+func (client *client) cachedPartitions(topic string, partitionSet partitionType) []int32 {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	partitions, exists := client.cachedPartitionsResults[topic]
+
+	if !exists {
+		return nil
+	}
+	return partitions[partitionSet]
+}
+
+func (client *client) setPartitionCache(topic string, partitionSet partitionType) []int32 {
+	partitions := client.metadata[topic]
+
+	if partitions == nil {
+		return nil
+	}
+
+	ret := make([]int32, 0, len(partitions))
+	for _, partition := range partitions {
+		if partitionSet == writablePartitions && partition.Err == ErrLeaderNotAvailable {
+			continue
+		}
+		ret = append(ret, partition.ID)
+	}
+
+	sort.Sort(int32Slice(ret))
+	return ret
+}
+
+func (client *client) cachedLeader(topic string, partitionID int32) (*Broker, error) {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	partitions := client.metadata[topic]
+	if partitions != nil {
+		metadata, ok := partitions[partitionID]
+		if ok {
+			if metadata.Err == ErrLeaderNotAvailable {
+				return nil, ErrLeaderNotAvailable
+			}
+			b := client.brokers[metadata.Leader]
+			if b == nil {
+				return nil, ErrLeaderNotAvailable
+			}
+			_ = b.Open(client.conf)
+			return b, nil
+		}
+	}
+
+	return nil, ErrUnknownTopicOrPartition
+}
+
+func (client *client) getOffset(topic string, partitionID int32, time int64) (int64, error) {
+	broker, err := client.Leader(topic, partitionID)
+	if err != nil {
+		return -1, err
+	}
+
+	request := &OffsetRequest{}
+	if client.conf.Version.IsAtLeast(V0_10_1_0) {
+		request.Version = 1
+	}
+	request.AddBlock(topic, partitionID, time, 1)
+
+	response, err := broker.GetAvailableOffsets(request)
+	if err != nil {
+		_ = broker.Close()
+		return -1, err
+	}
+
+	block := response.GetBlock(topic, partitionID)
+	if block == nil {
+		_ = broker.Close()
+		return -1, ErrIncompleteResponse
+	}
+	if block.Err != ErrNoError {
+		return -1, block.Err
+	}
+	if len(block.Offsets) != 1 {
+		return -1, ErrOffsetOutOfRange
+	}
+
+	return block.Offsets[0], nil
+}
+
+// core metadata update logic
+
+func (client *client) backgroundMetadataUpdater() {
+	defer close(client.closed)
+
+	if client.conf.Metadata.RefreshFrequency == time.Duration(0) {
+		return
+	}
+
+	ticker := time.NewTicker(client.conf.Metadata.RefreshFrequency)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ticker.C:
+			if err := client.refreshMetadata(); err != nil {
+				Logger.Println("Client background metadata update:", err)
+			}
+		case <-client.closer:
+			return
+		}
+	}
+}
+
+func (client *client) refreshMetadata() error {
+	topics := []string{}
+
+	if !client.conf.Metadata.Full {
+		if specificTopics, err := client.MetadataTopics(); err != nil {
+			return err
+		} else if len(specificTopics) == 0 {
+			return ErrNoTopicsToUpdateMetadata
+		} else {
+			topics = specificTopics
+		}
+	}
+
+	if err := client.RefreshMetadata(topics...); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (client *client) tryRefreshMetadata(topics []string, attemptsRemaining int) error {
+	retry := func(err error) error {
+		if attemptsRemaining > 0 {
+			backoff := client.computeBackoff(attemptsRemaining)
+			Logger.Printf("client/metadata retrying after %dms... (%d attempts remaining)\n", client.conf.Metadata.Retry.Backoff/time.Millisecond, attemptsRemaining)
+			if backoff > 0 {
+				time.Sleep(backoff)
+			}
+			return client.tryRefreshMetadata(topics, attemptsRemaining-1)
+		}
+		return err
+	}
+
+	for broker := client.any(); broker != nil; broker = client.any() {
+		if len(topics) > 0 {
+			Logger.Printf("client/metadata fetching metadata for %v from broker %s\n", topics, broker.addr)
+		} else {
+			Logger.Printf("client/metadata fetching metadata for all topics from broker %s\n", broker.addr)
+		}
+
+		req := &MetadataRequest{Topics: topics}
+		if client.conf.Version.IsAtLeast(V0_10_0_0) {
+			req.Version = 1
+		}
+		response, err := broker.GetMetadata(req)
+
+		switch err.(type) {
+		case nil:
+			allKnownMetaData := len(topics) == 0
+			// valid response, use it
+			shouldRetry, err := client.updateMetadata(response, allKnownMetaData)
+			if shouldRetry {
+				Logger.Println("client/metadata found some partitions to be leaderless")
+				return retry(err) // note: err can be nil
+			}
+			return err
+
+		case PacketEncodingError:
+			// didn't even send, return the error
+			return err
+		default:
+			// some other error, remove that broker and try again
+			Logger.Printf("client/metadata got error from broker %d while fetching metadata: %v\n", broker.ID(), err)
+			_ = broker.Close()
+			client.deregisterBroker(broker)
+		}
+	}
+
+	Logger.Println("client/metadata no available broker to send metadata request to")
+	client.resurrectDeadBrokers()
+	return retry(ErrOutOfBrokers)
+}
+
+// if no fatal error, returns a list of topics that need retrying due to ErrLeaderNotAvailable
+func (client *client) updateMetadata(data *MetadataResponse, allKnownMetaData bool) (retry bool, err error) {
+	client.lock.Lock()
+	defer client.lock.Unlock()
+
+	// For all the brokers we received:
+	// - if it is a new ID, save it
+	// - if it is an existing ID, but the address we have is stale, discard the old one and save it
+	// - otherwise ignore it, replacing our existing one would just bounce the connection
+	for _, broker := range data.Brokers {
+		client.registerBroker(broker)
+	}
+
+	client.controllerID = data.ControllerID
+
+	if allKnownMetaData {
+		client.metadata = make(map[string]map[int32]*PartitionMetadata)
+		client.metadataTopics = make(map[string]none)
+		client.cachedPartitionsResults = make(map[string][maxPartitionIndex][]int32)
+	}
+	for _, topic := range data.Topics {
+		// topics must be added firstly to `metadataTopics` to guarantee that all
+		// requested topics must be recorded to keep them trackable for periodically
+		// metadata refresh.
+		if _, exists := client.metadataTopics[topic.Name]; !exists {
+			client.metadataTopics[topic.Name] = none{}
+		}
+		delete(client.metadata, topic.Name)
+		delete(client.cachedPartitionsResults, topic.Name)
+
+		switch topic.Err {
+		case ErrNoError:
+			break
+		case ErrInvalidTopic, ErrTopicAuthorizationFailed: // don't retry, don't store partial results
+			err = topic.Err
+			continue
+		case ErrUnknownTopicOrPartition: // retry, do not store partial partition results
+			err = topic.Err
+			retry = true
+			continue
+		case ErrLeaderNotAvailable: // retry, but store partial partition results
+			retry = true
+			break
+		default: // don't retry, don't store partial results
+			Logger.Printf("Unexpected topic-level metadata error: %s", topic.Err)
+			err = topic.Err
+			continue
+		}
+
+		client.metadata[topic.Name] = make(map[int32]*PartitionMetadata, len(topic.Partitions))
+		for _, partition := range topic.Partitions {
+			client.metadata[topic.Name][partition.ID] = partition
+			if partition.Err == ErrLeaderNotAvailable {
+				retry = true
+			}
+		}
+
+		var partitionCache [maxPartitionIndex][]int32
+		partitionCache[allPartitions] = client.setPartitionCache(topic.Name, allPartitions)
+		partitionCache[writablePartitions] = client.setPartitionCache(topic.Name, writablePartitions)
+		client.cachedPartitionsResults[topic.Name] = partitionCache
+	}
+
+	return
+}
+
+func (client *client) cachedCoordinator(consumerGroup string) *Broker {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+	if coordinatorID, ok := client.coordinators[consumerGroup]; ok {
+		return client.brokers[coordinatorID]
+	}
+	return nil
+}
+
+func (client *client) cachedController() *Broker {
+	client.lock.RLock()
+	defer client.lock.RUnlock()
+
+	return client.brokers[client.controllerID]
+}
+
+func (client *client) computeBackoff(attemptsRemaining int) time.Duration {
+	if client.conf.Metadata.Retry.BackoffFunc != nil {
+		maxRetries := client.conf.Metadata.Retry.Max
+		retries := maxRetries - attemptsRemaining
+		return client.conf.Metadata.Retry.BackoffFunc(retries, maxRetries)
+	} else {
+		return client.conf.Metadata.Retry.Backoff
+	}
+}
+
+func (client *client) getConsumerMetadata(consumerGroup string, attemptsRemaining int) (*FindCoordinatorResponse, error) {
+	retry := func(err error) (*FindCoordinatorResponse, error) {
+		if attemptsRemaining > 0 {
+			backoff := client.computeBackoff(attemptsRemaining)
+			Logger.Printf("client/coordinator retrying after %dms... (%d attempts remaining)\n", backoff/time.Millisecond, attemptsRemaining)
+			time.Sleep(backoff)
+			return client.getConsumerMetadata(consumerGroup, attemptsRemaining-1)
+		}
+		return nil, err
+	}
+
+	for broker := client.any(); broker != nil; broker = client.any() {
+		Logger.Printf("client/coordinator requesting coordinator for consumergroup %s from %s\n", consumerGroup, broker.Addr())
+
+		request := new(FindCoordinatorRequest)
+		request.CoordinatorKey = consumerGroup
+		request.CoordinatorType = CoordinatorGroup
+
+		response, err := broker.FindCoordinator(request)
+
+		if err != nil {
+			Logger.Printf("client/coordinator request to broker %s failed: %s\n", broker.Addr(), err)
+
+			switch err.(type) {
+			case PacketEncodingError:
+				return nil, err
+			default:
+				_ = broker.Close()
+				client.deregisterBroker(broker)
+				continue
+			}
+		}
+
+		switch response.Err {
+		case ErrNoError:
+			Logger.Printf("client/coordinator coordinator for consumergroup %s is #%d (%s)\n", consumerGroup, response.Coordinator.ID(), response.Coordinator.Addr())
+			return response, nil
+
+		case ErrConsumerCoordinatorNotAvailable:
+			Logger.Printf("client/coordinator coordinator for consumer group %s is not available\n", consumerGroup)
+
+			// This is very ugly, but this scenario will only happen once per cluster.
+			// The __consumer_offsets topic only has to be created one time.
+			// The number of partitions not configurable, but partition 0 should always exist.
+			if _, err := client.Leader("__consumer_offsets", 0); err != nil {
+				Logger.Printf("client/coordinator the __consumer_offsets topic is not initialized completely yet. Waiting 2 seconds...\n")
+				time.Sleep(2 * time.Second)
+			}
+
+			return retry(ErrConsumerCoordinatorNotAvailable)
+		default:
+			return nil, response.Err
+		}
+	}
+
+	Logger.Println("client/coordinator no available broker to send consumer metadata request to")
+	client.resurrectDeadBrokers()
+	return retry(ErrOutOfBrokers)
+}
diff --git a/vendor/github.com/Shopify/sarama/compress.go b/vendor/github.com/Shopify/sarama/compress.go
new file mode 100644
index 0000000..94b716e
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/compress.go
@@ -0,0 +1,75 @@
+package sarama
+
+import (
+	"bytes"
+	"compress/gzip"
+	"fmt"
+	"sync"
+
+	"github.com/eapache/go-xerial-snappy"
+	"github.com/pierrec/lz4"
+)
+
+var (
+	lz4WriterPool = sync.Pool{
+		New: func() interface{} {
+			return lz4.NewWriter(nil)
+		},
+	}
+
+	gzipWriterPool = sync.Pool{
+		New: func() interface{} {
+			return gzip.NewWriter(nil)
+		},
+	}
+)
+
+func compress(cc CompressionCodec, level int, data []byte) ([]byte, error) {
+	switch cc {
+	case CompressionNone:
+		return data, nil
+	case CompressionGZIP:
+		var (
+			err    error
+			buf    bytes.Buffer
+			writer *gzip.Writer
+		)
+		if level != CompressionLevelDefault {
+			writer, err = gzip.NewWriterLevel(&buf, level)
+			if err != nil {
+				return nil, err
+			}
+		} else {
+			writer = gzipWriterPool.Get().(*gzip.Writer)
+			defer gzipWriterPool.Put(writer)
+			writer.Reset(&buf)
+		}
+		if _, err := writer.Write(data); err != nil {
+			return nil, err
+		}
+		if err := writer.Close(); err != nil {
+			return nil, err
+		}
+		return buf.Bytes(), nil
+	case CompressionSnappy:
+		return snappy.Encode(data), nil
+	case CompressionLZ4:
+		writer := lz4WriterPool.Get().(*lz4.Writer)
+		defer lz4WriterPool.Put(writer)
+
+		var buf bytes.Buffer
+		writer.Reset(&buf)
+
+		if _, err := writer.Write(data); err != nil {
+			return nil, err
+		}
+		if err := writer.Close(); err != nil {
+			return nil, err
+		}
+		return buf.Bytes(), nil
+	case CompressionZSTD:
+		return zstdCompressLevel(nil, data, level)
+	default:
+		return nil, PacketEncodingError{fmt.Sprintf("unsupported compression codec (%d)", cc)}
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/config.go b/vendor/github.com/Shopify/sarama/config.go
new file mode 100644
index 0000000..9495b7f
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/config.go
@@ -0,0 +1,616 @@
+package sarama
+
+import (
+	"compress/gzip"
+	"crypto/tls"
+	"fmt"
+	"io/ioutil"
+	"net"
+	"regexp"
+	"time"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+const defaultClientID = "sarama"
+
+var validID = regexp.MustCompile(`\A[A-Za-z0-9._-]+\z`)
+
+// Config is used to pass multiple configuration options to Sarama's constructors.
+type Config struct {
+	// Admin is the namespace for ClusterAdmin properties used by the administrative Kafka client.
+	Admin struct {
+		// The maximum duration the administrative Kafka client will wait for ClusterAdmin operations,
+		// including topics, brokers, configurations and ACLs (defaults to 3 seconds).
+		Timeout time.Duration
+	}
+
+	// Net is the namespace for network-level properties used by the Broker, and
+	// shared by the Client/Producer/Consumer.
+	Net struct {
+		// How many outstanding requests a connection is allowed to have before
+		// sending on it blocks (default 5).
+		MaxOpenRequests int
+
+		// All three of the below configurations are similar to the
+		// `socket.timeout.ms` setting in JVM kafka. All of them default
+		// to 30 seconds.
+		DialTimeout  time.Duration // How long to wait for the initial connection.
+		ReadTimeout  time.Duration // How long to wait for a response.
+		WriteTimeout time.Duration // How long to wait for a transmit.
+
+		TLS struct {
+			// Whether or not to use TLS when connecting to the broker
+			// (defaults to false).
+			Enable bool
+			// The TLS configuration to use for secure connections if
+			// enabled (defaults to nil).
+			Config *tls.Config
+		}
+
+		// SASL based authentication with broker. While there are multiple SASL authentication methods
+		// the current implementation is limited to plaintext (SASL/PLAIN) authentication
+		SASL struct {
+			// Whether or not to use SASL authentication when connecting to the broker
+			// (defaults to false).
+			Enable bool
+			// SASLMechanism is the name of the enabled SASL mechanism.
+			// Possible values: OAUTHBEARER, PLAIN (defaults to PLAIN).
+			Mechanism SASLMechanism
+			// Whether or not to send the Kafka SASL handshake first if enabled
+			// (defaults to true). You should only set this to false if you're using
+			// a non-Kafka SASL proxy.
+			Handshake bool
+			//username and password for SASL/PLAIN authentication
+			User     string
+			Password string
+			// TokenProvider is a user-defined callback for generating
+			// access tokens for SASL/OAUTHBEARER auth. See the
+			// AccessTokenProvider interface docs for proper implementation
+			// guidelines.
+			TokenProvider AccessTokenProvider
+		}
+
+		// KeepAlive specifies the keep-alive period for an active network connection.
+		// If zero, keep-alives are disabled. (default is 0: disabled).
+		KeepAlive time.Duration
+
+		// LocalAddr is the local address to use when dialing an
+		// address. The address must be of a compatible type for the
+		// network being dialed.
+		// If nil, a local address is automatically chosen.
+		LocalAddr net.Addr
+	}
+
+	// Metadata is the namespace for metadata management properties used by the
+	// Client, and shared by the Producer/Consumer.
+	Metadata struct {
+		Retry struct {
+			// The total number of times to retry a metadata request when the
+			// cluster is in the middle of a leader election (default 3).
+			Max int
+			// How long to wait for leader election to occur before retrying
+			// (default 250ms). Similar to the JVM's `retry.backoff.ms`.
+			Backoff time.Duration
+			// Called to compute backoff time dynamically. Useful for implementing
+			// more sophisticated backoff strategies. This takes precedence over
+			// `Backoff` if set.
+			BackoffFunc func(retries, maxRetries int) time.Duration
+		}
+		// How frequently to refresh the cluster metadata in the background.
+		// Defaults to 10 minutes. Set to 0 to disable. Similar to
+		// `topic.metadata.refresh.interval.ms` in the JVM version.
+		RefreshFrequency time.Duration
+
+		// Whether to maintain a full set of metadata for all topics, or just
+		// the minimal set that has been necessary so far. The full set is simpler
+		// and usually more convenient, but can take up a substantial amount of
+		// memory if you have many topics and partitions. Defaults to true.
+		Full bool
+	}
+
+	// Producer is the namespace for configuration related to producing messages,
+	// used by the Producer.
+	Producer struct {
+		// The maximum permitted size of a message (defaults to 1000000). Should be
+		// set equal to or smaller than the broker's `message.max.bytes`.
+		MaxMessageBytes int
+		// The level of acknowledgement reliability needed from the broker (defaults
+		// to WaitForLocal). Equivalent to the `request.required.acks` setting of the
+		// JVM producer.
+		RequiredAcks RequiredAcks
+		// The maximum duration the broker will wait the receipt of the number of
+		// RequiredAcks (defaults to 10 seconds). This is only relevant when
+		// RequiredAcks is set to WaitForAll or a number > 1. Only supports
+		// millisecond resolution, nanoseconds will be truncated. Equivalent to
+		// the JVM producer's `request.timeout.ms` setting.
+		Timeout time.Duration
+		// The type of compression to use on messages (defaults to no compression).
+		// Similar to `compression.codec` setting of the JVM producer.
+		Compression CompressionCodec
+		// The level of compression to use on messages. The meaning depends
+		// on the actual compression type used and defaults to default compression
+		// level for the codec.
+		CompressionLevel int
+		// Generates partitioners for choosing the partition to send messages to
+		// (defaults to hashing the message key). Similar to the `partitioner.class`
+		// setting for the JVM producer.
+		Partitioner PartitionerConstructor
+		// If enabled, the producer will ensure that exactly one copy of each message is
+		// written.
+		Idempotent bool
+
+		// Return specifies what channels will be populated. If they are set to true,
+		// you must read from the respective channels to prevent deadlock. If,
+		// however, this config is used to create a `SyncProducer`, both must be set
+		// to true and you shall not read from the channels since the producer does
+		// this internally.
+		Return struct {
+			// If enabled, successfully delivered messages will be returned on the
+			// Successes channel (default disabled).
+			Successes bool
+
+			// If enabled, messages that failed to deliver will be returned on the
+			// Errors channel, including error (default enabled).
+			Errors bool
+		}
+
+		// The following config options control how often messages are batched up and
+		// sent to the broker. By default, messages are sent as fast as possible, and
+		// all messages received while the current batch is in-flight are placed
+		// into the subsequent batch.
+		Flush struct {
+			// The best-effort number of bytes needed to trigger a flush. Use the
+			// global sarama.MaxRequestSize to set a hard upper limit.
+			Bytes int
+			// The best-effort number of messages needed to trigger a flush. Use
+			// `MaxMessages` to set a hard upper limit.
+			Messages int
+			// The best-effort frequency of flushes. Equivalent to
+			// `queue.buffering.max.ms` setting of JVM producer.
+			Frequency time.Duration
+			// The maximum number of messages the producer will send in a single
+			// broker request. Defaults to 0 for unlimited. Similar to
+			// `queue.buffering.max.messages` in the JVM producer.
+			MaxMessages int
+		}
+
+		Retry struct {
+			// The total number of times to retry sending a message (default 3).
+			// Similar to the `message.send.max.retries` setting of the JVM producer.
+			Max int
+			// How long to wait for the cluster to settle between retries
+			// (default 100ms). Similar to the `retry.backoff.ms` setting of the
+			// JVM producer.
+			Backoff time.Duration
+			// Called to compute backoff time dynamically. Useful for implementing
+			// more sophisticated backoff strategies. This takes precedence over
+			// `Backoff` if set.
+			BackoffFunc func(retries, maxRetries int) time.Duration
+		}
+	}
+
+	// Consumer is the namespace for configuration related to consuming messages,
+	// used by the Consumer.
+	Consumer struct {
+
+		// Group is the namespace for configuring consumer group.
+		Group struct {
+			Session struct {
+				// The timeout used to detect consumer failures when using Kafka's group management facility.
+				// The consumer sends periodic heartbeats to indicate its liveness to the broker.
+				// If no heartbeats are received by the broker before the expiration of this session timeout,
+				// then the broker will remove this consumer from the group and initiate a rebalance.
+				// Note that the value must be in the allowable range as configured in the broker configuration
+				// by `group.min.session.timeout.ms` and `group.max.session.timeout.ms` (default 10s)
+				Timeout time.Duration
+			}
+			Heartbeat struct {
+				// The expected time between heartbeats to the consumer coordinator when using Kafka's group
+				// management facilities. Heartbeats are used to ensure that the consumer's session stays active and
+				// to facilitate rebalancing when new consumers join or leave the group.
+				// The value must be set lower than Consumer.Group.Session.Timeout, but typically should be set no
+				// higher than 1/3 of that value.
+				// It can be adjusted even lower to control the expected time for normal rebalances (default 3s)
+				Interval time.Duration
+			}
+			Rebalance struct {
+				// Strategy for allocating topic partitions to members (default BalanceStrategyRange)
+				Strategy BalanceStrategy
+				// The maximum allowed time for each worker to join the group once a rebalance has begun.
+				// This is basically a limit on the amount of time needed for all tasks to flush any pending
+				// data and commit offsets. If the timeout is exceeded, then the worker will be removed from
+				// the group, which will cause offset commit failures (default 60s).
+				Timeout time.Duration
+
+				Retry struct {
+					// When a new consumer joins a consumer group the set of consumers attempt to "rebalance"
+					// the load to assign partitions to each consumer. If the set of consumers changes while
+					// this assignment is taking place the rebalance will fail and retry. This setting controls
+					// the maximum number of attempts before giving up (default 4).
+					Max int
+					// Backoff time between retries during rebalance (default 2s)
+					Backoff time.Duration
+				}
+			}
+			Member struct {
+				// Custom metadata to include when joining the group. The user data for all joined members
+				// can be retrieved by sending a DescribeGroupRequest to the broker that is the
+				// coordinator for the group.
+				UserData []byte
+			}
+		}
+
+		Retry struct {
+			// How long to wait after a failing to read from a partition before
+			// trying again (default 2s).
+			Backoff time.Duration
+			// Called to compute backoff time dynamically. Useful for implementing
+			// more sophisticated backoff strategies. This takes precedence over
+			// `Backoff` if set.
+			BackoffFunc func(retries int) time.Duration
+		}
+
+		// Fetch is the namespace for controlling how many bytes are retrieved by any
+		// given request.
+		Fetch struct {
+			// The minimum number of message bytes to fetch in a request - the broker
+			// will wait until at least this many are available. The default is 1,
+			// as 0 causes the consumer to spin when no messages are available.
+			// Equivalent to the JVM's `fetch.min.bytes`.
+			Min int32
+			// The default number of message bytes to fetch from the broker in each
+			// request (default 1MB). This should be larger than the majority of
+			// your messages, or else the consumer will spend a lot of time
+			// negotiating sizes and not actually consuming. Similar to the JVM's
+			// `fetch.message.max.bytes`.
+			Default int32
+			// The maximum number of message bytes to fetch from the broker in a
+			// single request. Messages larger than this will return
+			// ErrMessageTooLarge and will not be consumable, so you must be sure
+			// this is at least as large as your largest message. Defaults to 0
+			// (no limit). Similar to the JVM's `fetch.message.max.bytes`. The
+			// global `sarama.MaxResponseSize` still applies.
+			Max int32
+		}
+		// The maximum amount of time the broker will wait for Consumer.Fetch.Min
+		// bytes to become available before it returns fewer than that anyways. The
+		// default is 250ms, since 0 causes the consumer to spin when no events are
+		// available. 100-500ms is a reasonable range for most cases. Kafka only
+		// supports precision up to milliseconds; nanoseconds will be truncated.
+		// Equivalent to the JVM's `fetch.wait.max.ms`.
+		MaxWaitTime time.Duration
+
+		// The maximum amount of time the consumer expects a message takes to
+		// process for the user. If writing to the Messages channel takes longer
+		// than this, that partition will stop fetching more messages until it
+		// can proceed again.
+		// Note that, since the Messages channel is buffered, the actual grace time is
+		// (MaxProcessingTime * ChanneBufferSize). Defaults to 100ms.
+		// If a message is not written to the Messages channel between two ticks
+		// of the expiryTicker then a timeout is detected.
+		// Using a ticker instead of a timer to detect timeouts should typically
+		// result in many fewer calls to Timer functions which may result in a
+		// significant performance improvement if many messages are being sent
+		// and timeouts are infrequent.
+		// The disadvantage of using a ticker instead of a timer is that
+		// timeouts will be less accurate. That is, the effective timeout could
+		// be between `MaxProcessingTime` and `2 * MaxProcessingTime`. For
+		// example, if `MaxProcessingTime` is 100ms then a delay of 180ms
+		// between two messages being sent may not be recognized as a timeout.
+		MaxProcessingTime time.Duration
+
+		// Return specifies what channels will be populated. If they are set to true,
+		// you must read from them to prevent deadlock.
+		Return struct {
+			// If enabled, any errors that occurred while consuming are returned on
+			// the Errors channel (default disabled).
+			Errors bool
+		}
+
+		// Offsets specifies configuration for how and when to commit consumed
+		// offsets. This currently requires the manual use of an OffsetManager
+		// but will eventually be automated.
+		Offsets struct {
+			// How frequently to commit updated offsets. Defaults to 1s.
+			CommitInterval time.Duration
+
+			// The initial offset to use if no offset was previously committed.
+			// Should be OffsetNewest or OffsetOldest. Defaults to OffsetNewest.
+			Initial int64
+
+			// The retention duration for committed offsets. If zero, disabled
+			// (in which case the `offsets.retention.minutes` option on the
+			// broker will be used).  Kafka only supports precision up to
+			// milliseconds; nanoseconds will be truncated. Requires Kafka
+			// broker version 0.9.0 or later.
+			// (default is 0: disabled).
+			Retention time.Duration
+
+			Retry struct {
+				// The total number of times to retry failing commit
+				// requests during OffsetManager shutdown (default 3).
+				Max int
+			}
+		}
+	}
+
+	// A user-provided string sent with every request to the brokers for logging,
+	// debugging, and auditing purposes. Defaults to "sarama", but you should
+	// probably set it to something specific to your application.
+	ClientID string
+	// The number of events to buffer in internal and external channels. This
+	// permits the producer and consumer to continue processing some messages
+	// in the background while user code is working, greatly improving throughput.
+	// Defaults to 256.
+	ChannelBufferSize int
+	// The version of Kafka that Sarama will assume it is running against.
+	// Defaults to the oldest supported stable version. Since Kafka provides
+	// backwards-compatibility, setting it to a version older than you have
+	// will not break anything, although it may prevent you from using the
+	// latest features. Setting it to a version greater than you are actually
+	// running may lead to random breakage.
+	Version KafkaVersion
+	// The registry to define metrics into.
+	// Defaults to a local registry.
+	// If you want to disable metrics gathering, set "metrics.UseNilMetrics" to "true"
+	// prior to starting Sarama.
+	// See Examples on how to use the metrics registry
+	MetricRegistry metrics.Registry
+}
+
+// NewConfig returns a new configuration instance with sane defaults.
+func NewConfig() *Config {
+	c := &Config{}
+
+	c.Admin.Timeout = 3 * time.Second
+
+	c.Net.MaxOpenRequests = 5
+	c.Net.DialTimeout = 30 * time.Second
+	c.Net.ReadTimeout = 30 * time.Second
+	c.Net.WriteTimeout = 30 * time.Second
+	c.Net.SASL.Handshake = true
+
+	c.Metadata.Retry.Max = 3
+	c.Metadata.Retry.Backoff = 250 * time.Millisecond
+	c.Metadata.RefreshFrequency = 10 * time.Minute
+	c.Metadata.Full = true
+
+	c.Producer.MaxMessageBytes = 1000000
+	c.Producer.RequiredAcks = WaitForLocal
+	c.Producer.Timeout = 10 * time.Second
+	c.Producer.Partitioner = NewHashPartitioner
+	c.Producer.Retry.Max = 3
+	c.Producer.Retry.Backoff = 100 * time.Millisecond
+	c.Producer.Return.Errors = true
+	c.Producer.CompressionLevel = CompressionLevelDefault
+
+	c.Consumer.Fetch.Min = 1
+	c.Consumer.Fetch.Default = 1024 * 1024
+	c.Consumer.Retry.Backoff = 2 * time.Second
+	c.Consumer.MaxWaitTime = 250 * time.Millisecond
+	c.Consumer.MaxProcessingTime = 100 * time.Millisecond
+	c.Consumer.Return.Errors = false
+	c.Consumer.Offsets.CommitInterval = 1 * time.Second
+	c.Consumer.Offsets.Initial = OffsetNewest
+	c.Consumer.Offsets.Retry.Max = 3
+
+	c.Consumer.Group.Session.Timeout = 10 * time.Second
+	c.Consumer.Group.Heartbeat.Interval = 3 * time.Second
+	c.Consumer.Group.Rebalance.Strategy = BalanceStrategyRange
+	c.Consumer.Group.Rebalance.Timeout = 60 * time.Second
+	c.Consumer.Group.Rebalance.Retry.Max = 4
+	c.Consumer.Group.Rebalance.Retry.Backoff = 2 * time.Second
+
+	c.ClientID = defaultClientID
+	c.ChannelBufferSize = 256
+	c.Version = MinVersion
+	c.MetricRegistry = metrics.NewRegistry()
+
+	return c
+}
+
+// Validate checks a Config instance. It will return a
+// ConfigurationError if the specified values don't make sense.
+func (c *Config) Validate() error {
+	// some configuration values should be warned on but not fail completely, do those first
+	if c.Net.TLS.Enable == false && c.Net.TLS.Config != nil {
+		Logger.Println("Net.TLS is disabled but a non-nil configuration was provided.")
+	}
+	if c.Net.SASL.Enable == false {
+		if c.Net.SASL.User != "" {
+			Logger.Println("Net.SASL is disabled but a non-empty username was provided.")
+		}
+		if c.Net.SASL.Password != "" {
+			Logger.Println("Net.SASL is disabled but a non-empty password was provided.")
+		}
+	}
+	if c.Producer.RequiredAcks > 1 {
+		Logger.Println("Producer.RequiredAcks > 1 is deprecated and will raise an exception with kafka >= 0.8.2.0.")
+	}
+	if c.Producer.MaxMessageBytes >= int(MaxRequestSize) {
+		Logger.Println("Producer.MaxMessageBytes must be smaller than MaxRequestSize; it will be ignored.")
+	}
+	if c.Producer.Flush.Bytes >= int(MaxRequestSize) {
+		Logger.Println("Producer.Flush.Bytes must be smaller than MaxRequestSize; it will be ignored.")
+	}
+	if (c.Producer.Flush.Bytes > 0 || c.Producer.Flush.Messages > 0) && c.Producer.Flush.Frequency == 0 {
+		Logger.Println("Producer.Flush: Bytes or Messages are set, but Frequency is not; messages may not get flushed.")
+	}
+	if c.Producer.Timeout%time.Millisecond != 0 {
+		Logger.Println("Producer.Timeout only supports millisecond resolution; nanoseconds will be truncated.")
+	}
+	if c.Consumer.MaxWaitTime < 100*time.Millisecond {
+		Logger.Println("Consumer.MaxWaitTime is very low, which can cause high CPU and network usage. See documentation for details.")
+	}
+	if c.Consumer.MaxWaitTime%time.Millisecond != 0 {
+		Logger.Println("Consumer.MaxWaitTime only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Consumer.Offsets.Retention%time.Millisecond != 0 {
+		Logger.Println("Consumer.Offsets.Retention only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Consumer.Group.Session.Timeout%time.Millisecond != 0 {
+		Logger.Println("Consumer.Group.Session.Timeout only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Consumer.Group.Heartbeat.Interval%time.Millisecond != 0 {
+		Logger.Println("Consumer.Group.Heartbeat.Interval only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Consumer.Group.Rebalance.Timeout%time.Millisecond != 0 {
+		Logger.Println("Consumer.Group.Rebalance.Timeout only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.ClientID == defaultClientID {
+		Logger.Println("ClientID is the default of 'sarama', you should consider setting it to something application-specific.")
+	}
+
+	// validate Net values
+	switch {
+	case c.Net.MaxOpenRequests <= 0:
+		return ConfigurationError("Net.MaxOpenRequests must be > 0")
+	case c.Net.DialTimeout <= 0:
+		return ConfigurationError("Net.DialTimeout must be > 0")
+	case c.Net.ReadTimeout <= 0:
+		return ConfigurationError("Net.ReadTimeout must be > 0")
+	case c.Net.WriteTimeout <= 0:
+		return ConfigurationError("Net.WriteTimeout must be > 0")
+	case c.Net.KeepAlive < 0:
+		return ConfigurationError("Net.KeepAlive must be >= 0")
+	case c.Net.SASL.Enable:
+		// For backwards compatibility, empty mechanism value defaults to PLAIN
+		isSASLPlain := len(c.Net.SASL.Mechanism) == 0 || c.Net.SASL.Mechanism == SASLTypePlaintext
+		if isSASLPlain {
+			if c.Net.SASL.User == "" {
+				return ConfigurationError("Net.SASL.User must not be empty when SASL is enabled")
+			}
+			if c.Net.SASL.Password == "" {
+				return ConfigurationError("Net.SASL.Password must not be empty when SASL is enabled")
+			}
+		} else if c.Net.SASL.Mechanism == SASLTypeOAuth {
+			if c.Net.SASL.TokenProvider == nil {
+				return ConfigurationError("An AccessTokenProvider instance must be provided to Net.SASL.User.TokenProvider")
+			}
+		} else {
+			msg := fmt.Sprintf("The SASL mechanism configuration is invalid. Possible values are `%s` and `%s`",
+				SASLTypeOAuth, SASLTypePlaintext)
+			return ConfigurationError(msg)
+		}
+	}
+
+	// validate the Admin values
+	switch {
+	case c.Admin.Timeout <= 0:
+		return ConfigurationError("Admin.Timeout must be > 0")
+	}
+
+	// validate the Metadata values
+	switch {
+	case c.Metadata.Retry.Max < 0:
+		return ConfigurationError("Metadata.Retry.Max must be >= 0")
+	case c.Metadata.Retry.Backoff < 0:
+		return ConfigurationError("Metadata.Retry.Backoff must be >= 0")
+	case c.Metadata.RefreshFrequency < 0:
+		return ConfigurationError("Metadata.RefreshFrequency must be >= 0")
+	}
+
+	// validate the Producer values
+	switch {
+	case c.Producer.MaxMessageBytes <= 0:
+		return ConfigurationError("Producer.MaxMessageBytes must be > 0")
+	case c.Producer.RequiredAcks < -1:
+		return ConfigurationError("Producer.RequiredAcks must be >= -1")
+	case c.Producer.Timeout <= 0:
+		return ConfigurationError("Producer.Timeout must be > 0")
+	case c.Producer.Partitioner == nil:
+		return ConfigurationError("Producer.Partitioner must not be nil")
+	case c.Producer.Flush.Bytes < 0:
+		return ConfigurationError("Producer.Flush.Bytes must be >= 0")
+	case c.Producer.Flush.Messages < 0:
+		return ConfigurationError("Producer.Flush.Messages must be >= 0")
+	case c.Producer.Flush.Frequency < 0:
+		return ConfigurationError("Producer.Flush.Frequency must be >= 0")
+	case c.Producer.Flush.MaxMessages < 0:
+		return ConfigurationError("Producer.Flush.MaxMessages must be >= 0")
+	case c.Producer.Flush.MaxMessages > 0 && c.Producer.Flush.MaxMessages < c.Producer.Flush.Messages:
+		return ConfigurationError("Producer.Flush.MaxMessages must be >= Producer.Flush.Messages when set")
+	case c.Producer.Retry.Max < 0:
+		return ConfigurationError("Producer.Retry.Max must be >= 0")
+	case c.Producer.Retry.Backoff < 0:
+		return ConfigurationError("Producer.Retry.Backoff must be >= 0")
+	}
+
+	if c.Producer.Compression == CompressionLZ4 && !c.Version.IsAtLeast(V0_10_0_0) {
+		return ConfigurationError("lz4 compression requires Version >= V0_10_0_0")
+	}
+
+	if c.Producer.Compression == CompressionGZIP {
+		if c.Producer.CompressionLevel != CompressionLevelDefault {
+			if _, err := gzip.NewWriterLevel(ioutil.Discard, c.Producer.CompressionLevel); err != nil {
+				return ConfigurationError(fmt.Sprintf("gzip compression does not work with level %d: %v", c.Producer.CompressionLevel, err))
+			}
+		}
+	}
+
+	if c.Producer.Idempotent {
+		if !c.Version.IsAtLeast(V0_11_0_0) {
+			return ConfigurationError("Idempotent producer requires Version >= V0_11_0_0")
+		}
+		if c.Producer.Retry.Max == 0 {
+			return ConfigurationError("Idempotent producer requires Producer.Retry.Max >= 1")
+		}
+		if c.Producer.RequiredAcks != WaitForAll {
+			return ConfigurationError("Idempotent producer requires Producer.RequiredAcks to be WaitForAll")
+		}
+		if c.Net.MaxOpenRequests > 1 {
+			return ConfigurationError("Idempotent producer requires Net.MaxOpenRequests to be 1")
+		}
+	}
+
+	// validate the Consumer values
+	switch {
+	case c.Consumer.Fetch.Min <= 0:
+		return ConfigurationError("Consumer.Fetch.Min must be > 0")
+	case c.Consumer.Fetch.Default <= 0:
+		return ConfigurationError("Consumer.Fetch.Default must be > 0")
+	case c.Consumer.Fetch.Max < 0:
+		return ConfigurationError("Consumer.Fetch.Max must be >= 0")
+	case c.Consumer.MaxWaitTime < 1*time.Millisecond:
+		return ConfigurationError("Consumer.MaxWaitTime must be >= 1ms")
+	case c.Consumer.MaxProcessingTime <= 0:
+		return ConfigurationError("Consumer.MaxProcessingTime must be > 0")
+	case c.Consumer.Retry.Backoff < 0:
+		return ConfigurationError("Consumer.Retry.Backoff must be >= 0")
+	case c.Consumer.Offsets.CommitInterval <= 0:
+		return ConfigurationError("Consumer.Offsets.CommitInterval must be > 0")
+	case c.Consumer.Offsets.Initial != OffsetOldest && c.Consumer.Offsets.Initial != OffsetNewest:
+		return ConfigurationError("Consumer.Offsets.Initial must be OffsetOldest or OffsetNewest")
+	case c.Consumer.Offsets.Retry.Max < 0:
+		return ConfigurationError("Consumer.Offsets.Retry.Max must be >= 0")
+	}
+
+	// validate the Consumer Group values
+	switch {
+	case c.Consumer.Group.Session.Timeout <= 2*time.Millisecond:
+		return ConfigurationError("Consumer.Group.Session.Timeout must be >= 2ms")
+	case c.Consumer.Group.Heartbeat.Interval < 1*time.Millisecond:
+		return ConfigurationError("Consumer.Group.Heartbeat.Interval must be >= 1ms")
+	case c.Consumer.Group.Heartbeat.Interval >= c.Consumer.Group.Session.Timeout:
+		return ConfigurationError("Consumer.Group.Heartbeat.Interval must be < Consumer.Group.Session.Timeout")
+	case c.Consumer.Group.Rebalance.Strategy == nil:
+		return ConfigurationError("Consumer.Group.Rebalance.Strategy must not be empty")
+	case c.Consumer.Group.Rebalance.Timeout <= time.Millisecond:
+		return ConfigurationError("Consumer.Group.Rebalance.Timeout must be >= 1ms")
+	case c.Consumer.Group.Rebalance.Retry.Max < 0:
+		return ConfigurationError("Consumer.Group.Rebalance.Retry.Max must be >= 0")
+	case c.Consumer.Group.Rebalance.Retry.Backoff < 0:
+		return ConfigurationError("Consumer.Group.Rebalance.Retry.Backoff must be >= 0")
+	}
+
+	// validate misc shared values
+	switch {
+	case c.ChannelBufferSize < 0:
+		return ConfigurationError("ChannelBufferSize must be >= 0")
+	case !validID.MatchString(c.ClientID):
+		return ConfigurationError("ClientID is invalid")
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/config_resource_type.go b/vendor/github.com/Shopify/sarama/config_resource_type.go
new file mode 100644
index 0000000..848cc9c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/config_resource_type.go
@@ -0,0 +1,15 @@
+package sarama
+
+type ConfigResourceType int8
+
+// Taken from :
+// https://cwiki.apache.org/confluence/display/KAFKA/KIP-133%3A+Describe+and+Alter+Configs+Admin+APIs#KIP-133:DescribeandAlterConfigsAdminAPIs-WireFormattypes
+
+const (
+	UnknownResource ConfigResourceType = 0
+	AnyResource     ConfigResourceType = 1
+	TopicResource   ConfigResourceType = 2
+	GroupResource   ConfigResourceType = 3
+	ClusterResource ConfigResourceType = 4
+	BrokerResource  ConfigResourceType = 5
+)
diff --git a/vendor/github.com/Shopify/sarama/consumer.go b/vendor/github.com/Shopify/sarama/consumer.go
new file mode 100644
index 0000000..ce72ff1
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/consumer.go
@@ -0,0 +1,833 @@
+package sarama
+
+import (
+	"errors"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// ConsumerMessage encapsulates a Kafka message returned by the consumer.
+type ConsumerMessage struct {
+	Key, Value     []byte
+	Topic          string
+	Partition      int32
+	Offset         int64
+	Timestamp      time.Time       // only set if kafka is version 0.10+, inner message timestamp
+	BlockTimestamp time.Time       // only set if kafka is version 0.10+, outer (compressed) block timestamp
+	Headers        []*RecordHeader // only set if kafka is version 0.11+
+}
+
+// ConsumerError is what is provided to the user when an error occurs.
+// It wraps an error and includes the topic and partition.
+type ConsumerError struct {
+	Topic     string
+	Partition int32
+	Err       error
+}
+
+func (ce ConsumerError) Error() string {
+	return fmt.Sprintf("kafka: error while consuming %s/%d: %s", ce.Topic, ce.Partition, ce.Err)
+}
+
+// ConsumerErrors is a type that wraps a batch of errors and implements the Error interface.
+// It can be returned from the PartitionConsumer's Close methods to avoid the need to manually drain errors
+// when stopping.
+type ConsumerErrors []*ConsumerError
+
+func (ce ConsumerErrors) Error() string {
+	return fmt.Sprintf("kafka: %d errors while consuming", len(ce))
+}
+
+// Consumer manages PartitionConsumers which process Kafka messages from brokers. You MUST call Close()
+// on a consumer to avoid leaks, it will not be garbage-collected automatically when it passes out of
+// scope.
+//
+// Sarama's Consumer type does not currently support automatic consumer-group rebalancing and offset tracking.
+// For Zookeeper-based tracking (Kafka 0.8.2 and earlier), the https://github.com/wvanbergen/kafka library
+// builds on Sarama to add this support. For Kafka-based tracking (Kafka 0.9 and later), the
+// https://github.com/bsm/sarama-cluster library builds on Sarama to add this support.
+type Consumer interface {
+
+	// Topics returns the set of available topics as retrieved from the cluster
+	// metadata. This method is the same as Client.Topics(), and is provided for
+	// convenience.
+	Topics() ([]string, error)
+
+	// Partitions returns the sorted list of all partition IDs for the given topic.
+	// This method is the same as Client.Partitions(), and is provided for convenience.
+	Partitions(topic string) ([]int32, error)
+
+	// ConsumePartition creates a PartitionConsumer on the given topic/partition with
+	// the given offset. It will return an error if this Consumer is already consuming
+	// on the given topic/partition. Offset can be a literal offset, or OffsetNewest
+	// or OffsetOldest
+	ConsumePartition(topic string, partition int32, offset int64) (PartitionConsumer, error)
+
+	// HighWaterMarks returns the current high water marks for each topic and partition.
+	// Consistency between partitions is not guaranteed since high water marks are updated separately.
+	HighWaterMarks() map[string]map[int32]int64
+
+	// Close shuts down the consumer. It must be called after all child
+	// PartitionConsumers have already been closed.
+	Close() error
+}
+
+type consumer struct {
+	client    Client
+	conf      *Config
+	ownClient bool
+
+	lock            sync.Mutex
+	children        map[string]map[int32]*partitionConsumer
+	brokerConsumers map[*Broker]*brokerConsumer
+}
+
+// NewConsumer creates a new consumer using the given broker addresses and configuration.
+func NewConsumer(addrs []string, config *Config) (Consumer, error) {
+	client, err := NewClient(addrs, config)
+	if err != nil {
+		return nil, err
+	}
+
+	c, err := NewConsumerFromClient(client)
+	if err != nil {
+		return nil, err
+	}
+	c.(*consumer).ownClient = true
+	return c, nil
+}
+
+// NewConsumerFromClient creates a new consumer using the given client. It is still
+// necessary to call Close() on the underlying client when shutting down this consumer.
+func NewConsumerFromClient(client Client) (Consumer, error) {
+	// Check that we are not dealing with a closed Client before processing any other arguments
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	c := &consumer{
+		client:          client,
+		conf:            client.Config(),
+		children:        make(map[string]map[int32]*partitionConsumer),
+		brokerConsumers: make(map[*Broker]*brokerConsumer),
+	}
+
+	return c, nil
+}
+
+func (c *consumer) Close() error {
+	if c.ownClient {
+		return c.client.Close()
+	}
+	return nil
+}
+
+func (c *consumer) Topics() ([]string, error) {
+	return c.client.Topics()
+}
+
+func (c *consumer) Partitions(topic string) ([]int32, error) {
+	return c.client.Partitions(topic)
+}
+
+func (c *consumer) ConsumePartition(topic string, partition int32, offset int64) (PartitionConsumer, error) {
+	child := &partitionConsumer{
+		consumer:  c,
+		conf:      c.conf,
+		topic:     topic,
+		partition: partition,
+		messages:  make(chan *ConsumerMessage, c.conf.ChannelBufferSize),
+		errors:    make(chan *ConsumerError, c.conf.ChannelBufferSize),
+		feeder:    make(chan *FetchResponse, 1),
+		trigger:   make(chan none, 1),
+		dying:     make(chan none),
+		fetchSize: c.conf.Consumer.Fetch.Default,
+	}
+
+	if err := child.chooseStartingOffset(offset); err != nil {
+		return nil, err
+	}
+
+	var leader *Broker
+	var err error
+	if leader, err = c.client.Leader(child.topic, child.partition); err != nil {
+		return nil, err
+	}
+
+	if err := c.addChild(child); err != nil {
+		return nil, err
+	}
+
+	go withRecover(child.dispatcher)
+	go withRecover(child.responseFeeder)
+
+	child.broker = c.refBrokerConsumer(leader)
+	child.broker.input <- child
+
+	return child, nil
+}
+
+func (c *consumer) HighWaterMarks() map[string]map[int32]int64 {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	hwms := make(map[string]map[int32]int64)
+	for topic, p := range c.children {
+		hwm := make(map[int32]int64, len(p))
+		for partition, pc := range p {
+			hwm[partition] = pc.HighWaterMarkOffset()
+		}
+		hwms[topic] = hwm
+	}
+
+	return hwms
+}
+
+func (c *consumer) addChild(child *partitionConsumer) error {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	topicChildren := c.children[child.topic]
+	if topicChildren == nil {
+		topicChildren = make(map[int32]*partitionConsumer)
+		c.children[child.topic] = topicChildren
+	}
+
+	if topicChildren[child.partition] != nil {
+		return ConfigurationError("That topic/partition is already being consumed")
+	}
+
+	topicChildren[child.partition] = child
+	return nil
+}
+
+func (c *consumer) removeChild(child *partitionConsumer) {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	delete(c.children[child.topic], child.partition)
+}
+
+func (c *consumer) refBrokerConsumer(broker *Broker) *brokerConsumer {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	bc := c.brokerConsumers[broker]
+	if bc == nil {
+		bc = c.newBrokerConsumer(broker)
+		c.brokerConsumers[broker] = bc
+	}
+
+	bc.refs++
+
+	return bc
+}
+
+func (c *consumer) unrefBrokerConsumer(brokerWorker *brokerConsumer) {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	brokerWorker.refs--
+
+	if brokerWorker.refs == 0 {
+		close(brokerWorker.input)
+		if c.brokerConsumers[brokerWorker.broker] == brokerWorker {
+			delete(c.brokerConsumers, brokerWorker.broker)
+		}
+	}
+}
+
+func (c *consumer) abandonBrokerConsumer(brokerWorker *brokerConsumer) {
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	delete(c.brokerConsumers, brokerWorker.broker)
+}
+
+// PartitionConsumer
+
+// PartitionConsumer processes Kafka messages from a given topic and partition. You MUST call one of Close() or
+// AsyncClose() on a PartitionConsumer to avoid leaks; it will not be garbage-collected automatically when it passes out
+// of scope.
+//
+// The simplest way of using a PartitionConsumer is to loop over its Messages channel using a for/range
+// loop. The PartitionConsumer will only stop itself in one case: when the offset being consumed is reported
+// as out of range by the brokers. In this case you should decide what you want to do (try a different offset,
+// notify a human, etc) and handle it appropriately. For all other error cases, it will just keep retrying.
+// By default, it logs these errors to sarama.Logger; if you want to be notified directly of all errors, set
+// your config's Consumer.Return.Errors to true and read from the Errors channel, using a select statement
+// or a separate goroutine. Check out the Consumer examples to see implementations of these different approaches.
+//
+// To terminate such a for/range loop while the loop is executing, call AsyncClose. This will kick off the process of
+// consumer tear-down & return imediately. Continue to loop, servicing the Messages channel until the teardown process
+// AsyncClose initiated closes it (thus terminating the for/range loop). If you've already ceased reading Messages, call
+// Close; this will signal the PartitionConsumer's goroutines to begin shutting down (just like AsyncClose), but will
+// also drain the Messages channel, harvest all errors & return them once cleanup has completed.
+type PartitionConsumer interface {
+
+	// AsyncClose initiates a shutdown of the PartitionConsumer. This method will return immediately, after which you
+	// should continue to service the 'Messages' and 'Errors' channels until they are empty. It is required to call this
+	// function, or Close before a consumer object passes out of scope, as it will otherwise leak memory. You must call
+	// this before calling Close on the underlying client.
+	AsyncClose()
+
+	// Close stops the PartitionConsumer from fetching messages. It will initiate a shutdown just like AsyncClose, drain
+	// the Messages channel, harvest any errors & return them to the caller. Note that if you are continuing to service
+	// the Messages channel when this function is called, you will be competing with Close for messages; consider
+	// calling AsyncClose, instead. It is required to call this function (or AsyncClose) before a consumer object passes
+	// out of scope, as it will otherwise leak memory. You must call this before calling Close on the underlying client.
+	Close() error
+
+	// Messages returns the read channel for the messages that are returned by
+	// the broker.
+	Messages() <-chan *ConsumerMessage
+
+	// Errors returns a read channel of errors that occurred during consuming, if
+	// enabled. By default, errors are logged and not returned over this channel.
+	// If you want to implement any custom error handling, set your config's
+	// Consumer.Return.Errors setting to true, and read from this channel.
+	Errors() <-chan *ConsumerError
+
+	// HighWaterMarkOffset returns the high water mark offset of the partition,
+	// i.e. the offset that will be used for the next message that will be produced.
+	// You can use this to determine how far behind the processing is.
+	HighWaterMarkOffset() int64
+}
+
+type partitionConsumer struct {
+	highWaterMarkOffset int64 // must be at the top of the struct because https://golang.org/pkg/sync/atomic/#pkg-note-BUG
+	consumer            *consumer
+	conf                *Config
+	topic               string
+	partition           int32
+
+	broker   *brokerConsumer
+	messages chan *ConsumerMessage
+	errors   chan *ConsumerError
+	feeder   chan *FetchResponse
+
+	trigger, dying chan none
+	responseResult error
+	closeOnce      sync.Once
+
+	fetchSize int32
+	offset    int64
+
+	retries int32
+}
+
+var errTimedOut = errors.New("timed out feeding messages to the user") // not user-facing
+
+func (child *partitionConsumer) sendError(err error) {
+	cErr := &ConsumerError{
+		Topic:     child.topic,
+		Partition: child.partition,
+		Err:       err,
+	}
+
+	if child.conf.Consumer.Return.Errors {
+		child.errors <- cErr
+	} else {
+		Logger.Println(cErr)
+	}
+}
+
+func (child *partitionConsumer) computeBackoff() time.Duration {
+	if child.conf.Consumer.Retry.BackoffFunc != nil {
+		retries := atomic.AddInt32(&child.retries, 1)
+		return child.conf.Consumer.Retry.BackoffFunc(int(retries))
+	} else {
+		return child.conf.Consumer.Retry.Backoff
+	}
+}
+
+func (child *partitionConsumer) dispatcher() {
+	for range child.trigger {
+		select {
+		case <-child.dying:
+			close(child.trigger)
+		case <-time.After(child.computeBackoff()):
+			if child.broker != nil {
+				child.consumer.unrefBrokerConsumer(child.broker)
+				child.broker = nil
+			}
+
+			Logger.Printf("consumer/%s/%d finding new broker\n", child.topic, child.partition)
+			if err := child.dispatch(); err != nil {
+				child.sendError(err)
+				child.trigger <- none{}
+			}
+		}
+	}
+
+	if child.broker != nil {
+		child.consumer.unrefBrokerConsumer(child.broker)
+	}
+	child.consumer.removeChild(child)
+	close(child.feeder)
+}
+
+func (child *partitionConsumer) dispatch() error {
+	if err := child.consumer.client.RefreshMetadata(child.topic); err != nil {
+		return err
+	}
+
+	var leader *Broker
+	var err error
+	if leader, err = child.consumer.client.Leader(child.topic, child.partition); err != nil {
+		return err
+	}
+
+	child.broker = child.consumer.refBrokerConsumer(leader)
+
+	child.broker.input <- child
+
+	return nil
+}
+
+func (child *partitionConsumer) chooseStartingOffset(offset int64) error {
+	newestOffset, err := child.consumer.client.GetOffset(child.topic, child.partition, OffsetNewest)
+	if err != nil {
+		return err
+	}
+	oldestOffset, err := child.consumer.client.GetOffset(child.topic, child.partition, OffsetOldest)
+	if err != nil {
+		return err
+	}
+
+	switch {
+	case offset == OffsetNewest:
+		child.offset = newestOffset
+	case offset == OffsetOldest:
+		child.offset = oldestOffset
+	case offset >= oldestOffset && offset <= newestOffset:
+		child.offset = offset
+	default:
+		return ErrOffsetOutOfRange
+	}
+
+	return nil
+}
+
+func (child *partitionConsumer) Messages() <-chan *ConsumerMessage {
+	return child.messages
+}
+
+func (child *partitionConsumer) Errors() <-chan *ConsumerError {
+	return child.errors
+}
+
+func (child *partitionConsumer) AsyncClose() {
+	// this triggers whatever broker owns this child to abandon it and close its trigger channel, which causes
+	// the dispatcher to exit its loop, which removes it from the consumer then closes its 'messages' and
+	// 'errors' channel (alternatively, if the child is already at the dispatcher for some reason, that will
+	// also just close itself)
+	child.closeOnce.Do(func() {
+		close(child.dying)
+	})
+}
+
+func (child *partitionConsumer) Close() error {
+	child.AsyncClose()
+
+	go withRecover(func() {
+		for range child.messages {
+			// drain
+		}
+	})
+
+	var errors ConsumerErrors
+	for err := range child.errors {
+		errors = append(errors, err)
+	}
+
+	if len(errors) > 0 {
+		return errors
+	}
+	return nil
+}
+
+func (child *partitionConsumer) HighWaterMarkOffset() int64 {
+	return atomic.LoadInt64(&child.highWaterMarkOffset)
+}
+
+func (child *partitionConsumer) responseFeeder() {
+	var msgs []*ConsumerMessage
+	expiryTicker := time.NewTicker(child.conf.Consumer.MaxProcessingTime)
+	firstAttempt := true
+
+feederLoop:
+	for response := range child.feeder {
+		msgs, child.responseResult = child.parseResponse(response)
+
+		if child.responseResult == nil {
+			atomic.StoreInt32(&child.retries, 0)
+		}
+
+		for i, msg := range msgs {
+		messageSelect:
+			select {
+			case child.messages <- msg:
+				firstAttempt = true
+			case <-expiryTicker.C:
+				if !firstAttempt {
+					child.responseResult = errTimedOut
+					child.broker.acks.Done()
+					for _, msg = range msgs[i:] {
+						child.messages <- msg
+					}
+					child.broker.input <- child
+					continue feederLoop
+				} else {
+					// current message has not been sent, return to select
+					// statement
+					firstAttempt = false
+					goto messageSelect
+				}
+			}
+		}
+
+		child.broker.acks.Done()
+	}
+
+	expiryTicker.Stop()
+	close(child.messages)
+	close(child.errors)
+}
+
+func (child *partitionConsumer) parseMessages(msgSet *MessageSet) ([]*ConsumerMessage, error) {
+	var messages []*ConsumerMessage
+	for _, msgBlock := range msgSet.Messages {
+		for _, msg := range msgBlock.Messages() {
+			offset := msg.Offset
+			timestamp := msg.Msg.Timestamp
+			if msg.Msg.Version >= 1 {
+				baseOffset := msgBlock.Offset - msgBlock.Messages()[len(msgBlock.Messages())-1].Offset
+				offset += baseOffset
+				if msg.Msg.LogAppendTime {
+					timestamp = msgBlock.Msg.Timestamp
+				}
+			}
+			if offset < child.offset {
+				continue
+			}
+			messages = append(messages, &ConsumerMessage{
+				Topic:          child.topic,
+				Partition:      child.partition,
+				Key:            msg.Msg.Key,
+				Value:          msg.Msg.Value,
+				Offset:         offset,
+				Timestamp:      timestamp,
+				BlockTimestamp: msgBlock.Msg.Timestamp,
+			})
+			child.offset = offset + 1
+		}
+	}
+	if len(messages) == 0 {
+		child.offset++
+	}
+	return messages, nil
+}
+
+func (child *partitionConsumer) parseRecords(batch *RecordBatch) ([]*ConsumerMessage, error) {
+	var messages []*ConsumerMessage
+	for _, rec := range batch.Records {
+		offset := batch.FirstOffset + rec.OffsetDelta
+		if offset < child.offset {
+			continue
+		}
+		timestamp := batch.FirstTimestamp.Add(rec.TimestampDelta)
+		if batch.LogAppendTime {
+			timestamp = batch.MaxTimestamp
+		}
+		messages = append(messages, &ConsumerMessage{
+			Topic:     child.topic,
+			Partition: child.partition,
+			Key:       rec.Key,
+			Value:     rec.Value,
+			Offset:    offset,
+			Timestamp: timestamp,
+			Headers:   rec.Headers,
+		})
+		child.offset = offset + 1
+	}
+	if len(messages) == 0 {
+		child.offset++
+	}
+	return messages, nil
+}
+
+func (child *partitionConsumer) parseResponse(response *FetchResponse) ([]*ConsumerMessage, error) {
+	block := response.GetBlock(child.topic, child.partition)
+	if block == nil {
+		return nil, ErrIncompleteResponse
+	}
+
+	if block.Err != ErrNoError {
+		return nil, block.Err
+	}
+
+	nRecs, err := block.numRecords()
+	if err != nil {
+		return nil, err
+	}
+	if nRecs == 0 {
+		partialTrailingMessage, err := block.isPartial()
+		if err != nil {
+			return nil, err
+		}
+		// We got no messages. If we got a trailing one then we need to ask for more data.
+		// Otherwise we just poll again and wait for one to be produced...
+		if partialTrailingMessage {
+			if child.conf.Consumer.Fetch.Max > 0 && child.fetchSize == child.conf.Consumer.Fetch.Max {
+				// we can't ask for more data, we've hit the configured limit
+				child.sendError(ErrMessageTooLarge)
+				child.offset++ // skip this one so we can keep processing future messages
+			} else {
+				child.fetchSize *= 2
+				if child.conf.Consumer.Fetch.Max > 0 && child.fetchSize > child.conf.Consumer.Fetch.Max {
+					child.fetchSize = child.conf.Consumer.Fetch.Max
+				}
+			}
+		}
+
+		return nil, nil
+	}
+
+	// we got messages, reset our fetch size in case it was increased for a previous request
+	child.fetchSize = child.conf.Consumer.Fetch.Default
+	atomic.StoreInt64(&child.highWaterMarkOffset, block.HighWaterMarkOffset)
+
+	messages := []*ConsumerMessage{}
+	for _, records := range block.RecordsSet {
+		switch records.recordsType {
+		case legacyRecords:
+			messageSetMessages, err := child.parseMessages(records.MsgSet)
+			if err != nil {
+				return nil, err
+			}
+
+			messages = append(messages, messageSetMessages...)
+		case defaultRecords:
+			recordBatchMessages, err := child.parseRecords(records.RecordBatch)
+			if err != nil {
+				return nil, err
+			}
+			if control, err := records.isControl(); err != nil || control {
+				continue
+			}
+
+			messages = append(messages, recordBatchMessages...)
+		default:
+			return nil, fmt.Errorf("unknown records type: %v", records.recordsType)
+		}
+	}
+
+	return messages, nil
+}
+
+// brokerConsumer
+
+type brokerConsumer struct {
+	consumer         *consumer
+	broker           *Broker
+	input            chan *partitionConsumer
+	newSubscriptions chan []*partitionConsumer
+	wait             chan none
+	subscriptions    map[*partitionConsumer]none
+	acks             sync.WaitGroup
+	refs             int
+}
+
+func (c *consumer) newBrokerConsumer(broker *Broker) *brokerConsumer {
+	bc := &brokerConsumer{
+		consumer:         c,
+		broker:           broker,
+		input:            make(chan *partitionConsumer),
+		newSubscriptions: make(chan []*partitionConsumer),
+		wait:             make(chan none),
+		subscriptions:    make(map[*partitionConsumer]none),
+		refs:             0,
+	}
+
+	go withRecover(bc.subscriptionManager)
+	go withRecover(bc.subscriptionConsumer)
+
+	return bc
+}
+
+func (bc *brokerConsumer) subscriptionManager() {
+	var buffer []*partitionConsumer
+
+	// The subscriptionManager constantly accepts new subscriptions on `input` (even when the main subscriptionConsumer
+	// goroutine is in the middle of a network request) and batches it up. The main worker goroutine picks
+	// up a batch of new subscriptions between every network request by reading from `newSubscriptions`, so we give
+	// it nil if no new subscriptions are available. We also write to `wait` only when new subscriptions is available,
+	// so the main goroutine can block waiting for work if it has none.
+	for {
+		if len(buffer) > 0 {
+			select {
+			case event, ok := <-bc.input:
+				if !ok {
+					goto done
+				}
+				buffer = append(buffer, event)
+			case bc.newSubscriptions <- buffer:
+				buffer = nil
+			case bc.wait <- none{}:
+			}
+		} else {
+			select {
+			case event, ok := <-bc.input:
+				if !ok {
+					goto done
+				}
+				buffer = append(buffer, event)
+			case bc.newSubscriptions <- nil:
+			}
+		}
+	}
+
+done:
+	close(bc.wait)
+	if len(buffer) > 0 {
+		bc.newSubscriptions <- buffer
+	}
+	close(bc.newSubscriptions)
+}
+
+func (bc *brokerConsumer) subscriptionConsumer() {
+	<-bc.wait // wait for our first piece of work
+
+	// the subscriptionConsumer ensures we will get nil right away if no new subscriptions is available
+	for newSubscriptions := range bc.newSubscriptions {
+		bc.updateSubscriptions(newSubscriptions)
+
+		if len(bc.subscriptions) == 0 {
+			// We're about to be shut down or we're about to receive more subscriptions.
+			// Either way, the signal just hasn't propagated to our goroutine yet.
+			<-bc.wait
+			continue
+		}
+
+		response, err := bc.fetchNewMessages()
+
+		if err != nil {
+			Logger.Printf("consumer/broker/%d disconnecting due to error processing FetchRequest: %s\n", bc.broker.ID(), err)
+			bc.abort(err)
+			return
+		}
+
+		bc.acks.Add(len(bc.subscriptions))
+		for child := range bc.subscriptions {
+			child.feeder <- response
+		}
+		bc.acks.Wait()
+		bc.handleResponses()
+	}
+}
+
+func (bc *brokerConsumer) updateSubscriptions(newSubscriptions []*partitionConsumer) {
+	for _, child := range newSubscriptions {
+		bc.subscriptions[child] = none{}
+		Logger.Printf("consumer/broker/%d added subscription to %s/%d\n", bc.broker.ID(), child.topic, child.partition)
+	}
+
+	for child := range bc.subscriptions {
+		select {
+		case <-child.dying:
+			Logger.Printf("consumer/broker/%d closed dead subscription to %s/%d\n", bc.broker.ID(), child.topic, child.partition)
+			close(child.trigger)
+			delete(bc.subscriptions, child)
+		default:
+			break
+		}
+	}
+}
+
+func (bc *brokerConsumer) handleResponses() {
+	// handles the response codes left for us by our subscriptions, and abandons ones that have been closed
+	for child := range bc.subscriptions {
+		result := child.responseResult
+		child.responseResult = nil
+
+		switch result {
+		case nil:
+			break
+		case errTimedOut:
+			Logger.Printf("consumer/broker/%d abandoned subscription to %s/%d because consuming was taking too long\n",
+				bc.broker.ID(), child.topic, child.partition)
+			delete(bc.subscriptions, child)
+		case ErrOffsetOutOfRange:
+			// there's no point in retrying this it will just fail the same way again
+			// shut it down and force the user to choose what to do
+			child.sendError(result)
+			Logger.Printf("consumer/%s/%d shutting down because %s\n", child.topic, child.partition, result)
+			close(child.trigger)
+			delete(bc.subscriptions, child)
+		case ErrUnknownTopicOrPartition, ErrNotLeaderForPartition, ErrLeaderNotAvailable, ErrReplicaNotAvailable:
+			// not an error, but does need redispatching
+			Logger.Printf("consumer/broker/%d abandoned subscription to %s/%d because %s\n",
+				bc.broker.ID(), child.topic, child.partition, result)
+			child.trigger <- none{}
+			delete(bc.subscriptions, child)
+		default:
+			// dunno, tell the user and try redispatching
+			child.sendError(result)
+			Logger.Printf("consumer/broker/%d abandoned subscription to %s/%d because %s\n",
+				bc.broker.ID(), child.topic, child.partition, result)
+			child.trigger <- none{}
+			delete(bc.subscriptions, child)
+		}
+	}
+}
+
+func (bc *brokerConsumer) abort(err error) {
+	bc.consumer.abandonBrokerConsumer(bc)
+	_ = bc.broker.Close() // we don't care about the error this might return, we already have one
+
+	for child := range bc.subscriptions {
+		child.sendError(err)
+		child.trigger <- none{}
+	}
+
+	for newSubscriptions := range bc.newSubscriptions {
+		if len(newSubscriptions) == 0 {
+			<-bc.wait
+			continue
+		}
+		for _, child := range newSubscriptions {
+			child.sendError(err)
+			child.trigger <- none{}
+		}
+	}
+}
+
+func (bc *brokerConsumer) fetchNewMessages() (*FetchResponse, error) {
+	request := &FetchRequest{
+		MinBytes:    bc.consumer.conf.Consumer.Fetch.Min,
+		MaxWaitTime: int32(bc.consumer.conf.Consumer.MaxWaitTime / time.Millisecond),
+	}
+	if bc.consumer.conf.Version.IsAtLeast(V0_9_0_0) {
+		request.Version = 1
+	}
+	if bc.consumer.conf.Version.IsAtLeast(V0_10_0_0) {
+		request.Version = 2
+	}
+	if bc.consumer.conf.Version.IsAtLeast(V0_10_1_0) {
+		request.Version = 3
+		request.MaxBytes = MaxResponseSize
+	}
+	if bc.consumer.conf.Version.IsAtLeast(V0_11_0_0) {
+		request.Version = 4
+		request.Isolation = ReadUncommitted // We don't support yet transactions.
+	}
+
+	for child := range bc.subscriptions {
+		request.AddBlock(child.topic, child.partition, child.offset, child.fetchSize)
+	}
+
+	return bc.broker.Fetch(request)
+}
diff --git a/vendor/github.com/Shopify/sarama/consumer_group.go b/vendor/github.com/Shopify/sarama/consumer_group.go
new file mode 100644
index 0000000..8c8babc
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/consumer_group.go
@@ -0,0 +1,780 @@
+package sarama
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sort"
+	"sync"
+	"time"
+)
+
+// ErrClosedConsumerGroup is the error returned when a method is called on a consumer group that has been closed.
+var ErrClosedConsumerGroup = errors.New("kafka: tried to use a consumer group that was closed")
+
+// ConsumerGroup is responsible for dividing up processing of topics and partitions
+// over a collection of processes (the members of the consumer group).
+type ConsumerGroup interface {
+	// Consume joins a cluster of consumers for a given list of topics and
+	// starts a blocking ConsumerGroupSession through the ConsumerGroupHandler.
+	//
+	// The life-cycle of a session is represented by the following steps:
+	//
+	// 1. The consumers join the group (as explained in https://kafka.apache.org/documentation/#intro_consumers)
+	//    and is assigned their "fair share" of partitions, aka 'claims'.
+	// 2. Before processing starts, the handler's Setup() hook is called to notify the user
+	//    of the claims and allow any necessary preparation or alteration of state.
+	// 3. For each of the assigned claims the handler's ConsumeClaim() function is then called
+	//    in a separate goroutine which requires it to be thread-safe. Any state must be carefully protected
+	//    from concurrent reads/writes.
+	// 4. The session will persist until one of the ConsumeClaim() functions exits. This can be either when the
+	//    parent context is cancelled or when a server-side rebalance cycle is initiated.
+	// 5. Once all the ConsumeClaim() loops have exited, the handler's Cleanup() hook is called
+	//    to allow the user to perform any final tasks before a rebalance.
+	// 6. Finally, marked offsets are committed one last time before claims are released.
+	//
+	// Please note, that once a rebalance is triggered, sessions must be completed within
+	// Config.Consumer.Group.Rebalance.Timeout. This means that ConsumeClaim() functions must exit
+	// as quickly as possible to allow time for Cleanup() and the final offset commit. If the timeout
+	// is exceeded, the consumer will be removed from the group by Kafka, which will cause offset
+	// commit failures.
+	Consume(ctx context.Context, topics []string, handler ConsumerGroupHandler) error
+
+	// Errors returns a read channel of errors that occurred during the consumer life-cycle.
+	// By default, errors are logged and not returned over this channel.
+	// If you want to implement any custom error handling, set your config's
+	// Consumer.Return.Errors setting to true, and read from this channel.
+	Errors() <-chan error
+
+	// Close stops the ConsumerGroup and detaches any running sessions. It is required to call
+	// this function before the object passes out of scope, as it will otherwise leak memory.
+	Close() error
+}
+
+type consumerGroup struct {
+	client    Client
+	ownClient bool
+
+	config   *Config
+	consumer Consumer
+	groupID  string
+	memberID string
+	errors   chan error
+
+	lock      sync.Mutex
+	closed    chan none
+	closeOnce sync.Once
+}
+
+// NewConsumerGroup creates a new consumer group the given broker addresses and configuration.
+func NewConsumerGroup(addrs []string, groupID string, config *Config) (ConsumerGroup, error) {
+	client, err := NewClient(addrs, config)
+	if err != nil {
+		return nil, err
+	}
+
+	c, err := NewConsumerGroupFromClient(groupID, client)
+	if err != nil {
+		_ = client.Close()
+		return nil, err
+	}
+
+	c.(*consumerGroup).ownClient = true
+	return c, nil
+}
+
+// NewConsumerGroupFromClient creates a new consumer group using the given client. It is still
+// necessary to call Close() on the underlying client when shutting down this consumer.
+// PLEASE NOTE: consumer groups can only re-use but not share clients.
+func NewConsumerGroupFromClient(groupID string, client Client) (ConsumerGroup, error) {
+	config := client.Config()
+	if !config.Version.IsAtLeast(V0_10_2_0) {
+		return nil, ConfigurationError("consumer groups require Version to be >= V0_10_2_0")
+	}
+
+	consumer, err := NewConsumerFromClient(client)
+	if err != nil {
+		return nil, err
+	}
+
+	return &consumerGroup{
+		client:   client,
+		consumer: consumer,
+		config:   config,
+		groupID:  groupID,
+		errors:   make(chan error, config.ChannelBufferSize),
+		closed:   make(chan none),
+	}, nil
+}
+
+// Errors implements ConsumerGroup.
+func (c *consumerGroup) Errors() <-chan error { return c.errors }
+
+// Close implements ConsumerGroup.
+func (c *consumerGroup) Close() (err error) {
+	c.closeOnce.Do(func() {
+		close(c.closed)
+
+		c.lock.Lock()
+		defer c.lock.Unlock()
+
+		// leave group
+		if e := c.leave(); e != nil {
+			err = e
+		}
+
+		// drain errors
+		go func() {
+			close(c.errors)
+		}()
+		for e := range c.errors {
+			err = e
+		}
+
+		if c.ownClient {
+			if e := c.client.Close(); e != nil {
+				err = e
+			}
+		}
+	})
+	return
+}
+
+// Consume implements ConsumerGroup.
+func (c *consumerGroup) Consume(ctx context.Context, topics []string, handler ConsumerGroupHandler) error {
+	// Ensure group is not closed
+	select {
+	case <-c.closed:
+		return ErrClosedConsumerGroup
+	default:
+	}
+
+	c.lock.Lock()
+	defer c.lock.Unlock()
+
+	// Quick exit when no topics are provided
+	if len(topics) == 0 {
+		return fmt.Errorf("no topics provided")
+	}
+
+	// Refresh metadata for requested topics
+	if err := c.client.RefreshMetadata(topics...); err != nil {
+		return err
+	}
+
+	// Get coordinator
+	coordinator, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		return err
+	}
+
+	// Init session
+	sess, err := c.newSession(ctx, coordinator, topics, handler, c.config.Consumer.Group.Rebalance.Retry.Max)
+	if err == ErrClosedClient {
+		return ErrClosedConsumerGroup
+	} else if err != nil {
+		return err
+	}
+
+	// Wait for session exit signal
+	<-sess.ctx.Done()
+
+	// Gracefully release session claims
+	return sess.release(true)
+}
+
+func (c *consumerGroup) newSession(ctx context.Context, coordinator *Broker, topics []string, handler ConsumerGroupHandler, retries int) (*consumerGroupSession, error) {
+	// Join consumer group
+	join, err := c.joinGroupRequest(coordinator, topics)
+	if err != nil {
+		_ = coordinator.Close()
+		return nil, err
+	}
+	switch join.Err {
+	case ErrNoError:
+		c.memberID = join.MemberId
+	case ErrUnknownMemberId, ErrIllegalGeneration: // reset member ID and retry immediately
+		c.memberID = ""
+		return c.newSession(ctx, coordinator, topics, handler, retries)
+	case ErrRebalanceInProgress: // retry after backoff
+		if retries <= 0 {
+			return nil, join.Err
+		}
+
+		select {
+		case <-c.closed:
+			return nil, ErrClosedConsumerGroup
+		case <-time.After(c.config.Consumer.Group.Rebalance.Retry.Backoff):
+		}
+
+		return c.newSession(ctx, coordinator, topics, handler, retries-1)
+	default:
+		return nil, join.Err
+	}
+
+	// Prepare distribution plan if we joined as the leader
+	var plan BalanceStrategyPlan
+	if join.LeaderId == join.MemberId {
+		members, err := join.GetMembers()
+		if err != nil {
+			return nil, err
+		}
+
+		plan, err = c.balance(members)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	// Sync consumer group
+	sync, err := c.syncGroupRequest(coordinator, plan, join.GenerationId)
+	if err != nil {
+		_ = coordinator.Close()
+		return nil, err
+	}
+	switch sync.Err {
+	case ErrNoError:
+	case ErrUnknownMemberId, ErrIllegalGeneration: // reset member ID and retry immediately
+		c.memberID = ""
+		return c.newSession(ctx, coordinator, topics, handler, retries)
+	case ErrRebalanceInProgress: // retry after backoff
+		if retries <= 0 {
+			return nil, sync.Err
+		}
+
+		select {
+		case <-c.closed:
+			return nil, ErrClosedConsumerGroup
+		case <-time.After(c.config.Consumer.Group.Rebalance.Retry.Backoff):
+		}
+
+		return c.newSession(ctx, coordinator, topics, handler, retries-1)
+	default:
+		return nil, sync.Err
+	}
+
+	// Retrieve and sort claims
+	var claims map[string][]int32
+	if len(sync.MemberAssignment) > 0 {
+		members, err := sync.GetMemberAssignment()
+		if err != nil {
+			return nil, err
+		}
+		claims = members.Topics
+
+		for _, partitions := range claims {
+			sort.Sort(int32Slice(partitions))
+		}
+	}
+
+	return newConsumerGroupSession(ctx, c, claims, join.MemberId, join.GenerationId, handler)
+}
+
+func (c *consumerGroup) joinGroupRequest(coordinator *Broker, topics []string) (*JoinGroupResponse, error) {
+	req := &JoinGroupRequest{
+		GroupId:        c.groupID,
+		MemberId:       c.memberID,
+		SessionTimeout: int32(c.config.Consumer.Group.Session.Timeout / time.Millisecond),
+		ProtocolType:   "consumer",
+	}
+	if c.config.Version.IsAtLeast(V0_10_1_0) {
+		req.Version = 1
+		req.RebalanceTimeout = int32(c.config.Consumer.Group.Rebalance.Timeout / time.Millisecond)
+	}
+
+	meta := &ConsumerGroupMemberMetadata{
+		Topics:   topics,
+		UserData: c.config.Consumer.Group.Member.UserData,
+	}
+	strategy := c.config.Consumer.Group.Rebalance.Strategy
+	if err := req.AddGroupProtocolMetadata(strategy.Name(), meta); err != nil {
+		return nil, err
+	}
+
+	return coordinator.JoinGroup(req)
+}
+
+func (c *consumerGroup) syncGroupRequest(coordinator *Broker, plan BalanceStrategyPlan, generationID int32) (*SyncGroupResponse, error) {
+	req := &SyncGroupRequest{
+		GroupId:      c.groupID,
+		MemberId:     c.memberID,
+		GenerationId: generationID,
+	}
+	for memberID, topics := range plan {
+		err := req.AddGroupAssignmentMember(memberID, &ConsumerGroupMemberAssignment{
+			Topics: topics,
+		})
+		if err != nil {
+			return nil, err
+		}
+	}
+	return coordinator.SyncGroup(req)
+}
+
+func (c *consumerGroup) heartbeatRequest(coordinator *Broker, memberID string, generationID int32) (*HeartbeatResponse, error) {
+	req := &HeartbeatRequest{
+		GroupId:      c.groupID,
+		MemberId:     memberID,
+		GenerationId: generationID,
+	}
+
+	return coordinator.Heartbeat(req)
+}
+
+func (c *consumerGroup) balance(members map[string]ConsumerGroupMemberMetadata) (BalanceStrategyPlan, error) {
+	topics := make(map[string][]int32)
+	for _, meta := range members {
+		for _, topic := range meta.Topics {
+			topics[topic] = nil
+		}
+	}
+
+	for topic := range topics {
+		partitions, err := c.client.Partitions(topic)
+		if err != nil {
+			return nil, err
+		}
+		topics[topic] = partitions
+	}
+
+	strategy := c.config.Consumer.Group.Rebalance.Strategy
+	return strategy.Plan(members, topics)
+}
+
+// Leaves the cluster, called by Close, protected by lock.
+func (c *consumerGroup) leave() error {
+	if c.memberID == "" {
+		return nil
+	}
+
+	coordinator, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		return err
+	}
+
+	resp, err := coordinator.LeaveGroup(&LeaveGroupRequest{
+		GroupId:  c.groupID,
+		MemberId: c.memberID,
+	})
+	if err != nil {
+		_ = coordinator.Close()
+		return err
+	}
+
+	// Unset memberID
+	c.memberID = ""
+
+	// Check response
+	switch resp.Err {
+	case ErrRebalanceInProgress, ErrUnknownMemberId, ErrNoError:
+		return nil
+	default:
+		return resp.Err
+	}
+}
+
+func (c *consumerGroup) handleError(err error, topic string, partition int32) {
+	select {
+	case <-c.closed:
+		return
+	default:
+	}
+
+	if _, ok := err.(*ConsumerError); !ok && topic != "" && partition > -1 {
+		err = &ConsumerError{
+			Topic:     topic,
+			Partition: partition,
+			Err:       err,
+		}
+	}
+
+	if c.config.Consumer.Return.Errors {
+		select {
+		case c.errors <- err:
+		default:
+		}
+	} else {
+		Logger.Println(err)
+	}
+}
+
+// --------------------------------------------------------------------
+
+// ConsumerGroupSession represents a consumer group member session.
+type ConsumerGroupSession interface {
+	// Claims returns information about the claimed partitions by topic.
+	Claims() map[string][]int32
+
+	// MemberID returns the cluster member ID.
+	MemberID() string
+
+	// GenerationID returns the current generation ID.
+	GenerationID() int32
+
+	// MarkOffset marks the provided offset, alongside a metadata string
+	// that represents the state of the partition consumer at that point in time. The
+	// metadata string can be used by another consumer to restore that state, so it
+	// can resume consumption.
+	//
+	// To follow upstream conventions, you are expected to mark the offset of the
+	// next message to read, not the last message read. Thus, when calling `MarkOffset`
+	// you should typically add one to the offset of the last consumed message.
+	//
+	// Note: calling MarkOffset does not necessarily commit the offset to the backend
+	// store immediately for efficiency reasons, and it may never be committed if
+	// your application crashes. This means that you may end up processing the same
+	// message twice, and your processing should ideally be idempotent.
+	MarkOffset(topic string, partition int32, offset int64, metadata string)
+
+	// ResetOffset resets to the provided offset, alongside a metadata string that
+	// represents the state of the partition consumer at that point in time. Reset
+	// acts as a counterpart to MarkOffset, the difference being that it allows to
+	// reset an offset to an earlier or smaller value, where MarkOffset only
+	// allows incrementing the offset. cf MarkOffset for more details.
+	ResetOffset(topic string, partition int32, offset int64, metadata string)
+
+	// MarkMessage marks a message as consumed.
+	MarkMessage(msg *ConsumerMessage, metadata string)
+
+	// Context returns the session context.
+	Context() context.Context
+}
+
+type consumerGroupSession struct {
+	parent       *consumerGroup
+	memberID     string
+	generationID int32
+	handler      ConsumerGroupHandler
+
+	claims  map[string][]int32
+	offsets *offsetManager
+	ctx     context.Context
+	cancel  func()
+
+	waitGroup       sync.WaitGroup
+	releaseOnce     sync.Once
+	hbDying, hbDead chan none
+}
+
+func newConsumerGroupSession(ctx context.Context, parent *consumerGroup, claims map[string][]int32, memberID string, generationID int32, handler ConsumerGroupHandler) (*consumerGroupSession, error) {
+	// init offset manager
+	offsets, err := newOffsetManagerFromClient(parent.groupID, memberID, generationID, parent.client)
+	if err != nil {
+		return nil, err
+	}
+
+	// init context
+	ctx, cancel := context.WithCancel(ctx)
+
+	// init session
+	sess := &consumerGroupSession{
+		parent:       parent,
+		memberID:     memberID,
+		generationID: generationID,
+		handler:      handler,
+		offsets:      offsets,
+		claims:       claims,
+		ctx:          ctx,
+		cancel:       cancel,
+		hbDying:      make(chan none),
+		hbDead:       make(chan none),
+	}
+
+	// start heartbeat loop
+	go sess.heartbeatLoop()
+
+	// create a POM for each claim
+	for topic, partitions := range claims {
+		for _, partition := range partitions {
+			pom, err := offsets.ManagePartition(topic, partition)
+			if err != nil {
+				_ = sess.release(false)
+				return nil, err
+			}
+
+			// handle POM errors
+			go func(topic string, partition int32) {
+				for err := range pom.Errors() {
+					sess.parent.handleError(err, topic, partition)
+				}
+			}(topic, partition)
+		}
+	}
+
+	// perform setup
+	if err := handler.Setup(sess); err != nil {
+		_ = sess.release(true)
+		return nil, err
+	}
+
+	// start consuming
+	for topic, partitions := range claims {
+		for _, partition := range partitions {
+			sess.waitGroup.Add(1)
+
+			go func(topic string, partition int32) {
+				defer sess.waitGroup.Done()
+
+				// cancel the as session as soon as the first
+				// goroutine exits
+				defer sess.cancel()
+
+				// consume a single topic/partition, blocking
+				sess.consume(topic, partition)
+			}(topic, partition)
+		}
+	}
+	return sess, nil
+}
+
+func (s *consumerGroupSession) Claims() map[string][]int32 { return s.claims }
+func (s *consumerGroupSession) MemberID() string           { return s.memberID }
+func (s *consumerGroupSession) GenerationID() int32        { return s.generationID }
+
+func (s *consumerGroupSession) MarkOffset(topic string, partition int32, offset int64, metadata string) {
+	if pom := s.offsets.findPOM(topic, partition); pom != nil {
+		pom.MarkOffset(offset, metadata)
+	}
+}
+
+func (s *consumerGroupSession) ResetOffset(topic string, partition int32, offset int64, metadata string) {
+	if pom := s.offsets.findPOM(topic, partition); pom != nil {
+		pom.ResetOffset(offset, metadata)
+	}
+}
+
+func (s *consumerGroupSession) MarkMessage(msg *ConsumerMessage, metadata string) {
+	s.MarkOffset(msg.Topic, msg.Partition, msg.Offset+1, metadata)
+}
+
+func (s *consumerGroupSession) Context() context.Context {
+	return s.ctx
+}
+
+func (s *consumerGroupSession) consume(topic string, partition int32) {
+	// quick exit if rebalance is due
+	select {
+	case <-s.ctx.Done():
+		return
+	case <-s.parent.closed:
+		return
+	default:
+	}
+
+	// get next offset
+	offset := s.parent.config.Consumer.Offsets.Initial
+	if pom := s.offsets.findPOM(topic, partition); pom != nil {
+		offset, _ = pom.NextOffset()
+	}
+
+	// create new claim
+	claim, err := newConsumerGroupClaim(s, topic, partition, offset)
+	if err != nil {
+		s.parent.handleError(err, topic, partition)
+		return
+	}
+
+	// handle errors
+	go func() {
+		for err := range claim.Errors() {
+			s.parent.handleError(err, topic, partition)
+		}
+	}()
+
+	// trigger close when session is done
+	go func() {
+		select {
+		case <-s.ctx.Done():
+		case <-s.parent.closed:
+		}
+		claim.AsyncClose()
+	}()
+
+	// start processing
+	if err := s.handler.ConsumeClaim(s, claim); err != nil {
+		s.parent.handleError(err, topic, partition)
+	}
+
+	// ensure consumer is closed & drained
+	claim.AsyncClose()
+	for _, err := range claim.waitClosed() {
+		s.parent.handleError(err, topic, partition)
+	}
+}
+
+func (s *consumerGroupSession) release(withCleanup bool) (err error) {
+	// signal release, stop heartbeat
+	s.cancel()
+
+	// wait for consumers to exit
+	s.waitGroup.Wait()
+
+	// perform release
+	s.releaseOnce.Do(func() {
+		if withCleanup {
+			if e := s.handler.Cleanup(s); e != nil {
+				s.parent.handleError(err, "", -1)
+				err = e
+			}
+		}
+
+		if e := s.offsets.Close(); e != nil {
+			err = e
+		}
+
+		close(s.hbDying)
+		<-s.hbDead
+	})
+
+	return
+}
+
+func (s *consumerGroupSession) heartbeatLoop() {
+	defer close(s.hbDead)
+	defer s.cancel() // trigger the end of the session on exit
+
+	pause := time.NewTicker(s.parent.config.Consumer.Group.Heartbeat.Interval)
+	defer pause.Stop()
+
+	retries := s.parent.config.Metadata.Retry.Max
+	for {
+		coordinator, err := s.parent.client.Coordinator(s.parent.groupID)
+		if err != nil {
+			if retries <= 0 {
+				s.parent.handleError(err, "", -1)
+				return
+			}
+
+			select {
+			case <-s.hbDying:
+				return
+			case <-time.After(s.parent.config.Metadata.Retry.Backoff):
+				retries--
+			}
+			continue
+		}
+
+		resp, err := s.parent.heartbeatRequest(coordinator, s.memberID, s.generationID)
+		if err != nil {
+			_ = coordinator.Close()
+
+			if retries <= 0 {
+				s.parent.handleError(err, "", -1)
+				return
+			}
+
+			retries--
+			continue
+		}
+
+		switch resp.Err {
+		case ErrNoError:
+			retries = s.parent.config.Metadata.Retry.Max
+		case ErrRebalanceInProgress, ErrUnknownMemberId, ErrIllegalGeneration:
+			return
+		default:
+			s.parent.handleError(err, "", -1)
+			return
+		}
+
+		select {
+		case <-pause.C:
+		case <-s.hbDying:
+			return
+		}
+	}
+}
+
+// --------------------------------------------------------------------
+
+// ConsumerGroupHandler instances are used to handle individual topic/partition claims.
+// It also provides hooks for your consumer group session life-cycle and allow you to
+// trigger logic before or after the consume loop(s).
+//
+// PLEASE NOTE that handlers are likely be called from several goroutines concurrently,
+// ensure that all state is safely protected against race conditions.
+type ConsumerGroupHandler interface {
+	// Setup is run at the beginning of a new session, before ConsumeClaim.
+	Setup(ConsumerGroupSession) error
+
+	// Cleanup is run at the end of a session, once all ConsumeClaim goroutines have exited
+	// but before the offsets are committed for the very last time.
+	Cleanup(ConsumerGroupSession) error
+
+	// ConsumeClaim must start a consumer loop of ConsumerGroupClaim's Messages().
+	// Once the Messages() channel is closed, the Handler must finish its processing
+	// loop and exit.
+	ConsumeClaim(ConsumerGroupSession, ConsumerGroupClaim) error
+}
+
+// ConsumerGroupClaim processes Kafka messages from a given topic and partition within a consumer group.
+type ConsumerGroupClaim interface {
+	// Topic returns the consumed topic name.
+	Topic() string
+
+	// Partition returns the consumed partition.
+	Partition() int32
+
+	// InitialOffset returns the initial offset that was used as a starting point for this claim.
+	InitialOffset() int64
+
+	// HighWaterMarkOffset returns the high water mark offset of the partition,
+	// i.e. the offset that will be used for the next message that will be produced.
+	// You can use this to determine how far behind the processing is.
+	HighWaterMarkOffset() int64
+
+	// Messages returns the read channel for the messages that are returned by
+	// the broker. The messages channel will be closed when a new rebalance cycle
+	// is due. You must finish processing and mark offsets within
+	// Config.Consumer.Group.Session.Timeout before the topic/partition is eventually
+	// re-assigned to another group member.
+	Messages() <-chan *ConsumerMessage
+}
+
+type consumerGroupClaim struct {
+	topic     string
+	partition int32
+	offset    int64
+	PartitionConsumer
+}
+
+func newConsumerGroupClaim(sess *consumerGroupSession, topic string, partition int32, offset int64) (*consumerGroupClaim, error) {
+	pcm, err := sess.parent.consumer.ConsumePartition(topic, partition, offset)
+	if err == ErrOffsetOutOfRange {
+		offset = sess.parent.config.Consumer.Offsets.Initial
+		pcm, err = sess.parent.consumer.ConsumePartition(topic, partition, offset)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	go func() {
+		for err := range pcm.Errors() {
+			sess.parent.handleError(err, topic, partition)
+		}
+	}()
+
+	return &consumerGroupClaim{
+		topic:             topic,
+		partition:         partition,
+		offset:            offset,
+		PartitionConsumer: pcm,
+	}, nil
+}
+
+func (c *consumerGroupClaim) Topic() string        { return c.topic }
+func (c *consumerGroupClaim) Partition() int32     { return c.partition }
+func (c *consumerGroupClaim) InitialOffset() int64 { return c.offset }
+
+// Drains messages and errors, ensures the claim is fully closed.
+func (c *consumerGroupClaim) waitClosed() (errs ConsumerErrors) {
+	go func() {
+		for range c.Messages() {
+		}
+	}()
+
+	for err := range c.Errors() {
+		errs = append(errs, err)
+	}
+	return
+}
diff --git a/vendor/github.com/Shopify/sarama/consumer_group_members.go b/vendor/github.com/Shopify/sarama/consumer_group_members.go
new file mode 100644
index 0000000..2d02cc3
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/consumer_group_members.go
@@ -0,0 +1,96 @@
+package sarama
+
+//ConsumerGroupMemberMetadata holds the metadata for consumer group
+type ConsumerGroupMemberMetadata struct {
+	Version  int16
+	Topics   []string
+	UserData []byte
+}
+
+func (m *ConsumerGroupMemberMetadata) encode(pe packetEncoder) error {
+	pe.putInt16(m.Version)
+
+	if err := pe.putStringArray(m.Topics); err != nil {
+		return err
+	}
+
+	if err := pe.putBytes(m.UserData); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m *ConsumerGroupMemberMetadata) decode(pd packetDecoder) (err error) {
+	if m.Version, err = pd.getInt16(); err != nil {
+		return
+	}
+
+	if m.Topics, err = pd.getStringArray(); err != nil {
+		return
+	}
+
+	if m.UserData, err = pd.getBytes(); err != nil {
+		return
+	}
+
+	return nil
+}
+
+//ConsumerGroupMemberAssignment holds the member assignment for a consume group
+type ConsumerGroupMemberAssignment struct {
+	Version  int16
+	Topics   map[string][]int32
+	UserData []byte
+}
+
+func (m *ConsumerGroupMemberAssignment) encode(pe packetEncoder) error {
+	pe.putInt16(m.Version)
+
+	if err := pe.putArrayLength(len(m.Topics)); err != nil {
+		return err
+	}
+
+	for topic, partitions := range m.Topics {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putInt32Array(partitions); err != nil {
+			return err
+		}
+	}
+
+	if err := pe.putBytes(m.UserData); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m *ConsumerGroupMemberAssignment) decode(pd packetDecoder) (err error) {
+	if m.Version, err = pd.getInt16(); err != nil {
+		return
+	}
+
+	var topicLen int
+	if topicLen, err = pd.getArrayLength(); err != nil {
+		return
+	}
+
+	m.Topics = make(map[string][]int32, topicLen)
+	for i := 0; i < topicLen; i++ {
+		var topic string
+		if topic, err = pd.getString(); err != nil {
+			return
+		}
+		if m.Topics[topic], err = pd.getInt32Array(); err != nil {
+			return
+		}
+	}
+
+	if m.UserData, err = pd.getBytes(); err != nil {
+		return
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/consumer_metadata_request.go b/vendor/github.com/Shopify/sarama/consumer_metadata_request.go
new file mode 100644
index 0000000..a8dcaef
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/consumer_metadata_request.go
@@ -0,0 +1,34 @@
+package sarama
+
+//ConsumerMetadataRequest is used for metadata requests
+type ConsumerMetadataRequest struct {
+	ConsumerGroup string
+}
+
+func (r *ConsumerMetadataRequest) encode(pe packetEncoder) error {
+	tmp := new(FindCoordinatorRequest)
+	tmp.CoordinatorKey = r.ConsumerGroup
+	tmp.CoordinatorType = CoordinatorGroup
+	return tmp.encode(pe)
+}
+
+func (r *ConsumerMetadataRequest) decode(pd packetDecoder, version int16) (err error) {
+	tmp := new(FindCoordinatorRequest)
+	if err := tmp.decode(pd, version); err != nil {
+		return err
+	}
+	r.ConsumerGroup = tmp.CoordinatorKey
+	return nil
+}
+
+func (r *ConsumerMetadataRequest) key() int16 {
+	return 10
+}
+
+func (r *ConsumerMetadataRequest) version() int16 {
+	return 0
+}
+
+func (r *ConsumerMetadataRequest) requiredVersion() KafkaVersion {
+	return V0_8_2_0
+}
diff --git a/vendor/github.com/Shopify/sarama/consumer_metadata_response.go b/vendor/github.com/Shopify/sarama/consumer_metadata_response.go
new file mode 100644
index 0000000..f39a871
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/consumer_metadata_response.go
@@ -0,0 +1,78 @@
+package sarama
+
+import (
+	"net"
+	"strconv"
+)
+
+//ConsumerMetadataResponse holds the response for a consumer group meta data requests
+type ConsumerMetadataResponse struct {
+	Err             KError
+	Coordinator     *Broker
+	CoordinatorID   int32  // deprecated: use Coordinator.ID()
+	CoordinatorHost string // deprecated: use Coordinator.Addr()
+	CoordinatorPort int32  // deprecated: use Coordinator.Addr()
+}
+
+func (r *ConsumerMetadataResponse) decode(pd packetDecoder, version int16) (err error) {
+	tmp := new(FindCoordinatorResponse)
+
+	if err := tmp.decode(pd, version); err != nil {
+		return err
+	}
+
+	r.Err = tmp.Err
+
+	r.Coordinator = tmp.Coordinator
+	if tmp.Coordinator == nil {
+		return nil
+	}
+
+	// this can all go away in 2.0, but we have to fill in deprecated fields to maintain
+	// backwards compatibility
+	host, portstr, err := net.SplitHostPort(r.Coordinator.Addr())
+	if err != nil {
+		return err
+	}
+	port, err := strconv.ParseInt(portstr, 10, 32)
+	if err != nil {
+		return err
+	}
+	r.CoordinatorID = r.Coordinator.ID()
+	r.CoordinatorHost = host
+	r.CoordinatorPort = int32(port)
+
+	return nil
+}
+
+func (r *ConsumerMetadataResponse) encode(pe packetEncoder) error {
+	if r.Coordinator == nil {
+		r.Coordinator = new(Broker)
+		r.Coordinator.id = r.CoordinatorID
+		r.Coordinator.addr = net.JoinHostPort(r.CoordinatorHost, strconv.Itoa(int(r.CoordinatorPort)))
+	}
+
+	tmp := &FindCoordinatorResponse{
+		Version:     0,
+		Err:         r.Err,
+		Coordinator: r.Coordinator,
+	}
+
+	if err := tmp.encode(pe); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *ConsumerMetadataResponse) key() int16 {
+	return 10
+}
+
+func (r *ConsumerMetadataResponse) version() int16 {
+	return 0
+}
+
+func (r *ConsumerMetadataResponse) requiredVersion() KafkaVersion {
+	return V0_8_2_0
+}
diff --git a/vendor/github.com/Shopify/sarama/crc32_field.go b/vendor/github.com/Shopify/sarama/crc32_field.go
new file mode 100644
index 0000000..1f14443
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/crc32_field.go
@@ -0,0 +1,69 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+)
+
+type crcPolynomial int8
+
+const (
+	crcIEEE crcPolynomial = iota
+	crcCastagnoli
+)
+
+var castagnoliTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc32Field implements the pushEncoder and pushDecoder interfaces for calculating CRC32s.
+type crc32Field struct {
+	startOffset int
+	polynomial  crcPolynomial
+}
+
+func (c *crc32Field) saveOffset(in int) {
+	c.startOffset = in
+}
+
+func (c *crc32Field) reserveLength() int {
+	return 4
+}
+
+func newCRC32Field(polynomial crcPolynomial) *crc32Field {
+	return &crc32Field{polynomial: polynomial}
+}
+
+func (c *crc32Field) run(curOffset int, buf []byte) error {
+	crc, err := c.crc(curOffset, buf)
+	if err != nil {
+		return err
+	}
+	binary.BigEndian.PutUint32(buf[c.startOffset:], crc)
+	return nil
+}
+
+func (c *crc32Field) check(curOffset int, buf []byte) error {
+	crc, err := c.crc(curOffset, buf)
+	if err != nil {
+		return err
+	}
+
+	expected := binary.BigEndian.Uint32(buf[c.startOffset:])
+	if crc != expected {
+		return PacketDecodingError{fmt.Sprintf("CRC didn't match expected %#x got %#x", expected, crc)}
+	}
+
+	return nil
+}
+func (c *crc32Field) crc(curOffset int, buf []byte) (uint32, error) {
+	var tab *crc32.Table
+	switch c.polynomial {
+	case crcIEEE:
+		tab = crc32.IEEETable
+	case crcCastagnoli:
+		tab = castagnoliTable
+	default:
+		return 0, PacketDecodingError{"invalid CRC type"}
+	}
+	return crc32.Checksum(buf[c.startOffset+4:curOffset], tab), nil
+}
diff --git a/vendor/github.com/Shopify/sarama/create_partitions_request.go b/vendor/github.com/Shopify/sarama/create_partitions_request.go
new file mode 100644
index 0000000..af321e9
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/create_partitions_request.go
@@ -0,0 +1,121 @@
+package sarama
+
+import "time"
+
+type CreatePartitionsRequest struct {
+	TopicPartitions map[string]*TopicPartition
+	Timeout         time.Duration
+	ValidateOnly    bool
+}
+
+func (c *CreatePartitionsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(c.TopicPartitions)); err != nil {
+		return err
+	}
+
+	for topic, partition := range c.TopicPartitions {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := partition.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	pe.putInt32(int32(c.Timeout / time.Millisecond))
+
+	pe.putBool(c.ValidateOnly)
+
+	return nil
+}
+
+func (c *CreatePartitionsRequest) decode(pd packetDecoder, version int16) (err error) {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	c.TopicPartitions = make(map[string]*TopicPartition, n)
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		c.TopicPartitions[topic] = new(TopicPartition)
+		if err := c.TopicPartitions[topic].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	timeout, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	c.Timeout = time.Duration(timeout) * time.Millisecond
+
+	if c.ValidateOnly, err = pd.getBool(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *CreatePartitionsRequest) key() int16 {
+	return 37
+}
+
+func (r *CreatePartitionsRequest) version() int16 {
+	return 0
+}
+
+func (r *CreatePartitionsRequest) requiredVersion() KafkaVersion {
+	return V1_0_0_0
+}
+
+type TopicPartition struct {
+	Count      int32
+	Assignment [][]int32
+}
+
+func (t *TopicPartition) encode(pe packetEncoder) error {
+	pe.putInt32(t.Count)
+
+	if len(t.Assignment) == 0 {
+		pe.putInt32(-1)
+		return nil
+	}
+
+	if err := pe.putArrayLength(len(t.Assignment)); err != nil {
+		return err
+	}
+
+	for _, assign := range t.Assignment {
+		if err := pe.putInt32Array(assign); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (t *TopicPartition) decode(pd packetDecoder, version int16) (err error) {
+	if t.Count, err = pd.getInt32(); err != nil {
+		return err
+	}
+
+	n, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	if n <= 0 {
+		return nil
+	}
+	t.Assignment = make([][]int32, n)
+
+	for i := 0; i < int(n); i++ {
+		if t.Assignment[i], err = pd.getInt32Array(); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/create_partitions_response.go b/vendor/github.com/Shopify/sarama/create_partitions_response.go
new file mode 100644
index 0000000..abd621c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/create_partitions_response.go
@@ -0,0 +1,94 @@
+package sarama
+
+import "time"
+
+type CreatePartitionsResponse struct {
+	ThrottleTime         time.Duration
+	TopicPartitionErrors map[string]*TopicPartitionError
+}
+
+func (c *CreatePartitionsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(c.ThrottleTime / time.Millisecond))
+	if err := pe.putArrayLength(len(c.TopicPartitionErrors)); err != nil {
+		return err
+	}
+
+	for topic, partitionError := range c.TopicPartitionErrors {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := partitionError.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *CreatePartitionsResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	c.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	c.TopicPartitionErrors = make(map[string]*TopicPartitionError, n)
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		c.TopicPartitionErrors[topic] = new(TopicPartitionError)
+		if err := c.TopicPartitionErrors[topic].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *CreatePartitionsResponse) key() int16 {
+	return 37
+}
+
+func (r *CreatePartitionsResponse) version() int16 {
+	return 0
+}
+
+func (r *CreatePartitionsResponse) requiredVersion() KafkaVersion {
+	return V1_0_0_0
+}
+
+type TopicPartitionError struct {
+	Err    KError
+	ErrMsg *string
+}
+
+func (t *TopicPartitionError) encode(pe packetEncoder) error {
+	pe.putInt16(int16(t.Err))
+
+	if err := pe.putNullableString(t.ErrMsg); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (t *TopicPartitionError) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	t.Err = KError(kerr)
+
+	if t.ErrMsg, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/create_topics_request.go b/vendor/github.com/Shopify/sarama/create_topics_request.go
new file mode 100644
index 0000000..709c0a4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/create_topics_request.go
@@ -0,0 +1,174 @@
+package sarama
+
+import (
+	"time"
+)
+
+type CreateTopicsRequest struct {
+	Version int16
+
+	TopicDetails map[string]*TopicDetail
+	Timeout      time.Duration
+	ValidateOnly bool
+}
+
+func (c *CreateTopicsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(c.TopicDetails)); err != nil {
+		return err
+	}
+	for topic, detail := range c.TopicDetails {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := detail.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	pe.putInt32(int32(c.Timeout / time.Millisecond))
+
+	if c.Version >= 1 {
+		pe.putBool(c.ValidateOnly)
+	}
+
+	return nil
+}
+
+func (c *CreateTopicsRequest) decode(pd packetDecoder, version int16) (err error) {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	c.TopicDetails = make(map[string]*TopicDetail, n)
+
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		c.TopicDetails[topic] = new(TopicDetail)
+		if err = c.TopicDetails[topic].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	timeout, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	c.Timeout = time.Duration(timeout) * time.Millisecond
+
+	if version >= 1 {
+		c.ValidateOnly, err = pd.getBool()
+		if err != nil {
+			return err
+		}
+
+		c.Version = version
+	}
+
+	return nil
+}
+
+func (c *CreateTopicsRequest) key() int16 {
+	return 19
+}
+
+func (c *CreateTopicsRequest) version() int16 {
+	return c.Version
+}
+
+func (c *CreateTopicsRequest) requiredVersion() KafkaVersion {
+	switch c.Version {
+	case 2:
+		return V1_0_0_0
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_10_1_0
+	}
+}
+
+type TopicDetail struct {
+	NumPartitions     int32
+	ReplicationFactor int16
+	ReplicaAssignment map[int32][]int32
+	ConfigEntries     map[string]*string
+}
+
+func (t *TopicDetail) encode(pe packetEncoder) error {
+	pe.putInt32(t.NumPartitions)
+	pe.putInt16(t.ReplicationFactor)
+
+	if err := pe.putArrayLength(len(t.ReplicaAssignment)); err != nil {
+		return err
+	}
+	for partition, assignment := range t.ReplicaAssignment {
+		pe.putInt32(partition)
+		if err := pe.putInt32Array(assignment); err != nil {
+			return err
+		}
+	}
+
+	if err := pe.putArrayLength(len(t.ConfigEntries)); err != nil {
+		return err
+	}
+	for configKey, configValue := range t.ConfigEntries {
+		if err := pe.putString(configKey); err != nil {
+			return err
+		}
+		if err := pe.putNullableString(configValue); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (t *TopicDetail) decode(pd packetDecoder, version int16) (err error) {
+	if t.NumPartitions, err = pd.getInt32(); err != nil {
+		return err
+	}
+	if t.ReplicationFactor, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		t.ReplicaAssignment = make(map[int32][]int32, n)
+		for i := 0; i < n; i++ {
+			replica, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			if t.ReplicaAssignment[replica], err = pd.getInt32Array(); err != nil {
+				return err
+			}
+		}
+	}
+
+	n, err = pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		t.ConfigEntries = make(map[string]*string, n)
+		for i := 0; i < n; i++ {
+			configKey, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			if t.ConfigEntries[configKey], err = pd.getNullableString(); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/create_topics_response.go b/vendor/github.com/Shopify/sarama/create_topics_response.go
new file mode 100644
index 0000000..66207e0
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/create_topics_response.go
@@ -0,0 +1,112 @@
+package sarama
+
+import "time"
+
+type CreateTopicsResponse struct {
+	Version      int16
+	ThrottleTime time.Duration
+	TopicErrors  map[string]*TopicError
+}
+
+func (c *CreateTopicsResponse) encode(pe packetEncoder) error {
+	if c.Version >= 2 {
+		pe.putInt32(int32(c.ThrottleTime / time.Millisecond))
+	}
+
+	if err := pe.putArrayLength(len(c.TopicErrors)); err != nil {
+		return err
+	}
+	for topic, topicError := range c.TopicErrors {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := topicError.encode(pe, c.Version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *CreateTopicsResponse) decode(pd packetDecoder, version int16) (err error) {
+	c.Version = version
+
+	if version >= 2 {
+		throttleTime, err := pd.getInt32()
+		if err != nil {
+			return err
+		}
+		c.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	c.TopicErrors = make(map[string]*TopicError, n)
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		c.TopicErrors[topic] = new(TopicError)
+		if err := c.TopicErrors[topic].decode(pd, version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (c *CreateTopicsResponse) key() int16 {
+	return 19
+}
+
+func (c *CreateTopicsResponse) version() int16 {
+	return c.Version
+}
+
+func (c *CreateTopicsResponse) requiredVersion() KafkaVersion {
+	switch c.Version {
+	case 2:
+		return V1_0_0_0
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_10_1_0
+	}
+}
+
+type TopicError struct {
+	Err    KError
+	ErrMsg *string
+}
+
+func (t *TopicError) encode(pe packetEncoder, version int16) error {
+	pe.putInt16(int16(t.Err))
+
+	if version >= 1 {
+		if err := pe.putNullableString(t.ErrMsg); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (t *TopicError) decode(pd packetDecoder, version int16) (err error) {
+	kErr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	t.Err = KError(kErr)
+
+	if version >= 1 {
+		if t.ErrMsg, err = pd.getNullableString(); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/decompress.go b/vendor/github.com/Shopify/sarama/decompress.go
new file mode 100644
index 0000000..eaccbfc
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/decompress.go
@@ -0,0 +1,63 @@
+package sarama
+
+import (
+	"bytes"
+	"compress/gzip"
+	"fmt"
+	"io/ioutil"
+	"sync"
+
+	"github.com/eapache/go-xerial-snappy"
+	"github.com/pierrec/lz4"
+)
+
+var (
+	lz4ReaderPool = sync.Pool{
+		New: func() interface{} {
+			return lz4.NewReader(nil)
+		},
+	}
+
+	gzipReaderPool sync.Pool
+)
+
+func decompress(cc CompressionCodec, data []byte) ([]byte, error) {
+	switch cc {
+	case CompressionNone:
+		return data, nil
+	case CompressionGZIP:
+		var (
+			err        error
+			reader     *gzip.Reader
+			readerIntf = gzipReaderPool.Get()
+		)
+		if readerIntf != nil {
+			reader = readerIntf.(*gzip.Reader)
+		} else {
+			reader, err = gzip.NewReader(bytes.NewReader(data))
+			if err != nil {
+				return nil, err
+			}
+		}
+
+		defer gzipReaderPool.Put(reader)
+
+		if err := reader.Reset(bytes.NewReader(data)); err != nil {
+			return nil, err
+		}
+
+		return ioutil.ReadAll(reader)
+	case CompressionSnappy:
+		return snappy.Decode(data)
+	case CompressionLZ4:
+		reader := lz4ReaderPool.Get().(*lz4.Reader)
+		defer lz4ReaderPool.Put(reader)
+
+		reader.Reset(bytes.NewReader(data))
+		return ioutil.ReadAll(reader)
+	case CompressionZSTD:
+		return zstdDecompress(nil, data)
+	default:
+		return nil, PacketDecodingError{fmt.Sprintf("invalid compression specified (%d)", cc)}
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_groups_request.go b/vendor/github.com/Shopify/sarama/delete_groups_request.go
new file mode 100644
index 0000000..305a324
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_groups_request.go
@@ -0,0 +1,30 @@
+package sarama
+
+type DeleteGroupsRequest struct {
+	Groups []string
+}
+
+func (r *DeleteGroupsRequest) encode(pe packetEncoder) error {
+	return pe.putStringArray(r.Groups)
+}
+
+func (r *DeleteGroupsRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Groups, err = pd.getStringArray()
+	return
+}
+
+func (r *DeleteGroupsRequest) key() int16 {
+	return 42
+}
+
+func (r *DeleteGroupsRequest) version() int16 {
+	return 0
+}
+
+func (r *DeleteGroupsRequest) requiredVersion() KafkaVersion {
+	return V1_1_0_0
+}
+
+func (r *DeleteGroupsRequest) AddGroup(group string) {
+	r.Groups = append(r.Groups, group)
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_groups_response.go b/vendor/github.com/Shopify/sarama/delete_groups_response.go
new file mode 100644
index 0000000..c067ebb
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_groups_response.go
@@ -0,0 +1,70 @@
+package sarama
+
+import (
+	"time"
+)
+
+type DeleteGroupsResponse struct {
+	ThrottleTime    time.Duration
+	GroupErrorCodes map[string]KError
+}
+
+func (r *DeleteGroupsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(r.ThrottleTime / time.Millisecond))
+
+	if err := pe.putArrayLength(len(r.GroupErrorCodes)); err != nil {
+		return err
+	}
+	for groupID, errorCode := range r.GroupErrorCodes {
+		if err := pe.putString(groupID); err != nil {
+			return err
+		}
+		pe.putInt16(int16(errorCode))
+	}
+
+	return nil
+}
+
+func (r *DeleteGroupsResponse) decode(pd packetDecoder, version int16) error {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	r.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	r.GroupErrorCodes = make(map[string]KError, n)
+	for i := 0; i < n; i++ {
+		groupID, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		errorCode, err := pd.getInt16()
+		if err != nil {
+			return err
+		}
+
+		r.GroupErrorCodes[groupID] = KError(errorCode)
+	}
+
+	return nil
+}
+
+func (r *DeleteGroupsResponse) key() int16 {
+	return 42
+}
+
+func (r *DeleteGroupsResponse) version() int16 {
+	return 0
+}
+
+func (r *DeleteGroupsResponse) requiredVersion() KafkaVersion {
+	return V1_1_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_records_request.go b/vendor/github.com/Shopify/sarama/delete_records_request.go
new file mode 100644
index 0000000..93efafd
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_records_request.go
@@ -0,0 +1,126 @@
+package sarama
+
+import (
+	"sort"
+	"time"
+)
+
+// request message format is:
+// [topic] timeout(int32)
+// where topic is:
+//  name(string) [partition]
+// where partition is:
+//  id(int32) offset(int64)
+
+type DeleteRecordsRequest struct {
+	Topics  map[string]*DeleteRecordsRequestTopic
+	Timeout time.Duration
+}
+
+func (d *DeleteRecordsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(d.Topics)); err != nil {
+		return err
+	}
+	keys := make([]string, 0, len(d.Topics))
+	for topic := range d.Topics {
+		keys = append(keys, topic)
+	}
+	sort.Strings(keys)
+	for _, topic := range keys {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := d.Topics[topic].encode(pe); err != nil {
+			return err
+		}
+	}
+	pe.putInt32(int32(d.Timeout / time.Millisecond))
+
+	return nil
+}
+
+func (d *DeleteRecordsRequest) decode(pd packetDecoder, version int16) error {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		d.Topics = make(map[string]*DeleteRecordsRequestTopic, n)
+		for i := 0; i < n; i++ {
+			topic, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			details := new(DeleteRecordsRequestTopic)
+			if err = details.decode(pd, version); err != nil {
+				return err
+			}
+			d.Topics[topic] = details
+		}
+	}
+
+	timeout, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	d.Timeout = time.Duration(timeout) * time.Millisecond
+
+	return nil
+}
+
+func (d *DeleteRecordsRequest) key() int16 {
+	return 21
+}
+
+func (d *DeleteRecordsRequest) version() int16 {
+	return 0
+}
+
+func (d *DeleteRecordsRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type DeleteRecordsRequestTopic struct {
+	PartitionOffsets map[int32]int64 // partition => offset
+}
+
+func (t *DeleteRecordsRequestTopic) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(t.PartitionOffsets)); err != nil {
+		return err
+	}
+	keys := make([]int32, 0, len(t.PartitionOffsets))
+	for partition := range t.PartitionOffsets {
+		keys = append(keys, partition)
+	}
+	sort.Slice(keys, func(i, j int) bool { return keys[i] < keys[j] })
+	for _, partition := range keys {
+		pe.putInt32(partition)
+		pe.putInt64(t.PartitionOffsets[partition])
+	}
+	return nil
+}
+
+func (t *DeleteRecordsRequestTopic) decode(pd packetDecoder, version int16) error {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		t.PartitionOffsets = make(map[int32]int64, n)
+		for i := 0; i < n; i++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			offset, err := pd.getInt64()
+			if err != nil {
+				return err
+			}
+			t.PartitionOffsets[partition] = offset
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_records_response.go b/vendor/github.com/Shopify/sarama/delete_records_response.go
new file mode 100644
index 0000000..733a58b
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_records_response.go
@@ -0,0 +1,158 @@
+package sarama
+
+import (
+	"sort"
+	"time"
+)
+
+// response message format is:
+// throttleMs(int32) [topic]
+// where topic is:
+//  name(string) [partition]
+// where partition is:
+//  id(int32) low_watermark(int64) error_code(int16)
+
+type DeleteRecordsResponse struct {
+	Version      int16
+	ThrottleTime time.Duration
+	Topics       map[string]*DeleteRecordsResponseTopic
+}
+
+func (d *DeleteRecordsResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(d.ThrottleTime / time.Millisecond))
+
+	if err := pe.putArrayLength(len(d.Topics)); err != nil {
+		return err
+	}
+	keys := make([]string, 0, len(d.Topics))
+	for topic := range d.Topics {
+		keys = append(keys, topic)
+	}
+	sort.Strings(keys)
+	for _, topic := range keys {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := d.Topics[topic].encode(pe); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (d *DeleteRecordsResponse) decode(pd packetDecoder, version int16) error {
+	d.Version = version
+
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	d.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		d.Topics = make(map[string]*DeleteRecordsResponseTopic, n)
+		for i := 0; i < n; i++ {
+			topic, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			details := new(DeleteRecordsResponseTopic)
+			if err = details.decode(pd, version); err != nil {
+				return err
+			}
+			d.Topics[topic] = details
+		}
+	}
+
+	return nil
+}
+
+func (d *DeleteRecordsResponse) key() int16 {
+	return 21
+}
+
+func (d *DeleteRecordsResponse) version() int16 {
+	return 0
+}
+
+func (d *DeleteRecordsResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type DeleteRecordsResponseTopic struct {
+	Partitions map[int32]*DeleteRecordsResponsePartition
+}
+
+func (t *DeleteRecordsResponseTopic) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(t.Partitions)); err != nil {
+		return err
+	}
+	keys := make([]int32, 0, len(t.Partitions))
+	for partition := range t.Partitions {
+		keys = append(keys, partition)
+	}
+	sort.Slice(keys, func(i, j int) bool { return keys[i] < keys[j] })
+	for _, partition := range keys {
+		pe.putInt32(partition)
+		if err := t.Partitions[partition].encode(pe); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (t *DeleteRecordsResponseTopic) decode(pd packetDecoder, version int16) error {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if n > 0 {
+		t.Partitions = make(map[int32]*DeleteRecordsResponsePartition, n)
+		for i := 0; i < n; i++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			details := new(DeleteRecordsResponsePartition)
+			if err = details.decode(pd, version); err != nil {
+				return err
+			}
+			t.Partitions[partition] = details
+		}
+	}
+
+	return nil
+}
+
+type DeleteRecordsResponsePartition struct {
+	LowWatermark int64
+	Err          KError
+}
+
+func (t *DeleteRecordsResponsePartition) encode(pe packetEncoder) error {
+	pe.putInt64(t.LowWatermark)
+	pe.putInt16(int16(t.Err))
+	return nil
+}
+
+func (t *DeleteRecordsResponsePartition) decode(pd packetDecoder, version int16) error {
+	lowWatermark, err := pd.getInt64()
+	if err != nil {
+		return err
+	}
+	t.LowWatermark = lowWatermark
+
+	kErr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	t.Err = KError(kErr)
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_topics_request.go b/vendor/github.com/Shopify/sarama/delete_topics_request.go
new file mode 100644
index 0000000..911f67d
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_topics_request.go
@@ -0,0 +1,48 @@
+package sarama
+
+import "time"
+
+type DeleteTopicsRequest struct {
+	Version int16
+	Topics  []string
+	Timeout time.Duration
+}
+
+func (d *DeleteTopicsRequest) encode(pe packetEncoder) error {
+	if err := pe.putStringArray(d.Topics); err != nil {
+		return err
+	}
+	pe.putInt32(int32(d.Timeout / time.Millisecond))
+
+	return nil
+}
+
+func (d *DeleteTopicsRequest) decode(pd packetDecoder, version int16) (err error) {
+	if d.Topics, err = pd.getStringArray(); err != nil {
+		return err
+	}
+	timeout, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	d.Timeout = time.Duration(timeout) * time.Millisecond
+	d.Version = version
+	return nil
+}
+
+func (d *DeleteTopicsRequest) key() int16 {
+	return 20
+}
+
+func (d *DeleteTopicsRequest) version() int16 {
+	return d.Version
+}
+
+func (d *DeleteTopicsRequest) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_10_1_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/delete_topics_response.go b/vendor/github.com/Shopify/sarama/delete_topics_response.go
new file mode 100644
index 0000000..3422546
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/delete_topics_response.go
@@ -0,0 +1,78 @@
+package sarama
+
+import "time"
+
+type DeleteTopicsResponse struct {
+	Version         int16
+	ThrottleTime    time.Duration
+	TopicErrorCodes map[string]KError
+}
+
+func (d *DeleteTopicsResponse) encode(pe packetEncoder) error {
+	if d.Version >= 1 {
+		pe.putInt32(int32(d.ThrottleTime / time.Millisecond))
+	}
+
+	if err := pe.putArrayLength(len(d.TopicErrorCodes)); err != nil {
+		return err
+	}
+	for topic, errorCode := range d.TopicErrorCodes {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		pe.putInt16(int16(errorCode))
+	}
+
+	return nil
+}
+
+func (d *DeleteTopicsResponse) decode(pd packetDecoder, version int16) (err error) {
+	if version >= 1 {
+		throttleTime, err := pd.getInt32()
+		if err != nil {
+			return err
+		}
+		d.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+		d.Version = version
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	d.TopicErrorCodes = make(map[string]KError, n)
+
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		errorCode, err := pd.getInt16()
+		if err != nil {
+			return err
+		}
+
+		d.TopicErrorCodes[topic] = KError(errorCode)
+	}
+
+	return nil
+}
+
+func (d *DeleteTopicsResponse) key() int16 {
+	return 20
+}
+
+func (d *DeleteTopicsResponse) version() int16 {
+	return d.Version
+}
+
+func (d *DeleteTopicsResponse) requiredVersion() KafkaVersion {
+	switch d.Version {
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_10_1_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/describe_configs_request.go b/vendor/github.com/Shopify/sarama/describe_configs_request.go
new file mode 100644
index 0000000..ccb587b
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/describe_configs_request.go
@@ -0,0 +1,112 @@
+package sarama
+
+type DescribeConfigsRequest struct {
+	Version         int16
+	Resources       []*ConfigResource
+	IncludeSynonyms bool
+}
+
+type ConfigResource struct {
+	Type        ConfigResourceType
+	Name        string
+	ConfigNames []string
+}
+
+func (r *DescribeConfigsRequest) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(r.Resources)); err != nil {
+		return err
+	}
+
+	for _, c := range r.Resources {
+		pe.putInt8(int8(c.Type))
+		if err := pe.putString(c.Name); err != nil {
+			return err
+		}
+
+		if len(c.ConfigNames) == 0 {
+			pe.putInt32(-1)
+			continue
+		}
+		if err := pe.putStringArray(c.ConfigNames); err != nil {
+			return err
+		}
+	}
+
+	if r.Version >= 1 {
+		pe.putBool(r.IncludeSynonyms)
+	}
+
+	return nil
+}
+
+func (r *DescribeConfigsRequest) decode(pd packetDecoder, version int16) (err error) {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Resources = make([]*ConfigResource, n)
+
+	for i := 0; i < n; i++ {
+		r.Resources[i] = &ConfigResource{}
+		t, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+		r.Resources[i].Type = ConfigResourceType(t)
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		r.Resources[i].Name = name
+
+		confLength, err := pd.getArrayLength()
+
+		if err != nil {
+			return err
+		}
+
+		if confLength == -1 {
+			continue
+		}
+
+		cfnames := make([]string, confLength)
+		for i := 0; i < confLength; i++ {
+			s, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			cfnames[i] = s
+		}
+		r.Resources[i].ConfigNames = cfnames
+	}
+	r.Version = version
+	if r.Version >= 1 {
+		b, err := pd.getBool()
+		if err != nil {
+			return err
+		}
+		r.IncludeSynonyms = b
+	}
+
+	return nil
+}
+
+func (r *DescribeConfigsRequest) key() int16 {
+	return 32
+}
+
+func (r *DescribeConfigsRequest) version() int16 {
+	return r.Version
+}
+
+func (r *DescribeConfigsRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V1_1_0_0
+	case 2:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/describe_configs_response.go b/vendor/github.com/Shopify/sarama/describe_configs_response.go
new file mode 100644
index 0000000..63fb6ea
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/describe_configs_response.go
@@ -0,0 +1,320 @@
+package sarama
+
+import (
+	"fmt"
+	"time"
+)
+
+type ConfigSource int8
+
+func (s ConfigSource) String() string {
+	switch s {
+	case SourceUnknown:
+		return "Unknown"
+	case SourceTopic:
+		return "Topic"
+	case SourceDynamicBroker:
+		return "DynamicBroker"
+	case SourceDynamicDefaultBroker:
+		return "DynamicDefaultBroker"
+	case SourceStaticBroker:
+		return "StaticBroker"
+	case SourceDefault:
+		return "Default"
+	}
+	return fmt.Sprintf("Source Invalid: %d", int(s))
+}
+
+const (
+	SourceUnknown              ConfigSource = 0
+	SourceTopic                ConfigSource = 1
+	SourceDynamicBroker        ConfigSource = 2
+	SourceDynamicDefaultBroker ConfigSource = 3
+	SourceStaticBroker         ConfigSource = 4
+	SourceDefault              ConfigSource = 5
+)
+
+type DescribeConfigsResponse struct {
+	Version      int16
+	ThrottleTime time.Duration
+	Resources    []*ResourceResponse
+}
+
+type ResourceResponse struct {
+	ErrorCode int16
+	ErrorMsg  string
+	Type      ConfigResourceType
+	Name      string
+	Configs   []*ConfigEntry
+}
+
+type ConfigEntry struct {
+	Name      string
+	Value     string
+	ReadOnly  bool
+	Default   bool
+	Source    ConfigSource
+	Sensitive bool
+	Synonyms  []*ConfigSynonym
+}
+
+type ConfigSynonym struct {
+	ConfigName  string
+	ConfigValue string
+	Source      ConfigSource
+}
+
+func (r *DescribeConfigsResponse) encode(pe packetEncoder) (err error) {
+	pe.putInt32(int32(r.ThrottleTime / time.Millisecond))
+	if err = pe.putArrayLength(len(r.Resources)); err != nil {
+		return err
+	}
+
+	for _, c := range r.Resources {
+		if err = c.encode(pe, r.Version); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *DescribeConfigsResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	r.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Resources = make([]*ResourceResponse, n)
+	for i := 0; i < n; i++ {
+		rr := &ResourceResponse{}
+		if err := rr.decode(pd, version); err != nil {
+			return err
+		}
+		r.Resources[i] = rr
+	}
+
+	return nil
+}
+
+func (r *DescribeConfigsResponse) key() int16 {
+	return 32
+}
+
+func (r *DescribeConfigsResponse) version() int16 {
+	return r.Version
+}
+
+func (r *DescribeConfigsResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V1_0_0_0
+	case 2:
+		return V2_0_0_0
+	default:
+		return V0_11_0_0
+	}
+}
+
+func (r *ResourceResponse) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(r.ErrorCode)
+
+	if err = pe.putString(r.ErrorMsg); err != nil {
+		return err
+	}
+
+	pe.putInt8(int8(r.Type))
+
+	if err = pe.putString(r.Name); err != nil {
+		return err
+	}
+
+	if err = pe.putArrayLength(len(r.Configs)); err != nil {
+		return err
+	}
+
+	for _, c := range r.Configs {
+		if err = c.encode(pe, version); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (r *ResourceResponse) decode(pd packetDecoder, version int16) (err error) {
+	ec, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	r.ErrorCode = ec
+
+	em, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	r.ErrorMsg = em
+
+	t, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	r.Type = ConfigResourceType(t)
+
+	name, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	r.Name = name
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Configs = make([]*ConfigEntry, n)
+	for i := 0; i < n; i++ {
+		c := &ConfigEntry{}
+		if err := c.decode(pd, version); err != nil {
+			return err
+		}
+		r.Configs[i] = c
+	}
+	return nil
+}
+
+func (r *ConfigEntry) encode(pe packetEncoder, version int16) (err error) {
+	if err = pe.putString(r.Name); err != nil {
+		return err
+	}
+
+	if err = pe.putString(r.Value); err != nil {
+		return err
+	}
+
+	pe.putBool(r.ReadOnly)
+
+	if version <= 0 {
+		pe.putBool(r.Default)
+		pe.putBool(r.Sensitive)
+	} else {
+		pe.putInt8(int8(r.Source))
+		pe.putBool(r.Sensitive)
+
+		if err := pe.putArrayLength(len(r.Synonyms)); err != nil {
+			return err
+		}
+		for _, c := range r.Synonyms {
+			if err = c.encode(pe, version); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+//https://cwiki.apache.org/confluence/display/KAFKA/KIP-226+-+Dynamic+Broker+Configuration
+func (r *ConfigEntry) decode(pd packetDecoder, version int16) (err error) {
+	if version == 0 {
+		r.Source = SourceUnknown
+	}
+	name, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	r.Name = name
+
+	value, err := pd.getString()
+	if err != nil {
+		return err
+	}
+	r.Value = value
+
+	read, err := pd.getBool()
+	if err != nil {
+		return err
+	}
+	r.ReadOnly = read
+
+	if version == 0 {
+		defaultB, err := pd.getBool()
+		if err != nil {
+			return err
+		}
+		r.Default = defaultB
+	} else {
+		source, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+		r.Source = ConfigSource(source)
+	}
+
+	sensitive, err := pd.getBool()
+	if err != nil {
+		return err
+	}
+	r.Sensitive = sensitive
+
+	if version > 0 {
+		n, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+		r.Synonyms = make([]*ConfigSynonym, n)
+
+		for i := 0; i < n; i++ {
+			s := &ConfigSynonym{}
+			if err := s.decode(pd, version); err != nil {
+				return err
+			}
+			r.Synonyms[i] = s
+		}
+
+	}
+	return nil
+}
+
+func (c *ConfigSynonym) encode(pe packetEncoder, version int16) (err error) {
+	err = pe.putString(c.ConfigName)
+	if err != nil {
+		return err
+	}
+
+	err = pe.putString(c.ConfigValue)
+	if err != nil {
+		return err
+	}
+
+	pe.putInt8(int8(c.Source))
+
+	return nil
+}
+
+func (c *ConfigSynonym) decode(pd packetDecoder, version int16) error {
+	name, err := pd.getString()
+	if err != nil {
+		return nil
+	}
+	c.ConfigName = name
+
+	value, err := pd.getString()
+	if err != nil {
+		return nil
+	}
+	c.ConfigValue = value
+
+	source, err := pd.getInt8()
+	if err != nil {
+		return nil
+	}
+	c.Source = ConfigSource(source)
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/describe_groups_request.go b/vendor/github.com/Shopify/sarama/describe_groups_request.go
new file mode 100644
index 0000000..1fb3567
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/describe_groups_request.go
@@ -0,0 +1,30 @@
+package sarama
+
+type DescribeGroupsRequest struct {
+	Groups []string
+}
+
+func (r *DescribeGroupsRequest) encode(pe packetEncoder) error {
+	return pe.putStringArray(r.Groups)
+}
+
+func (r *DescribeGroupsRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Groups, err = pd.getStringArray()
+	return
+}
+
+func (r *DescribeGroupsRequest) key() int16 {
+	return 15
+}
+
+func (r *DescribeGroupsRequest) version() int16 {
+	return 0
+}
+
+func (r *DescribeGroupsRequest) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
+
+func (r *DescribeGroupsRequest) AddGroup(group string) {
+	r.Groups = append(r.Groups, group)
+}
diff --git a/vendor/github.com/Shopify/sarama/describe_groups_response.go b/vendor/github.com/Shopify/sarama/describe_groups_response.go
new file mode 100644
index 0000000..542b3a9
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/describe_groups_response.go
@@ -0,0 +1,187 @@
+package sarama
+
+type DescribeGroupsResponse struct {
+	Groups []*GroupDescription
+}
+
+func (r *DescribeGroupsResponse) encode(pe packetEncoder) error {
+	if err := pe.putArrayLength(len(r.Groups)); err != nil {
+		return err
+	}
+
+	for _, groupDescription := range r.Groups {
+		if err := groupDescription.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *DescribeGroupsResponse) decode(pd packetDecoder, version int16) (err error) {
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Groups = make([]*GroupDescription, n)
+	for i := 0; i < n; i++ {
+		r.Groups[i] = new(GroupDescription)
+		if err := r.Groups[i].decode(pd); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *DescribeGroupsResponse) key() int16 {
+	return 15
+}
+
+func (r *DescribeGroupsResponse) version() int16 {
+	return 0
+}
+
+func (r *DescribeGroupsResponse) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
+
+type GroupDescription struct {
+	Err          KError
+	GroupId      string
+	State        string
+	ProtocolType string
+	Protocol     string
+	Members      map[string]*GroupMemberDescription
+}
+
+func (gd *GroupDescription) encode(pe packetEncoder) error {
+	pe.putInt16(int16(gd.Err))
+
+	if err := pe.putString(gd.GroupId); err != nil {
+		return err
+	}
+	if err := pe.putString(gd.State); err != nil {
+		return err
+	}
+	if err := pe.putString(gd.ProtocolType); err != nil {
+		return err
+	}
+	if err := pe.putString(gd.Protocol); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(gd.Members)); err != nil {
+		return err
+	}
+
+	for memberId, groupMemberDescription := range gd.Members {
+		if err := pe.putString(memberId); err != nil {
+			return err
+		}
+		if err := groupMemberDescription.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (gd *GroupDescription) decode(pd packetDecoder) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	gd.Err = KError(kerr)
+
+	if gd.GroupId, err = pd.getString(); err != nil {
+		return
+	}
+	if gd.State, err = pd.getString(); err != nil {
+		return
+	}
+	if gd.ProtocolType, err = pd.getString(); err != nil {
+		return
+	}
+	if gd.Protocol, err = pd.getString(); err != nil {
+		return
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	gd.Members = make(map[string]*GroupMemberDescription)
+	for i := 0; i < n; i++ {
+		memberId, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		gd.Members[memberId] = new(GroupMemberDescription)
+		if err := gd.Members[memberId].decode(pd); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+type GroupMemberDescription struct {
+	ClientId         string
+	ClientHost       string
+	MemberMetadata   []byte
+	MemberAssignment []byte
+}
+
+func (gmd *GroupMemberDescription) encode(pe packetEncoder) error {
+	if err := pe.putString(gmd.ClientId); err != nil {
+		return err
+	}
+	if err := pe.putString(gmd.ClientHost); err != nil {
+		return err
+	}
+	if err := pe.putBytes(gmd.MemberMetadata); err != nil {
+		return err
+	}
+	if err := pe.putBytes(gmd.MemberAssignment); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (gmd *GroupMemberDescription) decode(pd packetDecoder) (err error) {
+	if gmd.ClientId, err = pd.getString(); err != nil {
+		return
+	}
+	if gmd.ClientHost, err = pd.getString(); err != nil {
+		return
+	}
+	if gmd.MemberMetadata, err = pd.getBytes(); err != nil {
+		return
+	}
+	if gmd.MemberAssignment, err = pd.getBytes(); err != nil {
+		return
+	}
+
+	return nil
+}
+
+func (gmd *GroupMemberDescription) GetMemberAssignment() (*ConsumerGroupMemberAssignment, error) {
+	assignment := new(ConsumerGroupMemberAssignment)
+	err := decode(gmd.MemberAssignment, assignment)
+	return assignment, err
+}
+
+func (gmd *GroupMemberDescription) GetMemberMetadata() (*ConsumerGroupMemberMetadata, error) {
+	metadata := new(ConsumerGroupMemberMetadata)
+	err := decode(gmd.MemberMetadata, metadata)
+	return metadata, err
+}
diff --git a/vendor/github.com/Shopify/sarama/dev.yml b/vendor/github.com/Shopify/sarama/dev.yml
new file mode 100644
index 0000000..97eed3a
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/dev.yml
@@ -0,0 +1,10 @@
+name: sarama
+
+up:
+  - go:
+      version: '1.11'
+
+commands:
+  test:
+    run: make test
+    desc: 'run unit tests'
diff --git a/vendor/github.com/Shopify/sarama/encoder_decoder.go b/vendor/github.com/Shopify/sarama/encoder_decoder.go
new file mode 100644
index 0000000..7ce3bc0
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/encoder_decoder.go
@@ -0,0 +1,89 @@
+package sarama
+
+import (
+	"fmt"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+// Encoder is the interface that wraps the basic Encode method.
+// Anything implementing Encoder can be turned into bytes using Kafka's encoding rules.
+type encoder interface {
+	encode(pe packetEncoder) error
+}
+
+// Encode takes an Encoder and turns it into bytes while potentially recording metrics.
+func encode(e encoder, metricRegistry metrics.Registry) ([]byte, error) {
+	if e == nil {
+		return nil, nil
+	}
+
+	var prepEnc prepEncoder
+	var realEnc realEncoder
+
+	err := e.encode(&prepEnc)
+	if err != nil {
+		return nil, err
+	}
+
+	if prepEnc.length < 0 || prepEnc.length > int(MaxRequestSize) {
+		return nil, PacketEncodingError{fmt.Sprintf("invalid request size (%d)", prepEnc.length)}
+	}
+
+	realEnc.raw = make([]byte, prepEnc.length)
+	realEnc.registry = metricRegistry
+	err = e.encode(&realEnc)
+	if err != nil {
+		return nil, err
+	}
+
+	return realEnc.raw, nil
+}
+
+// Decoder is the interface that wraps the basic Decode method.
+// Anything implementing Decoder can be extracted from bytes using Kafka's encoding rules.
+type decoder interface {
+	decode(pd packetDecoder) error
+}
+
+type versionedDecoder interface {
+	decode(pd packetDecoder, version int16) error
+}
+
+// Decode takes bytes and a Decoder and fills the fields of the decoder from the bytes,
+// interpreted using Kafka's encoding rules.
+func decode(buf []byte, in decoder) error {
+	if buf == nil {
+		return nil
+	}
+
+	helper := realDecoder{raw: buf}
+	err := in.decode(&helper)
+	if err != nil {
+		return err
+	}
+
+	if helper.off != len(buf) {
+		return PacketDecodingError{"invalid length"}
+	}
+
+	return nil
+}
+
+func versionedDecode(buf []byte, in versionedDecoder, version int16) error {
+	if buf == nil {
+		return nil
+	}
+
+	helper := realDecoder{raw: buf}
+	err := in.decode(&helper, version)
+	if err != nil {
+		return err
+	}
+
+	if helper.off != len(buf) {
+		return PacketDecodingError{"invalid length"}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/end_txn_request.go b/vendor/github.com/Shopify/sarama/end_txn_request.go
new file mode 100644
index 0000000..2cd9b50
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/end_txn_request.go
@@ -0,0 +1,50 @@
+package sarama
+
+type EndTxnRequest struct {
+	TransactionalID   string
+	ProducerID        int64
+	ProducerEpoch     int16
+	TransactionResult bool
+}
+
+func (a *EndTxnRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(a.TransactionalID); err != nil {
+		return err
+	}
+
+	pe.putInt64(a.ProducerID)
+
+	pe.putInt16(a.ProducerEpoch)
+
+	pe.putBool(a.TransactionResult)
+
+	return nil
+}
+
+func (a *EndTxnRequest) decode(pd packetDecoder, version int16) (err error) {
+	if a.TransactionalID, err = pd.getString(); err != nil {
+		return err
+	}
+	if a.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if a.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+	if a.TransactionResult, err = pd.getBool(); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *EndTxnRequest) key() int16 {
+	return 26
+}
+
+func (a *EndTxnRequest) version() int16 {
+	return 0
+}
+
+func (a *EndTxnRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/end_txn_response.go b/vendor/github.com/Shopify/sarama/end_txn_response.go
new file mode 100644
index 0000000..33b27e3
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/end_txn_response.go
@@ -0,0 +1,44 @@
+package sarama
+
+import (
+	"time"
+)
+
+type EndTxnResponse struct {
+	ThrottleTime time.Duration
+	Err          KError
+}
+
+func (e *EndTxnResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(e.ThrottleTime / time.Millisecond))
+	pe.putInt16(int16(e.Err))
+	return nil
+}
+
+func (e *EndTxnResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	e.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	e.Err = KError(kerr)
+
+	return nil
+}
+
+func (e *EndTxnResponse) key() int16 {
+	return 25
+}
+
+func (e *EndTxnResponse) version() int16 {
+	return 0
+}
+
+func (e *EndTxnResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/errors.go b/vendor/github.com/Shopify/sarama/errors.go
new file mode 100644
index 0000000..87a4c61
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/errors.go
@@ -0,0 +1,329 @@
+package sarama
+
+import (
+	"errors"
+	"fmt"
+)
+
+// ErrOutOfBrokers is the error returned when the client has run out of brokers to talk to because all of them errored
+// or otherwise failed to respond.
+var ErrOutOfBrokers = errors.New("kafka: client has run out of available brokers to talk to (Is your cluster reachable?)")
+
+// ErrClosedClient is the error returned when a method is called on a client that has been closed.
+var ErrClosedClient = errors.New("kafka: tried to use a client that was closed")
+
+// ErrIncompleteResponse is the error returned when the server returns a syntactically valid response, but it does
+// not contain the expected information.
+var ErrIncompleteResponse = errors.New("kafka: response did not contain all the expected topic/partition blocks")
+
+// ErrInvalidPartition is the error returned when a partitioner returns an invalid partition index
+// (meaning one outside of the range [0...numPartitions-1]).
+var ErrInvalidPartition = errors.New("kafka: partitioner returned an invalid partition index")
+
+// ErrAlreadyConnected is the error returned when calling Open() on a Broker that is already connected or connecting.
+var ErrAlreadyConnected = errors.New("kafka: broker connection already initiated")
+
+// ErrNotConnected is the error returned when trying to send or call Close() on a Broker that is not connected.
+var ErrNotConnected = errors.New("kafka: broker not connected")
+
+// ErrInsufficientData is returned when decoding and the packet is truncated. This can be expected
+// when requesting messages, since as an optimization the server is allowed to return a partial message at the end
+// of the message set.
+var ErrInsufficientData = errors.New("kafka: insufficient data to decode packet, more bytes expected")
+
+// ErrShuttingDown is returned when a producer receives a message during shutdown.
+var ErrShuttingDown = errors.New("kafka: message received by producer in process of shutting down")
+
+// ErrMessageTooLarge is returned when the next message to consume is larger than the configured Consumer.Fetch.Max
+var ErrMessageTooLarge = errors.New("kafka: message is larger than Consumer.Fetch.Max")
+
+// ErrConsumerOffsetNotAdvanced is returned when a partition consumer didn't advance its offset after parsing
+// a RecordBatch.
+var ErrConsumerOffsetNotAdvanced = errors.New("kafka: consumer offset was not advanced after a RecordBatch")
+
+// ErrControllerNotAvailable is returned when server didn't give correct controller id. May be kafka server's version
+// is lower than 0.10.0.0.
+var ErrControllerNotAvailable = errors.New("kafka: controller is not available")
+
+// ErrNoTopicsToUpdateMetadata is returned when Meta.Full is set to false but no specific topics were found to update
+// the metadata.
+var ErrNoTopicsToUpdateMetadata = errors.New("kafka: no specific topics to update metadata")
+
+// PacketEncodingError is returned from a failure while encoding a Kafka packet. This can happen, for example,
+// if you try to encode a string over 2^15 characters in length, since Kafka's encoding rules do not permit that.
+type PacketEncodingError struct {
+	Info string
+}
+
+func (err PacketEncodingError) Error() string {
+	return fmt.Sprintf("kafka: error encoding packet: %s", err.Info)
+}
+
+// PacketDecodingError is returned when there was an error (other than truncated data) decoding the Kafka broker's response.
+// This can be a bad CRC or length field, or any other invalid value.
+type PacketDecodingError struct {
+	Info string
+}
+
+func (err PacketDecodingError) Error() string {
+	return fmt.Sprintf("kafka: error decoding packet: %s", err.Info)
+}
+
+// ConfigurationError is the type of error returned from a constructor (e.g. NewClient, or NewConsumer)
+// when the specified configuration is invalid.
+type ConfigurationError string
+
+func (err ConfigurationError) Error() string {
+	return "kafka: invalid configuration (" + string(err) + ")"
+}
+
+// KError is the type of error that can be returned directly by the Kafka broker.
+// See https://cwiki.apache.org/confluence/display/KAFKA/A+Guide+To+The+Kafka+Protocol#AGuideToTheKafkaProtocol-ErrorCodes
+type KError int16
+
+// Numeric error codes returned by the Kafka server.
+const (
+	ErrNoError                            KError = 0
+	ErrUnknown                            KError = -1
+	ErrOffsetOutOfRange                   KError = 1
+	ErrInvalidMessage                     KError = 2
+	ErrUnknownTopicOrPartition            KError = 3
+	ErrInvalidMessageSize                 KError = 4
+	ErrLeaderNotAvailable                 KError = 5
+	ErrNotLeaderForPartition              KError = 6
+	ErrRequestTimedOut                    KError = 7
+	ErrBrokerNotAvailable                 KError = 8
+	ErrReplicaNotAvailable                KError = 9
+	ErrMessageSizeTooLarge                KError = 10
+	ErrStaleControllerEpochCode           KError = 11
+	ErrOffsetMetadataTooLarge             KError = 12
+	ErrNetworkException                   KError = 13
+	ErrOffsetsLoadInProgress              KError = 14
+	ErrConsumerCoordinatorNotAvailable    KError = 15
+	ErrNotCoordinatorForConsumer          KError = 16
+	ErrInvalidTopic                       KError = 17
+	ErrMessageSetSizeTooLarge             KError = 18
+	ErrNotEnoughReplicas                  KError = 19
+	ErrNotEnoughReplicasAfterAppend       KError = 20
+	ErrInvalidRequiredAcks                KError = 21
+	ErrIllegalGeneration                  KError = 22
+	ErrInconsistentGroupProtocol          KError = 23
+	ErrInvalidGroupId                     KError = 24
+	ErrUnknownMemberId                    KError = 25
+	ErrInvalidSessionTimeout              KError = 26
+	ErrRebalanceInProgress                KError = 27
+	ErrInvalidCommitOffsetSize            KError = 28
+	ErrTopicAuthorizationFailed           KError = 29
+	ErrGroupAuthorizationFailed           KError = 30
+	ErrClusterAuthorizationFailed         KError = 31
+	ErrInvalidTimestamp                   KError = 32
+	ErrUnsupportedSASLMechanism           KError = 33
+	ErrIllegalSASLState                   KError = 34
+	ErrUnsupportedVersion                 KError = 35
+	ErrTopicAlreadyExists                 KError = 36
+	ErrInvalidPartitions                  KError = 37
+	ErrInvalidReplicationFactor           KError = 38
+	ErrInvalidReplicaAssignment           KError = 39
+	ErrInvalidConfig                      KError = 40
+	ErrNotController                      KError = 41
+	ErrInvalidRequest                     KError = 42
+	ErrUnsupportedForMessageFormat        KError = 43
+	ErrPolicyViolation                    KError = 44
+	ErrOutOfOrderSequenceNumber           KError = 45
+	ErrDuplicateSequenceNumber            KError = 46
+	ErrInvalidProducerEpoch               KError = 47
+	ErrInvalidTxnState                    KError = 48
+	ErrInvalidProducerIDMapping           KError = 49
+	ErrInvalidTransactionTimeout          KError = 50
+	ErrConcurrentTransactions             KError = 51
+	ErrTransactionCoordinatorFenced       KError = 52
+	ErrTransactionalIDAuthorizationFailed KError = 53
+	ErrSecurityDisabled                   KError = 54
+	ErrOperationNotAttempted              KError = 55
+	ErrKafkaStorageError                  KError = 56
+	ErrLogDirNotFound                     KError = 57
+	ErrSASLAuthenticationFailed           KError = 58
+	ErrUnknownProducerID                  KError = 59
+	ErrReassignmentInProgress             KError = 60
+	ErrDelegationTokenAuthDisabled        KError = 61
+	ErrDelegationTokenNotFound            KError = 62
+	ErrDelegationTokenOwnerMismatch       KError = 63
+	ErrDelegationTokenRequestNotAllowed   KError = 64
+	ErrDelegationTokenAuthorizationFailed KError = 65
+	ErrDelegationTokenExpired             KError = 66
+	ErrInvalidPrincipalType               KError = 67
+	ErrNonEmptyGroup                      KError = 68
+	ErrGroupIDNotFound                    KError = 69
+	ErrFetchSessionIDNotFound             KError = 70
+	ErrInvalidFetchSessionEpoch           KError = 71
+	ErrListenerNotFound                   KError = 72
+	ErrTopicDeletionDisabled              KError = 73
+	ErrFencedLeaderEpoch                  KError = 74
+	ErrUnknownLeaderEpoch                 KError = 75
+	ErrUnsupportedCompressionType         KError = 76
+)
+
+func (err KError) Error() string {
+	// Error messages stolen/adapted from
+	// https://kafka.apache.org/protocol#protocol_error_codes
+	switch err {
+	case ErrNoError:
+		return "kafka server: Not an error, why are you printing me?"
+	case ErrUnknown:
+		return "kafka server: Unexpected (unknown?) server error."
+	case ErrOffsetOutOfRange:
+		return "kafka server: The requested offset is outside the range of offsets maintained by the server for the given topic/partition."
+	case ErrInvalidMessage:
+		return "kafka server: Message contents does not match its CRC."
+	case ErrUnknownTopicOrPartition:
+		return "kafka server: Request was for a topic or partition that does not exist on this broker."
+	case ErrInvalidMessageSize:
+		return "kafka server: The message has a negative size."
+	case ErrLeaderNotAvailable:
+		return "kafka server: In the middle of a leadership election, there is currently no leader for this partition and hence it is unavailable for writes."
+	case ErrNotLeaderForPartition:
+		return "kafka server: Tried to send a message to a replica that is not the leader for some partition. Your metadata is out of date."
+	case ErrRequestTimedOut:
+		return "kafka server: Request exceeded the user-specified time limit in the request."
+	case ErrBrokerNotAvailable:
+		return "kafka server: Broker not available. Not a client facing error, we should never receive this!!!"
+	case ErrReplicaNotAvailable:
+		return "kafka server: Replica information not available, one or more brokers are down."
+	case ErrMessageSizeTooLarge:
+		return "kafka server: Message was too large, server rejected it to avoid allocation error."
+	case ErrStaleControllerEpochCode:
+		return "kafka server: StaleControllerEpochCode (internal error code for broker-to-broker communication)."
+	case ErrOffsetMetadataTooLarge:
+		return "kafka server: Specified a string larger than the configured maximum for offset metadata."
+	case ErrNetworkException:
+		return "kafka server: The server disconnected before a response was received."
+	case ErrOffsetsLoadInProgress:
+		return "kafka server: The broker is still loading offsets after a leader change for that offset's topic partition."
+	case ErrConsumerCoordinatorNotAvailable:
+		return "kafka server: Offset's topic has not yet been created."
+	case ErrNotCoordinatorForConsumer:
+		return "kafka server: Request was for a consumer group that is not coordinated by this broker."
+	case ErrInvalidTopic:
+		return "kafka server: The request attempted to perform an operation on an invalid topic."
+	case ErrMessageSetSizeTooLarge:
+		return "kafka server: The request included message batch larger than the configured segment size on the server."
+	case ErrNotEnoughReplicas:
+		return "kafka server: Messages are rejected since there are fewer in-sync replicas than required."
+	case ErrNotEnoughReplicasAfterAppend:
+		return "kafka server: Messages are written to the log, but to fewer in-sync replicas than required."
+	case ErrInvalidRequiredAcks:
+		return "kafka server: The number of required acks is invalid (should be either -1, 0, or 1)."
+	case ErrIllegalGeneration:
+		return "kafka server: The provided generation id is not the current generation."
+	case ErrInconsistentGroupProtocol:
+		return "kafka server: The provider group protocol type is incompatible with the other members."
+	case ErrInvalidGroupId:
+		return "kafka server: The provided group id was empty."
+	case ErrUnknownMemberId:
+		return "kafka server: The provided member is not known in the current generation."
+	case ErrInvalidSessionTimeout:
+		return "kafka server: The provided session timeout is outside the allowed range."
+	case ErrRebalanceInProgress:
+		return "kafka server: A rebalance for the group is in progress. Please re-join the group."
+	case ErrInvalidCommitOffsetSize:
+		return "kafka server: The provided commit metadata was too large."
+	case ErrTopicAuthorizationFailed:
+		return "kafka server: The client is not authorized to access this topic."
+	case ErrGroupAuthorizationFailed:
+		return "kafka server: The client is not authorized to access this group."
+	case ErrClusterAuthorizationFailed:
+		return "kafka server: The client is not authorized to send this request type."
+	case ErrInvalidTimestamp:
+		return "kafka server: The timestamp of the message is out of acceptable range."
+	case ErrUnsupportedSASLMechanism:
+		return "kafka server: The broker does not support the requested SASL mechanism."
+	case ErrIllegalSASLState:
+		return "kafka server: Request is not valid given the current SASL state."
+	case ErrUnsupportedVersion:
+		return "kafka server: The version of API is not supported."
+	case ErrTopicAlreadyExists:
+		return "kafka server: Topic with this name already exists."
+	case ErrInvalidPartitions:
+		return "kafka server: Number of partitions is invalid."
+	case ErrInvalidReplicationFactor:
+		return "kafka server: Replication-factor is invalid."
+	case ErrInvalidReplicaAssignment:
+		return "kafka server: Replica assignment is invalid."
+	case ErrInvalidConfig:
+		return "kafka server: Configuration is invalid."
+	case ErrNotController:
+		return "kafka server: This is not the correct controller for this cluster."
+	case ErrInvalidRequest:
+		return "kafka server: This most likely occurs because of a request being malformed by the client library or the message was sent to an incompatible broker. See the broker logs for more details."
+	case ErrUnsupportedForMessageFormat:
+		return "kafka server: The requested operation is not supported by the message format version."
+	case ErrPolicyViolation:
+		return "kafka server: Request parameters do not satisfy the configured policy."
+	case ErrOutOfOrderSequenceNumber:
+		return "kafka server: The broker received an out of order sequence number."
+	case ErrDuplicateSequenceNumber:
+		return "kafka server: The broker received a duplicate sequence number."
+	case ErrInvalidProducerEpoch:
+		return "kafka server: Producer attempted an operation with an old epoch."
+	case ErrInvalidTxnState:
+		return "kafka server: The producer attempted a transactional operation in an invalid state."
+	case ErrInvalidProducerIDMapping:
+		return "kafka server: The producer attempted to use a producer id which is not currently assigned to its transactional id."
+	case ErrInvalidTransactionTimeout:
+		return "kafka server: The transaction timeout is larger than the maximum value allowed by the broker (as configured by max.transaction.timeout.ms)."
+	case ErrConcurrentTransactions:
+		return "kafka server: The producer attempted to update a transaction while another concurrent operation on the same transaction was ongoing."
+	case ErrTransactionCoordinatorFenced:
+		return "kafka server: The transaction coordinator sending a WriteTxnMarker is no longer the current coordinator for a given producer."
+	case ErrTransactionalIDAuthorizationFailed:
+		return "kafka server: Transactional ID authorization failed."
+	case ErrSecurityDisabled:
+		return "kafka server: Security features are disabled."
+	case ErrOperationNotAttempted:
+		return "kafka server: The broker did not attempt to execute this operation."
+	case ErrKafkaStorageError:
+		return "kafka server: Disk error when trying to access log file on the disk."
+	case ErrLogDirNotFound:
+		return "kafka server: The specified log directory is not found in the broker config."
+	case ErrSASLAuthenticationFailed:
+		return "kafka server: SASL Authentication failed."
+	case ErrUnknownProducerID:
+		return "kafka server: The broker could not locate the producer metadata associated with the Producer ID."
+	case ErrReassignmentInProgress:
+		return "kafka server: A partition reassignment is in progress."
+	case ErrDelegationTokenAuthDisabled:
+		return "kafka server: Delegation Token feature is not enabled."
+	case ErrDelegationTokenNotFound:
+		return "kafka server: Delegation Token is not found on server."
+	case ErrDelegationTokenOwnerMismatch:
+		return "kafka server: Specified Principal is not valid Owner/Renewer."
+	case ErrDelegationTokenRequestNotAllowed:
+		return "kafka server: Delegation Token requests are not allowed on PLAINTEXT/1-way SSL channels and on delegation token authenticated channels."
+	case ErrDelegationTokenAuthorizationFailed:
+		return "kafka server: Delegation Token authorization failed."
+	case ErrDelegationTokenExpired:
+		return "kafka server: Delegation Token is expired."
+	case ErrInvalidPrincipalType:
+		return "kafka server: Supplied principalType is not supported."
+	case ErrNonEmptyGroup:
+		return "kafka server: The group is not empty."
+	case ErrGroupIDNotFound:
+		return "kafka server: The group id does not exist."
+	case ErrFetchSessionIDNotFound:
+		return "kafka server: The fetch session ID was not found."
+	case ErrInvalidFetchSessionEpoch:
+		return "kafka server: The fetch session epoch is invalid."
+	case ErrListenerNotFound:
+		return "kafka server: There is no listener on the leader broker that matches the listener on which metadata request was processed."
+	case ErrTopicDeletionDisabled:
+		return "kafka server: Topic deletion is disabled."
+	case ErrFencedLeaderEpoch:
+		return "kafka server: The leader epoch in the request is older than the epoch on the broker."
+	case ErrUnknownLeaderEpoch:
+		return "kafka server: The leader epoch in the request is newer than the epoch on the broker."
+	case ErrUnsupportedCompressionType:
+		return "kafka server: The requesting client does not support the compression type of given partition."
+	}
+
+	return fmt.Sprintf("Unknown error, how did this happen? Error code = %d", err)
+}
diff --git a/vendor/github.com/Shopify/sarama/fetch_request.go b/vendor/github.com/Shopify/sarama/fetch_request.go
new file mode 100644
index 0000000..462ab8a
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/fetch_request.go
@@ -0,0 +1,170 @@
+package sarama
+
+type fetchRequestBlock struct {
+	fetchOffset int64
+	maxBytes    int32
+}
+
+func (b *fetchRequestBlock) encode(pe packetEncoder) error {
+	pe.putInt64(b.fetchOffset)
+	pe.putInt32(b.maxBytes)
+	return nil
+}
+
+func (b *fetchRequestBlock) decode(pd packetDecoder) (err error) {
+	if b.fetchOffset, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if b.maxBytes, err = pd.getInt32(); err != nil {
+		return err
+	}
+	return nil
+}
+
+// FetchRequest (API key 1) will fetch Kafka messages. Version 3 introduced the MaxBytes field. See
+// https://issues.apache.org/jira/browse/KAFKA-2063 for a discussion of the issues leading up to that.  The KIP is at
+// https://cwiki.apache.org/confluence/display/KAFKA/KIP-74%3A+Add+Fetch+Response+Size+Limit+in+Bytes
+type FetchRequest struct {
+	MaxWaitTime int32
+	MinBytes    int32
+	MaxBytes    int32
+	Version     int16
+	Isolation   IsolationLevel
+	blocks      map[string]map[int32]*fetchRequestBlock
+}
+
+type IsolationLevel int8
+
+const (
+	ReadUncommitted IsolationLevel = 0
+	ReadCommitted   IsolationLevel = 1
+)
+
+func (r *FetchRequest) encode(pe packetEncoder) (err error) {
+	pe.putInt32(-1) // replica ID is always -1 for clients
+	pe.putInt32(r.MaxWaitTime)
+	pe.putInt32(r.MinBytes)
+	if r.Version >= 3 {
+		pe.putInt32(r.MaxBytes)
+	}
+	if r.Version >= 4 {
+		pe.putInt8(int8(r.Isolation))
+	}
+	err = pe.putArrayLength(len(r.blocks))
+	if err != nil {
+		return err
+	}
+	for topic, blocks := range r.blocks {
+		err = pe.putString(topic)
+		if err != nil {
+			return err
+		}
+		err = pe.putArrayLength(len(blocks))
+		if err != nil {
+			return err
+		}
+		for partition, block := range blocks {
+			pe.putInt32(partition)
+			err = block.encode(pe)
+			if err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (r *FetchRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+	if _, err = pd.getInt32(); err != nil {
+		return err
+	}
+	if r.MaxWaitTime, err = pd.getInt32(); err != nil {
+		return err
+	}
+	if r.MinBytes, err = pd.getInt32(); err != nil {
+		return err
+	}
+	if r.Version >= 3 {
+		if r.MaxBytes, err = pd.getInt32(); err != nil {
+			return err
+		}
+	}
+	if r.Version >= 4 {
+		isolation, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+		r.Isolation = IsolationLevel(isolation)
+	}
+	topicCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if topicCount == 0 {
+		return nil
+	}
+	r.blocks = make(map[string]map[int32]*fetchRequestBlock)
+	for i := 0; i < topicCount; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		partitionCount, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+		r.blocks[topic] = make(map[int32]*fetchRequestBlock)
+		for j := 0; j < partitionCount; j++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			fetchBlock := &fetchRequestBlock{}
+			if err = fetchBlock.decode(pd); err != nil {
+				return err
+			}
+			r.blocks[topic][partition] = fetchBlock
+		}
+	}
+	return nil
+}
+
+func (r *FetchRequest) key() int16 {
+	return 1
+}
+
+func (r *FetchRequest) version() int16 {
+	return r.Version
+}
+
+func (r *FetchRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_9_0_0
+	case 2:
+		return V0_10_0_0
+	case 3:
+		return V0_10_1_0
+	case 4:
+		return V0_11_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *FetchRequest) AddBlock(topic string, partitionID int32, fetchOffset int64, maxBytes int32) {
+	if r.blocks == nil {
+		r.blocks = make(map[string]map[int32]*fetchRequestBlock)
+	}
+
+	if r.blocks[topic] == nil {
+		r.blocks[topic] = make(map[int32]*fetchRequestBlock)
+	}
+
+	tmp := new(fetchRequestBlock)
+	tmp.maxBytes = maxBytes
+	tmp.fetchOffset = fetchOffset
+
+	r.blocks[topic][partitionID] = tmp
+}
diff --git a/vendor/github.com/Shopify/sarama/fetch_response.go b/vendor/github.com/Shopify/sarama/fetch_response.go
new file mode 100644
index 0000000..9df99c1
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/fetch_response.go
@@ -0,0 +1,409 @@
+package sarama
+
+import (
+	"time"
+)
+
+type AbortedTransaction struct {
+	ProducerID  int64
+	FirstOffset int64
+}
+
+func (t *AbortedTransaction) decode(pd packetDecoder) (err error) {
+	if t.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	if t.FirstOffset, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (t *AbortedTransaction) encode(pe packetEncoder) (err error) {
+	pe.putInt64(t.ProducerID)
+	pe.putInt64(t.FirstOffset)
+
+	return nil
+}
+
+type FetchResponseBlock struct {
+	Err                 KError
+	HighWaterMarkOffset int64
+	LastStableOffset    int64
+	AbortedTransactions []*AbortedTransaction
+	Records             *Records // deprecated: use FetchResponseBlock.RecordsSet
+	RecordsSet          []*Records
+	Partial             bool
+}
+
+func (b *FetchResponseBlock) decode(pd packetDecoder, version int16) (err error) {
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	b.Err = KError(tmp)
+
+	b.HighWaterMarkOffset, err = pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	if version >= 4 {
+		b.LastStableOffset, err = pd.getInt64()
+		if err != nil {
+			return err
+		}
+
+		numTransact, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		if numTransact >= 0 {
+			b.AbortedTransactions = make([]*AbortedTransaction, numTransact)
+		}
+
+		for i := 0; i < numTransact; i++ {
+			transact := new(AbortedTransaction)
+			if err = transact.decode(pd); err != nil {
+				return err
+			}
+			b.AbortedTransactions[i] = transact
+		}
+	}
+
+	recordsSize, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	recordsDecoder, err := pd.getSubset(int(recordsSize))
+	if err != nil {
+		return err
+	}
+
+	b.RecordsSet = []*Records{}
+
+	for recordsDecoder.remaining() > 0 {
+		records := &Records{}
+		if err := records.decode(recordsDecoder); err != nil {
+			// If we have at least one decoded records, this is not an error
+			if err == ErrInsufficientData {
+				if len(b.RecordsSet) == 0 {
+					b.Partial = true
+				}
+				break
+			}
+			return err
+		}
+
+		partial, err := records.isPartial()
+		if err != nil {
+			return err
+		}
+
+		n, err := records.numRecords()
+		if err != nil {
+			return err
+		}
+
+		if n > 0 || (partial && len(b.RecordsSet) == 0) {
+			b.RecordsSet = append(b.RecordsSet, records)
+
+			if b.Records == nil {
+				b.Records = records
+			}
+		}
+
+		overflow, err := records.isOverflow()
+		if err != nil {
+			return err
+		}
+
+		if partial || overflow {
+			break
+		}
+	}
+
+	return nil
+}
+
+func (b *FetchResponseBlock) numRecords() (int, error) {
+	sum := 0
+
+	for _, records := range b.RecordsSet {
+		count, err := records.numRecords()
+		if err != nil {
+			return 0, err
+		}
+
+		sum += count
+	}
+
+	return sum, nil
+}
+
+func (b *FetchResponseBlock) isPartial() (bool, error) {
+	if b.Partial {
+		return true, nil
+	}
+
+	if len(b.RecordsSet) == 1 {
+		return b.RecordsSet[0].isPartial()
+	}
+
+	return false, nil
+}
+
+func (b *FetchResponseBlock) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(int16(b.Err))
+
+	pe.putInt64(b.HighWaterMarkOffset)
+
+	if version >= 4 {
+		pe.putInt64(b.LastStableOffset)
+
+		if err = pe.putArrayLength(len(b.AbortedTransactions)); err != nil {
+			return err
+		}
+		for _, transact := range b.AbortedTransactions {
+			if err = transact.encode(pe); err != nil {
+				return err
+			}
+		}
+	}
+
+	pe.push(&lengthField{})
+	for _, records := range b.RecordsSet {
+		err = records.encode(pe)
+		if err != nil {
+			return err
+		}
+	}
+	return pe.pop()
+}
+
+type FetchResponse struct {
+	Blocks        map[string]map[int32]*FetchResponseBlock
+	ThrottleTime  time.Duration
+	Version       int16 // v1 requires 0.9+, v2 requires 0.10+
+	LogAppendTime bool
+	Timestamp     time.Time
+}
+
+func (r *FetchResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if r.Version >= 1 {
+		throttle, err := pd.getInt32()
+		if err != nil {
+			return err
+		}
+		r.ThrottleTime = time.Duration(throttle) * time.Millisecond
+	}
+
+	numTopics, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Blocks = make(map[string]map[int32]*FetchResponseBlock, numTopics)
+	for i := 0; i < numTopics; i++ {
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		numBlocks, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		r.Blocks[name] = make(map[int32]*FetchResponseBlock, numBlocks)
+
+		for j := 0; j < numBlocks; j++ {
+			id, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+
+			block := new(FetchResponseBlock)
+			err = block.decode(pd, version)
+			if err != nil {
+				return err
+			}
+			r.Blocks[name][id] = block
+		}
+	}
+
+	return nil
+}
+
+func (r *FetchResponse) encode(pe packetEncoder) (err error) {
+	if r.Version >= 1 {
+		pe.putInt32(int32(r.ThrottleTime / time.Millisecond))
+	}
+
+	err = pe.putArrayLength(len(r.Blocks))
+	if err != nil {
+		return err
+	}
+
+	for topic, partitions := range r.Blocks {
+		err = pe.putString(topic)
+		if err != nil {
+			return err
+		}
+
+		err = pe.putArrayLength(len(partitions))
+		if err != nil {
+			return err
+		}
+
+		for id, block := range partitions {
+			pe.putInt32(id)
+			err = block.encode(pe, r.Version)
+			if err != nil {
+				return err
+			}
+		}
+
+	}
+	return nil
+}
+
+func (r *FetchResponse) key() int16 {
+	return 1
+}
+
+func (r *FetchResponse) version() int16 {
+	return r.Version
+}
+
+func (r *FetchResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_9_0_0
+	case 2:
+		return V0_10_0_0
+	case 3:
+		return V0_10_1_0
+	case 4:
+		return V0_11_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *FetchResponse) GetBlock(topic string, partition int32) *FetchResponseBlock {
+	if r.Blocks == nil {
+		return nil
+	}
+
+	if r.Blocks[topic] == nil {
+		return nil
+	}
+
+	return r.Blocks[topic][partition]
+}
+
+func (r *FetchResponse) AddError(topic string, partition int32, err KError) {
+	if r.Blocks == nil {
+		r.Blocks = make(map[string]map[int32]*FetchResponseBlock)
+	}
+	partitions, ok := r.Blocks[topic]
+	if !ok {
+		partitions = make(map[int32]*FetchResponseBlock)
+		r.Blocks[topic] = partitions
+	}
+	frb, ok := partitions[partition]
+	if !ok {
+		frb = new(FetchResponseBlock)
+		partitions[partition] = frb
+	}
+	frb.Err = err
+}
+
+func (r *FetchResponse) getOrCreateBlock(topic string, partition int32) *FetchResponseBlock {
+	if r.Blocks == nil {
+		r.Blocks = make(map[string]map[int32]*FetchResponseBlock)
+	}
+	partitions, ok := r.Blocks[topic]
+	if !ok {
+		partitions = make(map[int32]*FetchResponseBlock)
+		r.Blocks[topic] = partitions
+	}
+	frb, ok := partitions[partition]
+	if !ok {
+		frb = new(FetchResponseBlock)
+		partitions[partition] = frb
+	}
+
+	return frb
+}
+
+func encodeKV(key, value Encoder) ([]byte, []byte) {
+	var kb []byte
+	var vb []byte
+	if key != nil {
+		kb, _ = key.Encode()
+	}
+	if value != nil {
+		vb, _ = value.Encode()
+	}
+
+	return kb, vb
+}
+
+func (r *FetchResponse) AddMessageWithTimestamp(topic string, partition int32, key, value Encoder, offset int64, timestamp time.Time, version int8) {
+	frb := r.getOrCreateBlock(topic, partition)
+	kb, vb := encodeKV(key, value)
+	if r.LogAppendTime {
+		timestamp = r.Timestamp
+	}
+	msg := &Message{Key: kb, Value: vb, LogAppendTime: r.LogAppendTime, Timestamp: timestamp, Version: version}
+	msgBlock := &MessageBlock{Msg: msg, Offset: offset}
+	if len(frb.RecordsSet) == 0 {
+		records := newLegacyRecords(&MessageSet{})
+		frb.RecordsSet = []*Records{&records}
+	}
+	set := frb.RecordsSet[0].MsgSet
+	set.Messages = append(set.Messages, msgBlock)
+}
+
+func (r *FetchResponse) AddRecordWithTimestamp(topic string, partition int32, key, value Encoder, offset int64, timestamp time.Time) {
+	frb := r.getOrCreateBlock(topic, partition)
+	kb, vb := encodeKV(key, value)
+	if len(frb.RecordsSet) == 0 {
+		records := newDefaultRecords(&RecordBatch{Version: 2, LogAppendTime: r.LogAppendTime, FirstTimestamp: timestamp, MaxTimestamp: r.Timestamp})
+		frb.RecordsSet = []*Records{&records}
+	}
+	batch := frb.RecordsSet[0].RecordBatch
+	rec := &Record{Key: kb, Value: vb, OffsetDelta: offset, TimestampDelta: timestamp.Sub(batch.FirstTimestamp)}
+	batch.addRecord(rec)
+}
+
+func (r *FetchResponse) AddMessage(topic string, partition int32, key, value Encoder, offset int64) {
+	r.AddMessageWithTimestamp(topic, partition, key, value, offset, time.Time{}, 0)
+}
+
+func (r *FetchResponse) AddRecord(topic string, partition int32, key, value Encoder, offset int64) {
+	r.AddRecordWithTimestamp(topic, partition, key, value, offset, time.Time{})
+}
+
+func (r *FetchResponse) SetLastOffsetDelta(topic string, partition int32, offset int32) {
+	frb := r.getOrCreateBlock(topic, partition)
+	if len(frb.RecordsSet) == 0 {
+		records := newDefaultRecords(&RecordBatch{Version: 2})
+		frb.RecordsSet = []*Records{&records}
+	}
+	batch := frb.RecordsSet[0].RecordBatch
+	batch.LastOffsetDelta = offset
+}
+
+func (r *FetchResponse) SetLastStableOffset(topic string, partition int32, offset int64) {
+	frb := r.getOrCreateBlock(topic, partition)
+	frb.LastStableOffset = offset
+}
diff --git a/vendor/github.com/Shopify/sarama/find_coordinator_request.go b/vendor/github.com/Shopify/sarama/find_coordinator_request.go
new file mode 100644
index 0000000..0ab5cb5
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/find_coordinator_request.go
@@ -0,0 +1,61 @@
+package sarama
+
+type CoordinatorType int8
+
+const (
+	CoordinatorGroup       CoordinatorType = 0
+	CoordinatorTransaction CoordinatorType = 1
+)
+
+type FindCoordinatorRequest struct {
+	Version         int16
+	CoordinatorKey  string
+	CoordinatorType CoordinatorType
+}
+
+func (f *FindCoordinatorRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(f.CoordinatorKey); err != nil {
+		return err
+	}
+
+	if f.Version >= 1 {
+		pe.putInt8(int8(f.CoordinatorType))
+	}
+
+	return nil
+}
+
+func (f *FindCoordinatorRequest) decode(pd packetDecoder, version int16) (err error) {
+	if f.CoordinatorKey, err = pd.getString(); err != nil {
+		return err
+	}
+
+	if version >= 1 {
+		f.Version = version
+		coordinatorType, err := pd.getInt8()
+		if err != nil {
+			return err
+		}
+
+		f.CoordinatorType = CoordinatorType(coordinatorType)
+	}
+
+	return nil
+}
+
+func (f *FindCoordinatorRequest) key() int16 {
+	return 10
+}
+
+func (f *FindCoordinatorRequest) version() int16 {
+	return f.Version
+}
+
+func (f *FindCoordinatorRequest) requiredVersion() KafkaVersion {
+	switch f.Version {
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_8_2_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/find_coordinator_response.go b/vendor/github.com/Shopify/sarama/find_coordinator_response.go
new file mode 100644
index 0000000..9c900e8
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/find_coordinator_response.go
@@ -0,0 +1,92 @@
+package sarama
+
+import (
+	"time"
+)
+
+var NoNode = &Broker{id: -1, addr: ":-1"}
+
+type FindCoordinatorResponse struct {
+	Version      int16
+	ThrottleTime time.Duration
+	Err          KError
+	ErrMsg       *string
+	Coordinator  *Broker
+}
+
+func (f *FindCoordinatorResponse) decode(pd packetDecoder, version int16) (err error) {
+	if version >= 1 {
+		f.Version = version
+
+		throttleTime, err := pd.getInt32()
+		if err != nil {
+			return err
+		}
+		f.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+	}
+
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	f.Err = KError(tmp)
+
+	if version >= 1 {
+		if f.ErrMsg, err = pd.getNullableString(); err != nil {
+			return err
+		}
+	}
+
+	coordinator := new(Broker)
+	// The version is hardcoded to 0, as version 1 of the Broker-decode
+	// contains the rack-field which is not present in the FindCoordinatorResponse.
+	if err := coordinator.decode(pd, 0); err != nil {
+		return err
+	}
+	if coordinator.addr == ":0" {
+		return nil
+	}
+	f.Coordinator = coordinator
+
+	return nil
+}
+
+func (f *FindCoordinatorResponse) encode(pe packetEncoder) error {
+	if f.Version >= 1 {
+		pe.putInt32(int32(f.ThrottleTime / time.Millisecond))
+	}
+
+	pe.putInt16(int16(f.Err))
+
+	if f.Version >= 1 {
+		if err := pe.putNullableString(f.ErrMsg); err != nil {
+			return err
+		}
+	}
+
+	coordinator := f.Coordinator
+	if coordinator == nil {
+		coordinator = NoNode
+	}
+	if err := coordinator.encode(pe, 0); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (f *FindCoordinatorResponse) key() int16 {
+	return 10
+}
+
+func (f *FindCoordinatorResponse) version() int16 {
+	return f.Version
+}
+
+func (f *FindCoordinatorResponse) requiredVersion() KafkaVersion {
+	switch f.Version {
+	case 1:
+		return V0_11_0_0
+	default:
+		return V0_8_2_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/go.mod b/vendor/github.com/Shopify/sarama/go.mod
new file mode 100644
index 0000000..3715129
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/go.mod
@@ -0,0 +1,13 @@
+module github.com/Shopify/sarama
+
+require (
+	github.com/DataDog/zstd v1.3.5
+	github.com/Shopify/toxiproxy v2.1.4+incompatible
+	github.com/davecgh/go-spew v1.1.1
+	github.com/eapache/go-resiliency v1.1.0
+	github.com/eapache/go-xerial-snappy v0.0.0-20180814174437-776d5712da21
+	github.com/eapache/queue v1.1.0
+	github.com/golang/snappy v0.0.0-20180518054509-2e65f85255db // indirect
+	github.com/pierrec/lz4 v2.0.5+incompatible
+	github.com/rcrowley/go-metrics v0.0.0-20181016184325-3113b8401b8a
+)
diff --git a/vendor/github.com/Shopify/sarama/go.sum b/vendor/github.com/Shopify/sarama/go.sum
new file mode 100644
index 0000000..58e2e91
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/go.sum
@@ -0,0 +1,18 @@
+github.com/DataDog/zstd v1.3.5 h1:DtpNbljikUepEPD16hD4LvIcmhnhdLTiW/5pHgbmp14=
+github.com/DataDog/zstd v1.3.5/go.mod h1:1jcaCB/ufaK+sKp1NBhlGmpz41jOoPQ35bpF36t7BBo=
+github.com/Shopify/toxiproxy v2.1.4+incompatible h1:TKdv8HiTLgE5wdJuEML90aBgNWsokNbMijUGhmcoBJc=
+github.com/Shopify/toxiproxy v2.1.4+incompatible/go.mod h1:OXgGpZ6Cli1/URJOF1DMxUHB2q5Ap20/P/eIdh4G0pI=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/eapache/go-resiliency v1.1.0 h1:1NtRmCAqadE2FN4ZcN6g90TP3uk8cg9rn9eNK2197aU=
+github.com/eapache/go-resiliency v1.1.0/go.mod h1:kFI+JgMyC7bLPUVY133qvEBtVayf5mFgVsvEsIPBvNs=
+github.com/eapache/go-xerial-snappy v0.0.0-20180814174437-776d5712da21 h1:YEetp8/yCZMuEPMUDHG0CW/brkkEp8mzqk2+ODEitlw=
+github.com/eapache/go-xerial-snappy v0.0.0-20180814174437-776d5712da21/go.mod h1:+020luEh2TKB4/GOp8oxxtq0Daoen/Cii55CzbTV6DU=
+github.com/eapache/queue v1.1.0 h1:YOEu7KNc61ntiQlcEeUIoDTJ2o8mQznoNvUhiigpIqc=
+github.com/eapache/queue v1.1.0/go.mod h1:6eCeP0CKFpHLu8blIFXhExK/dRa7WDZfr6jVFPTqq+I=
+github.com/golang/snappy v0.0.0-20180518054509-2e65f85255db h1:woRePGFeVFfLKN/pOkfl+p/TAqKOfFu+7KPlMVpok/w=
+github.com/golang/snappy v0.0.0-20180518054509-2e65f85255db/go.mod h1:/XxbfmMg8lxefKM7IXC3fBNl/7bRcc72aCRzEWrmP2Q=
+github.com/pierrec/lz4 v2.0.5+incompatible h1:2xWsjqPFWcplujydGg4WmhC/6fZqK42wMM8aXeqhl0I=
+github.com/pierrec/lz4 v2.0.5+incompatible/go.mod h1:pdkljMzZIN41W+lC3N2tnIh5sFi+IEE17M5jbnwPHcY=
+github.com/rcrowley/go-metrics v0.0.0-20181016184325-3113b8401b8a h1:9ZKAASQSHhDYGoxY8uLVpewe1GDZ2vu2Tr/vTdVAkFQ=
+github.com/rcrowley/go-metrics v0.0.0-20181016184325-3113b8401b8a/go.mod h1:bCqnVzQkZxMG4s8nGwiZ5l3QUCyqpo9Y+/ZMZ9VjZe4=
diff --git a/vendor/github.com/Shopify/sarama/heartbeat_request.go b/vendor/github.com/Shopify/sarama/heartbeat_request.go
new file mode 100644
index 0000000..ce49c47
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/heartbeat_request.go
@@ -0,0 +1,47 @@
+package sarama
+
+type HeartbeatRequest struct {
+	GroupId      string
+	GenerationId int32
+	MemberId     string
+}
+
+func (r *HeartbeatRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(r.GroupId); err != nil {
+		return err
+	}
+
+	pe.putInt32(r.GenerationId)
+
+	if err := pe.putString(r.MemberId); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *HeartbeatRequest) decode(pd packetDecoder, version int16) (err error) {
+	if r.GroupId, err = pd.getString(); err != nil {
+		return
+	}
+	if r.GenerationId, err = pd.getInt32(); err != nil {
+		return
+	}
+	if r.MemberId, err = pd.getString(); err != nil {
+		return
+	}
+
+	return nil
+}
+
+func (r *HeartbeatRequest) key() int16 {
+	return 12
+}
+
+func (r *HeartbeatRequest) version() int16 {
+	return 0
+}
+
+func (r *HeartbeatRequest) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/heartbeat_response.go b/vendor/github.com/Shopify/sarama/heartbeat_response.go
new file mode 100644
index 0000000..766f5fd
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/heartbeat_response.go
@@ -0,0 +1,32 @@
+package sarama
+
+type HeartbeatResponse struct {
+	Err KError
+}
+
+func (r *HeartbeatResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	return nil
+}
+
+func (r *HeartbeatResponse) decode(pd packetDecoder, version int16) error {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	r.Err = KError(kerr)
+
+	return nil
+}
+
+func (r *HeartbeatResponse) key() int16 {
+	return 12
+}
+
+func (r *HeartbeatResponse) version() int16 {
+	return 0
+}
+
+func (r *HeartbeatResponse) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/init_producer_id_request.go b/vendor/github.com/Shopify/sarama/init_producer_id_request.go
new file mode 100644
index 0000000..8ceb6c2
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/init_producer_id_request.go
@@ -0,0 +1,43 @@
+package sarama
+
+import "time"
+
+type InitProducerIDRequest struct {
+	TransactionalID    *string
+	TransactionTimeout time.Duration
+}
+
+func (i *InitProducerIDRequest) encode(pe packetEncoder) error {
+	if err := pe.putNullableString(i.TransactionalID); err != nil {
+		return err
+	}
+	pe.putInt32(int32(i.TransactionTimeout / time.Millisecond))
+
+	return nil
+}
+
+func (i *InitProducerIDRequest) decode(pd packetDecoder, version int16) (err error) {
+	if i.TransactionalID, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	timeout, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	i.TransactionTimeout = time.Duration(timeout) * time.Millisecond
+
+	return nil
+}
+
+func (i *InitProducerIDRequest) key() int16 {
+	return 22
+}
+
+func (i *InitProducerIDRequest) version() int16 {
+	return 0
+}
+
+func (i *InitProducerIDRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/init_producer_id_response.go b/vendor/github.com/Shopify/sarama/init_producer_id_response.go
new file mode 100644
index 0000000..1b32eb0
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/init_producer_id_response.go
@@ -0,0 +1,55 @@
+package sarama
+
+import "time"
+
+type InitProducerIDResponse struct {
+	ThrottleTime  time.Duration
+	Err           KError
+	ProducerID    int64
+	ProducerEpoch int16
+}
+
+func (i *InitProducerIDResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(i.ThrottleTime / time.Millisecond))
+	pe.putInt16(int16(i.Err))
+	pe.putInt64(i.ProducerID)
+	pe.putInt16(i.ProducerEpoch)
+
+	return nil
+}
+
+func (i *InitProducerIDResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	i.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	i.Err = KError(kerr)
+
+	if i.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	if i.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (i *InitProducerIDResponse) key() int16 {
+	return 22
+}
+
+func (i *InitProducerIDResponse) version() int16 {
+	return 0
+}
+
+func (i *InitProducerIDResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/join_group_request.go b/vendor/github.com/Shopify/sarama/join_group_request.go
new file mode 100644
index 0000000..97e9299
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/join_group_request.go
@@ -0,0 +1,163 @@
+package sarama
+
+type GroupProtocol struct {
+	Name     string
+	Metadata []byte
+}
+
+func (p *GroupProtocol) decode(pd packetDecoder) (err error) {
+	p.Name, err = pd.getString()
+	if err != nil {
+		return err
+	}
+	p.Metadata, err = pd.getBytes()
+	return err
+}
+
+func (p *GroupProtocol) encode(pe packetEncoder) (err error) {
+	if err := pe.putString(p.Name); err != nil {
+		return err
+	}
+	if err := pe.putBytes(p.Metadata); err != nil {
+		return err
+	}
+	return nil
+}
+
+type JoinGroupRequest struct {
+	Version               int16
+	GroupId               string
+	SessionTimeout        int32
+	RebalanceTimeout      int32
+	MemberId              string
+	ProtocolType          string
+	GroupProtocols        map[string][]byte // deprecated; use OrderedGroupProtocols
+	OrderedGroupProtocols []*GroupProtocol
+}
+
+func (r *JoinGroupRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(r.GroupId); err != nil {
+		return err
+	}
+	pe.putInt32(r.SessionTimeout)
+	if r.Version >= 1 {
+		pe.putInt32(r.RebalanceTimeout)
+	}
+	if err := pe.putString(r.MemberId); err != nil {
+		return err
+	}
+	if err := pe.putString(r.ProtocolType); err != nil {
+		return err
+	}
+
+	if len(r.GroupProtocols) > 0 {
+		if len(r.OrderedGroupProtocols) > 0 {
+			return PacketDecodingError{"cannot specify both GroupProtocols and OrderedGroupProtocols on JoinGroupRequest"}
+		}
+
+		if err := pe.putArrayLength(len(r.GroupProtocols)); err != nil {
+			return err
+		}
+		for name, metadata := range r.GroupProtocols {
+			if err := pe.putString(name); err != nil {
+				return err
+			}
+			if err := pe.putBytes(metadata); err != nil {
+				return err
+			}
+		}
+	} else {
+		if err := pe.putArrayLength(len(r.OrderedGroupProtocols)); err != nil {
+			return err
+		}
+		for _, protocol := range r.OrderedGroupProtocols {
+			if err := protocol.encode(pe); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (r *JoinGroupRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if r.GroupId, err = pd.getString(); err != nil {
+		return
+	}
+
+	if r.SessionTimeout, err = pd.getInt32(); err != nil {
+		return
+	}
+
+	if version >= 1 {
+		if r.RebalanceTimeout, err = pd.getInt32(); err != nil {
+			return err
+		}
+	}
+
+	if r.MemberId, err = pd.getString(); err != nil {
+		return
+	}
+
+	if r.ProtocolType, err = pd.getString(); err != nil {
+		return
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	r.GroupProtocols = make(map[string][]byte)
+	for i := 0; i < n; i++ {
+		protocol := &GroupProtocol{}
+		if err := protocol.decode(pd); err != nil {
+			return err
+		}
+		r.GroupProtocols[protocol.Name] = protocol.Metadata
+		r.OrderedGroupProtocols = append(r.OrderedGroupProtocols, protocol)
+	}
+
+	return nil
+}
+
+func (r *JoinGroupRequest) key() int16 {
+	return 11
+}
+
+func (r *JoinGroupRequest) version() int16 {
+	return r.Version
+}
+
+func (r *JoinGroupRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 2:
+		return V0_11_0_0
+	case 1:
+		return V0_10_1_0
+	default:
+		return V0_9_0_0
+	}
+}
+
+func (r *JoinGroupRequest) AddGroupProtocol(name string, metadata []byte) {
+	r.OrderedGroupProtocols = append(r.OrderedGroupProtocols, &GroupProtocol{
+		Name:     name,
+		Metadata: metadata,
+	})
+}
+
+func (r *JoinGroupRequest) AddGroupProtocolMetadata(name string, metadata *ConsumerGroupMemberMetadata) error {
+	bin, err := encode(metadata, nil)
+	if err != nil {
+		return err
+	}
+
+	r.AddGroupProtocol(name, bin)
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/join_group_response.go b/vendor/github.com/Shopify/sarama/join_group_response.go
new file mode 100644
index 0000000..5752acc
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/join_group_response.go
@@ -0,0 +1,135 @@
+package sarama
+
+type JoinGroupResponse struct {
+	Version       int16
+	ThrottleTime  int32
+	Err           KError
+	GenerationId  int32
+	GroupProtocol string
+	LeaderId      string
+	MemberId      string
+	Members       map[string][]byte
+}
+
+func (r *JoinGroupResponse) GetMembers() (map[string]ConsumerGroupMemberMetadata, error) {
+	members := make(map[string]ConsumerGroupMemberMetadata, len(r.Members))
+	for id, bin := range r.Members {
+		meta := new(ConsumerGroupMemberMetadata)
+		if err := decode(bin, meta); err != nil {
+			return nil, err
+		}
+		members[id] = *meta
+	}
+	return members, nil
+}
+
+func (r *JoinGroupResponse) encode(pe packetEncoder) error {
+	if r.Version >= 2 {
+		pe.putInt32(r.ThrottleTime)
+	}
+	pe.putInt16(int16(r.Err))
+	pe.putInt32(r.GenerationId)
+
+	if err := pe.putString(r.GroupProtocol); err != nil {
+		return err
+	}
+	if err := pe.putString(r.LeaderId); err != nil {
+		return err
+	}
+	if err := pe.putString(r.MemberId); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(r.Members)); err != nil {
+		return err
+	}
+
+	for memberId, memberMetadata := range r.Members {
+		if err := pe.putString(memberId); err != nil {
+			return err
+		}
+
+		if err := pe.putBytes(memberMetadata); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *JoinGroupResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if version >= 2 {
+		if r.ThrottleTime, err = pd.getInt32(); err != nil {
+			return
+		}
+	}
+
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	if r.GenerationId, err = pd.getInt32(); err != nil {
+		return
+	}
+
+	if r.GroupProtocol, err = pd.getString(); err != nil {
+		return
+	}
+
+	if r.LeaderId, err = pd.getString(); err != nil {
+		return
+	}
+
+	if r.MemberId, err = pd.getString(); err != nil {
+		return
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	r.Members = make(map[string][]byte)
+	for i := 0; i < n; i++ {
+		memberId, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		memberMetadata, err := pd.getBytes()
+		if err != nil {
+			return err
+		}
+
+		r.Members[memberId] = memberMetadata
+	}
+
+	return nil
+}
+
+func (r *JoinGroupResponse) key() int16 {
+	return 11
+}
+
+func (r *JoinGroupResponse) version() int16 {
+	return r.Version
+}
+
+func (r *JoinGroupResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 2:
+		return V0_11_0_0
+	case 1:
+		return V0_10_1_0
+	default:
+		return V0_9_0_0
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/leave_group_request.go b/vendor/github.com/Shopify/sarama/leave_group_request.go
new file mode 100644
index 0000000..e177427
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/leave_group_request.go
@@ -0,0 +1,40 @@
+package sarama
+
+type LeaveGroupRequest struct {
+	GroupId  string
+	MemberId string
+}
+
+func (r *LeaveGroupRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(r.GroupId); err != nil {
+		return err
+	}
+	if err := pe.putString(r.MemberId); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *LeaveGroupRequest) decode(pd packetDecoder, version int16) (err error) {
+	if r.GroupId, err = pd.getString(); err != nil {
+		return
+	}
+	if r.MemberId, err = pd.getString(); err != nil {
+		return
+	}
+
+	return nil
+}
+
+func (r *LeaveGroupRequest) key() int16 {
+	return 13
+}
+
+func (r *LeaveGroupRequest) version() int16 {
+	return 0
+}
+
+func (r *LeaveGroupRequest) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/leave_group_response.go b/vendor/github.com/Shopify/sarama/leave_group_response.go
new file mode 100644
index 0000000..d60c626
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/leave_group_response.go
@@ -0,0 +1,32 @@
+package sarama
+
+type LeaveGroupResponse struct {
+	Err KError
+}
+
+func (r *LeaveGroupResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	return nil
+}
+
+func (r *LeaveGroupResponse) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	r.Err = KError(kerr)
+
+	return nil
+}
+
+func (r *LeaveGroupResponse) key() int16 {
+	return 13
+}
+
+func (r *LeaveGroupResponse) version() int16 {
+	return 0
+}
+
+func (r *LeaveGroupResponse) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/length_field.go b/vendor/github.com/Shopify/sarama/length_field.go
new file mode 100644
index 0000000..da199a7
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/length_field.go
@@ -0,0 +1,82 @@
+package sarama
+
+import "encoding/binary"
+
+// LengthField implements the PushEncoder and PushDecoder interfaces for calculating 4-byte lengths.
+type lengthField struct {
+	startOffset int
+	length      int32
+}
+
+func (l *lengthField) decode(pd packetDecoder) error {
+	var err error
+	l.length, err = pd.getInt32()
+	if err != nil {
+		return err
+	}
+	if l.length > int32(pd.remaining()) {
+		return ErrInsufficientData
+	}
+	return nil
+}
+
+func (l *lengthField) saveOffset(in int) {
+	l.startOffset = in
+}
+
+func (l *lengthField) reserveLength() int {
+	return 4
+}
+
+func (l *lengthField) run(curOffset int, buf []byte) error {
+	binary.BigEndian.PutUint32(buf[l.startOffset:], uint32(curOffset-l.startOffset-4))
+	return nil
+}
+
+func (l *lengthField) check(curOffset int, buf []byte) error {
+	if int32(curOffset-l.startOffset-4) != l.length {
+		return PacketDecodingError{"length field invalid"}
+	}
+
+	return nil
+}
+
+type varintLengthField struct {
+	startOffset int
+	length      int64
+}
+
+func (l *varintLengthField) decode(pd packetDecoder) error {
+	var err error
+	l.length, err = pd.getVarint()
+	return err
+}
+
+func (l *varintLengthField) saveOffset(in int) {
+	l.startOffset = in
+}
+
+func (l *varintLengthField) adjustLength(currOffset int) int {
+	oldFieldSize := l.reserveLength()
+	l.length = int64(currOffset - l.startOffset - oldFieldSize)
+
+	return l.reserveLength() - oldFieldSize
+}
+
+func (l *varintLengthField) reserveLength() int {
+	var tmp [binary.MaxVarintLen64]byte
+	return binary.PutVarint(tmp[:], l.length)
+}
+
+func (l *varintLengthField) run(curOffset int, buf []byte) error {
+	binary.PutVarint(buf[l.startOffset:], l.length)
+	return nil
+}
+
+func (l *varintLengthField) check(curOffset int, buf []byte) error {
+	if int64(curOffset-l.startOffset-l.reserveLength()) != l.length {
+		return PacketDecodingError{"length field invalid"}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/list_groups_request.go b/vendor/github.com/Shopify/sarama/list_groups_request.go
new file mode 100644
index 0000000..3b16abf
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/list_groups_request.go
@@ -0,0 +1,24 @@
+package sarama
+
+type ListGroupsRequest struct {
+}
+
+func (r *ListGroupsRequest) encode(pe packetEncoder) error {
+	return nil
+}
+
+func (r *ListGroupsRequest) decode(pd packetDecoder, version int16) (err error) {
+	return nil
+}
+
+func (r *ListGroupsRequest) key() int16 {
+	return 16
+}
+
+func (r *ListGroupsRequest) version() int16 {
+	return 0
+}
+
+func (r *ListGroupsRequest) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/list_groups_response.go b/vendor/github.com/Shopify/sarama/list_groups_response.go
new file mode 100644
index 0000000..56115d4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/list_groups_response.go
@@ -0,0 +1,69 @@
+package sarama
+
+type ListGroupsResponse struct {
+	Err    KError
+	Groups map[string]string
+}
+
+func (r *ListGroupsResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+
+	if err := pe.putArrayLength(len(r.Groups)); err != nil {
+		return err
+	}
+	for groupId, protocolType := range r.Groups {
+		if err := pe.putString(groupId); err != nil {
+			return err
+		}
+		if err := pe.putString(protocolType); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *ListGroupsResponse) decode(pd packetDecoder, version int16) error {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	r.Groups = make(map[string]string)
+	for i := 0; i < n; i++ {
+		groupId, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		protocolType, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		r.Groups[groupId] = protocolType
+	}
+
+	return nil
+}
+
+func (r *ListGroupsResponse) key() int16 {
+	return 16
+}
+
+func (r *ListGroupsResponse) version() int16 {
+	return 0
+}
+
+func (r *ListGroupsResponse) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/message.go b/vendor/github.com/Shopify/sarama/message.go
new file mode 100644
index 0000000..f64c79b
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/message.go
@@ -0,0 +1,165 @@
+package sarama
+
+import (
+	"fmt"
+	"time"
+)
+
+// The lowest 3 bits contain the compression codec used for the message
+const compressionCodecMask int8 = 0x07
+
+// Bit 3 set for "LogAppend" timestamps
+const timestampTypeMask = 0x08
+
+// CompressionCodec represents the various compression codecs recognized by Kafka in messages.
+type CompressionCodec int8
+
+const (
+	CompressionNone   CompressionCodec = 0
+	CompressionGZIP   CompressionCodec = 1
+	CompressionSnappy CompressionCodec = 2
+	CompressionLZ4    CompressionCodec = 3
+	CompressionZSTD   CompressionCodec = 4
+)
+
+func (cc CompressionCodec) String() string {
+	return []string{
+		"none",
+		"gzip",
+		"snappy",
+		"lz4",
+	}[int(cc)]
+}
+
+// CompressionLevelDefault is the constant to use in CompressionLevel
+// to have the default compression level for any codec. The value is picked
+// that we don't use any existing compression levels.
+const CompressionLevelDefault = -1000
+
+type Message struct {
+	Codec            CompressionCodec // codec used to compress the message contents
+	CompressionLevel int              // compression level
+	LogAppendTime    bool             // the used timestamp is LogAppendTime
+	Key              []byte           // the message key, may be nil
+	Value            []byte           // the message contents
+	Set              *MessageSet      // the message set a message might wrap
+	Version          int8             // v1 requires Kafka 0.10
+	Timestamp        time.Time        // the timestamp of the message (version 1+ only)
+
+	compressedCache []byte
+	compressedSize  int // used for computing the compression ratio metrics
+}
+
+func (m *Message) encode(pe packetEncoder) error {
+	pe.push(newCRC32Field(crcIEEE))
+
+	pe.putInt8(m.Version)
+
+	attributes := int8(m.Codec) & compressionCodecMask
+	if m.LogAppendTime {
+		attributes |= timestampTypeMask
+	}
+	pe.putInt8(attributes)
+
+	if m.Version >= 1 {
+		if err := (Timestamp{&m.Timestamp}).encode(pe); err != nil {
+			return err
+		}
+	}
+
+	err := pe.putBytes(m.Key)
+	if err != nil {
+		return err
+	}
+
+	var payload []byte
+
+	if m.compressedCache != nil {
+		payload = m.compressedCache
+		m.compressedCache = nil
+	} else if m.Value != nil {
+
+		payload, err = compress(m.Codec, m.CompressionLevel, m.Value)
+		if err != nil {
+			return err
+		}
+		m.compressedCache = payload
+		// Keep in mind the compressed payload size for metric gathering
+		m.compressedSize = len(payload)
+	}
+
+	if err = pe.putBytes(payload); err != nil {
+		return err
+	}
+
+	return pe.pop()
+}
+
+func (m *Message) decode(pd packetDecoder) (err error) {
+	err = pd.push(newCRC32Field(crcIEEE))
+	if err != nil {
+		return err
+	}
+
+	m.Version, err = pd.getInt8()
+	if err != nil {
+		return err
+	}
+
+	if m.Version > 1 {
+		return PacketDecodingError{fmt.Sprintf("unknown magic byte (%v)", m.Version)}
+	}
+
+	attribute, err := pd.getInt8()
+	if err != nil {
+		return err
+	}
+	m.Codec = CompressionCodec(attribute & compressionCodecMask)
+	m.LogAppendTime = attribute&timestampTypeMask == timestampTypeMask
+
+	if m.Version == 1 {
+		if err := (Timestamp{&m.Timestamp}).decode(pd); err != nil {
+			return err
+		}
+	}
+
+	m.Key, err = pd.getBytes()
+	if err != nil {
+		return err
+	}
+
+	m.Value, err = pd.getBytes()
+	if err != nil {
+		return err
+	}
+
+	// Required for deep equal assertion during tests but might be useful
+	// for future metrics about the compression ratio in fetch requests
+	m.compressedSize = len(m.Value)
+
+	switch m.Codec {
+	case CompressionNone:
+		// nothing to do
+	default:
+		if m.Value == nil {
+			break
+		}
+
+		m.Value, err = decompress(m.Codec, m.Value)
+		if err != nil {
+			return err
+		}
+		if err := m.decodeSet(); err != nil {
+			return err
+		}
+	}
+
+	return pd.pop()
+}
+
+// decodes a message set from a previously encoded bulk-message
+func (m *Message) decodeSet() (err error) {
+	pd := realDecoder{raw: m.Value}
+	m.Set = &MessageSet{}
+	return m.Set.decode(&pd)
+}
diff --git a/vendor/github.com/Shopify/sarama/message_set.go b/vendor/github.com/Shopify/sarama/message_set.go
new file mode 100644
index 0000000..600c7c4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/message_set.go
@@ -0,0 +1,108 @@
+package sarama
+
+type MessageBlock struct {
+	Offset int64
+	Msg    *Message
+}
+
+// Messages convenience helper which returns either all the
+// messages that are wrapped in this block
+func (msb *MessageBlock) Messages() []*MessageBlock {
+	if msb.Msg.Set != nil {
+		return msb.Msg.Set.Messages
+	}
+	return []*MessageBlock{msb}
+}
+
+func (msb *MessageBlock) encode(pe packetEncoder) error {
+	pe.putInt64(msb.Offset)
+	pe.push(&lengthField{})
+	err := msb.Msg.encode(pe)
+	if err != nil {
+		return err
+	}
+	return pe.pop()
+}
+
+func (msb *MessageBlock) decode(pd packetDecoder) (err error) {
+	if msb.Offset, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	if err = pd.push(&lengthField{}); err != nil {
+		return err
+	}
+
+	msb.Msg = new(Message)
+	if err = msb.Msg.decode(pd); err != nil {
+		return err
+	}
+
+	if err = pd.pop(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+type MessageSet struct {
+	PartialTrailingMessage bool // whether the set on the wire contained an incomplete trailing MessageBlock
+	OverflowMessage        bool // whether the set on the wire contained an overflow message
+	Messages               []*MessageBlock
+}
+
+func (ms *MessageSet) encode(pe packetEncoder) error {
+	for i := range ms.Messages {
+		err := ms.Messages[i].encode(pe)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (ms *MessageSet) decode(pd packetDecoder) (err error) {
+	ms.Messages = nil
+
+	for pd.remaining() > 0 {
+		magic, err := magicValue(pd)
+		if err != nil {
+			if err == ErrInsufficientData {
+				ms.PartialTrailingMessage = true
+				return nil
+			}
+			return err
+		}
+
+		if magic > 1 {
+			return nil
+		}
+
+		msb := new(MessageBlock)
+		err = msb.decode(pd)
+		switch err {
+		case nil:
+			ms.Messages = append(ms.Messages, msb)
+		case ErrInsufficientData:
+			// As an optimization the server is allowed to return a partial message at the
+			// end of the message set. Clients should handle this case. So we just ignore such things.
+			if msb.Offset == -1 {
+				// This is an overflow message caused by chunked down conversion
+				ms.OverflowMessage = true
+			} else {
+				ms.PartialTrailingMessage = true
+			}
+			return nil
+		default:
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (ms *MessageSet) addMessage(msg *Message) {
+	block := new(MessageBlock)
+	block.Msg = msg
+	ms.Messages = append(ms.Messages, block)
+}
diff --git a/vendor/github.com/Shopify/sarama/metadata_request.go b/vendor/github.com/Shopify/sarama/metadata_request.go
new file mode 100644
index 0000000..17dc428
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/metadata_request.go
@@ -0,0 +1,88 @@
+package sarama
+
+type MetadataRequest struct {
+	Version                int16
+	Topics                 []string
+	AllowAutoTopicCreation bool
+}
+
+func (r *MetadataRequest) encode(pe packetEncoder) error {
+	if r.Version < 0 || r.Version > 5 {
+		return PacketEncodingError{"invalid or unsupported MetadataRequest version field"}
+	}
+	if r.Version == 0 || len(r.Topics) > 0 {
+		err := pe.putArrayLength(len(r.Topics))
+		if err != nil {
+			return err
+		}
+
+		for i := range r.Topics {
+			err = pe.putString(r.Topics[i])
+			if err != nil {
+				return err
+			}
+		}
+	} else {
+		pe.putInt32(-1)
+	}
+	if r.Version > 3 {
+		pe.putBool(r.AllowAutoTopicCreation)
+	}
+	return nil
+}
+
+func (r *MetadataRequest) decode(pd packetDecoder, version int16) error {
+	r.Version = version
+	size, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	if size < 0 {
+		return nil
+	} else {
+		topicCount := size
+		if topicCount == 0 {
+			return nil
+		}
+
+		r.Topics = make([]string, topicCount)
+		for i := range r.Topics {
+			topic, err := pd.getString()
+			if err != nil {
+				return err
+			}
+			r.Topics[i] = topic
+		}
+	}
+	if r.Version > 3 {
+		autoCreation, err := pd.getBool()
+		if err != nil {
+			return err
+		}
+		r.AllowAutoTopicCreation = autoCreation
+	}
+	return nil
+}
+
+func (r *MetadataRequest) key() int16 {
+	return 3
+}
+
+func (r *MetadataRequest) version() int16 {
+	return r.Version
+}
+
+func (r *MetadataRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_10_0_0
+	case 2:
+		return V0_10_1_0
+	case 3, 4:
+		return V0_11_0_0
+	case 5:
+		return V1_0_0_0
+	default:
+		return MinVersion
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/metadata_response.go b/vendor/github.com/Shopify/sarama/metadata_response.go
new file mode 100644
index 0000000..c402d05
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/metadata_response.go
@@ -0,0 +1,321 @@
+package sarama
+
+type PartitionMetadata struct {
+	Err             KError
+	ID              int32
+	Leader          int32
+	Replicas        []int32
+	Isr             []int32
+	OfflineReplicas []int32
+}
+
+func (pm *PartitionMetadata) decode(pd packetDecoder, version int16) (err error) {
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	pm.Err = KError(tmp)
+
+	pm.ID, err = pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	pm.Leader, err = pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	pm.Replicas, err = pd.getInt32Array()
+	if err != nil {
+		return err
+	}
+
+	pm.Isr, err = pd.getInt32Array()
+	if err != nil {
+		return err
+	}
+
+	if version >= 5 {
+		pm.OfflineReplicas, err = pd.getInt32Array()
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (pm *PartitionMetadata) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(int16(pm.Err))
+	pe.putInt32(pm.ID)
+	pe.putInt32(pm.Leader)
+
+	err = pe.putInt32Array(pm.Replicas)
+	if err != nil {
+		return err
+	}
+
+	err = pe.putInt32Array(pm.Isr)
+	if err != nil {
+		return err
+	}
+
+	if version >= 5 {
+		err = pe.putInt32Array(pm.OfflineReplicas)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+type TopicMetadata struct {
+	Err        KError
+	Name       string
+	IsInternal bool // Only valid for Version >= 1
+	Partitions []*PartitionMetadata
+}
+
+func (tm *TopicMetadata) decode(pd packetDecoder, version int16) (err error) {
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	tm.Err = KError(tmp)
+
+	tm.Name, err = pd.getString()
+	if err != nil {
+		return err
+	}
+
+	if version >= 1 {
+		tm.IsInternal, err = pd.getBool()
+		if err != nil {
+			return err
+		}
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	tm.Partitions = make([]*PartitionMetadata, n)
+	for i := 0; i < n; i++ {
+		tm.Partitions[i] = new(PartitionMetadata)
+		err = tm.Partitions[i].decode(pd, version)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (tm *TopicMetadata) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(int16(tm.Err))
+
+	err = pe.putString(tm.Name)
+	if err != nil {
+		return err
+	}
+
+	if version >= 1 {
+		pe.putBool(tm.IsInternal)
+	}
+
+	err = pe.putArrayLength(len(tm.Partitions))
+	if err != nil {
+		return err
+	}
+
+	for _, pm := range tm.Partitions {
+		err = pm.encode(pe, version)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+type MetadataResponse struct {
+	Version        int16
+	ThrottleTimeMs int32
+	Brokers        []*Broker
+	ClusterID      *string
+	ControllerID   int32
+	Topics         []*TopicMetadata
+}
+
+func (r *MetadataResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if version >= 3 {
+		r.ThrottleTimeMs, err = pd.getInt32()
+		if err != nil {
+			return err
+		}
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Brokers = make([]*Broker, n)
+	for i := 0; i < n; i++ {
+		r.Brokers[i] = new(Broker)
+		err = r.Brokers[i].decode(pd, version)
+		if err != nil {
+			return err
+		}
+	}
+
+	if version >= 2 {
+		r.ClusterID, err = pd.getNullableString()
+		if err != nil {
+			return err
+		}
+	}
+
+	if version >= 1 {
+		r.ControllerID, err = pd.getInt32()
+		if err != nil {
+			return err
+		}
+	} else {
+		r.ControllerID = -1
+	}
+
+	n, err = pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Topics = make([]*TopicMetadata, n)
+	for i := 0; i < n; i++ {
+		r.Topics[i] = new(TopicMetadata)
+		err = r.Topics[i].decode(pd, version)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *MetadataResponse) encode(pe packetEncoder) error {
+	if r.Version >= 3 {
+		pe.putInt32(r.ThrottleTimeMs)
+	}
+
+	err := pe.putArrayLength(len(r.Brokers))
+	if err != nil {
+		return err
+	}
+	for _, broker := range r.Brokers {
+		err = broker.encode(pe, r.Version)
+		if err != nil {
+			return err
+		}
+	}
+
+	if r.Version >= 2 {
+		err := pe.putNullableString(r.ClusterID)
+		if err != nil {
+			return err
+		}
+	}
+
+	if r.Version >= 1 {
+		pe.putInt32(r.ControllerID)
+	}
+
+	err = pe.putArrayLength(len(r.Topics))
+	if err != nil {
+		return err
+	}
+	for _, tm := range r.Topics {
+		err = tm.encode(pe, r.Version)
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *MetadataResponse) key() int16 {
+	return 3
+}
+
+func (r *MetadataResponse) version() int16 {
+	return r.Version
+}
+
+func (r *MetadataResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_10_0_0
+	case 2:
+		return V0_10_1_0
+	case 3, 4:
+		return V0_11_0_0
+	case 5:
+		return V1_0_0_0
+	default:
+		return MinVersion
+	}
+}
+
+// testing API
+
+func (r *MetadataResponse) AddBroker(addr string, id int32) {
+	r.Brokers = append(r.Brokers, &Broker{id: id, addr: addr})
+}
+
+func (r *MetadataResponse) AddTopic(topic string, err KError) *TopicMetadata {
+	var tmatch *TopicMetadata
+
+	for _, tm := range r.Topics {
+		if tm.Name == topic {
+			tmatch = tm
+			goto foundTopic
+		}
+	}
+
+	tmatch = new(TopicMetadata)
+	tmatch.Name = topic
+	r.Topics = append(r.Topics, tmatch)
+
+foundTopic:
+
+	tmatch.Err = err
+	return tmatch
+}
+
+func (r *MetadataResponse) AddTopicPartition(topic string, partition, brokerID int32, replicas, isr []int32, err KError) {
+	tmatch := r.AddTopic(topic, ErrNoError)
+	var pmatch *PartitionMetadata
+
+	for _, pm := range tmatch.Partitions {
+		if pm.ID == partition {
+			pmatch = pm
+			goto foundPartition
+		}
+	}
+
+	pmatch = new(PartitionMetadata)
+	pmatch.ID = partition
+	tmatch.Partitions = append(tmatch.Partitions, pmatch)
+
+foundPartition:
+
+	pmatch.Leader = brokerID
+	pmatch.Replicas = replicas
+	pmatch.Isr = isr
+	pmatch.Err = err
+
+}
diff --git a/vendor/github.com/Shopify/sarama/metrics.go b/vendor/github.com/Shopify/sarama/metrics.go
new file mode 100644
index 0000000..4869708
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/metrics.go
@@ -0,0 +1,51 @@
+package sarama
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+// Use exponentially decaying reservoir for sampling histograms with the same defaults as the Java library:
+// 1028 elements, which offers a 99.9% confidence level with a 5% margin of error assuming a normal distribution,
+// and an alpha factor of 0.015, which heavily biases the reservoir to the past 5 minutes of measurements.
+// See https://github.com/dropwizard/metrics/blob/v3.1.0/metrics-core/src/main/java/com/codahale/metrics/ExponentiallyDecayingReservoir.java#L38
+const (
+	metricsReservoirSize = 1028
+	metricsAlphaFactor   = 0.015
+)
+
+func getOrRegisterHistogram(name string, r metrics.Registry) metrics.Histogram {
+	return r.GetOrRegister(name, func() metrics.Histogram {
+		return metrics.NewHistogram(metrics.NewExpDecaySample(metricsReservoirSize, metricsAlphaFactor))
+	}).(metrics.Histogram)
+}
+
+func getMetricNameForBroker(name string, broker *Broker) string {
+	// Use broker id like the Java client as it does not contain '.' or ':' characters that
+	// can be interpreted as special character by monitoring tool (e.g. Graphite)
+	return fmt.Sprintf(name+"-for-broker-%d", broker.ID())
+}
+
+func getOrRegisterBrokerMeter(name string, broker *Broker, r metrics.Registry) metrics.Meter {
+	return metrics.GetOrRegisterMeter(getMetricNameForBroker(name, broker), r)
+}
+
+func getOrRegisterBrokerHistogram(name string, broker *Broker, r metrics.Registry) metrics.Histogram {
+	return getOrRegisterHistogram(getMetricNameForBroker(name, broker), r)
+}
+
+func getMetricNameForTopic(name string, topic string) string {
+	// Convert dot to _ since reporters like Graphite typically use dot to represent hierarchy
+	// cf. KAFKA-1902 and KAFKA-2337
+	return fmt.Sprintf(name+"-for-topic-%s", strings.Replace(topic, ".", "_", -1))
+}
+
+func getOrRegisterTopicMeter(name string, topic string, r metrics.Registry) metrics.Meter {
+	return metrics.GetOrRegisterMeter(getMetricNameForTopic(name, topic), r)
+}
+
+func getOrRegisterTopicHistogram(name string, topic string, r metrics.Registry) metrics.Histogram {
+	return getOrRegisterHistogram(getMetricNameForTopic(name, topic), r)
+}
diff --git a/vendor/github.com/Shopify/sarama/mockbroker.go b/vendor/github.com/Shopify/sarama/mockbroker.go
new file mode 100644
index 0000000..55ef1e2
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/mockbroker.go
@@ -0,0 +1,330 @@
+package sarama
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"net"
+	"reflect"
+	"strconv"
+	"sync"
+	"time"
+
+	"github.com/davecgh/go-spew/spew"
+)
+
+const (
+	expectationTimeout = 500 * time.Millisecond
+)
+
+type requestHandlerFunc func(req *request) (res encoder)
+
+// RequestNotifierFunc is invoked when a mock broker processes a request successfully
+// and will provides the number of bytes read and written.
+type RequestNotifierFunc func(bytesRead, bytesWritten int)
+
+// MockBroker is a mock Kafka broker that is used in unit tests. It is exposed
+// to facilitate testing of higher level or specialized consumers and producers
+// built on top of Sarama. Note that it does not 'mimic' the Kafka API protocol,
+// but rather provides a facility to do that. It takes care of the TCP
+// transport, request unmarshaling, response marshaling, and makes it the test
+// writer responsibility to program correct according to the Kafka API protocol
+// MockBroker behaviour.
+//
+// MockBroker is implemented as a TCP server listening on a kernel-selected
+// localhost port that can accept many connections. It reads Kafka requests
+// from that connection and returns responses programmed by the SetHandlerByMap
+// function. If a MockBroker receives a request that it has no programmed
+// response for, then it returns nothing and the request times out.
+//
+// A set of MockRequest builders to define mappings used by MockBroker is
+// provided by Sarama. But users can develop MockRequests of their own and use
+// them along with or instead of the standard ones.
+//
+// When running tests with MockBroker it is strongly recommended to specify
+// a timeout to `go test` so that if the broker hangs waiting for a response,
+// the test panics.
+//
+// It is not necessary to prefix message length or correlation ID to your
+// response bytes, the server does that automatically as a convenience.
+type MockBroker struct {
+	brokerID     int32
+	port         int32
+	closing      chan none
+	stopper      chan none
+	expectations chan encoder
+	listener     net.Listener
+	t            TestReporter
+	latency      time.Duration
+	handler      requestHandlerFunc
+	notifier     RequestNotifierFunc
+	history      []RequestResponse
+	lock         sync.Mutex
+}
+
+// RequestResponse represents a Request/Response pair processed by MockBroker.
+type RequestResponse struct {
+	Request  protocolBody
+	Response encoder
+}
+
+// SetLatency makes broker pause for the specified period every time before
+// replying.
+func (b *MockBroker) SetLatency(latency time.Duration) {
+	b.latency = latency
+}
+
+// SetHandlerByMap defines mapping of Request types to MockResponses. When a
+// request is received by the broker, it looks up the request type in the map
+// and uses the found MockResponse instance to generate an appropriate reply.
+// If the request type is not found in the map then nothing is sent.
+func (b *MockBroker) SetHandlerByMap(handlerMap map[string]MockResponse) {
+	b.setHandler(func(req *request) (res encoder) {
+		reqTypeName := reflect.TypeOf(req.body).Elem().Name()
+		mockResponse := handlerMap[reqTypeName]
+		if mockResponse == nil {
+			return nil
+		}
+		return mockResponse.For(req.body)
+	})
+}
+
+// SetNotifier set a function that will get invoked whenever a request has been
+// processed successfully and will provide the number of bytes read and written
+func (b *MockBroker) SetNotifier(notifier RequestNotifierFunc) {
+	b.lock.Lock()
+	b.notifier = notifier
+	b.lock.Unlock()
+}
+
+// BrokerID returns broker ID assigned to the broker.
+func (b *MockBroker) BrokerID() int32 {
+	return b.brokerID
+}
+
+// History returns a slice of RequestResponse pairs in the order they were
+// processed by the broker. Note that in case of multiple connections to the
+// broker the order expected by a test can be different from the order recorded
+// in the history, unless some synchronization is implemented in the test.
+func (b *MockBroker) History() []RequestResponse {
+	b.lock.Lock()
+	history := make([]RequestResponse, len(b.history))
+	copy(history, b.history)
+	b.lock.Unlock()
+	return history
+}
+
+// Port returns the TCP port number the broker is listening for requests on.
+func (b *MockBroker) Port() int32 {
+	return b.port
+}
+
+// Addr returns the broker connection string in the form "<address>:<port>".
+func (b *MockBroker) Addr() string {
+	return b.listener.Addr().String()
+}
+
+// Close terminates the broker blocking until it stops internal goroutines and
+// releases all resources.
+func (b *MockBroker) Close() {
+	close(b.expectations)
+	if len(b.expectations) > 0 {
+		buf := bytes.NewBufferString(fmt.Sprintf("mockbroker/%d: not all expectations were satisfied! Still waiting on:\n", b.BrokerID()))
+		for e := range b.expectations {
+			_, _ = buf.WriteString(spew.Sdump(e))
+		}
+		b.t.Error(buf.String())
+	}
+	close(b.closing)
+	<-b.stopper
+}
+
+// setHandler sets the specified function as the request handler. Whenever
+// a mock broker reads a request from the wire it passes the request to the
+// function and sends back whatever the handler function returns.
+func (b *MockBroker) setHandler(handler requestHandlerFunc) {
+	b.lock.Lock()
+	b.handler = handler
+	b.lock.Unlock()
+}
+
+func (b *MockBroker) serverLoop() {
+	defer close(b.stopper)
+	var err error
+	var conn net.Conn
+
+	go func() {
+		<-b.closing
+		err := b.listener.Close()
+		if err != nil {
+			b.t.Error(err)
+		}
+	}()
+
+	wg := &sync.WaitGroup{}
+	i := 0
+	for conn, err = b.listener.Accept(); err == nil; conn, err = b.listener.Accept() {
+		wg.Add(1)
+		go b.handleRequests(conn, i, wg)
+		i++
+	}
+	wg.Wait()
+	Logger.Printf("*** mockbroker/%d: listener closed, err=%v", b.BrokerID(), err)
+}
+
+func (b *MockBroker) handleRequests(conn net.Conn, idx int, wg *sync.WaitGroup) {
+	defer wg.Done()
+	defer func() {
+		_ = conn.Close()
+	}()
+	Logger.Printf("*** mockbroker/%d/%d: connection opened", b.BrokerID(), idx)
+	var err error
+
+	abort := make(chan none)
+	defer close(abort)
+	go func() {
+		select {
+		case <-b.closing:
+			_ = conn.Close()
+		case <-abort:
+		}
+	}()
+
+	resHeader := make([]byte, 8)
+	for {
+		req, bytesRead, err := decodeRequest(conn)
+		if err != nil {
+			Logger.Printf("*** mockbroker/%d/%d: invalid request: err=%+v, %+v", b.brokerID, idx, err, spew.Sdump(req))
+			b.serverError(err)
+			break
+		}
+
+		if b.latency > 0 {
+			time.Sleep(b.latency)
+		}
+
+		b.lock.Lock()
+		res := b.handler(req)
+		b.history = append(b.history, RequestResponse{req.body, res})
+		b.lock.Unlock()
+
+		if res == nil {
+			Logger.Printf("*** mockbroker/%d/%d: ignored %v", b.brokerID, idx, spew.Sdump(req))
+			continue
+		}
+		Logger.Printf("*** mockbroker/%d/%d: served %v -> %v", b.brokerID, idx, req, res)
+
+		encodedRes, err := encode(res, nil)
+		if err != nil {
+			b.serverError(err)
+			break
+		}
+		if len(encodedRes) == 0 {
+			b.lock.Lock()
+			if b.notifier != nil {
+				b.notifier(bytesRead, 0)
+			}
+			b.lock.Unlock()
+			continue
+		}
+
+		binary.BigEndian.PutUint32(resHeader, uint32(len(encodedRes)+4))
+		binary.BigEndian.PutUint32(resHeader[4:], uint32(req.correlationID))
+		if _, err = conn.Write(resHeader); err != nil {
+			b.serverError(err)
+			break
+		}
+		if _, err = conn.Write(encodedRes); err != nil {
+			b.serverError(err)
+			break
+		}
+
+		b.lock.Lock()
+		if b.notifier != nil {
+			b.notifier(bytesRead, len(resHeader)+len(encodedRes))
+		}
+		b.lock.Unlock()
+	}
+	Logger.Printf("*** mockbroker/%d/%d: connection closed, err=%v", b.BrokerID(), idx, err)
+}
+
+func (b *MockBroker) defaultRequestHandler(req *request) (res encoder) {
+	select {
+	case res, ok := <-b.expectations:
+		if !ok {
+			return nil
+		}
+		return res
+	case <-time.After(expectationTimeout):
+		return nil
+	}
+}
+
+func (b *MockBroker) serverError(err error) {
+	isConnectionClosedError := false
+	if _, ok := err.(*net.OpError); ok {
+		isConnectionClosedError = true
+	} else if err == io.EOF {
+		isConnectionClosedError = true
+	} else if err.Error() == "use of closed network connection" {
+		isConnectionClosedError = true
+	}
+
+	if isConnectionClosedError {
+		return
+	}
+
+	b.t.Errorf(err.Error())
+}
+
+// NewMockBroker launches a fake Kafka broker. It takes a TestReporter as provided by the
+// test framework and a channel of responses to use.  If an error occurs it is
+// simply logged to the TestReporter and the broker exits.
+func NewMockBroker(t TestReporter, brokerID int32) *MockBroker {
+	return NewMockBrokerAddr(t, brokerID, "localhost:0")
+}
+
+// NewMockBrokerAddr behaves like newMockBroker but listens on the address you give
+// it rather than just some ephemeral port.
+func NewMockBrokerAddr(t TestReporter, brokerID int32, addr string) *MockBroker {
+	listener, err := net.Listen("tcp", addr)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return NewMockBrokerListener(t, brokerID, listener)
+}
+
+// NewMockBrokerListener behaves like newMockBrokerAddr but accepts connections on the listener specified.
+func NewMockBrokerListener(t TestReporter, brokerID int32, listener net.Listener) *MockBroker {
+	var err error
+
+	broker := &MockBroker{
+		closing:      make(chan none),
+		stopper:      make(chan none),
+		t:            t,
+		brokerID:     brokerID,
+		expectations: make(chan encoder, 512),
+		listener:     listener,
+	}
+	broker.handler = broker.defaultRequestHandler
+
+	Logger.Printf("*** mockbroker/%d listening on %s\n", brokerID, broker.listener.Addr().String())
+	_, portStr, err := net.SplitHostPort(broker.listener.Addr().String())
+	if err != nil {
+		t.Fatal(err)
+	}
+	tmp, err := strconv.ParseInt(portStr, 10, 32)
+	if err != nil {
+		t.Fatal(err)
+	}
+	broker.port = int32(tmp)
+
+	go broker.serverLoop()
+
+	return broker
+}
+
+func (b *MockBroker) Returns(e encoder) {
+	b.expectations <- e
+}
diff --git a/vendor/github.com/Shopify/sarama/mockresponses.go b/vendor/github.com/Shopify/sarama/mockresponses.go
new file mode 100644
index 0000000..348c223
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/mockresponses.go
@@ -0,0 +1,868 @@
+package sarama
+
+import (
+	"fmt"
+)
+
+// TestReporter has methods matching go's testing.T to avoid importing
+// `testing` in the main part of the library.
+type TestReporter interface {
+	Error(...interface{})
+	Errorf(string, ...interface{})
+	Fatal(...interface{})
+	Fatalf(string, ...interface{})
+}
+
+// MockResponse is a response builder interface it defines one method that
+// allows generating a response based on a request body. MockResponses are used
+// to program behavior of MockBroker in tests.
+type MockResponse interface {
+	For(reqBody versionedDecoder) (res encoder)
+}
+
+// MockWrapper is a mock response builder that returns a particular concrete
+// response regardless of the actual request passed to the `For` method.
+type MockWrapper struct {
+	res encoder
+}
+
+func (mw *MockWrapper) For(reqBody versionedDecoder) (res encoder) {
+	return mw.res
+}
+
+func NewMockWrapper(res encoder) *MockWrapper {
+	return &MockWrapper{res: res}
+}
+
+// MockSequence is a mock response builder that is created from a sequence of
+// concrete responses. Every time when a `MockBroker` calls its `For` method
+// the next response from the sequence is returned. When the end of the
+// sequence is reached the last element from the sequence is returned.
+type MockSequence struct {
+	responses []MockResponse
+}
+
+func NewMockSequence(responses ...interface{}) *MockSequence {
+	ms := &MockSequence{}
+	ms.responses = make([]MockResponse, len(responses))
+	for i, res := range responses {
+		switch res := res.(type) {
+		case MockResponse:
+			ms.responses[i] = res
+		case encoder:
+			ms.responses[i] = NewMockWrapper(res)
+		default:
+			panic(fmt.Sprintf("Unexpected response type: %T", res))
+		}
+	}
+	return ms
+}
+
+func (mc *MockSequence) For(reqBody versionedDecoder) (res encoder) {
+	res = mc.responses[0].For(reqBody)
+	if len(mc.responses) > 1 {
+		mc.responses = mc.responses[1:]
+	}
+	return res
+}
+
+type MockListGroupsResponse struct {
+	groups map[string]string
+	t      TestReporter
+}
+
+func NewMockListGroupsResponse(t TestReporter) *MockListGroupsResponse {
+	return &MockListGroupsResponse{
+		groups: make(map[string]string),
+		t:      t,
+	}
+}
+
+func (m *MockListGroupsResponse) For(reqBody versionedDecoder) encoder {
+	request := reqBody.(*ListGroupsRequest)
+	_ = request
+	response := &ListGroupsResponse{
+		Groups: m.groups,
+	}
+	return response
+}
+
+func (m *MockListGroupsResponse) AddGroup(groupID, protocolType string) *MockListGroupsResponse {
+	m.groups[groupID] = protocolType
+	return m
+}
+
+type MockDescribeGroupsResponse struct {
+	groups map[string]*GroupDescription
+	t      TestReporter
+}
+
+func NewMockDescribeGroupsResponse(t TestReporter) *MockDescribeGroupsResponse {
+	return &MockDescribeGroupsResponse{
+		t:      t,
+		groups: make(map[string]*GroupDescription),
+	}
+}
+
+func (m *MockDescribeGroupsResponse) AddGroupDescription(groupID string, description *GroupDescription) *MockDescribeGroupsResponse {
+	m.groups[groupID] = description
+	return m
+}
+
+func (m *MockDescribeGroupsResponse) For(reqBody versionedDecoder) encoder {
+	request := reqBody.(*DescribeGroupsRequest)
+
+	response := &DescribeGroupsResponse{}
+	for _, requestedGroup := range request.Groups {
+		if group, ok := m.groups[requestedGroup]; ok {
+			response.Groups = append(response.Groups, group)
+		} else {
+			// Mimic real kafka - if a group doesn't exist, return
+			// an entry with state "Dead"
+			response.Groups = append(response.Groups, &GroupDescription{
+				GroupId: requestedGroup,
+				State:   "Dead",
+			})
+		}
+	}
+
+	return response
+}
+
+// MockMetadataResponse is a `MetadataResponse` builder.
+type MockMetadataResponse struct {
+	controllerID int32
+	leaders      map[string]map[int32]int32
+	brokers      map[string]int32
+	t            TestReporter
+}
+
+func NewMockMetadataResponse(t TestReporter) *MockMetadataResponse {
+	return &MockMetadataResponse{
+		leaders: make(map[string]map[int32]int32),
+		brokers: make(map[string]int32),
+		t:       t,
+	}
+}
+
+func (mmr *MockMetadataResponse) SetLeader(topic string, partition, brokerID int32) *MockMetadataResponse {
+	partitions := mmr.leaders[topic]
+	if partitions == nil {
+		partitions = make(map[int32]int32)
+		mmr.leaders[topic] = partitions
+	}
+	partitions[partition] = brokerID
+	return mmr
+}
+
+func (mmr *MockMetadataResponse) SetBroker(addr string, brokerID int32) *MockMetadataResponse {
+	mmr.brokers[addr] = brokerID
+	return mmr
+}
+
+func (mmr *MockMetadataResponse) SetController(brokerID int32) *MockMetadataResponse {
+	mmr.controllerID = brokerID
+	return mmr
+}
+
+func (mmr *MockMetadataResponse) For(reqBody versionedDecoder) encoder {
+	metadataRequest := reqBody.(*MetadataRequest)
+	metadataResponse := &MetadataResponse{
+		Version:      metadataRequest.version(),
+		ControllerID: mmr.controllerID,
+	}
+	for addr, brokerID := range mmr.brokers {
+		metadataResponse.AddBroker(addr, brokerID)
+	}
+
+	// Generate set of replicas
+	replicas := []int32{}
+
+	for _, brokerID := range mmr.brokers {
+		replicas = append(replicas, brokerID)
+	}
+
+	if len(metadataRequest.Topics) == 0 {
+		for topic, partitions := range mmr.leaders {
+			for partition, brokerID := range partitions {
+				metadataResponse.AddTopicPartition(topic, partition, brokerID, replicas, replicas, ErrNoError)
+			}
+		}
+		return metadataResponse
+	}
+	for _, topic := range metadataRequest.Topics {
+		for partition, brokerID := range mmr.leaders[topic] {
+			metadataResponse.AddTopicPartition(topic, partition, brokerID, replicas, replicas, ErrNoError)
+		}
+	}
+	return metadataResponse
+}
+
+// MockOffsetResponse is an `OffsetResponse` builder.
+type MockOffsetResponse struct {
+	offsets map[string]map[int32]map[int64]int64
+	t       TestReporter
+	version int16
+}
+
+func NewMockOffsetResponse(t TestReporter) *MockOffsetResponse {
+	return &MockOffsetResponse{
+		offsets: make(map[string]map[int32]map[int64]int64),
+		t:       t,
+	}
+}
+
+func (mor *MockOffsetResponse) SetVersion(version int16) *MockOffsetResponse {
+	mor.version = version
+	return mor
+}
+
+func (mor *MockOffsetResponse) SetOffset(topic string, partition int32, time, offset int64) *MockOffsetResponse {
+	partitions := mor.offsets[topic]
+	if partitions == nil {
+		partitions = make(map[int32]map[int64]int64)
+		mor.offsets[topic] = partitions
+	}
+	times := partitions[partition]
+	if times == nil {
+		times = make(map[int64]int64)
+		partitions[partition] = times
+	}
+	times[time] = offset
+	return mor
+}
+
+func (mor *MockOffsetResponse) For(reqBody versionedDecoder) encoder {
+	offsetRequest := reqBody.(*OffsetRequest)
+	offsetResponse := &OffsetResponse{Version: mor.version}
+	for topic, partitions := range offsetRequest.blocks {
+		for partition, block := range partitions {
+			offset := mor.getOffset(topic, partition, block.time)
+			offsetResponse.AddTopicPartition(topic, partition, offset)
+		}
+	}
+	return offsetResponse
+}
+
+func (mor *MockOffsetResponse) getOffset(topic string, partition int32, time int64) int64 {
+	partitions := mor.offsets[topic]
+	if partitions == nil {
+		mor.t.Errorf("missing topic: %s", topic)
+	}
+	times := partitions[partition]
+	if times == nil {
+		mor.t.Errorf("missing partition: %d", partition)
+	}
+	offset, ok := times[time]
+	if !ok {
+		mor.t.Errorf("missing time: %d", time)
+	}
+	return offset
+}
+
+// MockFetchResponse is a `FetchResponse` builder.
+type MockFetchResponse struct {
+	messages       map[string]map[int32]map[int64]Encoder
+	highWaterMarks map[string]map[int32]int64
+	t              TestReporter
+	batchSize      int
+	version        int16
+}
+
+func NewMockFetchResponse(t TestReporter, batchSize int) *MockFetchResponse {
+	return &MockFetchResponse{
+		messages:       make(map[string]map[int32]map[int64]Encoder),
+		highWaterMarks: make(map[string]map[int32]int64),
+		t:              t,
+		batchSize:      batchSize,
+	}
+}
+
+func (mfr *MockFetchResponse) SetVersion(version int16) *MockFetchResponse {
+	mfr.version = version
+	return mfr
+}
+
+func (mfr *MockFetchResponse) SetMessage(topic string, partition int32, offset int64, msg Encoder) *MockFetchResponse {
+	partitions := mfr.messages[topic]
+	if partitions == nil {
+		partitions = make(map[int32]map[int64]Encoder)
+		mfr.messages[topic] = partitions
+	}
+	messages := partitions[partition]
+	if messages == nil {
+		messages = make(map[int64]Encoder)
+		partitions[partition] = messages
+	}
+	messages[offset] = msg
+	return mfr
+}
+
+func (mfr *MockFetchResponse) SetHighWaterMark(topic string, partition int32, offset int64) *MockFetchResponse {
+	partitions := mfr.highWaterMarks[topic]
+	if partitions == nil {
+		partitions = make(map[int32]int64)
+		mfr.highWaterMarks[topic] = partitions
+	}
+	partitions[partition] = offset
+	return mfr
+}
+
+func (mfr *MockFetchResponse) For(reqBody versionedDecoder) encoder {
+	fetchRequest := reqBody.(*FetchRequest)
+	res := &FetchResponse{
+		Version: mfr.version,
+	}
+	for topic, partitions := range fetchRequest.blocks {
+		for partition, block := range partitions {
+			initialOffset := block.fetchOffset
+			offset := initialOffset
+			maxOffset := initialOffset + int64(mfr.getMessageCount(topic, partition))
+			for i := 0; i < mfr.batchSize && offset < maxOffset; {
+				msg := mfr.getMessage(topic, partition, offset)
+				if msg != nil {
+					res.AddMessage(topic, partition, nil, msg, offset)
+					i++
+				}
+				offset++
+			}
+			fb := res.GetBlock(topic, partition)
+			if fb == nil {
+				res.AddError(topic, partition, ErrNoError)
+				fb = res.GetBlock(topic, partition)
+			}
+			fb.HighWaterMarkOffset = mfr.getHighWaterMark(topic, partition)
+		}
+	}
+	return res
+}
+
+func (mfr *MockFetchResponse) getMessage(topic string, partition int32, offset int64) Encoder {
+	partitions := mfr.messages[topic]
+	if partitions == nil {
+		return nil
+	}
+	messages := partitions[partition]
+	if messages == nil {
+		return nil
+	}
+	return messages[offset]
+}
+
+func (mfr *MockFetchResponse) getMessageCount(topic string, partition int32) int {
+	partitions := mfr.messages[topic]
+	if partitions == nil {
+		return 0
+	}
+	messages := partitions[partition]
+	if messages == nil {
+		return 0
+	}
+	return len(messages)
+}
+
+func (mfr *MockFetchResponse) getHighWaterMark(topic string, partition int32) int64 {
+	partitions := mfr.highWaterMarks[topic]
+	if partitions == nil {
+		return 0
+	}
+	return partitions[partition]
+}
+
+// MockConsumerMetadataResponse is a `ConsumerMetadataResponse` builder.
+type MockConsumerMetadataResponse struct {
+	coordinators map[string]interface{}
+	t            TestReporter
+}
+
+func NewMockConsumerMetadataResponse(t TestReporter) *MockConsumerMetadataResponse {
+	return &MockConsumerMetadataResponse{
+		coordinators: make(map[string]interface{}),
+		t:            t,
+	}
+}
+
+func (mr *MockConsumerMetadataResponse) SetCoordinator(group string, broker *MockBroker) *MockConsumerMetadataResponse {
+	mr.coordinators[group] = broker
+	return mr
+}
+
+func (mr *MockConsumerMetadataResponse) SetError(group string, kerror KError) *MockConsumerMetadataResponse {
+	mr.coordinators[group] = kerror
+	return mr
+}
+
+func (mr *MockConsumerMetadataResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*ConsumerMetadataRequest)
+	group := req.ConsumerGroup
+	res := &ConsumerMetadataResponse{}
+	v := mr.coordinators[group]
+	switch v := v.(type) {
+	case *MockBroker:
+		res.Coordinator = &Broker{id: v.BrokerID(), addr: v.Addr()}
+	case KError:
+		res.Err = v
+	}
+	return res
+}
+
+// MockFindCoordinatorResponse is a `FindCoordinatorResponse` builder.
+type MockFindCoordinatorResponse struct {
+	groupCoordinators map[string]interface{}
+	transCoordinators map[string]interface{}
+	t                 TestReporter
+}
+
+func NewMockFindCoordinatorResponse(t TestReporter) *MockFindCoordinatorResponse {
+	return &MockFindCoordinatorResponse{
+		groupCoordinators: make(map[string]interface{}),
+		transCoordinators: make(map[string]interface{}),
+		t:                 t,
+	}
+}
+
+func (mr *MockFindCoordinatorResponse) SetCoordinator(coordinatorType CoordinatorType, group string, broker *MockBroker) *MockFindCoordinatorResponse {
+	switch coordinatorType {
+	case CoordinatorGroup:
+		mr.groupCoordinators[group] = broker
+	case CoordinatorTransaction:
+		mr.transCoordinators[group] = broker
+	}
+	return mr
+}
+
+func (mr *MockFindCoordinatorResponse) SetError(coordinatorType CoordinatorType, group string, kerror KError) *MockFindCoordinatorResponse {
+	switch coordinatorType {
+	case CoordinatorGroup:
+		mr.groupCoordinators[group] = kerror
+	case CoordinatorTransaction:
+		mr.transCoordinators[group] = kerror
+	}
+	return mr
+}
+
+func (mr *MockFindCoordinatorResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*FindCoordinatorRequest)
+	res := &FindCoordinatorResponse{}
+	var v interface{}
+	switch req.CoordinatorType {
+	case CoordinatorGroup:
+		v = mr.groupCoordinators[req.CoordinatorKey]
+	case CoordinatorTransaction:
+		v = mr.transCoordinators[req.CoordinatorKey]
+	}
+	switch v := v.(type) {
+	case *MockBroker:
+		res.Coordinator = &Broker{id: v.BrokerID(), addr: v.Addr()}
+	case KError:
+		res.Err = v
+	}
+	return res
+}
+
+// MockOffsetCommitResponse is a `OffsetCommitResponse` builder.
+type MockOffsetCommitResponse struct {
+	errors map[string]map[string]map[int32]KError
+	t      TestReporter
+}
+
+func NewMockOffsetCommitResponse(t TestReporter) *MockOffsetCommitResponse {
+	return &MockOffsetCommitResponse{t: t}
+}
+
+func (mr *MockOffsetCommitResponse) SetError(group, topic string, partition int32, kerror KError) *MockOffsetCommitResponse {
+	if mr.errors == nil {
+		mr.errors = make(map[string]map[string]map[int32]KError)
+	}
+	topics := mr.errors[group]
+	if topics == nil {
+		topics = make(map[string]map[int32]KError)
+		mr.errors[group] = topics
+	}
+	partitions := topics[topic]
+	if partitions == nil {
+		partitions = make(map[int32]KError)
+		topics[topic] = partitions
+	}
+	partitions[partition] = kerror
+	return mr
+}
+
+func (mr *MockOffsetCommitResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*OffsetCommitRequest)
+	group := req.ConsumerGroup
+	res := &OffsetCommitResponse{}
+	for topic, partitions := range req.blocks {
+		for partition := range partitions {
+			res.AddError(topic, partition, mr.getError(group, topic, partition))
+		}
+	}
+	return res
+}
+
+func (mr *MockOffsetCommitResponse) getError(group, topic string, partition int32) KError {
+	topics := mr.errors[group]
+	if topics == nil {
+		return ErrNoError
+	}
+	partitions := topics[topic]
+	if partitions == nil {
+		return ErrNoError
+	}
+	kerror, ok := partitions[partition]
+	if !ok {
+		return ErrNoError
+	}
+	return kerror
+}
+
+// MockProduceResponse is a `ProduceResponse` builder.
+type MockProduceResponse struct {
+	version int16
+	errors  map[string]map[int32]KError
+	t       TestReporter
+}
+
+func NewMockProduceResponse(t TestReporter) *MockProduceResponse {
+	return &MockProduceResponse{t: t}
+}
+
+func (mr *MockProduceResponse) SetVersion(version int16) *MockProduceResponse {
+	mr.version = version
+	return mr
+}
+
+func (mr *MockProduceResponse) SetError(topic string, partition int32, kerror KError) *MockProduceResponse {
+	if mr.errors == nil {
+		mr.errors = make(map[string]map[int32]KError)
+	}
+	partitions := mr.errors[topic]
+	if partitions == nil {
+		partitions = make(map[int32]KError)
+		mr.errors[topic] = partitions
+	}
+	partitions[partition] = kerror
+	return mr
+}
+
+func (mr *MockProduceResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*ProduceRequest)
+	res := &ProduceResponse{
+		Version: mr.version,
+	}
+	for topic, partitions := range req.records {
+		for partition := range partitions {
+			res.AddTopicPartition(topic, partition, mr.getError(topic, partition))
+		}
+	}
+	return res
+}
+
+func (mr *MockProduceResponse) getError(topic string, partition int32) KError {
+	partitions := mr.errors[topic]
+	if partitions == nil {
+		return ErrNoError
+	}
+	kerror, ok := partitions[partition]
+	if !ok {
+		return ErrNoError
+	}
+	return kerror
+}
+
+// MockOffsetFetchResponse is a `OffsetFetchResponse` builder.
+type MockOffsetFetchResponse struct {
+	offsets map[string]map[string]map[int32]*OffsetFetchResponseBlock
+	t       TestReporter
+}
+
+func NewMockOffsetFetchResponse(t TestReporter) *MockOffsetFetchResponse {
+	return &MockOffsetFetchResponse{t: t}
+}
+
+func (mr *MockOffsetFetchResponse) SetOffset(group, topic string, partition int32, offset int64, metadata string, kerror KError) *MockOffsetFetchResponse {
+	if mr.offsets == nil {
+		mr.offsets = make(map[string]map[string]map[int32]*OffsetFetchResponseBlock)
+	}
+	topics := mr.offsets[group]
+	if topics == nil {
+		topics = make(map[string]map[int32]*OffsetFetchResponseBlock)
+		mr.offsets[group] = topics
+	}
+	partitions := topics[topic]
+	if partitions == nil {
+		partitions = make(map[int32]*OffsetFetchResponseBlock)
+		topics[topic] = partitions
+	}
+	partitions[partition] = &OffsetFetchResponseBlock{offset, 0, metadata, kerror}
+	return mr
+}
+
+func (mr *MockOffsetFetchResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*OffsetFetchRequest)
+	group := req.ConsumerGroup
+	res := &OffsetFetchResponse{}
+	for topic, partitions := range mr.offsets[group] {
+		for partition, block := range partitions {
+			res.AddBlock(topic, partition, block)
+		}
+	}
+	return res
+}
+
+type MockCreateTopicsResponse struct {
+	t TestReporter
+}
+
+func NewMockCreateTopicsResponse(t TestReporter) *MockCreateTopicsResponse {
+	return &MockCreateTopicsResponse{t: t}
+}
+
+func (mr *MockCreateTopicsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*CreateTopicsRequest)
+	res := &CreateTopicsResponse{}
+	res.TopicErrors = make(map[string]*TopicError)
+
+	for topic, _ := range req.TopicDetails {
+		res.TopicErrors[topic] = &TopicError{Err: ErrNoError}
+	}
+	return res
+}
+
+type MockDeleteTopicsResponse struct {
+	t TestReporter
+}
+
+func NewMockDeleteTopicsResponse(t TestReporter) *MockDeleteTopicsResponse {
+	return &MockDeleteTopicsResponse{t: t}
+}
+
+func (mr *MockDeleteTopicsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*DeleteTopicsRequest)
+	res := &DeleteTopicsResponse{}
+	res.TopicErrorCodes = make(map[string]KError)
+
+	for _, topic := range req.Topics {
+		res.TopicErrorCodes[topic] = ErrNoError
+	}
+	return res
+}
+
+type MockCreatePartitionsResponse struct {
+	t TestReporter
+}
+
+func NewMockCreatePartitionsResponse(t TestReporter) *MockCreatePartitionsResponse {
+	return &MockCreatePartitionsResponse{t: t}
+}
+
+func (mr *MockCreatePartitionsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*CreatePartitionsRequest)
+	res := &CreatePartitionsResponse{}
+	res.TopicPartitionErrors = make(map[string]*TopicPartitionError)
+
+	for topic, _ := range req.TopicPartitions {
+		res.TopicPartitionErrors[topic] = &TopicPartitionError{Err: ErrNoError}
+	}
+	return res
+}
+
+type MockDeleteRecordsResponse struct {
+	t TestReporter
+}
+
+func NewMockDeleteRecordsResponse(t TestReporter) *MockDeleteRecordsResponse {
+	return &MockDeleteRecordsResponse{t: t}
+}
+
+func (mr *MockDeleteRecordsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*DeleteRecordsRequest)
+	res := &DeleteRecordsResponse{}
+	res.Topics = make(map[string]*DeleteRecordsResponseTopic)
+
+	for topic, deleteRecordRequestTopic := range req.Topics {
+		partitions := make(map[int32]*DeleteRecordsResponsePartition)
+		for partition, _ := range deleteRecordRequestTopic.PartitionOffsets {
+			partitions[partition] = &DeleteRecordsResponsePartition{Err: ErrNoError}
+		}
+		res.Topics[topic] = &DeleteRecordsResponseTopic{Partitions: partitions}
+	}
+	return res
+}
+
+type MockDescribeConfigsResponse struct {
+	t TestReporter
+}
+
+func NewMockDescribeConfigsResponse(t TestReporter) *MockDescribeConfigsResponse {
+	return &MockDescribeConfigsResponse{t: t}
+}
+
+func (mr *MockDescribeConfigsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*DescribeConfigsRequest)
+	res := &DescribeConfigsResponse{}
+
+	for _, r := range req.Resources {
+		var configEntries []*ConfigEntry
+		switch r.Type {
+		case TopicResource:
+			configEntries = append(configEntries,
+				&ConfigEntry{Name: "max.message.bytes",
+					Value:     "1000000",
+					ReadOnly:  false,
+					Default:   true,
+					Sensitive: false,
+				}, &ConfigEntry{Name: "retention.ms",
+					Value:     "5000",
+					ReadOnly:  false,
+					Default:   false,
+					Sensitive: false,
+				}, &ConfigEntry{Name: "password",
+					Value:     "12345",
+					ReadOnly:  false,
+					Default:   false,
+					Sensitive: true,
+				})
+			res.Resources = append(res.Resources, &ResourceResponse{
+				Name:    r.Name,
+				Configs: configEntries,
+			})
+		}
+	}
+	return res
+}
+
+type MockAlterConfigsResponse struct {
+	t TestReporter
+}
+
+func NewMockAlterConfigsResponse(t TestReporter) *MockAlterConfigsResponse {
+	return &MockAlterConfigsResponse{t: t}
+}
+
+func (mr *MockAlterConfigsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*AlterConfigsRequest)
+	res := &AlterConfigsResponse{}
+
+	for _, r := range req.Resources {
+		res.Resources = append(res.Resources, &AlterConfigsResourceResponse{Name: r.Name,
+			Type:     TopicResource,
+			ErrorMsg: "",
+		})
+	}
+	return res
+}
+
+type MockCreateAclsResponse struct {
+	t TestReporter
+}
+
+func NewMockCreateAclsResponse(t TestReporter) *MockCreateAclsResponse {
+	return &MockCreateAclsResponse{t: t}
+}
+
+func (mr *MockCreateAclsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*CreateAclsRequest)
+	res := &CreateAclsResponse{}
+
+	for range req.AclCreations {
+		res.AclCreationResponses = append(res.AclCreationResponses, &AclCreationResponse{Err: ErrNoError})
+	}
+	return res
+}
+
+type MockListAclsResponse struct {
+	t TestReporter
+}
+
+func NewMockListAclsResponse(t TestReporter) *MockListAclsResponse {
+	return &MockListAclsResponse{t: t}
+}
+
+func (mr *MockListAclsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*DescribeAclsRequest)
+	res := &DescribeAclsResponse{}
+
+	res.Err = ErrNoError
+	acl := &ResourceAcls{}
+	acl.Resource.ResourceName = *req.ResourceName
+	acl.Resource.ResourceType = req.ResourceType
+	acl.Acls = append(acl.Acls, &Acl{})
+	res.ResourceAcls = append(res.ResourceAcls, acl)
+
+	return res
+}
+
+type MockSaslAuthenticateResponse struct {
+	t             TestReporter
+	kerror        KError
+	saslAuthBytes []byte
+}
+
+func NewMockSaslAuthenticateResponse(t TestReporter) *MockSaslAuthenticateResponse {
+	return &MockSaslAuthenticateResponse{t: t}
+}
+
+func (msar *MockSaslAuthenticateResponse) For(reqBody versionedDecoder) encoder {
+	res := &SaslAuthenticateResponse{}
+	res.Err = msar.kerror
+	res.SaslAuthBytes = msar.saslAuthBytes
+	return res
+}
+
+func (msar *MockSaslAuthenticateResponse) SetError(kerror KError) *MockSaslAuthenticateResponse {
+	msar.kerror = kerror
+	return msar
+}
+
+func (msar *MockSaslAuthenticateResponse) SetAuthBytes(saslAuthBytes []byte) *MockSaslAuthenticateResponse {
+	msar.saslAuthBytes = saslAuthBytes
+	return msar
+}
+
+type MockDeleteAclsResponse struct {
+	t TestReporter
+}
+
+type MockSaslHandshakeResponse struct {
+	enabledMechanisms []string
+	kerror            KError
+	t                 TestReporter
+}
+
+func NewMockSaslHandshakeResponse(t TestReporter) *MockSaslHandshakeResponse {
+	return &MockSaslHandshakeResponse{t: t}
+}
+
+func (mshr *MockSaslHandshakeResponse) For(reqBody versionedDecoder) encoder {
+	res := &SaslHandshakeResponse{}
+	res.Err = mshr.kerror
+	res.EnabledMechanisms = mshr.enabledMechanisms
+	return res
+}
+
+func (mshr *MockSaslHandshakeResponse) SetError(kerror KError) *MockSaslHandshakeResponse {
+	mshr.kerror = kerror
+	return mshr
+}
+
+func (mshr *MockSaslHandshakeResponse) SetEnabledMechanisms(enabledMechanisms []string) *MockSaslHandshakeResponse {
+	mshr.enabledMechanisms = enabledMechanisms
+	return mshr
+}
+
+func NewMockDeleteAclsResponse(t TestReporter) *MockDeleteAclsResponse {
+	return &MockDeleteAclsResponse{t: t}
+}
+
+func (mr *MockDeleteAclsResponse) For(reqBody versionedDecoder) encoder {
+	req := reqBody.(*DeleteAclsRequest)
+	res := &DeleteAclsResponse{}
+
+	for range req.Filters {
+		response := &FilterResponse{Err: ErrNoError}
+		response.MatchingAcls = append(response.MatchingAcls, &MatchingAcl{Err: ErrNoError})
+		res.FilterResponses = append(res.FilterResponses, response)
+	}
+	return res
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_commit_request.go b/vendor/github.com/Shopify/sarama/offset_commit_request.go
new file mode 100644
index 0000000..1ec583e
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_commit_request.go
@@ -0,0 +1,210 @@
+package sarama
+
+import "errors"
+
+// ReceiveTime is a special value for the timestamp field of Offset Commit Requests which
+// tells the broker to set the timestamp to the time at which the request was received.
+// The timestamp is only used if message version 1 is used, which requires kafka 0.8.2.
+const ReceiveTime int64 = -1
+
+// GroupGenerationUndefined is a special value for the group generation field of
+// Offset Commit Requests that should be used when a consumer group does not rely
+// on Kafka for partition management.
+const GroupGenerationUndefined = -1
+
+type offsetCommitRequestBlock struct {
+	offset    int64
+	timestamp int64
+	metadata  string
+}
+
+func (b *offsetCommitRequestBlock) encode(pe packetEncoder, version int16) error {
+	pe.putInt64(b.offset)
+	if version == 1 {
+		pe.putInt64(b.timestamp)
+	} else if b.timestamp != 0 {
+		Logger.Println("Non-zero timestamp specified for OffsetCommitRequest not v1, it will be ignored")
+	}
+
+	return pe.putString(b.metadata)
+}
+
+func (b *offsetCommitRequestBlock) decode(pd packetDecoder, version int16) (err error) {
+	if b.offset, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if version == 1 {
+		if b.timestamp, err = pd.getInt64(); err != nil {
+			return err
+		}
+	}
+	b.metadata, err = pd.getString()
+	return err
+}
+
+type OffsetCommitRequest struct {
+	ConsumerGroup           string
+	ConsumerGroupGeneration int32  // v1 or later
+	ConsumerID              string // v1 or later
+	RetentionTime           int64  // v2 or later
+
+	// Version can be:
+	// - 0 (kafka 0.8.1 and later)
+	// - 1 (kafka 0.8.2 and later)
+	// - 2 (kafka 0.9.0 and later)
+	// - 3 (kafka 0.11.0 and later)
+	// - 4 (kafka 2.0.0 and later)
+	Version int16
+	blocks  map[string]map[int32]*offsetCommitRequestBlock
+}
+
+func (r *OffsetCommitRequest) encode(pe packetEncoder) error {
+	if r.Version < 0 || r.Version > 4 {
+		return PacketEncodingError{"invalid or unsupported OffsetCommitRequest version field"}
+	}
+
+	if err := pe.putString(r.ConsumerGroup); err != nil {
+		return err
+	}
+
+	if r.Version >= 1 {
+		pe.putInt32(r.ConsumerGroupGeneration)
+		if err := pe.putString(r.ConsumerID); err != nil {
+			return err
+		}
+	} else {
+		if r.ConsumerGroupGeneration != 0 {
+			Logger.Println("Non-zero ConsumerGroupGeneration specified for OffsetCommitRequest v0, it will be ignored")
+		}
+		if r.ConsumerID != "" {
+			Logger.Println("Non-empty ConsumerID specified for OffsetCommitRequest v0, it will be ignored")
+		}
+	}
+
+	if r.Version >= 2 {
+		pe.putInt64(r.RetentionTime)
+	} else if r.RetentionTime != 0 {
+		Logger.Println("Non-zero RetentionTime specified for OffsetCommitRequest version <2, it will be ignored")
+	}
+
+	if err := pe.putArrayLength(len(r.blocks)); err != nil {
+		return err
+	}
+	for topic, partitions := range r.blocks {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(partitions)); err != nil {
+			return err
+		}
+		for partition, block := range partitions {
+			pe.putInt32(partition)
+			if err := block.encode(pe, r.Version); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (r *OffsetCommitRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if r.ConsumerGroup, err = pd.getString(); err != nil {
+		return err
+	}
+
+	if r.Version >= 1 {
+		if r.ConsumerGroupGeneration, err = pd.getInt32(); err != nil {
+			return err
+		}
+		if r.ConsumerID, err = pd.getString(); err != nil {
+			return err
+		}
+	}
+
+	if r.Version >= 2 {
+		if r.RetentionTime, err = pd.getInt64(); err != nil {
+			return err
+		}
+	}
+
+	topicCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if topicCount == 0 {
+		return nil
+	}
+	r.blocks = make(map[string]map[int32]*offsetCommitRequestBlock)
+	for i := 0; i < topicCount; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		partitionCount, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+		r.blocks[topic] = make(map[int32]*offsetCommitRequestBlock)
+		for j := 0; j < partitionCount; j++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			block := &offsetCommitRequestBlock{}
+			if err := block.decode(pd, r.Version); err != nil {
+				return err
+			}
+			r.blocks[topic][partition] = block
+		}
+	}
+	return nil
+}
+
+func (r *OffsetCommitRequest) key() int16 {
+	return 8
+}
+
+func (r *OffsetCommitRequest) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetCommitRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_8_2_0
+	case 2:
+		return V0_9_0_0
+	case 3:
+		return V0_11_0_0
+	case 4:
+		return V2_0_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *OffsetCommitRequest) AddBlock(topic string, partitionID int32, offset int64, timestamp int64, metadata string) {
+	if r.blocks == nil {
+		r.blocks = make(map[string]map[int32]*offsetCommitRequestBlock)
+	}
+
+	if r.blocks[topic] == nil {
+		r.blocks[topic] = make(map[int32]*offsetCommitRequestBlock)
+	}
+
+	r.blocks[topic][partitionID] = &offsetCommitRequestBlock{offset, timestamp, metadata}
+}
+
+func (r *OffsetCommitRequest) Offset(topic string, partitionID int32) (int64, string, error) {
+	partitions := r.blocks[topic]
+	if partitions == nil {
+		return 0, "", errors.New("No such offset")
+	}
+	block := partitions[partitionID]
+	if block == nil {
+		return 0, "", errors.New("No such offset")
+	}
+	return block.offset, block.metadata, nil
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_commit_response.go b/vendor/github.com/Shopify/sarama/offset_commit_response.go
new file mode 100644
index 0000000..e842298
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_commit_response.go
@@ -0,0 +1,110 @@
+package sarama
+
+type OffsetCommitResponse struct {
+	Version        int16
+	ThrottleTimeMs int32
+	Errors         map[string]map[int32]KError
+}
+
+func (r *OffsetCommitResponse) AddError(topic string, partition int32, kerror KError) {
+	if r.Errors == nil {
+		r.Errors = make(map[string]map[int32]KError)
+	}
+	partitions := r.Errors[topic]
+	if partitions == nil {
+		partitions = make(map[int32]KError)
+		r.Errors[topic] = partitions
+	}
+	partitions[partition] = kerror
+}
+
+func (r *OffsetCommitResponse) encode(pe packetEncoder) error {
+	if r.Version >= 3 {
+		pe.putInt32(r.ThrottleTimeMs)
+	}
+	if err := pe.putArrayLength(len(r.Errors)); err != nil {
+		return err
+	}
+	for topic, partitions := range r.Errors {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(partitions)); err != nil {
+			return err
+		}
+		for partition, kerror := range partitions {
+			pe.putInt32(partition)
+			pe.putInt16(int16(kerror))
+		}
+	}
+	return nil
+}
+
+func (r *OffsetCommitResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if version >= 3 {
+		r.ThrottleTimeMs, err = pd.getInt32()
+		if err != nil {
+			return err
+		}
+	}
+
+	numTopics, err := pd.getArrayLength()
+	if err != nil || numTopics == 0 {
+		return err
+	}
+
+	r.Errors = make(map[string]map[int32]KError, numTopics)
+	for i := 0; i < numTopics; i++ {
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		numErrors, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		r.Errors[name] = make(map[int32]KError, numErrors)
+
+		for j := 0; j < numErrors; j++ {
+			id, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+
+			tmp, err := pd.getInt16()
+			if err != nil {
+				return err
+			}
+			r.Errors[name][id] = KError(tmp)
+		}
+	}
+
+	return nil
+}
+
+func (r *OffsetCommitResponse) key() int16 {
+	return 8
+}
+
+func (r *OffsetCommitResponse) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetCommitResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_8_2_0
+	case 2:
+		return V0_9_0_0
+	case 3:
+		return V0_11_0_0
+	case 4:
+		return V2_0_0_0
+	default:
+		return MinVersion
+	}
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_fetch_request.go b/vendor/github.com/Shopify/sarama/offset_fetch_request.go
new file mode 100644
index 0000000..6860824
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_fetch_request.go
@@ -0,0 +1,100 @@
+package sarama
+
+type OffsetFetchRequest struct {
+	Version       int16
+	ConsumerGroup string
+	partitions    map[string][]int32
+}
+
+func (r *OffsetFetchRequest) encode(pe packetEncoder) (err error) {
+	if r.Version < 0 || r.Version > 5 {
+		return PacketEncodingError{"invalid or unsupported OffsetFetchRequest version field"}
+	}
+
+	if err = pe.putString(r.ConsumerGroup); err != nil {
+		return err
+	}
+
+	if r.Version >= 2 && r.partitions == nil {
+		pe.putInt32(-1)
+	} else {
+		if err = pe.putArrayLength(len(r.partitions)); err != nil {
+			return err
+		}
+		for topic, partitions := range r.partitions {
+			if err = pe.putString(topic); err != nil {
+				return err
+			}
+			if err = pe.putInt32Array(partitions); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (r *OffsetFetchRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+	if r.ConsumerGroup, err = pd.getString(); err != nil {
+		return err
+	}
+	partitionCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if (partitionCount == 0 && version < 2) || partitionCount < 0 {
+		return nil
+	}
+	r.partitions = make(map[string][]int32)
+	for i := 0; i < partitionCount; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		partitions, err := pd.getInt32Array()
+		if err != nil {
+			return err
+		}
+		r.partitions[topic] = partitions
+	}
+	return nil
+}
+
+func (r *OffsetFetchRequest) key() int16 {
+	return 9
+}
+
+func (r *OffsetFetchRequest) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetFetchRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_8_2_0
+	case 2:
+		return V0_10_2_0
+	case 3:
+		return V0_11_0_0
+	case 4:
+		return V2_0_0_0
+	case 5:
+		return V2_1_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *OffsetFetchRequest) ZeroPartitions() {
+	if r.partitions == nil && r.Version >= 2 {
+		r.partitions = make(map[string][]int32)
+	}
+}
+
+func (r *OffsetFetchRequest) AddPartition(topic string, partitionID int32) {
+	if r.partitions == nil {
+		r.partitions = make(map[string][]int32)
+	}
+
+	r.partitions[topic] = append(r.partitions[topic], partitionID)
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_fetch_response.go b/vendor/github.com/Shopify/sarama/offset_fetch_response.go
new file mode 100644
index 0000000..9e25702
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_fetch_response.go
@@ -0,0 +1,197 @@
+package sarama
+
+type OffsetFetchResponseBlock struct {
+	Offset      int64
+	LeaderEpoch int32
+	Metadata    string
+	Err         KError
+}
+
+func (b *OffsetFetchResponseBlock) decode(pd packetDecoder, version int16) (err error) {
+	b.Offset, err = pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	if version >= 5 {
+		b.LeaderEpoch, err = pd.getInt32()
+		if err != nil {
+			return err
+		}
+	}
+
+	b.Metadata, err = pd.getString()
+	if err != nil {
+		return err
+	}
+
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	b.Err = KError(tmp)
+
+	return nil
+}
+
+func (b *OffsetFetchResponseBlock) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt64(b.Offset)
+
+	if version >= 5 {
+		pe.putInt32(b.LeaderEpoch)
+	}
+
+	err = pe.putString(b.Metadata)
+	if err != nil {
+		return err
+	}
+
+	pe.putInt16(int16(b.Err))
+
+	return nil
+}
+
+type OffsetFetchResponse struct {
+	Version        int16
+	ThrottleTimeMs int32
+	Blocks         map[string]map[int32]*OffsetFetchResponseBlock
+	Err            KError
+}
+
+func (r *OffsetFetchResponse) encode(pe packetEncoder) error {
+	if r.Version >= 3 {
+		pe.putInt32(r.ThrottleTimeMs)
+	}
+
+	if err := pe.putArrayLength(len(r.Blocks)); err != nil {
+		return err
+	}
+	for topic, partitions := range r.Blocks {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(partitions)); err != nil {
+			return err
+		}
+		for partition, block := range partitions {
+			pe.putInt32(partition)
+			if err := block.encode(pe, r.Version); err != nil {
+				return err
+			}
+		}
+	}
+	if r.Version >= 2 {
+		pe.putInt16(int16(r.Err))
+	}
+	return nil
+}
+
+func (r *OffsetFetchResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	if version >= 3 {
+		r.ThrottleTimeMs, err = pd.getInt32()
+		if err != nil {
+			return err
+		}
+	}
+
+	numTopics, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	if numTopics > 0 {
+		r.Blocks = make(map[string]map[int32]*OffsetFetchResponseBlock, numTopics)
+		for i := 0; i < numTopics; i++ {
+			name, err := pd.getString()
+			if err != nil {
+				return err
+			}
+
+			numBlocks, err := pd.getArrayLength()
+			if err != nil {
+				return err
+			}
+
+			if numBlocks == 0 {
+				r.Blocks[name] = nil
+				continue
+			}
+			r.Blocks[name] = make(map[int32]*OffsetFetchResponseBlock, numBlocks)
+
+			for j := 0; j < numBlocks; j++ {
+				id, err := pd.getInt32()
+				if err != nil {
+					return err
+				}
+
+				block := new(OffsetFetchResponseBlock)
+				err = block.decode(pd, version)
+				if err != nil {
+					return err
+				}
+				r.Blocks[name][id] = block
+			}
+		}
+	}
+
+	if version >= 2 {
+		kerr, err := pd.getInt16()
+		if err != nil {
+			return err
+		}
+		r.Err = KError(kerr)
+	}
+
+	return nil
+}
+
+func (r *OffsetFetchResponse) key() int16 {
+	return 9
+}
+
+func (r *OffsetFetchResponse) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetFetchResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_8_2_0
+	case 2:
+		return V0_10_2_0
+	case 3:
+		return V0_11_0_0
+	case 4:
+		return V2_0_0_0
+	case 5:
+		return V2_1_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *OffsetFetchResponse) GetBlock(topic string, partition int32) *OffsetFetchResponseBlock {
+	if r.Blocks == nil {
+		return nil
+	}
+
+	if r.Blocks[topic] == nil {
+		return nil
+	}
+
+	return r.Blocks[topic][partition]
+}
+
+func (r *OffsetFetchResponse) AddBlock(topic string, partition int32, block *OffsetFetchResponseBlock) {
+	if r.Blocks == nil {
+		r.Blocks = make(map[string]map[int32]*OffsetFetchResponseBlock)
+	}
+	partitions := r.Blocks[topic]
+	if partitions == nil {
+		partitions = make(map[int32]*OffsetFetchResponseBlock)
+		r.Blocks[topic] = partitions
+	}
+	partitions[partition] = block
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_manager.go b/vendor/github.com/Shopify/sarama/offset_manager.go
new file mode 100644
index 0000000..2432f7b
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_manager.go
@@ -0,0 +1,581 @@
+package sarama
+
+import (
+	"sync"
+	"time"
+)
+
+// Offset Manager
+
+// OffsetManager uses Kafka to store and fetch consumed partition offsets.
+type OffsetManager interface {
+	// ManagePartition creates a PartitionOffsetManager on the given topic/partition.
+	// It will return an error if this OffsetManager is already managing the given
+	// topic/partition.
+	ManagePartition(topic string, partition int32) (PartitionOffsetManager, error)
+
+	// Close stops the OffsetManager from managing offsets. It is required to call
+	// this function before an OffsetManager object passes out of scope, as it
+	// will otherwise leak memory. You must call this after all the
+	// PartitionOffsetManagers are closed.
+	Close() error
+}
+
+type offsetManager struct {
+	client Client
+	conf   *Config
+	group  string
+	ticker *time.Ticker
+
+	memberID   string
+	generation int32
+
+	broker     *Broker
+	brokerLock sync.RWMutex
+
+	poms     map[string]map[int32]*partitionOffsetManager
+	pomsLock sync.RWMutex
+
+	closeOnce sync.Once
+	closing   chan none
+	closed    chan none
+}
+
+// NewOffsetManagerFromClient creates a new OffsetManager from the given client.
+// It is still necessary to call Close() on the underlying client when finished with the partition manager.
+func NewOffsetManagerFromClient(group string, client Client) (OffsetManager, error) {
+	return newOffsetManagerFromClient(group, "", GroupGenerationUndefined, client)
+}
+
+func newOffsetManagerFromClient(group, memberID string, generation int32, client Client) (*offsetManager, error) {
+	// Check that we are not dealing with a closed Client before processing any other arguments
+	if client.Closed() {
+		return nil, ErrClosedClient
+	}
+
+	conf := client.Config()
+	om := &offsetManager{
+		client: client,
+		conf:   conf,
+		group:  group,
+		ticker: time.NewTicker(conf.Consumer.Offsets.CommitInterval),
+		poms:   make(map[string]map[int32]*partitionOffsetManager),
+
+		memberID:   memberID,
+		generation: generation,
+
+		closing: make(chan none),
+		closed:  make(chan none),
+	}
+	go withRecover(om.mainLoop)
+
+	return om, nil
+}
+
+func (om *offsetManager) ManagePartition(topic string, partition int32) (PartitionOffsetManager, error) {
+	pom, err := om.newPartitionOffsetManager(topic, partition)
+	if err != nil {
+		return nil, err
+	}
+
+	om.pomsLock.Lock()
+	defer om.pomsLock.Unlock()
+
+	topicManagers := om.poms[topic]
+	if topicManagers == nil {
+		topicManagers = make(map[int32]*partitionOffsetManager)
+		om.poms[topic] = topicManagers
+	}
+
+	if topicManagers[partition] != nil {
+		return nil, ConfigurationError("That topic/partition is already being managed")
+	}
+
+	topicManagers[partition] = pom
+	return pom, nil
+}
+
+func (om *offsetManager) Close() error {
+	om.closeOnce.Do(func() {
+		// exit the mainLoop
+		close(om.closing)
+		<-om.closed
+
+		// mark all POMs as closed
+		om.asyncClosePOMs()
+
+		// flush one last time
+		for attempt := 0; attempt <= om.conf.Consumer.Offsets.Retry.Max; attempt++ {
+			om.flushToBroker()
+			if om.releasePOMs(false) == 0 {
+				break
+			}
+		}
+
+		om.releasePOMs(true)
+		om.brokerLock.Lock()
+		om.broker = nil
+		om.brokerLock.Unlock()
+	})
+	return nil
+}
+
+func (om *offsetManager) computeBackoff(retries int) time.Duration {
+	if om.conf.Metadata.Retry.BackoffFunc != nil {
+		return om.conf.Metadata.Retry.BackoffFunc(retries, om.conf.Metadata.Retry.Max)
+	} else {
+		return om.conf.Metadata.Retry.Backoff
+	}
+}
+
+func (om *offsetManager) fetchInitialOffset(topic string, partition int32, retries int) (int64, string, error) {
+	broker, err := om.coordinator()
+	if err != nil {
+		if retries <= 0 {
+			return 0, "", err
+		}
+		return om.fetchInitialOffset(topic, partition, retries-1)
+	}
+
+	req := new(OffsetFetchRequest)
+	req.Version = 1
+	req.ConsumerGroup = om.group
+	req.AddPartition(topic, partition)
+
+	resp, err := broker.FetchOffset(req)
+	if err != nil {
+		if retries <= 0 {
+			return 0, "", err
+		}
+		om.releaseCoordinator(broker)
+		return om.fetchInitialOffset(topic, partition, retries-1)
+	}
+
+	block := resp.GetBlock(topic, partition)
+	if block == nil {
+		return 0, "", ErrIncompleteResponse
+	}
+
+	switch block.Err {
+	case ErrNoError:
+		return block.Offset, block.Metadata, nil
+	case ErrNotCoordinatorForConsumer:
+		if retries <= 0 {
+			return 0, "", block.Err
+		}
+		om.releaseCoordinator(broker)
+		return om.fetchInitialOffset(topic, partition, retries-1)
+	case ErrOffsetsLoadInProgress:
+		if retries <= 0 {
+			return 0, "", block.Err
+		}
+		backoff := om.computeBackoff(retries)
+		select {
+		case <-om.closing:
+			return 0, "", block.Err
+		case <-time.After(backoff):
+		}
+		return om.fetchInitialOffset(topic, partition, retries-1)
+	default:
+		return 0, "", block.Err
+	}
+}
+
+func (om *offsetManager) coordinator() (*Broker, error) {
+	om.brokerLock.RLock()
+	broker := om.broker
+	om.brokerLock.RUnlock()
+
+	if broker != nil {
+		return broker, nil
+	}
+
+	om.brokerLock.Lock()
+	defer om.brokerLock.Unlock()
+
+	if broker := om.broker; broker != nil {
+		return broker, nil
+	}
+
+	if err := om.client.RefreshCoordinator(om.group); err != nil {
+		return nil, err
+	}
+
+	broker, err := om.client.Coordinator(om.group)
+	if err != nil {
+		return nil, err
+	}
+
+	om.broker = broker
+	return broker, nil
+}
+
+func (om *offsetManager) releaseCoordinator(b *Broker) {
+	om.brokerLock.Lock()
+	if om.broker == b {
+		om.broker = nil
+	}
+	om.brokerLock.Unlock()
+}
+
+func (om *offsetManager) mainLoop() {
+	defer om.ticker.Stop()
+	defer close(om.closed)
+
+	for {
+		select {
+		case <-om.ticker.C:
+			om.flushToBroker()
+			om.releasePOMs(false)
+		case <-om.closing:
+			return
+		}
+	}
+}
+
+func (om *offsetManager) flushToBroker() {
+	req := om.constructRequest()
+	if req == nil {
+		return
+	}
+
+	broker, err := om.coordinator()
+	if err != nil {
+		om.handleError(err)
+		return
+	}
+
+	resp, err := broker.CommitOffset(req)
+	if err != nil {
+		om.handleError(err)
+		om.releaseCoordinator(broker)
+		_ = broker.Close()
+		return
+	}
+
+	om.handleResponse(broker, req, resp)
+}
+
+func (om *offsetManager) constructRequest() *OffsetCommitRequest {
+	var r *OffsetCommitRequest
+	var perPartitionTimestamp int64
+	if om.conf.Consumer.Offsets.Retention == 0 {
+		perPartitionTimestamp = ReceiveTime
+		r = &OffsetCommitRequest{
+			Version:                 1,
+			ConsumerGroup:           om.group,
+			ConsumerID:              om.memberID,
+			ConsumerGroupGeneration: om.generation,
+		}
+	} else {
+		r = &OffsetCommitRequest{
+			Version:                 2,
+			RetentionTime:           int64(om.conf.Consumer.Offsets.Retention / time.Millisecond),
+			ConsumerGroup:           om.group,
+			ConsumerID:              om.memberID,
+			ConsumerGroupGeneration: om.generation,
+		}
+
+	}
+
+	om.pomsLock.RLock()
+	defer om.pomsLock.RUnlock()
+
+	for _, topicManagers := range om.poms {
+		for _, pom := range topicManagers {
+			pom.lock.Lock()
+			if pom.dirty {
+				r.AddBlock(pom.topic, pom.partition, pom.offset, perPartitionTimestamp, pom.metadata)
+			}
+			pom.lock.Unlock()
+		}
+	}
+
+	if len(r.blocks) > 0 {
+		return r
+	}
+
+	return nil
+}
+
+func (om *offsetManager) handleResponse(broker *Broker, req *OffsetCommitRequest, resp *OffsetCommitResponse) {
+	om.pomsLock.RLock()
+	defer om.pomsLock.RUnlock()
+
+	for _, topicManagers := range om.poms {
+		for _, pom := range topicManagers {
+			if req.blocks[pom.topic] == nil || req.blocks[pom.topic][pom.partition] == nil {
+				continue
+			}
+
+			var err KError
+			var ok bool
+
+			if resp.Errors[pom.topic] == nil {
+				pom.handleError(ErrIncompleteResponse)
+				continue
+			}
+			if err, ok = resp.Errors[pom.topic][pom.partition]; !ok {
+				pom.handleError(ErrIncompleteResponse)
+				continue
+			}
+
+			switch err {
+			case ErrNoError:
+				block := req.blocks[pom.topic][pom.partition]
+				pom.updateCommitted(block.offset, block.metadata)
+			case ErrNotLeaderForPartition, ErrLeaderNotAvailable,
+				ErrConsumerCoordinatorNotAvailable, ErrNotCoordinatorForConsumer:
+				// not a critical error, we just need to redispatch
+				om.releaseCoordinator(broker)
+			case ErrOffsetMetadataTooLarge, ErrInvalidCommitOffsetSize:
+				// nothing we can do about this, just tell the user and carry on
+				pom.handleError(err)
+			case ErrOffsetsLoadInProgress:
+				// nothing wrong but we didn't commit, we'll get it next time round
+				break
+			case ErrUnknownTopicOrPartition:
+				// let the user know *and* try redispatching - if topic-auto-create is
+				// enabled, redispatching should trigger a metadata req and create the
+				// topic; if not then re-dispatching won't help, but we've let the user
+				// know and it shouldn't hurt either (see https://github.com/Shopify/sarama/issues/706)
+				fallthrough
+			default:
+				// dunno, tell the user and try redispatching
+				pom.handleError(err)
+				om.releaseCoordinator(broker)
+			}
+		}
+	}
+}
+
+func (om *offsetManager) handleError(err error) {
+	om.pomsLock.RLock()
+	defer om.pomsLock.RUnlock()
+
+	for _, topicManagers := range om.poms {
+		for _, pom := range topicManagers {
+			pom.handleError(err)
+		}
+	}
+}
+
+func (om *offsetManager) asyncClosePOMs() {
+	om.pomsLock.RLock()
+	defer om.pomsLock.RUnlock()
+
+	for _, topicManagers := range om.poms {
+		for _, pom := range topicManagers {
+			pom.AsyncClose()
+		}
+	}
+}
+
+// Releases/removes closed POMs once they are clean (or when forced)
+func (om *offsetManager) releasePOMs(force bool) (remaining int) {
+	om.pomsLock.Lock()
+	defer om.pomsLock.Unlock()
+
+	for topic, topicManagers := range om.poms {
+		for partition, pom := range topicManagers {
+			pom.lock.Lock()
+			releaseDue := pom.done && (force || !pom.dirty)
+			pom.lock.Unlock()
+
+			if releaseDue {
+				pom.release()
+
+				delete(om.poms[topic], partition)
+				if len(om.poms[topic]) == 0 {
+					delete(om.poms, topic)
+				}
+			}
+		}
+		remaining += len(om.poms[topic])
+	}
+	return
+}
+
+func (om *offsetManager) findPOM(topic string, partition int32) *partitionOffsetManager {
+	om.pomsLock.RLock()
+	defer om.pomsLock.RUnlock()
+
+	if partitions, ok := om.poms[topic]; ok {
+		if pom, ok := partitions[partition]; ok {
+			return pom
+		}
+	}
+	return nil
+}
+
+// Partition Offset Manager
+
+// PartitionOffsetManager uses Kafka to store and fetch consumed partition offsets. You MUST call Close()
+// on a partition offset manager to avoid leaks, it will not be garbage-collected automatically when it passes
+// out of scope.
+type PartitionOffsetManager interface {
+	// NextOffset returns the next offset that should be consumed for the managed
+	// partition, accompanied by metadata which can be used to reconstruct the state
+	// of the partition consumer when it resumes. NextOffset() will return
+	// `config.Consumer.Offsets.Initial` and an empty metadata string if no offset
+	// was committed for this partition yet.
+	NextOffset() (int64, string)
+
+	// MarkOffset marks the provided offset, alongside a metadata string
+	// that represents the state of the partition consumer at that point in time. The
+	// metadata string can be used by another consumer to restore that state, so it
+	// can resume consumption.
+	//
+	// To follow upstream conventions, you are expected to mark the offset of the
+	// next message to read, not the last message read. Thus, when calling `MarkOffset`
+	// you should typically add one to the offset of the last consumed message.
+	//
+	// Note: calling MarkOffset does not necessarily commit the offset to the backend
+	// store immediately for efficiency reasons, and it may never be committed if
+	// your application crashes. This means that you may end up processing the same
+	// message twice, and your processing should ideally be idempotent.
+	MarkOffset(offset int64, metadata string)
+
+	// ResetOffset resets to the provided offset, alongside a metadata string that
+	// represents the state of the partition consumer at that point in time. Reset
+	// acts as a counterpart to MarkOffset, the difference being that it allows to
+	// reset an offset to an earlier or smaller value, where MarkOffset only
+	// allows incrementing the offset. cf MarkOffset for more details.
+	ResetOffset(offset int64, metadata string)
+
+	// Errors returns a read channel of errors that occur during offset management, if
+	// enabled. By default, errors are logged and not returned over this channel. If
+	// you want to implement any custom error handling, set your config's
+	// Consumer.Return.Errors setting to true, and read from this channel.
+	Errors() <-chan *ConsumerError
+
+	// AsyncClose initiates a shutdown of the PartitionOffsetManager. This method will
+	// return immediately, after which you should wait until the 'errors' channel has
+	// been drained and closed. It is required to call this function, or Close before
+	// a consumer object passes out of scope, as it will otherwise leak memory. You
+	// must call this before calling Close on the underlying client.
+	AsyncClose()
+
+	// Close stops the PartitionOffsetManager from managing offsets. It is required to
+	// call this function (or AsyncClose) before a PartitionOffsetManager object
+	// passes out of scope, as it will otherwise leak memory. You must call this
+	// before calling Close on the underlying client.
+	Close() error
+}
+
+type partitionOffsetManager struct {
+	parent    *offsetManager
+	topic     string
+	partition int32
+
+	lock     sync.Mutex
+	offset   int64
+	metadata string
+	dirty    bool
+	done     bool
+
+	releaseOnce sync.Once
+	errors      chan *ConsumerError
+}
+
+func (om *offsetManager) newPartitionOffsetManager(topic string, partition int32) (*partitionOffsetManager, error) {
+	offset, metadata, err := om.fetchInitialOffset(topic, partition, om.conf.Metadata.Retry.Max)
+	if err != nil {
+		return nil, err
+	}
+
+	return &partitionOffsetManager{
+		parent:    om,
+		topic:     topic,
+		partition: partition,
+		errors:    make(chan *ConsumerError, om.conf.ChannelBufferSize),
+		offset:    offset,
+		metadata:  metadata,
+	}, nil
+}
+
+func (pom *partitionOffsetManager) Errors() <-chan *ConsumerError {
+	return pom.errors
+}
+
+func (pom *partitionOffsetManager) MarkOffset(offset int64, metadata string) {
+	pom.lock.Lock()
+	defer pom.lock.Unlock()
+
+	if offset > pom.offset {
+		pom.offset = offset
+		pom.metadata = metadata
+		pom.dirty = true
+	}
+}
+
+func (pom *partitionOffsetManager) ResetOffset(offset int64, metadata string) {
+	pom.lock.Lock()
+	defer pom.lock.Unlock()
+
+	if offset <= pom.offset {
+		pom.offset = offset
+		pom.metadata = metadata
+		pom.dirty = true
+	}
+}
+
+func (pom *partitionOffsetManager) updateCommitted(offset int64, metadata string) {
+	pom.lock.Lock()
+	defer pom.lock.Unlock()
+
+	if pom.offset == offset && pom.metadata == metadata {
+		pom.dirty = false
+	}
+}
+
+func (pom *partitionOffsetManager) NextOffset() (int64, string) {
+	pom.lock.Lock()
+	defer pom.lock.Unlock()
+
+	if pom.offset >= 0 {
+		return pom.offset, pom.metadata
+	}
+
+	return pom.parent.conf.Consumer.Offsets.Initial, ""
+}
+
+func (pom *partitionOffsetManager) AsyncClose() {
+	pom.lock.Lock()
+	pom.done = true
+	pom.lock.Unlock()
+}
+
+func (pom *partitionOffsetManager) Close() error {
+	pom.AsyncClose()
+
+	var errors ConsumerErrors
+	for err := range pom.errors {
+		errors = append(errors, err)
+	}
+
+	if len(errors) > 0 {
+		return errors
+	}
+	return nil
+}
+
+func (pom *partitionOffsetManager) handleError(err error) {
+	cErr := &ConsumerError{
+		Topic:     pom.topic,
+		Partition: pom.partition,
+		Err:       err,
+	}
+
+	if pom.parent.conf.Consumer.Return.Errors {
+		pom.errors <- cErr
+	} else {
+		Logger.Println(cErr)
+	}
+}
+
+func (pom *partitionOffsetManager) release() {
+	pom.releaseOnce.Do(func() {
+		go close(pom.errors)
+	})
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_request.go b/vendor/github.com/Shopify/sarama/offset_request.go
new file mode 100644
index 0000000..326c372
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_request.go
@@ -0,0 +1,156 @@
+package sarama
+
+type offsetRequestBlock struct {
+	time       int64
+	maxOffsets int32 // Only used in version 0
+}
+
+func (b *offsetRequestBlock) encode(pe packetEncoder, version int16) error {
+	pe.putInt64(int64(b.time))
+	if version == 0 {
+		pe.putInt32(b.maxOffsets)
+	}
+
+	return nil
+}
+
+func (b *offsetRequestBlock) decode(pd packetDecoder, version int16) (err error) {
+	if b.time, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if version == 0 {
+		if b.maxOffsets, err = pd.getInt32(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+type OffsetRequest struct {
+	Version        int16
+	replicaID      int32
+	isReplicaIDSet bool
+	blocks         map[string]map[int32]*offsetRequestBlock
+}
+
+func (r *OffsetRequest) encode(pe packetEncoder) error {
+	if r.isReplicaIDSet {
+		pe.putInt32(r.replicaID)
+	} else {
+		// default replica ID is always -1 for clients
+		pe.putInt32(-1)
+	}
+
+	err := pe.putArrayLength(len(r.blocks))
+	if err != nil {
+		return err
+	}
+	for topic, partitions := range r.blocks {
+		err = pe.putString(topic)
+		if err != nil {
+			return err
+		}
+		err = pe.putArrayLength(len(partitions))
+		if err != nil {
+			return err
+		}
+		for partition, block := range partitions {
+			pe.putInt32(partition)
+			if err = block.encode(pe, r.Version); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (r *OffsetRequest) decode(pd packetDecoder, version int16) error {
+	r.Version = version
+
+	replicaID, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	if replicaID >= 0 {
+		r.SetReplicaID(replicaID)
+	}
+
+	blockCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if blockCount == 0 {
+		return nil
+	}
+	r.blocks = make(map[string]map[int32]*offsetRequestBlock)
+	for i := 0; i < blockCount; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		partitionCount, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+		r.blocks[topic] = make(map[int32]*offsetRequestBlock)
+		for j := 0; j < partitionCount; j++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			block := &offsetRequestBlock{}
+			if err := block.decode(pd, version); err != nil {
+				return err
+			}
+			r.blocks[topic][partition] = block
+		}
+	}
+	return nil
+}
+
+func (r *OffsetRequest) key() int16 {
+	return 2
+}
+
+func (r *OffsetRequest) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_10_1_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *OffsetRequest) SetReplicaID(id int32) {
+	r.replicaID = id
+	r.isReplicaIDSet = true
+}
+
+func (r *OffsetRequest) ReplicaID() int32 {
+	if r.isReplicaIDSet {
+		return r.replicaID
+	}
+	return -1
+}
+
+func (r *OffsetRequest) AddBlock(topic string, partitionID int32, time int64, maxOffsets int32) {
+	if r.blocks == nil {
+		r.blocks = make(map[string]map[int32]*offsetRequestBlock)
+	}
+
+	if r.blocks[topic] == nil {
+		r.blocks[topic] = make(map[int32]*offsetRequestBlock)
+	}
+
+	tmp := new(offsetRequestBlock)
+	tmp.time = time
+	if r.Version == 0 {
+		tmp.maxOffsets = maxOffsets
+	}
+
+	r.blocks[topic][partitionID] = tmp
+}
diff --git a/vendor/github.com/Shopify/sarama/offset_response.go b/vendor/github.com/Shopify/sarama/offset_response.go
new file mode 100644
index 0000000..8b2193f
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/offset_response.go
@@ -0,0 +1,174 @@
+package sarama
+
+type OffsetResponseBlock struct {
+	Err       KError
+	Offsets   []int64 // Version 0
+	Offset    int64   // Version 1
+	Timestamp int64   // Version 1
+}
+
+func (b *OffsetResponseBlock) decode(pd packetDecoder, version int16) (err error) {
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	b.Err = KError(tmp)
+
+	if version == 0 {
+		b.Offsets, err = pd.getInt64Array()
+
+		return err
+	}
+
+	b.Timestamp, err = pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	b.Offset, err = pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	// For backwards compatibility put the offset in the offsets array too
+	b.Offsets = []int64{b.Offset}
+
+	return nil
+}
+
+func (b *OffsetResponseBlock) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(int16(b.Err))
+
+	if version == 0 {
+		return pe.putInt64Array(b.Offsets)
+	}
+
+	pe.putInt64(b.Timestamp)
+	pe.putInt64(b.Offset)
+
+	return nil
+}
+
+type OffsetResponse struct {
+	Version int16
+	Blocks  map[string]map[int32]*OffsetResponseBlock
+}
+
+func (r *OffsetResponse) decode(pd packetDecoder, version int16) (err error) {
+	numTopics, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Blocks = make(map[string]map[int32]*OffsetResponseBlock, numTopics)
+	for i := 0; i < numTopics; i++ {
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		numBlocks, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		r.Blocks[name] = make(map[int32]*OffsetResponseBlock, numBlocks)
+
+		for j := 0; j < numBlocks; j++ {
+			id, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+
+			block := new(OffsetResponseBlock)
+			err = block.decode(pd, version)
+			if err != nil {
+				return err
+			}
+			r.Blocks[name][id] = block
+		}
+	}
+
+	return nil
+}
+
+func (r *OffsetResponse) GetBlock(topic string, partition int32) *OffsetResponseBlock {
+	if r.Blocks == nil {
+		return nil
+	}
+
+	if r.Blocks[topic] == nil {
+		return nil
+	}
+
+	return r.Blocks[topic][partition]
+}
+
+/*
+// [0 0 0 1 ntopics
+0 8 109 121 95 116 111 112 105 99 topic
+0 0 0 1 npartitions
+0 0 0 0 id
+0 0
+
+0 0 0 1 0 0 0 0
+0 1 1 1 0 0 0 1
+0 8 109 121 95 116 111 112
+105 99 0 0 0 1 0 0
+0 0 0 0 0 0 0 1
+0 0 0 0 0 1 1 1] <nil>
+
+*/
+func (r *OffsetResponse) encode(pe packetEncoder) (err error) {
+	if err = pe.putArrayLength(len(r.Blocks)); err != nil {
+		return err
+	}
+
+	for topic, partitions := range r.Blocks {
+		if err = pe.putString(topic); err != nil {
+			return err
+		}
+		if err = pe.putArrayLength(len(partitions)); err != nil {
+			return err
+		}
+		for partition, block := range partitions {
+			pe.putInt32(partition)
+			if err = block.encode(pe, r.version()); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (r *OffsetResponse) key() int16 {
+	return 2
+}
+
+func (r *OffsetResponse) version() int16 {
+	return r.Version
+}
+
+func (r *OffsetResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_10_1_0
+	default:
+		return MinVersion
+	}
+}
+
+// testing API
+
+func (r *OffsetResponse) AddTopicPartition(topic string, partition int32, offset int64) {
+	if r.Blocks == nil {
+		r.Blocks = make(map[string]map[int32]*OffsetResponseBlock)
+	}
+	byTopic, ok := r.Blocks[topic]
+	if !ok {
+		byTopic = make(map[int32]*OffsetResponseBlock)
+		r.Blocks[topic] = byTopic
+	}
+	byTopic[partition] = &OffsetResponseBlock{Offsets: []int64{offset}, Offset: offset}
+}
diff --git a/vendor/github.com/Shopify/sarama/packet_decoder.go b/vendor/github.com/Shopify/sarama/packet_decoder.go
new file mode 100644
index 0000000..74805cc
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/packet_decoder.go
@@ -0,0 +1,60 @@
+package sarama
+
+// PacketDecoder is the interface providing helpers for reading with Kafka's encoding rules.
+// Types implementing Decoder only need to worry about calling methods like GetString,
+// not about how a string is represented in Kafka.
+type packetDecoder interface {
+	// Primitives
+	getInt8() (int8, error)
+	getInt16() (int16, error)
+	getInt32() (int32, error)
+	getInt64() (int64, error)
+	getVarint() (int64, error)
+	getArrayLength() (int, error)
+	getBool() (bool, error)
+
+	// Collections
+	getBytes() ([]byte, error)
+	getVarintBytes() ([]byte, error)
+	getRawBytes(length int) ([]byte, error)
+	getString() (string, error)
+	getNullableString() (*string, error)
+	getInt32Array() ([]int32, error)
+	getInt64Array() ([]int64, error)
+	getStringArray() ([]string, error)
+
+	// Subsets
+	remaining() int
+	getSubset(length int) (packetDecoder, error)
+	peek(offset, length int) (packetDecoder, error) // similar to getSubset, but it doesn't advance the offset
+
+	// Stacks, see PushDecoder
+	push(in pushDecoder) error
+	pop() error
+}
+
+// PushDecoder is the interface for decoding fields like CRCs and lengths where the validity
+// of the field depends on what is after it in the packet. Start them with PacketDecoder.Push() where
+// the actual value is located in the packet, then PacketDecoder.Pop() them when all the bytes they
+// depend upon have been decoded.
+type pushDecoder interface {
+	// Saves the offset into the input buffer as the location to actually read the calculated value when able.
+	saveOffset(in int)
+
+	// Returns the length of data to reserve for the input of this encoder (eg 4 bytes for a CRC32).
+	reserveLength() int
+
+	// Indicates that all required data is now available to calculate and check the field.
+	// SaveOffset is guaranteed to have been called first. The implementation should read ReserveLength() bytes
+	// of data from the saved offset, and verify it based on the data between the saved offset and curOffset.
+	check(curOffset int, buf []byte) error
+}
+
+// dynamicPushDecoder extends the interface of pushDecoder for uses cases where the length of the
+// fields itself is unknown until its value was decoded (for instance varint encoded length
+// fields).
+// During push, dynamicPushDecoder.decode() method will be called instead of reserveLength()
+type dynamicPushDecoder interface {
+	pushDecoder
+	decoder
+}
diff --git a/vendor/github.com/Shopify/sarama/packet_encoder.go b/vendor/github.com/Shopify/sarama/packet_encoder.go
new file mode 100644
index 0000000..67b8dae
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/packet_encoder.go
@@ -0,0 +1,65 @@
+package sarama
+
+import "github.com/rcrowley/go-metrics"
+
+// PacketEncoder is the interface providing helpers for writing with Kafka's encoding rules.
+// Types implementing Encoder only need to worry about calling methods like PutString,
+// not about how a string is represented in Kafka.
+type packetEncoder interface {
+	// Primitives
+	putInt8(in int8)
+	putInt16(in int16)
+	putInt32(in int32)
+	putInt64(in int64)
+	putVarint(in int64)
+	putArrayLength(in int) error
+	putBool(in bool)
+
+	// Collections
+	putBytes(in []byte) error
+	putVarintBytes(in []byte) error
+	putRawBytes(in []byte) error
+	putString(in string) error
+	putNullableString(in *string) error
+	putStringArray(in []string) error
+	putInt32Array(in []int32) error
+	putInt64Array(in []int64) error
+
+	// Provide the current offset to record the batch size metric
+	offset() int
+
+	// Stacks, see PushEncoder
+	push(in pushEncoder)
+	pop() error
+
+	// To record metrics when provided
+	metricRegistry() metrics.Registry
+}
+
+// PushEncoder is the interface for encoding fields like CRCs and lengths where the value
+// of the field depends on what is encoded after it in the packet. Start them with PacketEncoder.Push() where
+// the actual value is located in the packet, then PacketEncoder.Pop() them when all the bytes they
+// depend upon have been written.
+type pushEncoder interface {
+	// Saves the offset into the input buffer as the location to actually write the calculated value when able.
+	saveOffset(in int)
+
+	// Returns the length of data to reserve for the output of this encoder (eg 4 bytes for a CRC32).
+	reserveLength() int
+
+	// Indicates that all required data is now available to calculate and write the field.
+	// SaveOffset is guaranteed to have been called first. The implementation should write ReserveLength() bytes
+	// of data to the saved offset, based on the data between the saved offset and curOffset.
+	run(curOffset int, buf []byte) error
+}
+
+// dynamicPushEncoder extends the interface of pushEncoder for uses cases where the length of the
+// fields itself is unknown until its value was computed (for instance varint encoded length
+// fields).
+type dynamicPushEncoder interface {
+	pushEncoder
+
+	// Called during pop() to adjust the length of the field.
+	// It should return the difference in bytes between the last computed length and current length.
+	adjustLength(currOffset int) int
+}
diff --git a/vendor/github.com/Shopify/sarama/partitioner.go b/vendor/github.com/Shopify/sarama/partitioner.go
new file mode 100644
index 0000000..6a708e7
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/partitioner.go
@@ -0,0 +1,217 @@
+package sarama
+
+import (
+	"hash"
+	"hash/fnv"
+	"math/rand"
+	"time"
+)
+
+// Partitioner is anything that, given a Kafka message and a number of partitions indexed [0...numPartitions-1],
+// decides to which partition to send the message. RandomPartitioner, RoundRobinPartitioner and HashPartitioner are provided
+// as simple default implementations.
+type Partitioner interface {
+	// Partition takes a message and partition count and chooses a partition
+	Partition(message *ProducerMessage, numPartitions int32) (int32, error)
+
+	// RequiresConsistency indicates to the user of the partitioner whether the
+	// mapping of key->partition is consistent or not. Specifically, if a
+	// partitioner requires consistency then it must be allowed to choose from all
+	// partitions (even ones known to be unavailable), and its choice must be
+	// respected by the caller. The obvious example is the HashPartitioner.
+	RequiresConsistency() bool
+}
+
+// DynamicConsistencyPartitioner can optionally be implemented by Partitioners
+// in order to allow more flexibility than is originally allowed by the
+// RequiresConsistency method in the Partitioner interface. This allows
+// partitioners to require consistency sometimes, but not all times. It's useful
+// for, e.g., the HashPartitioner, which does not require consistency if the
+// message key is nil.
+type DynamicConsistencyPartitioner interface {
+	Partitioner
+
+	// MessageRequiresConsistency is similar to Partitioner.RequiresConsistency,
+	// but takes in the message being partitioned so that the partitioner can
+	// make a per-message determination.
+	MessageRequiresConsistency(message *ProducerMessage) bool
+}
+
+// PartitionerConstructor is the type for a function capable of constructing new Partitioners.
+type PartitionerConstructor func(topic string) Partitioner
+
+type manualPartitioner struct{}
+
+// HashPartitionOption lets you modify default values of the partitioner
+type HashPartitionerOption func(*hashPartitioner)
+
+// WithAbsFirst means that the partitioner handles absolute values
+// in the same way as the reference Java implementation
+func WithAbsFirst() HashPartitionerOption {
+	return func(hp *hashPartitioner) {
+		hp.referenceAbs = true
+	}
+}
+
+// WithCustomHashFunction lets you specify what hash function to use for the partitioning
+func WithCustomHashFunction(hasher func() hash.Hash32) HashPartitionerOption {
+	return func(hp *hashPartitioner) {
+		hp.hasher = hasher()
+	}
+}
+
+// WithCustomFallbackPartitioner lets you specify what HashPartitioner should be used in case a Distribution Key is empty
+func WithCustomFallbackPartitioner(randomHP *hashPartitioner) HashPartitionerOption {
+	return func(hp *hashPartitioner) {
+		hp.random = hp
+	}
+}
+
+// NewManualPartitioner returns a Partitioner which uses the partition manually set in the provided
+// ProducerMessage's Partition field as the partition to produce to.
+func NewManualPartitioner(topic string) Partitioner {
+	return new(manualPartitioner)
+}
+
+func (p *manualPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) {
+	return message.Partition, nil
+}
+
+func (p *manualPartitioner) RequiresConsistency() bool {
+	return true
+}
+
+type randomPartitioner struct {
+	generator *rand.Rand
+}
+
+// NewRandomPartitioner returns a Partitioner which chooses a random partition each time.
+func NewRandomPartitioner(topic string) Partitioner {
+	p := new(randomPartitioner)
+	p.generator = rand.New(rand.NewSource(time.Now().UTC().UnixNano()))
+	return p
+}
+
+func (p *randomPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) {
+	return int32(p.generator.Intn(int(numPartitions))), nil
+}
+
+func (p *randomPartitioner) RequiresConsistency() bool {
+	return false
+}
+
+type roundRobinPartitioner struct {
+	partition int32
+}
+
+// NewRoundRobinPartitioner returns a Partitioner which walks through the available partitions one at a time.
+func NewRoundRobinPartitioner(topic string) Partitioner {
+	return &roundRobinPartitioner{}
+}
+
+func (p *roundRobinPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) {
+	if p.partition >= numPartitions {
+		p.partition = 0
+	}
+	ret := p.partition
+	p.partition++
+	return ret, nil
+}
+
+func (p *roundRobinPartitioner) RequiresConsistency() bool {
+	return false
+}
+
+type hashPartitioner struct {
+	random       Partitioner
+	hasher       hash.Hash32
+	referenceAbs bool
+}
+
+// NewCustomHashPartitioner is a wrapper around NewHashPartitioner, allowing the use of custom hasher.
+// The argument is a function providing the instance, implementing the hash.Hash32 interface. This is to ensure that
+// each partition dispatcher gets its own hasher, to avoid concurrency issues by sharing an instance.
+func NewCustomHashPartitioner(hasher func() hash.Hash32) PartitionerConstructor {
+	return func(topic string) Partitioner {
+		p := new(hashPartitioner)
+		p.random = NewRandomPartitioner(topic)
+		p.hasher = hasher()
+		p.referenceAbs = false
+		return p
+	}
+}
+
+// NewCustomPartitioner creates a default Partitioner but lets you specify the behavior of each component via options
+func NewCustomPartitioner(options ...HashPartitionerOption) PartitionerConstructor {
+	return func(topic string) Partitioner {
+		p := new(hashPartitioner)
+		p.random = NewRandomPartitioner(topic)
+		p.hasher = fnv.New32a()
+		p.referenceAbs = false
+		for _, option := range options {
+			option(p)
+		}
+		return p
+	}
+}
+
+// NewHashPartitioner returns a Partitioner which behaves as follows. If the message's key is nil then a
+// random partition is chosen. Otherwise the FNV-1a hash of the encoded bytes of the message key is used,
+// modulus the number of partitions. This ensures that messages with the same key always end up on the
+// same partition.
+func NewHashPartitioner(topic string) Partitioner {
+	p := new(hashPartitioner)
+	p.random = NewRandomPartitioner(topic)
+	p.hasher = fnv.New32a()
+	p.referenceAbs = false
+	return p
+}
+
+// NewReferenceHashPartitioner is like NewHashPartitioner except that it handles absolute values
+// in the same way as the reference Java implementation. NewHashPartitioner was supposed to do
+// that but it had a mistake and now there are people depending on both behaviours. This will
+// all go away on the next major version bump.
+func NewReferenceHashPartitioner(topic string) Partitioner {
+	p := new(hashPartitioner)
+	p.random = NewRandomPartitioner(topic)
+	p.hasher = fnv.New32a()
+	p.referenceAbs = true
+	return p
+}
+
+func (p *hashPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) {
+	if message.Key == nil {
+		return p.random.Partition(message, numPartitions)
+	}
+	bytes, err := message.Key.Encode()
+	if err != nil {
+		return -1, err
+	}
+	p.hasher.Reset()
+	_, err = p.hasher.Write(bytes)
+	if err != nil {
+		return -1, err
+	}
+	var partition int32
+	// Turns out we were doing our absolute value in a subtly different way from the upstream
+	// implementation, but now we need to maintain backwards compat for people who started using
+	// the old version; if referenceAbs is set we are compatible with the reference java client
+	// but not past Sarama versions
+	if p.referenceAbs {
+		partition = (int32(p.hasher.Sum32()) & 0x7fffffff) % numPartitions
+	} else {
+		partition = int32(p.hasher.Sum32()) % numPartitions
+		if partition < 0 {
+			partition = -partition
+		}
+	}
+	return partition, nil
+}
+
+func (p *hashPartitioner) RequiresConsistency() bool {
+	return true
+}
+
+func (p *hashPartitioner) MessageRequiresConsistency(message *ProducerMessage) bool {
+	return message.Key != nil
+}
diff --git a/vendor/github.com/Shopify/sarama/prep_encoder.go b/vendor/github.com/Shopify/sarama/prep_encoder.go
new file mode 100644
index 0000000..b633cd1
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/prep_encoder.go
@@ -0,0 +1,153 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"fmt"
+	"math"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+type prepEncoder struct {
+	stack  []pushEncoder
+	length int
+}
+
+// primitives
+
+func (pe *prepEncoder) putInt8(in int8) {
+	pe.length++
+}
+
+func (pe *prepEncoder) putInt16(in int16) {
+	pe.length += 2
+}
+
+func (pe *prepEncoder) putInt32(in int32) {
+	pe.length += 4
+}
+
+func (pe *prepEncoder) putInt64(in int64) {
+	pe.length += 8
+}
+
+func (pe *prepEncoder) putVarint(in int64) {
+	var buf [binary.MaxVarintLen64]byte
+	pe.length += binary.PutVarint(buf[:], in)
+}
+
+func (pe *prepEncoder) putArrayLength(in int) error {
+	if in > math.MaxInt32 {
+		return PacketEncodingError{fmt.Sprintf("array too long (%d)", in)}
+	}
+	pe.length += 4
+	return nil
+}
+
+func (pe *prepEncoder) putBool(in bool) {
+	pe.length++
+}
+
+// arrays
+
+func (pe *prepEncoder) putBytes(in []byte) error {
+	pe.length += 4
+	if in == nil {
+		return nil
+	}
+	return pe.putRawBytes(in)
+}
+
+func (pe *prepEncoder) putVarintBytes(in []byte) error {
+	if in == nil {
+		pe.putVarint(-1)
+		return nil
+	}
+	pe.putVarint(int64(len(in)))
+	return pe.putRawBytes(in)
+}
+
+func (pe *prepEncoder) putRawBytes(in []byte) error {
+	if len(in) > math.MaxInt32 {
+		return PacketEncodingError{fmt.Sprintf("byteslice too long (%d)", len(in))}
+	}
+	pe.length += len(in)
+	return nil
+}
+
+func (pe *prepEncoder) putNullableString(in *string) error {
+	if in == nil {
+		pe.length += 2
+		return nil
+	}
+	return pe.putString(*in)
+}
+
+func (pe *prepEncoder) putString(in string) error {
+	pe.length += 2
+	if len(in) > math.MaxInt16 {
+		return PacketEncodingError{fmt.Sprintf("string too long (%d)", len(in))}
+	}
+	pe.length += len(in)
+	return nil
+}
+
+func (pe *prepEncoder) putStringArray(in []string) error {
+	err := pe.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+
+	for _, str := range in {
+		if err := pe.putString(str); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (pe *prepEncoder) putInt32Array(in []int32) error {
+	err := pe.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+	pe.length += 4 * len(in)
+	return nil
+}
+
+func (pe *prepEncoder) putInt64Array(in []int64) error {
+	err := pe.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+	pe.length += 8 * len(in)
+	return nil
+}
+
+func (pe *prepEncoder) offset() int {
+	return pe.length
+}
+
+// stackable
+
+func (pe *prepEncoder) push(in pushEncoder) {
+	in.saveOffset(pe.length)
+	pe.length += in.reserveLength()
+	pe.stack = append(pe.stack, in)
+}
+
+func (pe *prepEncoder) pop() error {
+	in := pe.stack[len(pe.stack)-1]
+	pe.stack = pe.stack[:len(pe.stack)-1]
+	if dpe, ok := in.(dynamicPushEncoder); ok {
+		pe.length += dpe.adjustLength(pe.length)
+	}
+
+	return nil
+}
+
+// we do not record metrics during the prep encoder pass
+func (pe *prepEncoder) metricRegistry() metrics.Registry {
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/produce_request.go b/vendor/github.com/Shopify/sarama/produce_request.go
new file mode 100644
index 0000000..0c755d0
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/produce_request.go
@@ -0,0 +1,252 @@
+package sarama
+
+import "github.com/rcrowley/go-metrics"
+
+// RequiredAcks is used in Produce Requests to tell the broker how many replica acknowledgements
+// it must see before responding. Any of the constants defined here are valid. On broker versions
+// prior to 0.8.2.0 any other positive int16 is also valid (the broker will wait for that many
+// acknowledgements) but in 0.8.2.0 and later this will raise an exception (it has been replaced
+// by setting the `min.isr` value in the brokers configuration).
+type RequiredAcks int16
+
+const (
+	// NoResponse doesn't send any response, the TCP ACK is all you get.
+	NoResponse RequiredAcks = 0
+	// WaitForLocal waits for only the local commit to succeed before responding.
+	WaitForLocal RequiredAcks = 1
+	// WaitForAll waits for all in-sync replicas to commit before responding.
+	// The minimum number of in-sync replicas is configured on the broker via
+	// the `min.insync.replicas` configuration key.
+	WaitForAll RequiredAcks = -1
+)
+
+type ProduceRequest struct {
+	TransactionalID *string
+	RequiredAcks    RequiredAcks
+	Timeout         int32
+	Version         int16 // v1 requires Kafka 0.9, v2 requires Kafka 0.10, v3 requires Kafka 0.11
+	records         map[string]map[int32]Records
+}
+
+func updateMsgSetMetrics(msgSet *MessageSet, compressionRatioMetric metrics.Histogram,
+	topicCompressionRatioMetric metrics.Histogram) int64 {
+	var topicRecordCount int64
+	for _, messageBlock := range msgSet.Messages {
+		// Is this a fake "message" wrapping real messages?
+		if messageBlock.Msg.Set != nil {
+			topicRecordCount += int64(len(messageBlock.Msg.Set.Messages))
+		} else {
+			// A single uncompressed message
+			topicRecordCount++
+		}
+		// Better be safe than sorry when computing the compression ratio
+		if messageBlock.Msg.compressedSize != 0 {
+			compressionRatio := float64(len(messageBlock.Msg.Value)) /
+				float64(messageBlock.Msg.compressedSize)
+			// Histogram do not support decimal values, let's multiple it by 100 for better precision
+			intCompressionRatio := int64(100 * compressionRatio)
+			compressionRatioMetric.Update(intCompressionRatio)
+			topicCompressionRatioMetric.Update(intCompressionRatio)
+		}
+	}
+	return topicRecordCount
+}
+
+func updateBatchMetrics(recordBatch *RecordBatch, compressionRatioMetric metrics.Histogram,
+	topicCompressionRatioMetric metrics.Histogram) int64 {
+	if recordBatch.compressedRecords != nil {
+		compressionRatio := int64(float64(recordBatch.recordsLen) / float64(len(recordBatch.compressedRecords)) * 100)
+		compressionRatioMetric.Update(compressionRatio)
+		topicCompressionRatioMetric.Update(compressionRatio)
+	}
+
+	return int64(len(recordBatch.Records))
+}
+
+func (r *ProduceRequest) encode(pe packetEncoder) error {
+	if r.Version >= 3 {
+		if err := pe.putNullableString(r.TransactionalID); err != nil {
+			return err
+		}
+	}
+	pe.putInt16(int16(r.RequiredAcks))
+	pe.putInt32(r.Timeout)
+	metricRegistry := pe.metricRegistry()
+	var batchSizeMetric metrics.Histogram
+	var compressionRatioMetric metrics.Histogram
+	if metricRegistry != nil {
+		batchSizeMetric = getOrRegisterHistogram("batch-size", metricRegistry)
+		compressionRatioMetric = getOrRegisterHistogram("compression-ratio", metricRegistry)
+	}
+	totalRecordCount := int64(0)
+
+	err := pe.putArrayLength(len(r.records))
+	if err != nil {
+		return err
+	}
+
+	for topic, partitions := range r.records {
+		err = pe.putString(topic)
+		if err != nil {
+			return err
+		}
+		err = pe.putArrayLength(len(partitions))
+		if err != nil {
+			return err
+		}
+		topicRecordCount := int64(0)
+		var topicCompressionRatioMetric metrics.Histogram
+		if metricRegistry != nil {
+			topicCompressionRatioMetric = getOrRegisterTopicHistogram("compression-ratio", topic, metricRegistry)
+		}
+		for id, records := range partitions {
+			startOffset := pe.offset()
+			pe.putInt32(id)
+			pe.push(&lengthField{})
+			err = records.encode(pe)
+			if err != nil {
+				return err
+			}
+			err = pe.pop()
+			if err != nil {
+				return err
+			}
+			if metricRegistry != nil {
+				if r.Version >= 3 {
+					topicRecordCount += updateBatchMetrics(records.RecordBatch, compressionRatioMetric, topicCompressionRatioMetric)
+				} else {
+					topicRecordCount += updateMsgSetMetrics(records.MsgSet, compressionRatioMetric, topicCompressionRatioMetric)
+				}
+				batchSize := int64(pe.offset() - startOffset)
+				batchSizeMetric.Update(batchSize)
+				getOrRegisterTopicHistogram("batch-size", topic, metricRegistry).Update(batchSize)
+			}
+		}
+		if topicRecordCount > 0 {
+			getOrRegisterTopicMeter("record-send-rate", topic, metricRegistry).Mark(topicRecordCount)
+			getOrRegisterTopicHistogram("records-per-request", topic, metricRegistry).Update(topicRecordCount)
+			totalRecordCount += topicRecordCount
+		}
+	}
+	if totalRecordCount > 0 {
+		metrics.GetOrRegisterMeter("record-send-rate", metricRegistry).Mark(totalRecordCount)
+		getOrRegisterHistogram("records-per-request", metricRegistry).Update(totalRecordCount)
+	}
+
+	return nil
+}
+
+func (r *ProduceRequest) decode(pd packetDecoder, version int16) error {
+	r.Version = version
+
+	if version >= 3 {
+		id, err := pd.getNullableString()
+		if err != nil {
+			return err
+		}
+		r.TransactionalID = id
+	}
+	requiredAcks, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	r.RequiredAcks = RequiredAcks(requiredAcks)
+	if r.Timeout, err = pd.getInt32(); err != nil {
+		return err
+	}
+	topicCount, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if topicCount == 0 {
+		return nil
+	}
+
+	r.records = make(map[string]map[int32]Records)
+	for i := 0; i < topicCount; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		partitionCount, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+		r.records[topic] = make(map[int32]Records)
+
+		for j := 0; j < partitionCount; j++ {
+			partition, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			size, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+			recordsDecoder, err := pd.getSubset(int(size))
+			if err != nil {
+				return err
+			}
+			var records Records
+			if err := records.decode(recordsDecoder); err != nil {
+				return err
+			}
+			r.records[topic][partition] = records
+		}
+	}
+
+	return nil
+}
+
+func (r *ProduceRequest) key() int16 {
+	return 0
+}
+
+func (r *ProduceRequest) version() int16 {
+	return r.Version
+}
+
+func (r *ProduceRequest) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_9_0_0
+	case 2:
+		return V0_10_0_0
+	case 3:
+		return V0_11_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *ProduceRequest) ensureRecords(topic string, partition int32) {
+	if r.records == nil {
+		r.records = make(map[string]map[int32]Records)
+	}
+
+	if r.records[topic] == nil {
+		r.records[topic] = make(map[int32]Records)
+	}
+}
+
+func (r *ProduceRequest) AddMessage(topic string, partition int32, msg *Message) {
+	r.ensureRecords(topic, partition)
+	set := r.records[topic][partition].MsgSet
+
+	if set == nil {
+		set = new(MessageSet)
+		r.records[topic][partition] = newLegacyRecords(set)
+	}
+
+	set.addMessage(msg)
+}
+
+func (r *ProduceRequest) AddSet(topic string, partition int32, set *MessageSet) {
+	r.ensureRecords(topic, partition)
+	r.records[topic][partition] = newLegacyRecords(set)
+}
+
+func (r *ProduceRequest) AddBatch(topic string, partition int32, batch *RecordBatch) {
+	r.ensureRecords(topic, partition)
+	r.records[topic][partition] = newDefaultRecords(batch)
+}
diff --git a/vendor/github.com/Shopify/sarama/produce_response.go b/vendor/github.com/Shopify/sarama/produce_response.go
new file mode 100644
index 0000000..4c5cd35
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/produce_response.go
@@ -0,0 +1,189 @@
+package sarama
+
+import (
+	"fmt"
+	"time"
+)
+
+type ProduceResponseBlock struct {
+	Err    KError
+	Offset int64
+	// only provided if Version >= 2 and the broker is configured with `LogAppendTime`
+	Timestamp time.Time
+}
+
+func (b *ProduceResponseBlock) decode(pd packetDecoder, version int16) (err error) {
+	tmp, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	b.Err = KError(tmp)
+
+	b.Offset, err = pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	if version >= 2 {
+		if millis, err := pd.getInt64(); err != nil {
+			return err
+		} else if millis != -1 {
+			b.Timestamp = time.Unix(millis/1000, (millis%1000)*int64(time.Millisecond))
+		}
+	}
+
+	return nil
+}
+
+func (b *ProduceResponseBlock) encode(pe packetEncoder, version int16) (err error) {
+	pe.putInt16(int16(b.Err))
+	pe.putInt64(b.Offset)
+
+	if version >= 2 {
+		timestamp := int64(-1)
+		if !b.Timestamp.Before(time.Unix(0, 0)) {
+			timestamp = b.Timestamp.UnixNano() / int64(time.Millisecond)
+		} else if !b.Timestamp.IsZero() {
+			return PacketEncodingError{fmt.Sprintf("invalid timestamp (%v)", b.Timestamp)}
+		}
+		pe.putInt64(timestamp)
+	}
+
+	return nil
+}
+
+type ProduceResponse struct {
+	Blocks       map[string]map[int32]*ProduceResponseBlock
+	Version      int16
+	ThrottleTime time.Duration // only provided if Version >= 1
+}
+
+func (r *ProduceResponse) decode(pd packetDecoder, version int16) (err error) {
+	r.Version = version
+
+	numTopics, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	r.Blocks = make(map[string]map[int32]*ProduceResponseBlock, numTopics)
+	for i := 0; i < numTopics; i++ {
+		name, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		numBlocks, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		r.Blocks[name] = make(map[int32]*ProduceResponseBlock, numBlocks)
+
+		for j := 0; j < numBlocks; j++ {
+			id, err := pd.getInt32()
+			if err != nil {
+				return err
+			}
+
+			block := new(ProduceResponseBlock)
+			err = block.decode(pd, version)
+			if err != nil {
+				return err
+			}
+			r.Blocks[name][id] = block
+		}
+	}
+
+	if r.Version >= 1 {
+		millis, err := pd.getInt32()
+		if err != nil {
+			return err
+		}
+
+		r.ThrottleTime = time.Duration(millis) * time.Millisecond
+	}
+
+	return nil
+}
+
+func (r *ProduceResponse) encode(pe packetEncoder) error {
+	err := pe.putArrayLength(len(r.Blocks))
+	if err != nil {
+		return err
+	}
+	for topic, partitions := range r.Blocks {
+		err = pe.putString(topic)
+		if err != nil {
+			return err
+		}
+		err = pe.putArrayLength(len(partitions))
+		if err != nil {
+			return err
+		}
+		for id, prb := range partitions {
+			pe.putInt32(id)
+			err = prb.encode(pe, r.Version)
+			if err != nil {
+				return err
+			}
+		}
+	}
+	if r.Version >= 1 {
+		pe.putInt32(int32(r.ThrottleTime / time.Millisecond))
+	}
+	return nil
+}
+
+func (r *ProduceResponse) key() int16 {
+	return 0
+}
+
+func (r *ProduceResponse) version() int16 {
+	return r.Version
+}
+
+func (r *ProduceResponse) requiredVersion() KafkaVersion {
+	switch r.Version {
+	case 1:
+		return V0_9_0_0
+	case 2:
+		return V0_10_0_0
+	case 3:
+		return V0_11_0_0
+	default:
+		return MinVersion
+	}
+}
+
+func (r *ProduceResponse) GetBlock(topic string, partition int32) *ProduceResponseBlock {
+	if r.Blocks == nil {
+		return nil
+	}
+
+	if r.Blocks[topic] == nil {
+		return nil
+	}
+
+	return r.Blocks[topic][partition]
+}
+
+// Testing API
+
+func (r *ProduceResponse) AddTopicPartition(topic string, partition int32, err KError) {
+	if r.Blocks == nil {
+		r.Blocks = make(map[string]map[int32]*ProduceResponseBlock)
+	}
+	byTopic, ok := r.Blocks[topic]
+	if !ok {
+		byTopic = make(map[int32]*ProduceResponseBlock)
+		r.Blocks[topic] = byTopic
+	}
+	block := &ProduceResponseBlock{
+		Err: err,
+	}
+	if r.Version >= 2 {
+		block.Timestamp = time.Now()
+	}
+	byTopic[partition] = block
+}
diff --git a/vendor/github.com/Shopify/sarama/produce_set.go b/vendor/github.com/Shopify/sarama/produce_set.go
new file mode 100644
index 0000000..4b42d9c
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/produce_set.go
@@ -0,0 +1,258 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"errors"
+	"time"
+)
+
+type partitionSet struct {
+	msgs          []*ProducerMessage
+	recordsToSend Records
+	bufferBytes   int
+}
+
+type produceSet struct {
+	parent *asyncProducer
+	msgs   map[string]map[int32]*partitionSet
+
+	bufferBytes int
+	bufferCount int
+}
+
+func newProduceSet(parent *asyncProducer) *produceSet {
+	return &produceSet{
+		msgs:   make(map[string]map[int32]*partitionSet),
+		parent: parent,
+	}
+}
+
+func (ps *produceSet) add(msg *ProducerMessage) error {
+	var err error
+	var key, val []byte
+
+	if msg.Key != nil {
+		if key, err = msg.Key.Encode(); err != nil {
+			return err
+		}
+	}
+
+	if msg.Value != nil {
+		if val, err = msg.Value.Encode(); err != nil {
+			return err
+		}
+	}
+
+	timestamp := msg.Timestamp
+	if msg.Timestamp.IsZero() {
+		timestamp = time.Now()
+	}
+
+	partitions := ps.msgs[msg.Topic]
+	if partitions == nil {
+		partitions = make(map[int32]*partitionSet)
+		ps.msgs[msg.Topic] = partitions
+	}
+
+	var size int
+
+	set := partitions[msg.Partition]
+	if set == nil {
+		if ps.parent.conf.Version.IsAtLeast(V0_11_0_0) {
+			batch := &RecordBatch{
+				FirstTimestamp:   timestamp,
+				Version:          2,
+				Codec:            ps.parent.conf.Producer.Compression,
+				CompressionLevel: ps.parent.conf.Producer.CompressionLevel,
+				ProducerID:       ps.parent.txnmgr.producerID,
+				ProducerEpoch:    ps.parent.txnmgr.producerEpoch,
+			}
+			if ps.parent.conf.Producer.Idempotent {
+				batch.FirstSequence = msg.sequenceNumber
+			}
+			set = &partitionSet{recordsToSend: newDefaultRecords(batch)}
+			size = recordBatchOverhead
+		} else {
+			set = &partitionSet{recordsToSend: newLegacyRecords(new(MessageSet))}
+		}
+		partitions[msg.Partition] = set
+	}
+	set.msgs = append(set.msgs, msg)
+
+	if ps.parent.conf.Version.IsAtLeast(V0_11_0_0) {
+		if ps.parent.conf.Producer.Idempotent && msg.sequenceNumber < set.recordsToSend.RecordBatch.FirstSequence {
+			return errors.New("Assertion failed: Message out of sequence added to a batch")
+		}
+		// We are being conservative here to avoid having to prep encode the record
+		size += maximumRecordOverhead
+		rec := &Record{
+			Key:            key,
+			Value:          val,
+			TimestampDelta: timestamp.Sub(set.recordsToSend.RecordBatch.FirstTimestamp),
+		}
+		size += len(key) + len(val)
+		if len(msg.Headers) > 0 {
+			rec.Headers = make([]*RecordHeader, len(msg.Headers))
+			for i := range msg.Headers {
+				rec.Headers[i] = &msg.Headers[i]
+				size += len(rec.Headers[i].Key) + len(rec.Headers[i].Value) + 2*binary.MaxVarintLen32
+			}
+		}
+		set.recordsToSend.RecordBatch.addRecord(rec)
+	} else {
+		msgToSend := &Message{Codec: CompressionNone, Key: key, Value: val}
+		if ps.parent.conf.Version.IsAtLeast(V0_10_0_0) {
+			msgToSend.Timestamp = timestamp
+			msgToSend.Version = 1
+		}
+		set.recordsToSend.MsgSet.addMessage(msgToSend)
+		size = producerMessageOverhead + len(key) + len(val)
+	}
+
+	set.bufferBytes += size
+	ps.bufferBytes += size
+	ps.bufferCount++
+
+	return nil
+}
+
+func (ps *produceSet) buildRequest() *ProduceRequest {
+	req := &ProduceRequest{
+		RequiredAcks: ps.parent.conf.Producer.RequiredAcks,
+		Timeout:      int32(ps.parent.conf.Producer.Timeout / time.Millisecond),
+	}
+	if ps.parent.conf.Version.IsAtLeast(V0_10_0_0) {
+		req.Version = 2
+	}
+	if ps.parent.conf.Version.IsAtLeast(V0_11_0_0) {
+		req.Version = 3
+	}
+
+	for topic, partitionSets := range ps.msgs {
+		for partition, set := range partitionSets {
+			if req.Version >= 3 {
+				// If the API version we're hitting is 3 or greater, we need to calculate
+				// offsets for each record in the batch relative to FirstOffset.
+				// Additionally, we must set LastOffsetDelta to the value of the last offset
+				// in the batch. Since the OffsetDelta of the first record is 0, we know that the
+				// final record of any batch will have an offset of (# of records in batch) - 1.
+				// (See https://cwiki.apache.org/confluence/display/KAFKA/A+Guide+To+The+Kafka+Protocol#AGuideToTheKafkaProtocol-Messagesets
+				//  under the RecordBatch section for details.)
+				rb := set.recordsToSend.RecordBatch
+				if len(rb.Records) > 0 {
+					rb.LastOffsetDelta = int32(len(rb.Records) - 1)
+					for i, record := range rb.Records {
+						record.OffsetDelta = int64(i)
+					}
+				}
+				req.AddBatch(topic, partition, rb)
+				continue
+			}
+			if ps.parent.conf.Producer.Compression == CompressionNone {
+				req.AddSet(topic, partition, set.recordsToSend.MsgSet)
+			} else {
+				// When compression is enabled, the entire set for each partition is compressed
+				// and sent as the payload of a single fake "message" with the appropriate codec
+				// set and no key. When the server sees a message with a compression codec, it
+				// decompresses the payload and treats the result as its message set.
+
+				if ps.parent.conf.Version.IsAtLeast(V0_10_0_0) {
+					// If our version is 0.10 or later, assign relative offsets
+					// to the inner messages. This lets the broker avoid
+					// recompressing the message set.
+					// (See https://cwiki.apache.org/confluence/display/KAFKA/KIP-31+-+Move+to+relative+offsets+in+compressed+message+sets
+					// for details on relative offsets.)
+					for i, msg := range set.recordsToSend.MsgSet.Messages {
+						msg.Offset = int64(i)
+					}
+				}
+				payload, err := encode(set.recordsToSend.MsgSet, ps.parent.conf.MetricRegistry)
+				if err != nil {
+					Logger.Println(err) // if this happens, it's basically our fault.
+					panic(err)
+				}
+				compMsg := &Message{
+					Codec:            ps.parent.conf.Producer.Compression,
+					CompressionLevel: ps.parent.conf.Producer.CompressionLevel,
+					Key:              nil,
+					Value:            payload,
+					Set:              set.recordsToSend.MsgSet, // Provide the underlying message set for accurate metrics
+				}
+				if ps.parent.conf.Version.IsAtLeast(V0_10_0_0) {
+					compMsg.Version = 1
+					compMsg.Timestamp = set.recordsToSend.MsgSet.Messages[0].Msg.Timestamp
+				}
+				req.AddMessage(topic, partition, compMsg)
+			}
+		}
+	}
+
+	return req
+}
+
+func (ps *produceSet) eachPartition(cb func(topic string, partition int32, pSet *partitionSet)) {
+	for topic, partitionSet := range ps.msgs {
+		for partition, set := range partitionSet {
+			cb(topic, partition, set)
+		}
+	}
+}
+
+func (ps *produceSet) dropPartition(topic string, partition int32) []*ProducerMessage {
+	if ps.msgs[topic] == nil {
+		return nil
+	}
+	set := ps.msgs[topic][partition]
+	if set == nil {
+		return nil
+	}
+	ps.bufferBytes -= set.bufferBytes
+	ps.bufferCount -= len(set.msgs)
+	delete(ps.msgs[topic], partition)
+	return set.msgs
+}
+
+func (ps *produceSet) wouldOverflow(msg *ProducerMessage) bool {
+	version := 1
+	if ps.parent.conf.Version.IsAtLeast(V0_11_0_0) {
+		version = 2
+	}
+
+	switch {
+	// Would we overflow our maximum possible size-on-the-wire? 10KiB is arbitrary overhead for safety.
+	case ps.bufferBytes+msg.byteSize(version) >= int(MaxRequestSize-(10*1024)):
+		return true
+	// Would we overflow the size-limit of a message-batch for this partition?
+	case ps.msgs[msg.Topic] != nil && ps.msgs[msg.Topic][msg.Partition] != nil &&
+		ps.msgs[msg.Topic][msg.Partition].bufferBytes+msg.byteSize(version) >= ps.parent.conf.Producer.MaxMessageBytes:
+		return true
+	// Would we overflow simply in number of messages?
+	case ps.parent.conf.Producer.Flush.MaxMessages > 0 && ps.bufferCount >= ps.parent.conf.Producer.Flush.MaxMessages:
+		return true
+	default:
+		return false
+	}
+}
+
+func (ps *produceSet) readyToFlush() bool {
+	switch {
+	// If we don't have any messages, nothing else matters
+	case ps.empty():
+		return false
+	// If all three config values are 0, we always flush as-fast-as-possible
+	case ps.parent.conf.Producer.Flush.Frequency == 0 && ps.parent.conf.Producer.Flush.Bytes == 0 && ps.parent.conf.Producer.Flush.Messages == 0:
+		return true
+	// If we've passed the message trigger-point
+	case ps.parent.conf.Producer.Flush.Messages > 0 && ps.bufferCount >= ps.parent.conf.Producer.Flush.Messages:
+		return true
+	// If we've passed the byte trigger-point
+	case ps.parent.conf.Producer.Flush.Bytes > 0 && ps.bufferBytes >= ps.parent.conf.Producer.Flush.Bytes:
+		return true
+	default:
+		return false
+	}
+}
+
+func (ps *produceSet) empty() bool {
+	return ps.bufferCount == 0
+}
diff --git a/vendor/github.com/Shopify/sarama/real_decoder.go b/vendor/github.com/Shopify/sarama/real_decoder.go
new file mode 100644
index 0000000..23045e7
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/real_decoder.go
@@ -0,0 +1,324 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"math"
+)
+
+var errInvalidArrayLength = PacketDecodingError{"invalid array length"}
+var errInvalidByteSliceLength = PacketDecodingError{"invalid byteslice length"}
+var errInvalidByteSliceLengthType = PacketDecodingError{"invalid byteslice length type"}
+var errInvalidStringLength = PacketDecodingError{"invalid string length"}
+var errInvalidSubsetSize = PacketDecodingError{"invalid subset size"}
+var errVarintOverflow = PacketDecodingError{"varint overflow"}
+var errInvalidBool = PacketDecodingError{"invalid bool"}
+
+type realDecoder struct {
+	raw   []byte
+	off   int
+	stack []pushDecoder
+}
+
+// primitives
+
+func (rd *realDecoder) getInt8() (int8, error) {
+	if rd.remaining() < 1 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	tmp := int8(rd.raw[rd.off])
+	rd.off++
+	return tmp, nil
+}
+
+func (rd *realDecoder) getInt16() (int16, error) {
+	if rd.remaining() < 2 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	tmp := int16(binary.BigEndian.Uint16(rd.raw[rd.off:]))
+	rd.off += 2
+	return tmp, nil
+}
+
+func (rd *realDecoder) getInt32() (int32, error) {
+	if rd.remaining() < 4 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	tmp := int32(binary.BigEndian.Uint32(rd.raw[rd.off:]))
+	rd.off += 4
+	return tmp, nil
+}
+
+func (rd *realDecoder) getInt64() (int64, error) {
+	if rd.remaining() < 8 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	tmp := int64(binary.BigEndian.Uint64(rd.raw[rd.off:]))
+	rd.off += 8
+	return tmp, nil
+}
+
+func (rd *realDecoder) getVarint() (int64, error) {
+	tmp, n := binary.Varint(rd.raw[rd.off:])
+	if n == 0 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	if n < 0 {
+		rd.off -= n
+		return -1, errVarintOverflow
+	}
+	rd.off += n
+	return tmp, nil
+}
+
+func (rd *realDecoder) getArrayLength() (int, error) {
+	if rd.remaining() < 4 {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	}
+	tmp := int(int32(binary.BigEndian.Uint32(rd.raw[rd.off:])))
+	rd.off += 4
+	if tmp > rd.remaining() {
+		rd.off = len(rd.raw)
+		return -1, ErrInsufficientData
+	} else if tmp > 2*math.MaxUint16 {
+		return -1, errInvalidArrayLength
+	}
+	return tmp, nil
+}
+
+func (rd *realDecoder) getBool() (bool, error) {
+	b, err := rd.getInt8()
+	if err != nil || b == 0 {
+		return false, err
+	}
+	if b != 1 {
+		return false, errInvalidBool
+	}
+	return true, nil
+}
+
+// collections
+
+func (rd *realDecoder) getBytes() ([]byte, error) {
+	tmp, err := rd.getInt32()
+	if err != nil {
+		return nil, err
+	}
+	if tmp == -1 {
+		return nil, nil
+	}
+
+	return rd.getRawBytes(int(tmp))
+}
+
+func (rd *realDecoder) getVarintBytes() ([]byte, error) {
+	tmp, err := rd.getVarint()
+	if err != nil {
+		return nil, err
+	}
+	if tmp == -1 {
+		return nil, nil
+	}
+
+	return rd.getRawBytes(int(tmp))
+}
+
+func (rd *realDecoder) getStringLength() (int, error) {
+	length, err := rd.getInt16()
+	if err != nil {
+		return 0, err
+	}
+
+	n := int(length)
+
+	switch {
+	case n < -1:
+		return 0, errInvalidStringLength
+	case n > rd.remaining():
+		rd.off = len(rd.raw)
+		return 0, ErrInsufficientData
+	}
+
+	return n, nil
+}
+
+func (rd *realDecoder) getString() (string, error) {
+	n, err := rd.getStringLength()
+	if err != nil || n == -1 {
+		return "", err
+	}
+
+	tmpStr := string(rd.raw[rd.off : rd.off+n])
+	rd.off += n
+	return tmpStr, nil
+}
+
+func (rd *realDecoder) getNullableString() (*string, error) {
+	n, err := rd.getStringLength()
+	if err != nil || n == -1 {
+		return nil, err
+	}
+
+	tmpStr := string(rd.raw[rd.off : rd.off+n])
+	rd.off += n
+	return &tmpStr, err
+}
+
+func (rd *realDecoder) getInt32Array() ([]int32, error) {
+	if rd.remaining() < 4 {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+	n := int(binary.BigEndian.Uint32(rd.raw[rd.off:]))
+	rd.off += 4
+
+	if rd.remaining() < 4*n {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+
+	if n == 0 {
+		return nil, nil
+	}
+
+	if n < 0 {
+		return nil, errInvalidArrayLength
+	}
+
+	ret := make([]int32, n)
+	for i := range ret {
+		ret[i] = int32(binary.BigEndian.Uint32(rd.raw[rd.off:]))
+		rd.off += 4
+	}
+	return ret, nil
+}
+
+func (rd *realDecoder) getInt64Array() ([]int64, error) {
+	if rd.remaining() < 4 {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+	n := int(binary.BigEndian.Uint32(rd.raw[rd.off:]))
+	rd.off += 4
+
+	if rd.remaining() < 8*n {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+
+	if n == 0 {
+		return nil, nil
+	}
+
+	if n < 0 {
+		return nil, errInvalidArrayLength
+	}
+
+	ret := make([]int64, n)
+	for i := range ret {
+		ret[i] = int64(binary.BigEndian.Uint64(rd.raw[rd.off:]))
+		rd.off += 8
+	}
+	return ret, nil
+}
+
+func (rd *realDecoder) getStringArray() ([]string, error) {
+	if rd.remaining() < 4 {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+	n := int(binary.BigEndian.Uint32(rd.raw[rd.off:]))
+	rd.off += 4
+
+	if n == 0 {
+		return nil, nil
+	}
+
+	if n < 0 {
+		return nil, errInvalidArrayLength
+	}
+
+	ret := make([]string, n)
+	for i := range ret {
+		str, err := rd.getString()
+		if err != nil {
+			return nil, err
+		}
+
+		ret[i] = str
+	}
+	return ret, nil
+}
+
+// subsets
+
+func (rd *realDecoder) remaining() int {
+	return len(rd.raw) - rd.off
+}
+
+func (rd *realDecoder) getSubset(length int) (packetDecoder, error) {
+	buf, err := rd.getRawBytes(length)
+	if err != nil {
+		return nil, err
+	}
+	return &realDecoder{raw: buf}, nil
+}
+
+func (rd *realDecoder) getRawBytes(length int) ([]byte, error) {
+	if length < 0 {
+		return nil, errInvalidByteSliceLength
+	} else if length > rd.remaining() {
+		rd.off = len(rd.raw)
+		return nil, ErrInsufficientData
+	}
+
+	start := rd.off
+	rd.off += length
+	return rd.raw[start:rd.off], nil
+}
+
+func (rd *realDecoder) peek(offset, length int) (packetDecoder, error) {
+	if rd.remaining() < offset+length {
+		return nil, ErrInsufficientData
+	}
+	off := rd.off + offset
+	return &realDecoder{raw: rd.raw[off : off+length]}, nil
+}
+
+// stacks
+
+func (rd *realDecoder) push(in pushDecoder) error {
+	in.saveOffset(rd.off)
+
+	var reserve int
+	if dpd, ok := in.(dynamicPushDecoder); ok {
+		if err := dpd.decode(rd); err != nil {
+			return err
+		}
+	} else {
+		reserve = in.reserveLength()
+		if rd.remaining() < reserve {
+			rd.off = len(rd.raw)
+			return ErrInsufficientData
+		}
+	}
+
+	rd.stack = append(rd.stack, in)
+
+	rd.off += reserve
+
+	return nil
+}
+
+func (rd *realDecoder) pop() error {
+	// this is go's ugly pop pattern (the inverse of append)
+	in := rd.stack[len(rd.stack)-1]
+	rd.stack = rd.stack[:len(rd.stack)-1]
+
+	return in.check(rd.off, rd.raw)
+}
diff --git a/vendor/github.com/Shopify/sarama/real_encoder.go b/vendor/github.com/Shopify/sarama/real_encoder.go
new file mode 100644
index 0000000..3c75387
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/real_encoder.go
@@ -0,0 +1,156 @@
+package sarama
+
+import (
+	"encoding/binary"
+
+	"github.com/rcrowley/go-metrics"
+)
+
+type realEncoder struct {
+	raw      []byte
+	off      int
+	stack    []pushEncoder
+	registry metrics.Registry
+}
+
+// primitives
+
+func (re *realEncoder) putInt8(in int8) {
+	re.raw[re.off] = byte(in)
+	re.off++
+}
+
+func (re *realEncoder) putInt16(in int16) {
+	binary.BigEndian.PutUint16(re.raw[re.off:], uint16(in))
+	re.off += 2
+}
+
+func (re *realEncoder) putInt32(in int32) {
+	binary.BigEndian.PutUint32(re.raw[re.off:], uint32(in))
+	re.off += 4
+}
+
+func (re *realEncoder) putInt64(in int64) {
+	binary.BigEndian.PutUint64(re.raw[re.off:], uint64(in))
+	re.off += 8
+}
+
+func (re *realEncoder) putVarint(in int64) {
+	re.off += binary.PutVarint(re.raw[re.off:], in)
+}
+
+func (re *realEncoder) putArrayLength(in int) error {
+	re.putInt32(int32(in))
+	return nil
+}
+
+func (re *realEncoder) putBool(in bool) {
+	if in {
+		re.putInt8(1)
+		return
+	}
+	re.putInt8(0)
+}
+
+// collection
+
+func (re *realEncoder) putRawBytes(in []byte) error {
+	copy(re.raw[re.off:], in)
+	re.off += len(in)
+	return nil
+}
+
+func (re *realEncoder) putBytes(in []byte) error {
+	if in == nil {
+		re.putInt32(-1)
+		return nil
+	}
+	re.putInt32(int32(len(in)))
+	return re.putRawBytes(in)
+}
+
+func (re *realEncoder) putVarintBytes(in []byte) error {
+	if in == nil {
+		re.putVarint(-1)
+		return nil
+	}
+	re.putVarint(int64(len(in)))
+	return re.putRawBytes(in)
+}
+
+func (re *realEncoder) putString(in string) error {
+	re.putInt16(int16(len(in)))
+	copy(re.raw[re.off:], in)
+	re.off += len(in)
+	return nil
+}
+
+func (re *realEncoder) putNullableString(in *string) error {
+	if in == nil {
+		re.putInt16(-1)
+		return nil
+	}
+	return re.putString(*in)
+}
+
+func (re *realEncoder) putStringArray(in []string) error {
+	err := re.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+
+	for _, val := range in {
+		if err := re.putString(val); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (re *realEncoder) putInt32Array(in []int32) error {
+	err := re.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+	for _, val := range in {
+		re.putInt32(val)
+	}
+	return nil
+}
+
+func (re *realEncoder) putInt64Array(in []int64) error {
+	err := re.putArrayLength(len(in))
+	if err != nil {
+		return err
+	}
+	for _, val := range in {
+		re.putInt64(val)
+	}
+	return nil
+}
+
+func (re *realEncoder) offset() int {
+	return re.off
+}
+
+// stacks
+
+func (re *realEncoder) push(in pushEncoder) {
+	in.saveOffset(re.off)
+	re.off += in.reserveLength()
+	re.stack = append(re.stack, in)
+}
+
+func (re *realEncoder) pop() error {
+	// this is go's ugly pop pattern (the inverse of append)
+	in := re.stack[len(re.stack)-1]
+	re.stack = re.stack[:len(re.stack)-1]
+
+	return in.run(re.off, re.raw)
+}
+
+// we do record metrics during the real encoder pass
+func (re *realEncoder) metricRegistry() metrics.Registry {
+	return re.registry
+}
diff --git a/vendor/github.com/Shopify/sarama/record.go b/vendor/github.com/Shopify/sarama/record.go
new file mode 100644
index 0000000..cded308
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/record.go
@@ -0,0 +1,113 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"time"
+)
+
+const (
+	controlMask           = 0x20
+	maximumRecordOverhead = 5*binary.MaxVarintLen32 + binary.MaxVarintLen64 + 1
+)
+
+type RecordHeader struct {
+	Key   []byte
+	Value []byte
+}
+
+func (h *RecordHeader) encode(pe packetEncoder) error {
+	if err := pe.putVarintBytes(h.Key); err != nil {
+		return err
+	}
+	return pe.putVarintBytes(h.Value)
+}
+
+func (h *RecordHeader) decode(pd packetDecoder) (err error) {
+	if h.Key, err = pd.getVarintBytes(); err != nil {
+		return err
+	}
+
+	if h.Value, err = pd.getVarintBytes(); err != nil {
+		return err
+	}
+	return nil
+}
+
+type Record struct {
+	Attributes     int8
+	TimestampDelta time.Duration
+	OffsetDelta    int64
+	Key            []byte
+	Value          []byte
+	Headers        []*RecordHeader
+
+	length varintLengthField
+}
+
+func (r *Record) encode(pe packetEncoder) error {
+	pe.push(&r.length)
+	pe.putInt8(r.Attributes)
+	pe.putVarint(int64(r.TimestampDelta / time.Millisecond))
+	pe.putVarint(r.OffsetDelta)
+	if err := pe.putVarintBytes(r.Key); err != nil {
+		return err
+	}
+	if err := pe.putVarintBytes(r.Value); err != nil {
+		return err
+	}
+	pe.putVarint(int64(len(r.Headers)))
+
+	for _, h := range r.Headers {
+		if err := h.encode(pe); err != nil {
+			return err
+		}
+	}
+
+	return pe.pop()
+}
+
+func (r *Record) decode(pd packetDecoder) (err error) {
+	if err = pd.push(&r.length); err != nil {
+		return err
+	}
+
+	if r.Attributes, err = pd.getInt8(); err != nil {
+		return err
+	}
+
+	timestamp, err := pd.getVarint()
+	if err != nil {
+		return err
+	}
+	r.TimestampDelta = time.Duration(timestamp) * time.Millisecond
+
+	if r.OffsetDelta, err = pd.getVarint(); err != nil {
+		return err
+	}
+
+	if r.Key, err = pd.getVarintBytes(); err != nil {
+		return err
+	}
+
+	if r.Value, err = pd.getVarintBytes(); err != nil {
+		return err
+	}
+
+	numHeaders, err := pd.getVarint()
+	if err != nil {
+		return err
+	}
+
+	if numHeaders >= 0 {
+		r.Headers = make([]*RecordHeader, numHeaders)
+	}
+	for i := int64(0); i < numHeaders; i++ {
+		hdr := new(RecordHeader)
+		if err := hdr.decode(pd); err != nil {
+			return err
+		}
+		r.Headers[i] = hdr
+	}
+
+	return pd.pop()
+}
diff --git a/vendor/github.com/Shopify/sarama/record_batch.go b/vendor/github.com/Shopify/sarama/record_batch.go
new file mode 100644
index 0000000..a36f7e6
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/record_batch.go
@@ -0,0 +1,213 @@
+package sarama
+
+import (
+	"fmt"
+	"time"
+)
+
+const recordBatchOverhead = 49
+
+type recordsArray []*Record
+
+func (e recordsArray) encode(pe packetEncoder) error {
+	for _, r := range e {
+		if err := r.encode(pe); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (e recordsArray) decode(pd packetDecoder) error {
+	for i := range e {
+		rec := &Record{}
+		if err := rec.decode(pd); err != nil {
+			return err
+		}
+		e[i] = rec
+	}
+	return nil
+}
+
+type RecordBatch struct {
+	FirstOffset           int64
+	PartitionLeaderEpoch  int32
+	Version               int8
+	Codec                 CompressionCodec
+	CompressionLevel      int
+	Control               bool
+	LogAppendTime         bool
+	LastOffsetDelta       int32
+	FirstTimestamp        time.Time
+	MaxTimestamp          time.Time
+	ProducerID            int64
+	ProducerEpoch         int16
+	FirstSequence         int32
+	Records               []*Record
+	PartialTrailingRecord bool
+
+	compressedRecords []byte
+	recordsLen        int // uncompressed records size
+}
+
+func (b *RecordBatch) encode(pe packetEncoder) error {
+	if b.Version != 2 {
+		return PacketEncodingError{fmt.Sprintf("unsupported compression codec (%d)", b.Codec)}
+	}
+	pe.putInt64(b.FirstOffset)
+	pe.push(&lengthField{})
+	pe.putInt32(b.PartitionLeaderEpoch)
+	pe.putInt8(b.Version)
+	pe.push(newCRC32Field(crcCastagnoli))
+	pe.putInt16(b.computeAttributes())
+	pe.putInt32(b.LastOffsetDelta)
+
+	if err := (Timestamp{&b.FirstTimestamp}).encode(pe); err != nil {
+		return err
+	}
+
+	if err := (Timestamp{&b.MaxTimestamp}).encode(pe); err != nil {
+		return err
+	}
+
+	pe.putInt64(b.ProducerID)
+	pe.putInt16(b.ProducerEpoch)
+	pe.putInt32(b.FirstSequence)
+
+	if err := pe.putArrayLength(len(b.Records)); err != nil {
+		return err
+	}
+
+	if b.compressedRecords == nil {
+		if err := b.encodeRecords(pe); err != nil {
+			return err
+		}
+	}
+	if err := pe.putRawBytes(b.compressedRecords); err != nil {
+		return err
+	}
+
+	if err := pe.pop(); err != nil {
+		return err
+	}
+	return pe.pop()
+}
+
+func (b *RecordBatch) decode(pd packetDecoder) (err error) {
+	if b.FirstOffset, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	batchLen, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+
+	if b.PartitionLeaderEpoch, err = pd.getInt32(); err != nil {
+		return err
+	}
+
+	if b.Version, err = pd.getInt8(); err != nil {
+		return err
+	}
+
+	if err = pd.push(&crc32Field{polynomial: crcCastagnoli}); err != nil {
+		return err
+	}
+
+	attributes, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+	b.Codec = CompressionCodec(int8(attributes) & compressionCodecMask)
+	b.Control = attributes&controlMask == controlMask
+	b.LogAppendTime = attributes&timestampTypeMask == timestampTypeMask
+
+	if b.LastOffsetDelta, err = pd.getInt32(); err != nil {
+		return err
+	}
+
+	if err = (Timestamp{&b.FirstTimestamp}).decode(pd); err != nil {
+		return err
+	}
+
+	if err = (Timestamp{&b.MaxTimestamp}).decode(pd); err != nil {
+		return err
+	}
+
+	if b.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+
+	if b.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	if b.FirstSequence, err = pd.getInt32(); err != nil {
+		return err
+	}
+
+	numRecs, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if numRecs >= 0 {
+		b.Records = make([]*Record, numRecs)
+	}
+
+	bufSize := int(batchLen) - recordBatchOverhead
+	recBuffer, err := pd.getRawBytes(bufSize)
+	if err != nil {
+		if err == ErrInsufficientData {
+			b.PartialTrailingRecord = true
+			b.Records = nil
+			return nil
+		}
+		return err
+	}
+
+	if err = pd.pop(); err != nil {
+		return err
+	}
+
+	recBuffer, err = decompress(b.Codec, recBuffer)
+	if err != nil {
+		return err
+	}
+
+	b.recordsLen = len(recBuffer)
+	err = decode(recBuffer, recordsArray(b.Records))
+	if err == ErrInsufficientData {
+		b.PartialTrailingRecord = true
+		b.Records = nil
+		return nil
+	}
+	return err
+}
+
+func (b *RecordBatch) encodeRecords(pe packetEncoder) error {
+	var raw []byte
+	var err error
+	if raw, err = encode(recordsArray(b.Records), pe.metricRegistry()); err != nil {
+		return err
+	}
+	b.recordsLen = len(raw)
+
+	b.compressedRecords, err = compress(b.Codec, b.CompressionLevel, raw)
+	return err
+}
+
+func (b *RecordBatch) computeAttributes() int16 {
+	attr := int16(b.Codec) & int16(compressionCodecMask)
+	if b.Control {
+		attr |= controlMask
+	}
+	if b.LogAppendTime {
+		attr |= timestampTypeMask
+	}
+	return attr
+}
+
+func (b *RecordBatch) addRecord(r *Record) {
+	b.Records = append(b.Records, r)
+}
diff --git a/vendor/github.com/Shopify/sarama/records.go b/vendor/github.com/Shopify/sarama/records.go
new file mode 100644
index 0000000..192f592
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/records.go
@@ -0,0 +1,194 @@
+package sarama
+
+import "fmt"
+
+const (
+	unknownRecords = iota
+	legacyRecords
+	defaultRecords
+
+	magicOffset = 16
+	magicLength = 1
+)
+
+// Records implements a union type containing either a RecordBatch or a legacy MessageSet.
+type Records struct {
+	recordsType int
+	MsgSet      *MessageSet
+	RecordBatch *RecordBatch
+}
+
+func newLegacyRecords(msgSet *MessageSet) Records {
+	return Records{recordsType: legacyRecords, MsgSet: msgSet}
+}
+
+func newDefaultRecords(batch *RecordBatch) Records {
+	return Records{recordsType: defaultRecords, RecordBatch: batch}
+}
+
+// setTypeFromFields sets type of Records depending on which of MsgSet or RecordBatch is not nil.
+// The first return value indicates whether both fields are nil (and the type is not set).
+// If both fields are not nil, it returns an error.
+func (r *Records) setTypeFromFields() (bool, error) {
+	if r.MsgSet == nil && r.RecordBatch == nil {
+		return true, nil
+	}
+	if r.MsgSet != nil && r.RecordBatch != nil {
+		return false, fmt.Errorf("both MsgSet and RecordBatch are set, but record type is unknown")
+	}
+	r.recordsType = defaultRecords
+	if r.MsgSet != nil {
+		r.recordsType = legacyRecords
+	}
+	return false, nil
+}
+
+func (r *Records) encode(pe packetEncoder) error {
+	if r.recordsType == unknownRecords {
+		if empty, err := r.setTypeFromFields(); err != nil || empty {
+			return err
+		}
+	}
+
+	switch r.recordsType {
+	case legacyRecords:
+		if r.MsgSet == nil {
+			return nil
+		}
+		return r.MsgSet.encode(pe)
+	case defaultRecords:
+		if r.RecordBatch == nil {
+			return nil
+		}
+		return r.RecordBatch.encode(pe)
+	}
+
+	return fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func (r *Records) setTypeFromMagic(pd packetDecoder) error {
+	magic, err := magicValue(pd)
+	if err != nil {
+		return err
+	}
+
+	r.recordsType = defaultRecords
+	if magic < 2 {
+		r.recordsType = legacyRecords
+	}
+
+	return nil
+}
+
+func (r *Records) decode(pd packetDecoder) error {
+	if r.recordsType == unknownRecords {
+		if err := r.setTypeFromMagic(pd); err != nil {
+			return err
+		}
+	}
+
+	switch r.recordsType {
+	case legacyRecords:
+		r.MsgSet = &MessageSet{}
+		return r.MsgSet.decode(pd)
+	case defaultRecords:
+		r.RecordBatch = &RecordBatch{}
+		return r.RecordBatch.decode(pd)
+	}
+	return fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func (r *Records) numRecords() (int, error) {
+	if r.recordsType == unknownRecords {
+		if empty, err := r.setTypeFromFields(); err != nil || empty {
+			return 0, err
+		}
+	}
+
+	switch r.recordsType {
+	case legacyRecords:
+		if r.MsgSet == nil {
+			return 0, nil
+		}
+		return len(r.MsgSet.Messages), nil
+	case defaultRecords:
+		if r.RecordBatch == nil {
+			return 0, nil
+		}
+		return len(r.RecordBatch.Records), nil
+	}
+	return 0, fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func (r *Records) isPartial() (bool, error) {
+	if r.recordsType == unknownRecords {
+		if empty, err := r.setTypeFromFields(); err != nil || empty {
+			return false, err
+		}
+	}
+
+	switch r.recordsType {
+	case unknownRecords:
+		return false, nil
+	case legacyRecords:
+		if r.MsgSet == nil {
+			return false, nil
+		}
+		return r.MsgSet.PartialTrailingMessage, nil
+	case defaultRecords:
+		if r.RecordBatch == nil {
+			return false, nil
+		}
+		return r.RecordBatch.PartialTrailingRecord, nil
+	}
+	return false, fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func (r *Records) isControl() (bool, error) {
+	if r.recordsType == unknownRecords {
+		if empty, err := r.setTypeFromFields(); err != nil || empty {
+			return false, err
+		}
+	}
+
+	switch r.recordsType {
+	case legacyRecords:
+		return false, nil
+	case defaultRecords:
+		if r.RecordBatch == nil {
+			return false, nil
+		}
+		return r.RecordBatch.Control, nil
+	}
+	return false, fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func (r *Records) isOverflow() (bool, error) {
+	if r.recordsType == unknownRecords {
+		if empty, err := r.setTypeFromFields(); err != nil || empty {
+			return false, err
+		}
+	}
+
+	switch r.recordsType {
+	case unknownRecords:
+		return false, nil
+	case legacyRecords:
+		if r.MsgSet == nil {
+			return false, nil
+		}
+		return r.MsgSet.OverflowMessage, nil
+	case defaultRecords:
+		return false, nil
+	}
+	return false, fmt.Errorf("unknown records type: %v", r.recordsType)
+}
+
+func magicValue(pd packetDecoder) (int8, error) {
+	dec, err := pd.peek(magicOffset, magicLength)
+	if err != nil {
+		return 0, err
+	}
+
+	return dec.getInt8()
+}
diff --git a/vendor/github.com/Shopify/sarama/request.go b/vendor/github.com/Shopify/sarama/request.go
new file mode 100644
index 0000000..f7eea59
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/request.go
@@ -0,0 +1,151 @@
+package sarama
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+)
+
+type protocolBody interface {
+	encoder
+	versionedDecoder
+	key() int16
+	version() int16
+	requiredVersion() KafkaVersion
+}
+
+type request struct {
+	correlationID int32
+	clientID      string
+	body          protocolBody
+}
+
+func (r *request) encode(pe packetEncoder) (err error) {
+	pe.push(&lengthField{})
+	pe.putInt16(r.body.key())
+	pe.putInt16(r.body.version())
+	pe.putInt32(r.correlationID)
+	err = pe.putString(r.clientID)
+	if err != nil {
+		return err
+	}
+	err = r.body.encode(pe)
+	if err != nil {
+		return err
+	}
+	return pe.pop()
+}
+
+func (r *request) decode(pd packetDecoder) (err error) {
+	var key int16
+	if key, err = pd.getInt16(); err != nil {
+		return err
+	}
+	var version int16
+	if version, err = pd.getInt16(); err != nil {
+		return err
+	}
+	if r.correlationID, err = pd.getInt32(); err != nil {
+		return err
+	}
+	r.clientID, err = pd.getString()
+
+	r.body = allocateBody(key, version)
+	if r.body == nil {
+		return PacketDecodingError{fmt.Sprintf("unknown request key (%d)", key)}
+	}
+	return r.body.decode(pd, version)
+}
+
+func decodeRequest(r io.Reader) (req *request, bytesRead int, err error) {
+	lengthBytes := make([]byte, 4)
+	if _, err := io.ReadFull(r, lengthBytes); err != nil {
+		return nil, bytesRead, err
+	}
+	bytesRead += len(lengthBytes)
+
+	length := int32(binary.BigEndian.Uint32(lengthBytes))
+	if length <= 4 || length > MaxRequestSize {
+		return nil, bytesRead, PacketDecodingError{fmt.Sprintf("message of length %d too large or too small", length)}
+	}
+
+	encodedReq := make([]byte, length)
+	if _, err := io.ReadFull(r, encodedReq); err != nil {
+		return nil, bytesRead, err
+	}
+	bytesRead += len(encodedReq)
+
+	req = &request{}
+	if err := decode(encodedReq, req); err != nil {
+		return nil, bytesRead, err
+	}
+	return req, bytesRead, nil
+}
+
+func allocateBody(key, version int16) protocolBody {
+	switch key {
+	case 0:
+		return &ProduceRequest{}
+	case 1:
+		return &FetchRequest{}
+	case 2:
+		return &OffsetRequest{Version: version}
+	case 3:
+		return &MetadataRequest{}
+	case 8:
+		return &OffsetCommitRequest{Version: version}
+	case 9:
+		return &OffsetFetchRequest{}
+	case 10:
+		return &FindCoordinatorRequest{}
+	case 11:
+		return &JoinGroupRequest{}
+	case 12:
+		return &HeartbeatRequest{}
+	case 13:
+		return &LeaveGroupRequest{}
+	case 14:
+		return &SyncGroupRequest{}
+	case 15:
+		return &DescribeGroupsRequest{}
+	case 16:
+		return &ListGroupsRequest{}
+	case 17:
+		return &SaslHandshakeRequest{}
+	case 18:
+		return &ApiVersionsRequest{}
+	case 19:
+		return &CreateTopicsRequest{}
+	case 20:
+		return &DeleteTopicsRequest{}
+	case 21:
+		return &DeleteRecordsRequest{}
+	case 22:
+		return &InitProducerIDRequest{}
+	case 24:
+		return &AddPartitionsToTxnRequest{}
+	case 25:
+		return &AddOffsetsToTxnRequest{}
+	case 26:
+		return &EndTxnRequest{}
+	case 28:
+		return &TxnOffsetCommitRequest{}
+	case 29:
+		return &DescribeAclsRequest{}
+	case 30:
+		return &CreateAclsRequest{}
+	case 31:
+		return &DeleteAclsRequest{}
+	case 32:
+		return &DescribeConfigsRequest{}
+	case 33:
+		return &AlterConfigsRequest{}
+	case 36:
+		return &SaslAuthenticateRequest{}
+	case 37:
+		return &CreatePartitionsRequest{}
+	case 42:
+		return &DeleteGroupsRequest{}
+	}
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/response_header.go b/vendor/github.com/Shopify/sarama/response_header.go
new file mode 100644
index 0000000..f3f4d27
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/response_header.go
@@ -0,0 +1,21 @@
+package sarama
+
+import "fmt"
+
+type responseHeader struct {
+	length        int32
+	correlationID int32
+}
+
+func (r *responseHeader) decode(pd packetDecoder) (err error) {
+	r.length, err = pd.getInt32()
+	if err != nil {
+		return err
+	}
+	if r.length <= 4 || r.length > MaxResponseSize {
+		return PacketDecodingError{fmt.Sprintf("message of length %d too large or too small", r.length)}
+	}
+
+	r.correlationID, err = pd.getInt32()
+	return err
+}
diff --git a/vendor/github.com/Shopify/sarama/sarama.go b/vendor/github.com/Shopify/sarama/sarama.go
new file mode 100644
index 0000000..7d5dc60
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sarama.go
@@ -0,0 +1,99 @@
+/*
+Package sarama is a pure Go client library for dealing with Apache Kafka (versions 0.8 and later). It includes a high-level
+API for easily producing and consuming messages, and a low-level API for controlling bytes on the wire when the high-level
+API is insufficient. Usage examples for the high-level APIs are provided inline with their full documentation.
+
+To produce messages, use either the AsyncProducer or the SyncProducer. The AsyncProducer accepts messages on a channel
+and produces them asynchronously in the background as efficiently as possible; it is preferred in most cases.
+The SyncProducer provides a method which will block until Kafka acknowledges the message as produced. This can be
+useful but comes with two caveats: it will generally be less efficient, and the actual durability guarantees
+depend on the configured value of `Producer.RequiredAcks`. There are configurations where a message acknowledged by the
+SyncProducer can still sometimes be lost.
+
+To consume messages, use the Consumer. Note that Sarama's Consumer implementation does not currently support automatic
+consumer-group rebalancing and offset tracking. For Zookeeper-based tracking (Kafka 0.8.2 and earlier), the
+https://github.com/wvanbergen/kafka library builds on Sarama to add this support. For Kafka-based tracking (Kafka 0.9
+and later), the https://github.com/bsm/sarama-cluster library builds on Sarama to add this support.
+
+For lower-level needs, the Broker and Request/Response objects permit precise control over each connection
+and message sent on the wire; the Client provides higher-level metadata management that is shared between
+the producers and the consumer. The Request/Response objects and properties are mostly undocumented, as they line up
+exactly with the protocol fields documented by Kafka at
+https://cwiki.apache.org/confluence/display/KAFKA/A+Guide+To+The+Kafka+Protocol
+
+Metrics are exposed through https://github.com/rcrowley/go-metrics library in a local registry.
+
+Broker related metrics:
+
+	+----------------------------------------------+------------+---------------------------------------------------------------+
+	| Name                                         | Type       | Description                                                   |
+	+----------------------------------------------+------------+---------------------------------------------------------------+
+	| incoming-byte-rate                           | meter      | Bytes/second read off all brokers                             |
+	| incoming-byte-rate-for-broker-<broker-id>    | meter      | Bytes/second read off a given broker                          |
+	| outgoing-byte-rate                           | meter      | Bytes/second written off all brokers                          |
+	| outgoing-byte-rate-for-broker-<broker-id>    | meter      | Bytes/second written off a given broker                       |
+	| request-rate                                 | meter      | Requests/second sent to all brokers                           |
+	| request-rate-for-broker-<broker-id>          | meter      | Requests/second sent to a given broker                        |
+	| request-size                                 | histogram  | Distribution of the request size in bytes for all brokers     |
+	| request-size-for-broker-<broker-id>          | histogram  | Distribution of the request size in bytes for a given broker  |
+	| request-latency-in-ms                        | histogram  | Distribution of the request latency in ms for all brokers     |
+	| request-latency-in-ms-for-broker-<broker-id> | histogram  | Distribution of the request latency in ms for a given broker  |
+	| response-rate                                | meter      | Responses/second received from all brokers                    |
+	| response-rate-for-broker-<broker-id>         | meter      | Responses/second received from a given broker                 |
+	| response-size                                | histogram  | Distribution of the response size in bytes for all brokers    |
+	| response-size-for-broker-<broker-id>         | histogram  | Distribution of the response size in bytes for a given broker |
+	+----------------------------------------------+------------+---------------------------------------------------------------+
+
+Note that we do not gather specific metrics for seed brokers but they are part of the "all brokers" metrics.
+
+Producer related metrics:
+
+	+-------------------------------------------+------------+--------------------------------------------------------------------------------------+
+	| Name                                      | Type       | Description                                                                          |
+	+-------------------------------------------+------------+--------------------------------------------------------------------------------------+
+	| batch-size                                | histogram  | Distribution of the number of bytes sent per partition per request for all topics    |
+	| batch-size-for-topic-<topic>              | histogram  | Distribution of the number of bytes sent per partition per request for a given topic |
+	| record-send-rate                          | meter      | Records/second sent to all topics                                                    |
+	| record-send-rate-for-topic-<topic>        | meter      | Records/second sent to a given topic                                                 |
+	| records-per-request                       | histogram  | Distribution of the number of records sent per request for all topics                |
+	| records-per-request-for-topic-<topic>     | histogram  | Distribution of the number of records sent per request for a given topic             |
+	| compression-ratio                         | histogram  | Distribution of the compression ratio times 100 of record batches for all topics     |
+	| compression-ratio-for-topic-<topic>       | histogram  | Distribution of the compression ratio times 100 of record batches for a given topic  |
+	+-------------------------------------------+------------+--------------------------------------------------------------------------------------+
+
+*/
+package sarama
+
+import (
+	"io/ioutil"
+	"log"
+)
+
+// Logger is the instance of a StdLogger interface that Sarama writes connection
+// management events to. By default it is set to discard all log messages via ioutil.Discard,
+// but you can set it to redirect wherever you want.
+var Logger StdLogger = log.New(ioutil.Discard, "[Sarama] ", log.LstdFlags)
+
+// StdLogger is used to log error messages.
+type StdLogger interface {
+	Print(v ...interface{})
+	Printf(format string, v ...interface{})
+	Println(v ...interface{})
+}
+
+// PanicHandler is called for recovering from panics spawned internally to the library (and thus
+// not recoverable by the caller's goroutine). Defaults to nil, which means panics are not recovered.
+var PanicHandler func(interface{})
+
+// MaxRequestSize is the maximum size (in bytes) of any request that Sarama will attempt to send. Trying
+// to send a request larger than this will result in an PacketEncodingError. The default of 100 MiB is aligned
+// with Kafka's default `socket.request.max.bytes`, which is the largest request the broker will attempt
+// to process.
+var MaxRequestSize int32 = 100 * 1024 * 1024
+
+// MaxResponseSize is the maximum size (in bytes) of any response that Sarama will attempt to parse. If
+// a broker returns a response message larger than this value, Sarama will return a PacketDecodingError to
+// protect the client from running out of memory. Please note that brokers do not have any natural limit on
+// the size of responses they send. In particular, they can send arbitrarily large fetch responses to consumers
+// (see https://issues.apache.org/jira/browse/KAFKA-2063).
+var MaxResponseSize int32 = 100 * 1024 * 1024
diff --git a/vendor/github.com/Shopify/sarama/sasl_authenticate_request.go b/vendor/github.com/Shopify/sarama/sasl_authenticate_request.go
new file mode 100644
index 0000000..54c8b09
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sasl_authenticate_request.go
@@ -0,0 +1,29 @@
+package sarama
+
+type SaslAuthenticateRequest struct {
+	SaslAuthBytes []byte
+}
+
+// APIKeySASLAuth is the API key for the SaslAuthenticate Kafka API
+const APIKeySASLAuth = 36
+
+func (r *SaslAuthenticateRequest) encode(pe packetEncoder) error {
+	return pe.putBytes(r.SaslAuthBytes)
+}
+
+func (r *SaslAuthenticateRequest) decode(pd packetDecoder, version int16) (err error) {
+	r.SaslAuthBytes, err = pd.getBytes()
+	return err
+}
+
+func (r *SaslAuthenticateRequest) key() int16 {
+	return APIKeySASLAuth
+}
+
+func (r *SaslAuthenticateRequest) version() int16 {
+	return 0
+}
+
+func (r *SaslAuthenticateRequest) requiredVersion() KafkaVersion {
+	return V1_0_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/sasl_authenticate_response.go b/vendor/github.com/Shopify/sarama/sasl_authenticate_response.go
new file mode 100644
index 0000000..0038c3f
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sasl_authenticate_response.go
@@ -0,0 +1,44 @@
+package sarama
+
+type SaslAuthenticateResponse struct {
+	Err           KError
+	ErrorMessage  *string
+	SaslAuthBytes []byte
+}
+
+func (r *SaslAuthenticateResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	if err := pe.putNullableString(r.ErrorMessage); err != nil {
+		return err
+	}
+	return pe.putBytes(r.SaslAuthBytes)
+}
+
+func (r *SaslAuthenticateResponse) decode(pd packetDecoder, version int16) error {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	if r.ErrorMessage, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	r.SaslAuthBytes, err = pd.getBytes()
+
+	return err
+}
+
+func (r *SaslAuthenticateResponse) key() int16 {
+	return APIKeySASLAuth
+}
+
+func (r *SaslAuthenticateResponse) version() int16 {
+	return 0
+}
+
+func (r *SaslAuthenticateResponse) requiredVersion() KafkaVersion {
+	return V1_0_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/sasl_handshake_request.go b/vendor/github.com/Shopify/sarama/sasl_handshake_request.go
new file mode 100644
index 0000000..fe5ba05
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sasl_handshake_request.go
@@ -0,0 +1,34 @@
+package sarama
+
+type SaslHandshakeRequest struct {
+	Mechanism string
+	Version   int16
+}
+
+func (r *SaslHandshakeRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(r.Mechanism); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *SaslHandshakeRequest) decode(pd packetDecoder, version int16) (err error) {
+	if r.Mechanism, err = pd.getString(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *SaslHandshakeRequest) key() int16 {
+	return 17
+}
+
+func (r *SaslHandshakeRequest) version() int16 {
+	return r.Version
+}
+
+func (r *SaslHandshakeRequest) requiredVersion() KafkaVersion {
+	return V0_10_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/sasl_handshake_response.go b/vendor/github.com/Shopify/sarama/sasl_handshake_response.go
new file mode 100644
index 0000000..ef290d4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sasl_handshake_response.go
@@ -0,0 +1,38 @@
+package sarama
+
+type SaslHandshakeResponse struct {
+	Err               KError
+	EnabledMechanisms []string
+}
+
+func (r *SaslHandshakeResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	return pe.putStringArray(r.EnabledMechanisms)
+}
+
+func (r *SaslHandshakeResponse) decode(pd packetDecoder, version int16) error {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	if r.EnabledMechanisms, err = pd.getStringArray(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (r *SaslHandshakeResponse) key() int16 {
+	return 17
+}
+
+func (r *SaslHandshakeResponse) version() int16 {
+	return 0
+}
+
+func (r *SaslHandshakeResponse) requiredVersion() KafkaVersion {
+	return V0_10_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/sync_group_request.go b/vendor/github.com/Shopify/sarama/sync_group_request.go
new file mode 100644
index 0000000..fe20708
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sync_group_request.go
@@ -0,0 +1,100 @@
+package sarama
+
+type SyncGroupRequest struct {
+	GroupId          string
+	GenerationId     int32
+	MemberId         string
+	GroupAssignments map[string][]byte
+}
+
+func (r *SyncGroupRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(r.GroupId); err != nil {
+		return err
+	}
+
+	pe.putInt32(r.GenerationId)
+
+	if err := pe.putString(r.MemberId); err != nil {
+		return err
+	}
+
+	if err := pe.putArrayLength(len(r.GroupAssignments)); err != nil {
+		return err
+	}
+	for memberId, memberAssignment := range r.GroupAssignments {
+		if err := pe.putString(memberId); err != nil {
+			return err
+		}
+		if err := pe.putBytes(memberAssignment); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (r *SyncGroupRequest) decode(pd packetDecoder, version int16) (err error) {
+	if r.GroupId, err = pd.getString(); err != nil {
+		return
+	}
+	if r.GenerationId, err = pd.getInt32(); err != nil {
+		return
+	}
+	if r.MemberId, err = pd.getString(); err != nil {
+		return
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+	if n == 0 {
+		return nil
+	}
+
+	r.GroupAssignments = make(map[string][]byte)
+	for i := 0; i < n; i++ {
+		memberId, err := pd.getString()
+		if err != nil {
+			return err
+		}
+		memberAssignment, err := pd.getBytes()
+		if err != nil {
+			return err
+		}
+
+		r.GroupAssignments[memberId] = memberAssignment
+	}
+
+	return nil
+}
+
+func (r *SyncGroupRequest) key() int16 {
+	return 14
+}
+
+func (r *SyncGroupRequest) version() int16 {
+	return 0
+}
+
+func (r *SyncGroupRequest) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
+
+func (r *SyncGroupRequest) AddGroupAssignment(memberId string, memberAssignment []byte) {
+	if r.GroupAssignments == nil {
+		r.GroupAssignments = make(map[string][]byte)
+	}
+
+	r.GroupAssignments[memberId] = memberAssignment
+}
+
+func (r *SyncGroupRequest) AddGroupAssignmentMember(memberId string, memberAssignment *ConsumerGroupMemberAssignment) error {
+	bin, err := encode(memberAssignment, nil)
+	if err != nil {
+		return err
+	}
+
+	r.AddGroupAssignment(memberId, bin)
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/sync_group_response.go b/vendor/github.com/Shopify/sarama/sync_group_response.go
new file mode 100644
index 0000000..194b382
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sync_group_response.go
@@ -0,0 +1,41 @@
+package sarama
+
+type SyncGroupResponse struct {
+	Err              KError
+	MemberAssignment []byte
+}
+
+func (r *SyncGroupResponse) GetMemberAssignment() (*ConsumerGroupMemberAssignment, error) {
+	assignment := new(ConsumerGroupMemberAssignment)
+	err := decode(r.MemberAssignment, assignment)
+	return assignment, err
+}
+
+func (r *SyncGroupResponse) encode(pe packetEncoder) error {
+	pe.putInt16(int16(r.Err))
+	return pe.putBytes(r.MemberAssignment)
+}
+
+func (r *SyncGroupResponse) decode(pd packetDecoder, version int16) (err error) {
+	kerr, err := pd.getInt16()
+	if err != nil {
+		return err
+	}
+
+	r.Err = KError(kerr)
+
+	r.MemberAssignment, err = pd.getBytes()
+	return
+}
+
+func (r *SyncGroupResponse) key() int16 {
+	return 14
+}
+
+func (r *SyncGroupResponse) version() int16 {
+	return 0
+}
+
+func (r *SyncGroupResponse) requiredVersion() KafkaVersion {
+	return V0_9_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/sync_producer.go b/vendor/github.com/Shopify/sarama/sync_producer.go
new file mode 100644
index 0000000..021c5a0
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/sync_producer.go
@@ -0,0 +1,149 @@
+package sarama
+
+import "sync"
+
+// SyncProducer publishes Kafka messages, blocking until they have been acknowledged. It routes messages to the correct
+// broker, refreshing metadata as appropriate, and parses responses for errors. You must call Close() on a producer
+// to avoid leaks, it may not be garbage-collected automatically when it passes out of scope.
+//
+// The SyncProducer comes with two caveats: it will generally be less efficient than the AsyncProducer, and the actual
+// durability guarantee provided when a message is acknowledged depend on the configured value of `Producer.RequiredAcks`.
+// There are configurations where a message acknowledged by the SyncProducer can still sometimes be lost.
+//
+// For implementation reasons, the SyncProducer requires `Producer.Return.Errors` and `Producer.Return.Successes` to
+// be set to true in its configuration.
+type SyncProducer interface {
+
+	// SendMessage produces a given message, and returns only when it either has
+	// succeeded or failed to produce. It will return the partition and the offset
+	// of the produced message, or an error if the message failed to produce.
+	SendMessage(msg *ProducerMessage) (partition int32, offset int64, err error)
+
+	// SendMessages produces a given set of messages, and returns only when all
+	// messages in the set have either succeeded or failed. Note that messages
+	// can succeed and fail individually; if some succeed and some fail,
+	// SendMessages will return an error.
+	SendMessages(msgs []*ProducerMessage) error
+
+	// Close shuts down the producer and waits for any buffered messages to be
+	// flushed. You must call this function before a producer object passes out of
+	// scope, as it may otherwise leak memory. You must call this before calling
+	// Close on the underlying client.
+	Close() error
+}
+
+type syncProducer struct {
+	producer *asyncProducer
+	wg       sync.WaitGroup
+}
+
+// NewSyncProducer creates a new SyncProducer using the given broker addresses and configuration.
+func NewSyncProducer(addrs []string, config *Config) (SyncProducer, error) {
+	if config == nil {
+		config = NewConfig()
+		config.Producer.Return.Successes = true
+	}
+
+	if err := verifyProducerConfig(config); err != nil {
+		return nil, err
+	}
+
+	p, err := NewAsyncProducer(addrs, config)
+	if err != nil {
+		return nil, err
+	}
+	return newSyncProducerFromAsyncProducer(p.(*asyncProducer)), nil
+}
+
+// NewSyncProducerFromClient creates a new SyncProducer using the given client. It is still
+// necessary to call Close() on the underlying client when shutting down this producer.
+func NewSyncProducerFromClient(client Client) (SyncProducer, error) {
+	if err := verifyProducerConfig(client.Config()); err != nil {
+		return nil, err
+	}
+
+	p, err := NewAsyncProducerFromClient(client)
+	if err != nil {
+		return nil, err
+	}
+	return newSyncProducerFromAsyncProducer(p.(*asyncProducer)), nil
+}
+
+func newSyncProducerFromAsyncProducer(p *asyncProducer) *syncProducer {
+	sp := &syncProducer{producer: p}
+
+	sp.wg.Add(2)
+	go withRecover(sp.handleSuccesses)
+	go withRecover(sp.handleErrors)
+
+	return sp
+}
+
+func verifyProducerConfig(config *Config) error {
+	if !config.Producer.Return.Errors {
+		return ConfigurationError("Producer.Return.Errors must be true to be used in a SyncProducer")
+	}
+	if !config.Producer.Return.Successes {
+		return ConfigurationError("Producer.Return.Successes must be true to be used in a SyncProducer")
+	}
+	return nil
+}
+
+func (sp *syncProducer) SendMessage(msg *ProducerMessage) (partition int32, offset int64, err error) {
+	expectation := make(chan *ProducerError, 1)
+	msg.expectation = expectation
+	sp.producer.Input() <- msg
+
+	if err := <-expectation; err != nil {
+		return -1, -1, err.Err
+	}
+
+	return msg.Partition, msg.Offset, nil
+}
+
+func (sp *syncProducer) SendMessages(msgs []*ProducerMessage) error {
+	expectations := make(chan chan *ProducerError, len(msgs))
+	go func() {
+		for _, msg := range msgs {
+			expectation := make(chan *ProducerError, 1)
+			msg.expectation = expectation
+			sp.producer.Input() <- msg
+			expectations <- expectation
+		}
+		close(expectations)
+	}()
+
+	var errors ProducerErrors
+	for expectation := range expectations {
+		if err := <-expectation; err != nil {
+			errors = append(errors, err)
+		}
+	}
+
+	if len(errors) > 0 {
+		return errors
+	}
+	return nil
+}
+
+func (sp *syncProducer) handleSuccesses() {
+	defer sp.wg.Done()
+	for msg := range sp.producer.Successes() {
+		expectation := msg.expectation
+		expectation <- nil
+	}
+}
+
+func (sp *syncProducer) handleErrors() {
+	defer sp.wg.Done()
+	for err := range sp.producer.Errors() {
+		expectation := err.Msg.expectation
+		expectation <- err
+	}
+}
+
+func (sp *syncProducer) Close() error {
+	sp.producer.AsyncClose()
+	sp.wg.Wait()
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/timestamp.go b/vendor/github.com/Shopify/sarama/timestamp.go
new file mode 100644
index 0000000..372278d
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/timestamp.go
@@ -0,0 +1,40 @@
+package sarama
+
+import (
+	"fmt"
+	"time"
+)
+
+type Timestamp struct {
+	*time.Time
+}
+
+func (t Timestamp) encode(pe packetEncoder) error {
+	timestamp := int64(-1)
+
+	if !t.Before(time.Unix(0, 0)) {
+		timestamp = t.UnixNano() / int64(time.Millisecond)
+	} else if !t.IsZero() {
+		return PacketEncodingError{fmt.Sprintf("invalid timestamp (%v)", t)}
+	}
+
+	pe.putInt64(timestamp)
+	return nil
+}
+
+func (t Timestamp) decode(pd packetDecoder) error {
+	millis, err := pd.getInt64()
+	if err != nil {
+		return err
+	}
+
+	// negative timestamps are invalid, in these cases we should return
+	// a zero time
+	timestamp := time.Time{}
+	if millis >= 0 {
+		timestamp = time.Unix(millis/1000, (millis%1000)*int64(time.Millisecond))
+	}
+
+	*t.Time = timestamp
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/txn_offset_commit_request.go b/vendor/github.com/Shopify/sarama/txn_offset_commit_request.go
new file mode 100644
index 0000000..71e95b8
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/txn_offset_commit_request.go
@@ -0,0 +1,126 @@
+package sarama
+
+type TxnOffsetCommitRequest struct {
+	TransactionalID string
+	GroupID         string
+	ProducerID      int64
+	ProducerEpoch   int16
+	Topics          map[string][]*PartitionOffsetMetadata
+}
+
+func (t *TxnOffsetCommitRequest) encode(pe packetEncoder) error {
+	if err := pe.putString(t.TransactionalID); err != nil {
+		return err
+	}
+	if err := pe.putString(t.GroupID); err != nil {
+		return err
+	}
+	pe.putInt64(t.ProducerID)
+	pe.putInt16(t.ProducerEpoch)
+
+	if err := pe.putArrayLength(len(t.Topics)); err != nil {
+		return err
+	}
+	for topic, partitions := range t.Topics {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(partitions)); err != nil {
+			return err
+		}
+		for _, partition := range partitions {
+			if err := partition.encode(pe); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (t *TxnOffsetCommitRequest) decode(pd packetDecoder, version int16) (err error) {
+	if t.TransactionalID, err = pd.getString(); err != nil {
+		return err
+	}
+	if t.GroupID, err = pd.getString(); err != nil {
+		return err
+	}
+	if t.ProducerID, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if t.ProducerEpoch, err = pd.getInt16(); err != nil {
+		return err
+	}
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	t.Topics = make(map[string][]*PartitionOffsetMetadata)
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		m, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		t.Topics[topic] = make([]*PartitionOffsetMetadata, m)
+
+		for j := 0; j < m; j++ {
+			partitionOffsetMetadata := new(PartitionOffsetMetadata)
+			if err := partitionOffsetMetadata.decode(pd, version); err != nil {
+				return err
+			}
+			t.Topics[topic][j] = partitionOffsetMetadata
+		}
+	}
+
+	return nil
+}
+
+func (a *TxnOffsetCommitRequest) key() int16 {
+	return 28
+}
+
+func (a *TxnOffsetCommitRequest) version() int16 {
+	return 0
+}
+
+func (a *TxnOffsetCommitRequest) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
+
+type PartitionOffsetMetadata struct {
+	Partition int32
+	Offset    int64
+	Metadata  *string
+}
+
+func (p *PartitionOffsetMetadata) encode(pe packetEncoder) error {
+	pe.putInt32(p.Partition)
+	pe.putInt64(p.Offset)
+	if err := pe.putNullableString(p.Metadata); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (p *PartitionOffsetMetadata) decode(pd packetDecoder, version int16) (err error) {
+	if p.Partition, err = pd.getInt32(); err != nil {
+		return err
+	}
+	if p.Offset, err = pd.getInt64(); err != nil {
+		return err
+	}
+	if p.Metadata, err = pd.getNullableString(); err != nil {
+		return err
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/Shopify/sarama/txn_offset_commit_response.go b/vendor/github.com/Shopify/sarama/txn_offset_commit_response.go
new file mode 100644
index 0000000..6c980f4
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/txn_offset_commit_response.go
@@ -0,0 +1,83 @@
+package sarama
+
+import (
+	"time"
+)
+
+type TxnOffsetCommitResponse struct {
+	ThrottleTime time.Duration
+	Topics       map[string][]*PartitionError
+}
+
+func (t *TxnOffsetCommitResponse) encode(pe packetEncoder) error {
+	pe.putInt32(int32(t.ThrottleTime / time.Millisecond))
+	if err := pe.putArrayLength(len(t.Topics)); err != nil {
+		return err
+	}
+
+	for topic, e := range t.Topics {
+		if err := pe.putString(topic); err != nil {
+			return err
+		}
+		if err := pe.putArrayLength(len(e)); err != nil {
+			return err
+		}
+		for _, partitionError := range e {
+			if err := partitionError.encode(pe); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (t *TxnOffsetCommitResponse) decode(pd packetDecoder, version int16) (err error) {
+	throttleTime, err := pd.getInt32()
+	if err != nil {
+		return err
+	}
+	t.ThrottleTime = time.Duration(throttleTime) * time.Millisecond
+
+	n, err := pd.getArrayLength()
+	if err != nil {
+		return err
+	}
+
+	t.Topics = make(map[string][]*PartitionError)
+
+	for i := 0; i < n; i++ {
+		topic, err := pd.getString()
+		if err != nil {
+			return err
+		}
+
+		m, err := pd.getArrayLength()
+		if err != nil {
+			return err
+		}
+
+		t.Topics[topic] = make([]*PartitionError, m)
+
+		for j := 0; j < m; j++ {
+			t.Topics[topic][j] = new(PartitionError)
+			if err := t.Topics[topic][j].decode(pd, version); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+func (a *TxnOffsetCommitResponse) key() int16 {
+	return 28
+}
+
+func (a *TxnOffsetCommitResponse) version() int16 {
+	return 0
+}
+
+func (a *TxnOffsetCommitResponse) requiredVersion() KafkaVersion {
+	return V0_11_0_0
+}
diff --git a/vendor/github.com/Shopify/sarama/utils.go b/vendor/github.com/Shopify/sarama/utils.go
new file mode 100644
index 0000000..7dcbf03
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/utils.go
@@ -0,0 +1,220 @@
+package sarama
+
+import (
+	"bufio"
+	"fmt"
+	"net"
+	"regexp"
+)
+
+type none struct{}
+
+// make []int32 sortable so we can sort partition numbers
+type int32Slice []int32
+
+func (slice int32Slice) Len() int {
+	return len(slice)
+}
+
+func (slice int32Slice) Less(i, j int) bool {
+	return slice[i] < slice[j]
+}
+
+func (slice int32Slice) Swap(i, j int) {
+	slice[i], slice[j] = slice[j], slice[i]
+}
+
+func dupInt32Slice(input []int32) []int32 {
+	ret := make([]int32, 0, len(input))
+	for _, val := range input {
+		ret = append(ret, val)
+	}
+	return ret
+}
+
+func withRecover(fn func()) {
+	defer func() {
+		handler := PanicHandler
+		if handler != nil {
+			if err := recover(); err != nil {
+				handler(err)
+			}
+		}
+	}()
+
+	fn()
+}
+
+func safeAsyncClose(b *Broker) {
+	tmp := b // local var prevents clobbering in goroutine
+	go withRecover(func() {
+		if connected, _ := tmp.Connected(); connected {
+			if err := tmp.Close(); err != nil {
+				Logger.Println("Error closing broker", tmp.ID(), ":", err)
+			}
+		}
+	})
+}
+
+// Encoder is a simple interface for any type that can be encoded as an array of bytes
+// in order to be sent as the key or value of a Kafka message. Length() is provided as an
+// optimization, and must return the same as len() on the result of Encode().
+type Encoder interface {
+	Encode() ([]byte, error)
+	Length() int
+}
+
+// make strings and byte slices encodable for convenience so they can be used as keys
+// and/or values in kafka messages
+
+// StringEncoder implements the Encoder interface for Go strings so that they can be used
+// as the Key or Value in a ProducerMessage.
+type StringEncoder string
+
+func (s StringEncoder) Encode() ([]byte, error) {
+	return []byte(s), nil
+}
+
+func (s StringEncoder) Length() int {
+	return len(s)
+}
+
+// ByteEncoder implements the Encoder interface for Go byte slices so that they can be used
+// as the Key or Value in a ProducerMessage.
+type ByteEncoder []byte
+
+func (b ByteEncoder) Encode() ([]byte, error) {
+	return b, nil
+}
+
+func (b ByteEncoder) Length() int {
+	return len(b)
+}
+
+// bufConn wraps a net.Conn with a buffer for reads to reduce the number of
+// reads that trigger syscalls.
+type bufConn struct {
+	net.Conn
+	buf *bufio.Reader
+}
+
+func newBufConn(conn net.Conn) *bufConn {
+	return &bufConn{
+		Conn: conn,
+		buf:  bufio.NewReader(conn),
+	}
+}
+
+func (bc *bufConn) Read(b []byte) (n int, err error) {
+	return bc.buf.Read(b)
+}
+
+// KafkaVersion instances represent versions of the upstream Kafka broker.
+type KafkaVersion struct {
+	// it's a struct rather than just typing the array directly to make it opaque and stop people
+	// generating their own arbitrary versions
+	version [4]uint
+}
+
+func newKafkaVersion(major, minor, veryMinor, patch uint) KafkaVersion {
+	return KafkaVersion{
+		version: [4]uint{major, minor, veryMinor, patch},
+	}
+}
+
+// IsAtLeast return true if and only if the version it is called on is
+// greater than or equal to the version passed in:
+//    V1.IsAtLeast(V2) // false
+//    V2.IsAtLeast(V1) // true
+func (v KafkaVersion) IsAtLeast(other KafkaVersion) bool {
+	for i := range v.version {
+		if v.version[i] > other.version[i] {
+			return true
+		} else if v.version[i] < other.version[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// Effective constants defining the supported kafka versions.
+var (
+	V0_8_2_0  = newKafkaVersion(0, 8, 2, 0)
+	V0_8_2_1  = newKafkaVersion(0, 8, 2, 1)
+	V0_8_2_2  = newKafkaVersion(0, 8, 2, 2)
+	V0_9_0_0  = newKafkaVersion(0, 9, 0, 0)
+	V0_9_0_1  = newKafkaVersion(0, 9, 0, 1)
+	V0_10_0_0 = newKafkaVersion(0, 10, 0, 0)
+	V0_10_0_1 = newKafkaVersion(0, 10, 0, 1)
+	V0_10_1_0 = newKafkaVersion(0, 10, 1, 0)
+	V0_10_1_1 = newKafkaVersion(0, 10, 1, 1)
+	V0_10_2_0 = newKafkaVersion(0, 10, 2, 0)
+	V0_10_2_1 = newKafkaVersion(0, 10, 2, 1)
+	V0_11_0_0 = newKafkaVersion(0, 11, 0, 0)
+	V0_11_0_1 = newKafkaVersion(0, 11, 0, 1)
+	V0_11_0_2 = newKafkaVersion(0, 11, 0, 2)
+	V1_0_0_0  = newKafkaVersion(1, 0, 0, 0)
+	V1_1_0_0  = newKafkaVersion(1, 1, 0, 0)
+	V1_1_1_0  = newKafkaVersion(1, 1, 1, 0)
+	V2_0_0_0  = newKafkaVersion(2, 0, 0, 0)
+	V2_0_1_0  = newKafkaVersion(2, 0, 1, 0)
+	V2_1_0_0  = newKafkaVersion(2, 1, 0, 0)
+
+	SupportedVersions = []KafkaVersion{
+		V0_8_2_0,
+		V0_8_2_1,
+		V0_8_2_2,
+		V0_9_0_0,
+		V0_9_0_1,
+		V0_10_0_0,
+		V0_10_0_1,
+		V0_10_1_0,
+		V0_10_1_1,
+		V0_10_2_0,
+		V0_10_2_1,
+		V0_11_0_0,
+		V0_11_0_1,
+		V0_11_0_2,
+		V1_0_0_0,
+		V1_1_0_0,
+		V1_1_1_0,
+		V2_0_0_0,
+		V2_0_1_0,
+		V2_1_0_0,
+	}
+	MinVersion = V0_8_2_0
+	MaxVersion = V2_1_0_0
+)
+
+func ParseKafkaVersion(s string) (KafkaVersion, error) {
+	if len(s) < 5 {
+		return MinVersion, fmt.Errorf("invalid version `%s`", s)
+	}
+	var major, minor, veryMinor, patch uint
+	var err error
+	if s[0] == '0' {
+		err = scanKafkaVersion(s, `^0\.\d+\.\d+\.\d+$`, "0.%d.%d.%d", [3]*uint{&minor, &veryMinor, &patch})
+	} else {
+		err = scanKafkaVersion(s, `^\d+\.\d+\.\d+$`, "%d.%d.%d", [3]*uint{&major, &minor, &veryMinor})
+	}
+	if err != nil {
+		return MinVersion, err
+	}
+	return newKafkaVersion(major, minor, veryMinor, patch), nil
+}
+
+func scanKafkaVersion(s string, pattern string, format string, v [3]*uint) error {
+	if !regexp.MustCompile(pattern).MatchString(s) {
+		return fmt.Errorf("invalid version `%s`", s)
+	}
+	_, err := fmt.Sscanf(s, format, v[0], v[1], v[2])
+	return err
+}
+
+func (v KafkaVersion) String() string {
+	if v.version[0] == 0 {
+		return fmt.Sprintf("0.%d.%d.%d", v.version[1], v.version[2], v.version[3])
+	}
+
+	return fmt.Sprintf("%d.%d.%d", v.version[0], v.version[1], v.version[2])
+}
diff --git a/vendor/github.com/Shopify/sarama/zstd_cgo.go b/vendor/github.com/Shopify/sarama/zstd_cgo.go
new file mode 100644
index 0000000..f5ccb31
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/zstd_cgo.go
@@ -0,0 +1,13 @@
+// +build cgo
+
+package sarama
+
+import "github.com/DataDog/zstd"
+
+func zstdDecompress(dst, src []byte) ([]byte, error) {
+	return zstd.Decompress(dst, src)
+}
+
+func zstdCompressLevel(dst, src []byte, level int) ([]byte, error) {
+	return zstd.CompressLevel(dst, src, level)
+}
diff --git a/vendor/github.com/Shopify/sarama/zstd_fallback.go b/vendor/github.com/Shopify/sarama/zstd_fallback.go
new file mode 100644
index 0000000..381a56b
--- /dev/null
+++ b/vendor/github.com/Shopify/sarama/zstd_fallback.go
@@ -0,0 +1,17 @@
+// +build !cgo
+
+package sarama
+
+import (
+	"errors"
+)
+
+var errZstdCgo = errors.New("zstd compression requires building with cgo enabled")
+
+func zstdDecompress(dst, src []byte) ([]byte, error) {
+	return nil, errZstdCgo
+}
+
+func zstdCompressLevel(dst, src []byte, level int) ([]byte, error) {
+	return nil, errZstdCgo
+}
diff --git a/vendor/github.com/armon/go-metrics/.gitignore b/vendor/github.com/armon/go-metrics/.gitignore
new file mode 100644
index 0000000..8c03ec1
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/.gitignore
@@ -0,0 +1,24 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+
+# Folders
+_obj
+_test
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+
+/metrics.out
diff --git a/vendor/github.com/armon/go-metrics/LICENSE b/vendor/github.com/armon/go-metrics/LICENSE
new file mode 100644
index 0000000..106569e
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/LICENSE
@@ -0,0 +1,20 @@
+The MIT License (MIT)
+
+Copyright (c) 2013 Armon Dadgar
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+the Software, and to permit persons to whom the Software is furnished to do so,
+subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/vendor/github.com/armon/go-metrics/README.md b/vendor/github.com/armon/go-metrics/README.md
new file mode 100644
index 0000000..aa73348
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/README.md
@@ -0,0 +1,91 @@
+go-metrics
+==========
+
+This library provides a `metrics` package which can be used to instrument code,
+expose application metrics, and profile runtime performance in a flexible manner.
+
+Current API: [![GoDoc](https://godoc.org/github.com/armon/go-metrics?status.svg)](https://godoc.org/github.com/armon/go-metrics)
+
+Sinks
+-----
+
+The `metrics` package makes use of a `MetricSink` interface to support delivery
+to any type of backend. Currently the following sinks are provided:
+
+* StatsiteSink : Sinks to a [statsite](https://github.com/armon/statsite/) instance (TCP)
+* StatsdSink: Sinks to a [StatsD](https://github.com/etsy/statsd/) / statsite instance (UDP)
+* PrometheusSink: Sinks to a [Prometheus](http://prometheus.io/) metrics endpoint (exposed via HTTP for scrapes)
+* InmemSink : Provides in-memory aggregation, can be used to export stats
+* FanoutSink : Sinks to multiple sinks. Enables writing to multiple statsite instances for example.
+* BlackholeSink : Sinks to nowhere
+
+In addition to the sinks, the `InmemSignal` can be used to catch a signal,
+and dump a formatted output of recent metrics. For example, when a process gets
+a SIGUSR1, it can dump to stderr recent performance metrics for debugging.
+
+Labels
+------
+
+Most metrics do have an equivalent ending with `WithLabels`, such methods
+allow to push metrics with labels and use some features of underlying Sinks
+(ex: translated into Prometheus labels).
+
+Since some of these labels may increase greatly cardinality of metrics, the
+library allow to filter labels using a blacklist/whitelist filtering system
+which is global to all metrics.
+
+* If `Config.AllowedLabels` is not nil, then only labels specified in this value will be sent to underlying Sink, otherwise, all labels are sent by default.
+* If `Config.BlockedLabels` is not nil, any label specified in this value will not be sent to underlying Sinks.
+
+By default, both `Config.AllowedLabels` and `Config.BlockedLabels` are nil, meaning that
+no tags are filetered at all, but it allow to a user to globally block some tags with high
+cardinality at application level.
+
+Examples
+--------
+
+Here is an example of using the package:
+
+```go
+func SlowMethod() {
+    // Profiling the runtime of a method
+    defer metrics.MeasureSince([]string{"SlowMethod"}, time.Now())
+}
+
+// Configure a statsite sink as the global metrics sink
+sink, _ := metrics.NewStatsiteSink("statsite:8125")
+metrics.NewGlobal(metrics.DefaultConfig("service-name"), sink)
+
+// Emit a Key/Value pair
+metrics.EmitKey([]string{"questions", "meaning of life"}, 42)
+```
+
+Here is an example of setting up a signal handler:
+
+```go
+// Setup the inmem sink and signal handler
+inm := metrics.NewInmemSink(10*time.Second, time.Minute)
+sig := metrics.DefaultInmemSignal(inm)
+metrics.NewGlobal(metrics.DefaultConfig("service-name"), inm)
+
+// Run some code
+inm.SetGauge([]string{"foo"}, 42)
+inm.EmitKey([]string{"bar"}, 30)
+
+inm.IncrCounter([]string{"baz"}, 42)
+inm.IncrCounter([]string{"baz"}, 1)
+inm.IncrCounter([]string{"baz"}, 80)
+
+inm.AddSample([]string{"method", "wow"}, 42)
+inm.AddSample([]string{"method", "wow"}, 100)
+inm.AddSample([]string{"method", "wow"}, 22)
+
+....
+```
+
+When a signal comes in, output like the following will be dumped to stderr:
+
+    [2014-01-28 14:57:33.04 -0800 PST][G] 'foo': 42.000
+    [2014-01-28 14:57:33.04 -0800 PST][P] 'bar': 30.000
+    [2014-01-28 14:57:33.04 -0800 PST][C] 'baz': Count: 3 Min: 1.000 Mean: 41.000 Max: 80.000 Stddev: 39.509
+    [2014-01-28 14:57:33.04 -0800 PST][S] 'method.wow': Count: 3 Min: 22.000 Mean: 54.667 Max: 100.000 Stddev: 40.513
\ No newline at end of file
diff --git a/vendor/github.com/armon/go-metrics/const_unix.go b/vendor/github.com/armon/go-metrics/const_unix.go
new file mode 100644
index 0000000..31098dd
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/const_unix.go
@@ -0,0 +1,12 @@
+// +build !windows
+
+package metrics
+
+import (
+	"syscall"
+)
+
+const (
+	// DefaultSignal is used with DefaultInmemSignal
+	DefaultSignal = syscall.SIGUSR1
+)
diff --git a/vendor/github.com/armon/go-metrics/const_windows.go b/vendor/github.com/armon/go-metrics/const_windows.go
new file mode 100644
index 0000000..38136af
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/const_windows.go
@@ -0,0 +1,13 @@
+// +build windows
+
+package metrics
+
+import (
+	"syscall"
+)
+
+const (
+	// DefaultSignal is used with DefaultInmemSignal
+	// Windows has no SIGUSR1, use SIGBREAK
+	DefaultSignal = syscall.Signal(21)
+)
diff --git a/vendor/github.com/armon/go-metrics/inmem.go b/vendor/github.com/armon/go-metrics/inmem.go
new file mode 100644
index 0000000..4e2d6a7
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/inmem.go
@@ -0,0 +1,348 @@
+package metrics
+
+import (
+	"bytes"
+	"fmt"
+	"math"
+	"net/url"
+	"strings"
+	"sync"
+	"time"
+)
+
+// InmemSink provides a MetricSink that does in-memory aggregation
+// without sending metrics over a network. It can be embedded within
+// an application to provide profiling information.
+type InmemSink struct {
+	// How long is each aggregation interval
+	interval time.Duration
+
+	// Retain controls how many metrics interval we keep
+	retain time.Duration
+
+	// maxIntervals is the maximum length of intervals.
+	// It is retain / interval.
+	maxIntervals int
+
+	// intervals is a slice of the retained intervals
+	intervals    []*IntervalMetrics
+	intervalLock sync.RWMutex
+
+	rateDenom float64
+}
+
+// IntervalMetrics stores the aggregated metrics
+// for a specific interval
+type IntervalMetrics struct {
+	sync.RWMutex
+
+	// The start time of the interval
+	Interval time.Time
+
+	// Gauges maps the key to the last set value
+	Gauges map[string]GaugeValue
+
+	// Points maps the string to the list of emitted values
+	// from EmitKey
+	Points map[string][]float32
+
+	// Counters maps the string key to a sum of the counter
+	// values
+	Counters map[string]SampledValue
+
+	// Samples maps the key to an AggregateSample,
+	// which has the rolled up view of a sample
+	Samples map[string]SampledValue
+}
+
+// NewIntervalMetrics creates a new IntervalMetrics for a given interval
+func NewIntervalMetrics(intv time.Time) *IntervalMetrics {
+	return &IntervalMetrics{
+		Interval: intv,
+		Gauges:   make(map[string]GaugeValue),
+		Points:   make(map[string][]float32),
+		Counters: make(map[string]SampledValue),
+		Samples:  make(map[string]SampledValue),
+	}
+}
+
+// AggregateSample is used to hold aggregate metrics
+// about a sample
+type AggregateSample struct {
+	Count       int       // The count of emitted pairs
+	Rate        float64   // The values rate per time unit (usually 1 second)
+	Sum         float64   // The sum of values
+	SumSq       float64   `json:"-"` // The sum of squared values
+	Min         float64   // Minimum value
+	Max         float64   // Maximum value
+	LastUpdated time.Time `json:"-"` // When value was last updated
+}
+
+// Computes a Stddev of the values
+func (a *AggregateSample) Stddev() float64 {
+	num := (float64(a.Count) * a.SumSq) - math.Pow(a.Sum, 2)
+	div := float64(a.Count * (a.Count - 1))
+	if div == 0 {
+		return 0
+	}
+	return math.Sqrt(num / div)
+}
+
+// Computes a mean of the values
+func (a *AggregateSample) Mean() float64 {
+	if a.Count == 0 {
+		return 0
+	}
+	return a.Sum / float64(a.Count)
+}
+
+// Ingest is used to update a sample
+func (a *AggregateSample) Ingest(v float64, rateDenom float64) {
+	a.Count++
+	a.Sum += v
+	a.SumSq += (v * v)
+	if v < a.Min || a.Count == 1 {
+		a.Min = v
+	}
+	if v > a.Max || a.Count == 1 {
+		a.Max = v
+	}
+	a.Rate = float64(a.Sum) / rateDenom
+	a.LastUpdated = time.Now()
+}
+
+func (a *AggregateSample) String() string {
+	if a.Count == 0 {
+		return "Count: 0"
+	} else if a.Stddev() == 0 {
+		return fmt.Sprintf("Count: %d Sum: %0.3f LastUpdated: %s", a.Count, a.Sum, a.LastUpdated)
+	} else {
+		return fmt.Sprintf("Count: %d Min: %0.3f Mean: %0.3f Max: %0.3f Stddev: %0.3f Sum: %0.3f LastUpdated: %s",
+			a.Count, a.Min, a.Mean(), a.Max, a.Stddev(), a.Sum, a.LastUpdated)
+	}
+}
+
+// NewInmemSinkFromURL creates an InmemSink from a URL. It is used
+// (and tested) from NewMetricSinkFromURL.
+func NewInmemSinkFromURL(u *url.URL) (MetricSink, error) {
+	params := u.Query()
+
+	interval, err := time.ParseDuration(params.Get("interval"))
+	if err != nil {
+		return nil, fmt.Errorf("Bad 'interval' param: %s", err)
+	}
+
+	retain, err := time.ParseDuration(params.Get("retain"))
+	if err != nil {
+		return nil, fmt.Errorf("Bad 'retain' param: %s", err)
+	}
+
+	return NewInmemSink(interval, retain), nil
+}
+
+// NewInmemSink is used to construct a new in-memory sink.
+// Uses an aggregation interval and maximum retention period.
+func NewInmemSink(interval, retain time.Duration) *InmemSink {
+	rateTimeUnit := time.Second
+	i := &InmemSink{
+		interval:     interval,
+		retain:       retain,
+		maxIntervals: int(retain / interval),
+		rateDenom:    float64(interval.Nanoseconds()) / float64(rateTimeUnit.Nanoseconds()),
+	}
+	i.intervals = make([]*IntervalMetrics, 0, i.maxIntervals)
+	return i
+}
+
+func (i *InmemSink) SetGauge(key []string, val float32) {
+	i.SetGaugeWithLabels(key, val, nil)
+}
+
+func (i *InmemSink) SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	k, name := i.flattenKeyLabels(key, labels)
+	intv := i.getInterval()
+
+	intv.Lock()
+	defer intv.Unlock()
+	intv.Gauges[k] = GaugeValue{Name: name, Value: val, Labels: labels}
+}
+
+func (i *InmemSink) EmitKey(key []string, val float32) {
+	k := i.flattenKey(key)
+	intv := i.getInterval()
+
+	intv.Lock()
+	defer intv.Unlock()
+	vals := intv.Points[k]
+	intv.Points[k] = append(vals, val)
+}
+
+func (i *InmemSink) IncrCounter(key []string, val float32) {
+	i.IncrCounterWithLabels(key, val, nil)
+}
+
+func (i *InmemSink) IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	k, name := i.flattenKeyLabels(key, labels)
+	intv := i.getInterval()
+
+	intv.Lock()
+	defer intv.Unlock()
+
+	agg, ok := intv.Counters[k]
+	if !ok {
+		agg = SampledValue{
+			Name:            name,
+			AggregateSample: &AggregateSample{},
+			Labels:          labels,
+		}
+		intv.Counters[k] = agg
+	}
+	agg.Ingest(float64(val), i.rateDenom)
+}
+
+func (i *InmemSink) AddSample(key []string, val float32) {
+	i.AddSampleWithLabels(key, val, nil)
+}
+
+func (i *InmemSink) AddSampleWithLabels(key []string, val float32, labels []Label) {
+	k, name := i.flattenKeyLabels(key, labels)
+	intv := i.getInterval()
+
+	intv.Lock()
+	defer intv.Unlock()
+
+	agg, ok := intv.Samples[k]
+	if !ok {
+		agg = SampledValue{
+			Name:            name,
+			AggregateSample: &AggregateSample{},
+			Labels:          labels,
+		}
+		intv.Samples[k] = agg
+	}
+	agg.Ingest(float64(val), i.rateDenom)
+}
+
+// Data is used to retrieve all the aggregated metrics
+// Intervals may be in use, and a read lock should be acquired
+func (i *InmemSink) Data() []*IntervalMetrics {
+	// Get the current interval, forces creation
+	i.getInterval()
+
+	i.intervalLock.RLock()
+	defer i.intervalLock.RUnlock()
+
+	n := len(i.intervals)
+	intervals := make([]*IntervalMetrics, n)
+
+	copy(intervals[:n-1], i.intervals[:n-1])
+	current := i.intervals[n-1]
+
+	// make its own copy for current interval
+	intervals[n-1] = &IntervalMetrics{}
+	copyCurrent := intervals[n-1]
+	current.RLock()
+	*copyCurrent = *current
+
+	copyCurrent.Gauges = make(map[string]GaugeValue, len(current.Gauges))
+	for k, v := range current.Gauges {
+		copyCurrent.Gauges[k] = v
+	}
+	// saved values will be not change, just copy its link
+	copyCurrent.Points = make(map[string][]float32, len(current.Points))
+	for k, v := range current.Points {
+		copyCurrent.Points[k] = v
+	}
+	copyCurrent.Counters = make(map[string]SampledValue, len(current.Counters))
+	for k, v := range current.Counters {
+		copyCurrent.Counters[k] = v
+	}
+	copyCurrent.Samples = make(map[string]SampledValue, len(current.Samples))
+	for k, v := range current.Samples {
+		copyCurrent.Samples[k] = v
+	}
+	current.RUnlock()
+
+	return intervals
+}
+
+func (i *InmemSink) getExistingInterval(intv time.Time) *IntervalMetrics {
+	i.intervalLock.RLock()
+	defer i.intervalLock.RUnlock()
+
+	n := len(i.intervals)
+	if n > 0 && i.intervals[n-1].Interval == intv {
+		return i.intervals[n-1]
+	}
+	return nil
+}
+
+func (i *InmemSink) createInterval(intv time.Time) *IntervalMetrics {
+	i.intervalLock.Lock()
+	defer i.intervalLock.Unlock()
+
+	// Check for an existing interval
+	n := len(i.intervals)
+	if n > 0 && i.intervals[n-1].Interval == intv {
+		return i.intervals[n-1]
+	}
+
+	// Add the current interval
+	current := NewIntervalMetrics(intv)
+	i.intervals = append(i.intervals, current)
+	n++
+
+	// Truncate the intervals if they are too long
+	if n >= i.maxIntervals {
+		copy(i.intervals[0:], i.intervals[n-i.maxIntervals:])
+		i.intervals = i.intervals[:i.maxIntervals]
+	}
+	return current
+}
+
+// getInterval returns the current interval to write to
+func (i *InmemSink) getInterval() *IntervalMetrics {
+	intv := time.Now().Truncate(i.interval)
+	if m := i.getExistingInterval(intv); m != nil {
+		return m
+	}
+	return i.createInterval(intv)
+}
+
+// Flattens the key for formatting, removes spaces
+func (i *InmemSink) flattenKey(parts []string) string {
+	buf := &bytes.Buffer{}
+	replacer := strings.NewReplacer(" ", "_")
+
+	if len(parts) > 0 {
+		replacer.WriteString(buf, parts[0])
+	}
+	for _, part := range parts[1:] {
+		replacer.WriteString(buf, ".")
+		replacer.WriteString(buf, part)
+	}
+
+	return buf.String()
+}
+
+// Flattens the key for formatting along with its labels, removes spaces
+func (i *InmemSink) flattenKeyLabels(parts []string, labels []Label) (string, string) {
+	buf := &bytes.Buffer{}
+	replacer := strings.NewReplacer(" ", "_")
+
+	if len(parts) > 0 {
+		replacer.WriteString(buf, parts[0])
+	}
+	for _, part := range parts[1:] {
+		replacer.WriteString(buf, ".")
+		replacer.WriteString(buf, part)
+	}
+
+	key := buf.String()
+
+	for _, label := range labels {
+		replacer.WriteString(buf, fmt.Sprintf(";%s=%s", label.Name, label.Value))
+	}
+
+	return buf.String(), key
+}
diff --git a/vendor/github.com/armon/go-metrics/inmem_endpoint.go b/vendor/github.com/armon/go-metrics/inmem_endpoint.go
new file mode 100644
index 0000000..504f1b3
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/inmem_endpoint.go
@@ -0,0 +1,118 @@
+package metrics
+
+import (
+	"fmt"
+	"net/http"
+	"sort"
+	"time"
+)
+
+// MetricsSummary holds a roll-up of metrics info for a given interval
+type MetricsSummary struct {
+	Timestamp string
+	Gauges    []GaugeValue
+	Points    []PointValue
+	Counters  []SampledValue
+	Samples   []SampledValue
+}
+
+type GaugeValue struct {
+	Name  string
+	Hash  string `json:"-"`
+	Value float32
+
+	Labels        []Label           `json:"-"`
+	DisplayLabels map[string]string `json:"Labels"`
+}
+
+type PointValue struct {
+	Name   string
+	Points []float32
+}
+
+type SampledValue struct {
+	Name string
+	Hash string `json:"-"`
+	*AggregateSample
+	Mean   float64
+	Stddev float64
+
+	Labels        []Label           `json:"-"`
+	DisplayLabels map[string]string `json:"Labels"`
+}
+
+// DisplayMetrics returns a summary of the metrics from the most recent finished interval.
+func (i *InmemSink) DisplayMetrics(resp http.ResponseWriter, req *http.Request) (interface{}, error) {
+	data := i.Data()
+
+	var interval *IntervalMetrics
+	n := len(data)
+	switch {
+	case n == 0:
+		return nil, fmt.Errorf("no metric intervals have been initialized yet")
+	case n == 1:
+		// Show the current interval if it's all we have
+		interval = i.intervals[0]
+	default:
+		// Show the most recent finished interval if we have one
+		interval = i.intervals[n-2]
+	}
+
+	summary := MetricsSummary{
+		Timestamp: interval.Interval.Round(time.Second).UTC().String(),
+		Gauges:    make([]GaugeValue, 0, len(interval.Gauges)),
+		Points:    make([]PointValue, 0, len(interval.Points)),
+	}
+
+	// Format and sort the output of each metric type, so it gets displayed in a
+	// deterministic order.
+	for name, points := range interval.Points {
+		summary.Points = append(summary.Points, PointValue{name, points})
+	}
+	sort.Slice(summary.Points, func(i, j int) bool {
+		return summary.Points[i].Name < summary.Points[j].Name
+	})
+
+	for hash, value := range interval.Gauges {
+		value.Hash = hash
+		value.DisplayLabels = make(map[string]string)
+		for _, label := range value.Labels {
+			value.DisplayLabels[label.Name] = label.Value
+		}
+		value.Labels = nil
+
+		summary.Gauges = append(summary.Gauges, value)
+	}
+	sort.Slice(summary.Gauges, func(i, j int) bool {
+		return summary.Gauges[i].Hash < summary.Gauges[j].Hash
+	})
+
+	summary.Counters = formatSamples(interval.Counters)
+	summary.Samples = formatSamples(interval.Samples)
+
+	return summary, nil
+}
+
+func formatSamples(source map[string]SampledValue) []SampledValue {
+	output := make([]SampledValue, 0, len(source))
+	for hash, sample := range source {
+		displayLabels := make(map[string]string)
+		for _, label := range sample.Labels {
+			displayLabels[label.Name] = label.Value
+		}
+
+		output = append(output, SampledValue{
+			Name:            sample.Name,
+			Hash:            hash,
+			AggregateSample: sample.AggregateSample,
+			Mean:            sample.AggregateSample.Mean(),
+			Stddev:          sample.AggregateSample.Stddev(),
+			DisplayLabels:   displayLabels,
+		})
+	}
+	sort.Slice(output, func(i, j int) bool {
+		return output[i].Hash < output[j].Hash
+	})
+
+	return output
+}
diff --git a/vendor/github.com/armon/go-metrics/inmem_signal.go b/vendor/github.com/armon/go-metrics/inmem_signal.go
new file mode 100644
index 0000000..0937f4a
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/inmem_signal.go
@@ -0,0 +1,117 @@
+package metrics
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+	"os/signal"
+	"strings"
+	"sync"
+	"syscall"
+)
+
+// InmemSignal is used to listen for a given signal, and when received,
+// to dump the current metrics from the InmemSink to an io.Writer
+type InmemSignal struct {
+	signal syscall.Signal
+	inm    *InmemSink
+	w      io.Writer
+	sigCh  chan os.Signal
+
+	stop     bool
+	stopCh   chan struct{}
+	stopLock sync.Mutex
+}
+
+// NewInmemSignal creates a new InmemSignal which listens for a given signal,
+// and dumps the current metrics out to a writer
+func NewInmemSignal(inmem *InmemSink, sig syscall.Signal, w io.Writer) *InmemSignal {
+	i := &InmemSignal{
+		signal: sig,
+		inm:    inmem,
+		w:      w,
+		sigCh:  make(chan os.Signal, 1),
+		stopCh: make(chan struct{}),
+	}
+	signal.Notify(i.sigCh, sig)
+	go i.run()
+	return i
+}
+
+// DefaultInmemSignal returns a new InmemSignal that responds to SIGUSR1
+// and writes output to stderr. Windows uses SIGBREAK
+func DefaultInmemSignal(inmem *InmemSink) *InmemSignal {
+	return NewInmemSignal(inmem, DefaultSignal, os.Stderr)
+}
+
+// Stop is used to stop the InmemSignal from listening
+func (i *InmemSignal) Stop() {
+	i.stopLock.Lock()
+	defer i.stopLock.Unlock()
+
+	if i.stop {
+		return
+	}
+	i.stop = true
+	close(i.stopCh)
+	signal.Stop(i.sigCh)
+}
+
+// run is a long running routine that handles signals
+func (i *InmemSignal) run() {
+	for {
+		select {
+		case <-i.sigCh:
+			i.dumpStats()
+		case <-i.stopCh:
+			return
+		}
+	}
+}
+
+// dumpStats is used to dump the data to output writer
+func (i *InmemSignal) dumpStats() {
+	buf := bytes.NewBuffer(nil)
+
+	data := i.inm.Data()
+	// Skip the last period which is still being aggregated
+	for j := 0; j < len(data)-1; j++ {
+		intv := data[j]
+		intv.RLock()
+		for _, val := range intv.Gauges {
+			name := i.flattenLabels(val.Name, val.Labels)
+			fmt.Fprintf(buf, "[%v][G] '%s': %0.3f\n", intv.Interval, name, val.Value)
+		}
+		for name, vals := range intv.Points {
+			for _, val := range vals {
+				fmt.Fprintf(buf, "[%v][P] '%s': %0.3f\n", intv.Interval, name, val)
+			}
+		}
+		for _, agg := range intv.Counters {
+			name := i.flattenLabels(agg.Name, agg.Labels)
+			fmt.Fprintf(buf, "[%v][C] '%s': %s\n", intv.Interval, name, agg.AggregateSample)
+		}
+		for _, agg := range intv.Samples {
+			name := i.flattenLabels(agg.Name, agg.Labels)
+			fmt.Fprintf(buf, "[%v][S] '%s': %s\n", intv.Interval, name, agg.AggregateSample)
+		}
+		intv.RUnlock()
+	}
+
+	// Write out the bytes
+	i.w.Write(buf.Bytes())
+}
+
+// Flattens the key for formatting along with its labels, removes spaces
+func (i *InmemSignal) flattenLabels(name string, labels []Label) string {
+	buf := bytes.NewBufferString(name)
+	replacer := strings.NewReplacer(" ", "_", ":", "_")
+
+	for _, label := range labels {
+		replacer.WriteString(buf, ".")
+		replacer.WriteString(buf, label.Value)
+	}
+
+	return buf.String()
+}
diff --git a/vendor/github.com/armon/go-metrics/metrics.go b/vendor/github.com/armon/go-metrics/metrics.go
new file mode 100644
index 0000000..cf9def7
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/metrics.go
@@ -0,0 +1,278 @@
+package metrics
+
+import (
+	"runtime"
+	"strings"
+	"time"
+
+	"github.com/hashicorp/go-immutable-radix"
+)
+
+type Label struct {
+	Name  string
+	Value string
+}
+
+func (m *Metrics) SetGauge(key []string, val float32) {
+	m.SetGaugeWithLabels(key, val, nil)
+}
+
+func (m *Metrics) SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	if m.HostName != "" {
+		if m.EnableHostnameLabel {
+			labels = append(labels, Label{"host", m.HostName})
+		} else if m.EnableHostname {
+			key = insert(0, m.HostName, key)
+		}
+	}
+	if m.EnableTypePrefix {
+		key = insert(0, "gauge", key)
+	}
+	if m.ServiceName != "" {
+		if m.EnableServiceLabel {
+			labels = append(labels, Label{"service", m.ServiceName})
+		} else {
+			key = insert(0, m.ServiceName, key)
+		}
+	}
+	allowed, labelsFiltered := m.allowMetric(key, labels)
+	if !allowed {
+		return
+	}
+	m.sink.SetGaugeWithLabels(key, val, labelsFiltered)
+}
+
+func (m *Metrics) EmitKey(key []string, val float32) {
+	if m.EnableTypePrefix {
+		key = insert(0, "kv", key)
+	}
+	if m.ServiceName != "" {
+		key = insert(0, m.ServiceName, key)
+	}
+	allowed, _ := m.allowMetric(key, nil)
+	if !allowed {
+		return
+	}
+	m.sink.EmitKey(key, val)
+}
+
+func (m *Metrics) IncrCounter(key []string, val float32) {
+	m.IncrCounterWithLabels(key, val, nil)
+}
+
+func (m *Metrics) IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	if m.HostName != "" && m.EnableHostnameLabel {
+		labels = append(labels, Label{"host", m.HostName})
+	}
+	if m.EnableTypePrefix {
+		key = insert(0, "counter", key)
+	}
+	if m.ServiceName != "" {
+		if m.EnableServiceLabel {
+			labels = append(labels, Label{"service", m.ServiceName})
+		} else {
+			key = insert(0, m.ServiceName, key)
+		}
+	}
+	allowed, labelsFiltered := m.allowMetric(key, labels)
+	if !allowed {
+		return
+	}
+	m.sink.IncrCounterWithLabels(key, val, labelsFiltered)
+}
+
+func (m *Metrics) AddSample(key []string, val float32) {
+	m.AddSampleWithLabels(key, val, nil)
+}
+
+func (m *Metrics) AddSampleWithLabels(key []string, val float32, labels []Label) {
+	if m.HostName != "" && m.EnableHostnameLabel {
+		labels = append(labels, Label{"host", m.HostName})
+	}
+	if m.EnableTypePrefix {
+		key = insert(0, "sample", key)
+	}
+	if m.ServiceName != "" {
+		if m.EnableServiceLabel {
+			labels = append(labels, Label{"service", m.ServiceName})
+		} else {
+			key = insert(0, m.ServiceName, key)
+		}
+	}
+	allowed, labelsFiltered := m.allowMetric(key, labels)
+	if !allowed {
+		return
+	}
+	m.sink.AddSampleWithLabels(key, val, labelsFiltered)
+}
+
+func (m *Metrics) MeasureSince(key []string, start time.Time) {
+	m.MeasureSinceWithLabels(key, start, nil)
+}
+
+func (m *Metrics) MeasureSinceWithLabels(key []string, start time.Time, labels []Label) {
+	if m.HostName != "" && m.EnableHostnameLabel {
+		labels = append(labels, Label{"host", m.HostName})
+	}
+	if m.EnableTypePrefix {
+		key = insert(0, "timer", key)
+	}
+	if m.ServiceName != "" {
+		if m.EnableServiceLabel {
+			labels = append(labels, Label{"service", m.ServiceName})
+		} else {
+			key = insert(0, m.ServiceName, key)
+		}
+	}
+	allowed, labelsFiltered := m.allowMetric(key, labels)
+	if !allowed {
+		return
+	}
+	now := time.Now()
+	elapsed := now.Sub(start)
+	msec := float32(elapsed.Nanoseconds()) / float32(m.TimerGranularity)
+	m.sink.AddSampleWithLabels(key, msec, labelsFiltered)
+}
+
+// UpdateFilter overwrites the existing filter with the given rules.
+func (m *Metrics) UpdateFilter(allow, block []string) {
+	m.UpdateFilterAndLabels(allow, block, m.AllowedLabels, m.BlockedLabels)
+}
+
+// UpdateFilterAndLabels overwrites the existing filter with the given rules.
+func (m *Metrics) UpdateFilterAndLabels(allow, block, allowedLabels, blockedLabels []string) {
+	m.filterLock.Lock()
+	defer m.filterLock.Unlock()
+
+	m.AllowedPrefixes = allow
+	m.BlockedPrefixes = block
+
+	if allowedLabels == nil {
+		// Having a white list means we take only elements from it
+		m.allowedLabels = nil
+	} else {
+		m.allowedLabels = make(map[string]bool)
+		for _, v := range allowedLabels {
+			m.allowedLabels[v] = true
+		}
+	}
+	m.blockedLabels = make(map[string]bool)
+	for _, v := range blockedLabels {
+		m.blockedLabels[v] = true
+	}
+	m.AllowedLabels = allowedLabels
+	m.BlockedLabels = blockedLabels
+
+	m.filter = iradix.New()
+	for _, prefix := range m.AllowedPrefixes {
+		m.filter, _, _ = m.filter.Insert([]byte(prefix), true)
+	}
+	for _, prefix := range m.BlockedPrefixes {
+		m.filter, _, _ = m.filter.Insert([]byte(prefix), false)
+	}
+}
+
+// labelIsAllowed return true if a should be included in metric
+// the caller should lock m.filterLock while calling this method
+func (m *Metrics) labelIsAllowed(label *Label) bool {
+	labelName := (*label).Name
+	if m.blockedLabels != nil {
+		_, ok := m.blockedLabels[labelName]
+		if ok {
+			// If present, let's remove this label
+			return false
+		}
+	}
+	if m.allowedLabels != nil {
+		_, ok := m.allowedLabels[labelName]
+		return ok
+	}
+	// Allow by default
+	return true
+}
+
+// filterLabels return only allowed labels
+// the caller should lock m.filterLock while calling this method
+func (m *Metrics) filterLabels(labels []Label) []Label {
+	if labels == nil {
+		return nil
+	}
+	toReturn := labels[:0]
+	for _, label := range labels {
+		if m.labelIsAllowed(&label) {
+			toReturn = append(toReturn, label)
+		}
+	}
+	return toReturn
+}
+
+// Returns whether the metric should be allowed based on configured prefix filters
+// Also return the applicable labels
+func (m *Metrics) allowMetric(key []string, labels []Label) (bool, []Label) {
+	m.filterLock.RLock()
+	defer m.filterLock.RUnlock()
+
+	if m.filter == nil || m.filter.Len() == 0 {
+		return m.Config.FilterDefault, m.filterLabels(labels)
+	}
+
+	_, allowed, ok := m.filter.Root().LongestPrefix([]byte(strings.Join(key, ".")))
+	if !ok {
+		return m.Config.FilterDefault, m.filterLabels(labels)
+	}
+
+	return allowed.(bool), m.filterLabels(labels)
+}
+
+// Periodically collects runtime stats to publish
+func (m *Metrics) collectStats() {
+	for {
+		time.Sleep(m.ProfileInterval)
+		m.emitRuntimeStats()
+	}
+}
+
+// Emits various runtime statsitics
+func (m *Metrics) emitRuntimeStats() {
+	// Export number of Goroutines
+	numRoutines := runtime.NumGoroutine()
+	m.SetGauge([]string{"runtime", "num_goroutines"}, float32(numRoutines))
+
+	// Export memory stats
+	var stats runtime.MemStats
+	runtime.ReadMemStats(&stats)
+	m.SetGauge([]string{"runtime", "alloc_bytes"}, float32(stats.Alloc))
+	m.SetGauge([]string{"runtime", "sys_bytes"}, float32(stats.Sys))
+	m.SetGauge([]string{"runtime", "malloc_count"}, float32(stats.Mallocs))
+	m.SetGauge([]string{"runtime", "free_count"}, float32(stats.Frees))
+	m.SetGauge([]string{"runtime", "heap_objects"}, float32(stats.HeapObjects))
+	m.SetGauge([]string{"runtime", "total_gc_pause_ns"}, float32(stats.PauseTotalNs))
+	m.SetGauge([]string{"runtime", "total_gc_runs"}, float32(stats.NumGC))
+
+	// Export info about the last few GC runs
+	num := stats.NumGC
+
+	// Handle wrap around
+	if num < m.lastNumGC {
+		m.lastNumGC = 0
+	}
+
+	// Ensure we don't scan more than 256
+	if num-m.lastNumGC >= 256 {
+		m.lastNumGC = num - 255
+	}
+
+	for i := m.lastNumGC; i < num; i++ {
+		pause := stats.PauseNs[i%256]
+		m.AddSample([]string{"runtime", "gc_pause_ns"}, float32(pause))
+	}
+	m.lastNumGC = num
+}
+
+// Inserts a string value at an index into the slice
+func insert(i int, v string, s []string) []string {
+	s = append(s, "")
+	copy(s[i+1:], s[i:])
+	s[i] = v
+	return s
+}
diff --git a/vendor/github.com/armon/go-metrics/sink.go b/vendor/github.com/armon/go-metrics/sink.go
new file mode 100644
index 0000000..0b7d6e4
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/sink.go
@@ -0,0 +1,115 @@
+package metrics
+
+import (
+	"fmt"
+	"net/url"
+)
+
+// The MetricSink interface is used to transmit metrics information
+// to an external system
+type MetricSink interface {
+	// A Gauge should retain the last value it is set to
+	SetGauge(key []string, val float32)
+	SetGaugeWithLabels(key []string, val float32, labels []Label)
+
+	// Should emit a Key/Value pair for each call
+	EmitKey(key []string, val float32)
+
+	// Counters should accumulate values
+	IncrCounter(key []string, val float32)
+	IncrCounterWithLabels(key []string, val float32, labels []Label)
+
+	// Samples are for timing information, where quantiles are used
+	AddSample(key []string, val float32)
+	AddSampleWithLabels(key []string, val float32, labels []Label)
+}
+
+// BlackholeSink is used to just blackhole messages
+type BlackholeSink struct{}
+
+func (*BlackholeSink) SetGauge(key []string, val float32)                              {}
+func (*BlackholeSink) SetGaugeWithLabels(key []string, val float32, labels []Label)    {}
+func (*BlackholeSink) EmitKey(key []string, val float32)                               {}
+func (*BlackholeSink) IncrCounter(key []string, val float32)                           {}
+func (*BlackholeSink) IncrCounterWithLabels(key []string, val float32, labels []Label) {}
+func (*BlackholeSink) AddSample(key []string, val float32)                             {}
+func (*BlackholeSink) AddSampleWithLabels(key []string, val float32, labels []Label)   {}
+
+// FanoutSink is used to sink to fanout values to multiple sinks
+type FanoutSink []MetricSink
+
+func (fh FanoutSink) SetGauge(key []string, val float32) {
+	fh.SetGaugeWithLabels(key, val, nil)
+}
+
+func (fh FanoutSink) SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	for _, s := range fh {
+		s.SetGaugeWithLabels(key, val, labels)
+	}
+}
+
+func (fh FanoutSink) EmitKey(key []string, val float32) {
+	for _, s := range fh {
+		s.EmitKey(key, val)
+	}
+}
+
+func (fh FanoutSink) IncrCounter(key []string, val float32) {
+	fh.IncrCounterWithLabels(key, val, nil)
+}
+
+func (fh FanoutSink) IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	for _, s := range fh {
+		s.IncrCounterWithLabels(key, val, labels)
+	}
+}
+
+func (fh FanoutSink) AddSample(key []string, val float32) {
+	fh.AddSampleWithLabels(key, val, nil)
+}
+
+func (fh FanoutSink) AddSampleWithLabels(key []string, val float32, labels []Label) {
+	for _, s := range fh {
+		s.AddSampleWithLabels(key, val, labels)
+	}
+}
+
+// sinkURLFactoryFunc is an generic interface around the *SinkFromURL() function provided
+// by each sink type
+type sinkURLFactoryFunc func(*url.URL) (MetricSink, error)
+
+// sinkRegistry supports the generic NewMetricSink function by mapping URL
+// schemes to metric sink factory functions
+var sinkRegistry = map[string]sinkURLFactoryFunc{
+	"statsd":   NewStatsdSinkFromURL,
+	"statsite": NewStatsiteSinkFromURL,
+	"inmem":    NewInmemSinkFromURL,
+}
+
+// NewMetricSinkFromURL allows a generic URL input to configure any of the
+// supported sinks. The scheme of the URL identifies the type of the sink, the
+// and query parameters are used to set options.
+//
+// "statsd://" - Initializes a StatsdSink. The host and port are passed through
+// as the "addr" of the sink
+//
+// "statsite://" - Initializes a StatsiteSink. The host and port become the
+// "addr" of the sink
+//
+// "inmem://" - Initializes an InmemSink. The host and port are ignored. The
+// "interval" and "duration" query parameters must be specified with valid
+// durations, see NewInmemSink for details.
+func NewMetricSinkFromURL(urlStr string) (MetricSink, error) {
+	u, err := url.Parse(urlStr)
+	if err != nil {
+		return nil, err
+	}
+
+	sinkURLFactoryFunc := sinkRegistry[u.Scheme]
+	if sinkURLFactoryFunc == nil {
+		return nil, fmt.Errorf(
+			"cannot create metric sink, unrecognized sink name: %q", u.Scheme)
+	}
+
+	return sinkURLFactoryFunc(u)
+}
diff --git a/vendor/github.com/armon/go-metrics/start.go b/vendor/github.com/armon/go-metrics/start.go
new file mode 100644
index 0000000..32a28c4
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/start.go
@@ -0,0 +1,141 @@
+package metrics
+
+import (
+	"os"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/hashicorp/go-immutable-radix"
+)
+
+// Config is used to configure metrics settings
+type Config struct {
+	ServiceName          string        // Prefixed with keys to separate services
+	HostName             string        // Hostname to use. If not provided and EnableHostname, it will be os.Hostname
+	EnableHostname       bool          // Enable prefixing gauge values with hostname
+	EnableHostnameLabel  bool          // Enable adding hostname to labels
+	EnableServiceLabel   bool          // Enable adding service to labels
+	EnableRuntimeMetrics bool          // Enables profiling of runtime metrics (GC, Goroutines, Memory)
+	EnableTypePrefix     bool          // Prefixes key with a type ("counter", "gauge", "timer")
+	TimerGranularity     time.Duration // Granularity of timers.
+	ProfileInterval      time.Duration // Interval to profile runtime metrics
+
+	AllowedPrefixes []string // A list of metric prefixes to allow, with '.' as the separator
+	BlockedPrefixes []string // A list of metric prefixes to block, with '.' as the separator
+	AllowedLabels   []string // A list of metric labels to allow, with '.' as the separator
+	BlockedLabels   []string // A list of metric labels to block, with '.' as the separator
+	FilterDefault   bool     // Whether to allow metrics by default
+}
+
+// Metrics represents an instance of a metrics sink that can
+// be used to emit
+type Metrics struct {
+	Config
+	lastNumGC     uint32
+	sink          MetricSink
+	filter        *iradix.Tree
+	allowedLabels map[string]bool
+	blockedLabels map[string]bool
+	filterLock    sync.RWMutex // Lock filters and allowedLabels/blockedLabels access
+}
+
+// Shared global metrics instance
+var globalMetrics atomic.Value // *Metrics
+
+func init() {
+	// Initialize to a blackhole sink to avoid errors
+	globalMetrics.Store(&Metrics{sink: &BlackholeSink{}})
+}
+
+// DefaultConfig provides a sane default configuration
+func DefaultConfig(serviceName string) *Config {
+	c := &Config{
+		ServiceName:          serviceName, // Use client provided service
+		HostName:             "",
+		EnableHostname:       true,             // Enable hostname prefix
+		EnableRuntimeMetrics: true,             // Enable runtime profiling
+		EnableTypePrefix:     false,            // Disable type prefix
+		TimerGranularity:     time.Millisecond, // Timers are in milliseconds
+		ProfileInterval:      time.Second,      // Poll runtime every second
+		FilterDefault:        true,             // Don't filter metrics by default
+	}
+
+	// Try to get the hostname
+	name, _ := os.Hostname()
+	c.HostName = name
+	return c
+}
+
+// New is used to create a new instance of Metrics
+func New(conf *Config, sink MetricSink) (*Metrics, error) {
+	met := &Metrics{}
+	met.Config = *conf
+	met.sink = sink
+	met.UpdateFilterAndLabels(conf.AllowedPrefixes, conf.BlockedPrefixes, conf.AllowedLabels, conf.BlockedLabels)
+
+	// Start the runtime collector
+	if conf.EnableRuntimeMetrics {
+		go met.collectStats()
+	}
+	return met, nil
+}
+
+// NewGlobal is the same as New, but it assigns the metrics object to be
+// used globally as well as returning it.
+func NewGlobal(conf *Config, sink MetricSink) (*Metrics, error) {
+	metrics, err := New(conf, sink)
+	if err == nil {
+		globalMetrics.Store(metrics)
+	}
+	return metrics, err
+}
+
+// Proxy all the methods to the globalMetrics instance
+func SetGauge(key []string, val float32) {
+	globalMetrics.Load().(*Metrics).SetGauge(key, val)
+}
+
+func SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	globalMetrics.Load().(*Metrics).SetGaugeWithLabels(key, val, labels)
+}
+
+func EmitKey(key []string, val float32) {
+	globalMetrics.Load().(*Metrics).EmitKey(key, val)
+}
+
+func IncrCounter(key []string, val float32) {
+	globalMetrics.Load().(*Metrics).IncrCounter(key, val)
+}
+
+func IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	globalMetrics.Load().(*Metrics).IncrCounterWithLabels(key, val, labels)
+}
+
+func AddSample(key []string, val float32) {
+	globalMetrics.Load().(*Metrics).AddSample(key, val)
+}
+
+func AddSampleWithLabels(key []string, val float32, labels []Label) {
+	globalMetrics.Load().(*Metrics).AddSampleWithLabels(key, val, labels)
+}
+
+func MeasureSince(key []string, start time.Time) {
+	globalMetrics.Load().(*Metrics).MeasureSince(key, start)
+}
+
+func MeasureSinceWithLabels(key []string, start time.Time, labels []Label) {
+	globalMetrics.Load().(*Metrics).MeasureSinceWithLabels(key, start, labels)
+}
+
+func UpdateFilter(allow, block []string) {
+	globalMetrics.Load().(*Metrics).UpdateFilter(allow, block)
+}
+
+// UpdateFilterAndLabels set allow/block prefixes of metrics while allowedLabels
+// and blockedLabels - when not nil - allow filtering of labels in order to
+// block/allow globally labels (especially useful when having large number of
+// values for a given label). See README.md for more information about usage.
+func UpdateFilterAndLabels(allow, block, allowedLabels, blockedLabels []string) {
+	globalMetrics.Load().(*Metrics).UpdateFilterAndLabels(allow, block, allowedLabels, blockedLabels)
+}
diff --git a/vendor/github.com/armon/go-metrics/statsd.go b/vendor/github.com/armon/go-metrics/statsd.go
new file mode 100644
index 0000000..1bfffce
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/statsd.go
@@ -0,0 +1,184 @@
+package metrics
+
+import (
+	"bytes"
+	"fmt"
+	"log"
+	"net"
+	"net/url"
+	"strings"
+	"time"
+)
+
+const (
+	// statsdMaxLen is the maximum size of a packet
+	// to send to statsd
+	statsdMaxLen = 1400
+)
+
+// StatsdSink provides a MetricSink that can be used
+// with a statsite or statsd metrics server. It uses
+// only UDP packets, while StatsiteSink uses TCP.
+type StatsdSink struct {
+	addr        string
+	metricQueue chan string
+}
+
+// NewStatsdSinkFromURL creates an StatsdSink from a URL. It is used
+// (and tested) from NewMetricSinkFromURL.
+func NewStatsdSinkFromURL(u *url.URL) (MetricSink, error) {
+	return NewStatsdSink(u.Host)
+}
+
+// NewStatsdSink is used to create a new StatsdSink
+func NewStatsdSink(addr string) (*StatsdSink, error) {
+	s := &StatsdSink{
+		addr:        addr,
+		metricQueue: make(chan string, 4096),
+	}
+	go s.flushMetrics()
+	return s, nil
+}
+
+// Close is used to stop flushing to statsd
+func (s *StatsdSink) Shutdown() {
+	close(s.metricQueue)
+}
+
+func (s *StatsdSink) SetGauge(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|g\n", flatKey, val))
+}
+
+func (s *StatsdSink) SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|g\n", flatKey, val))
+}
+
+func (s *StatsdSink) EmitKey(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|kv\n", flatKey, val))
+}
+
+func (s *StatsdSink) IncrCounter(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|c\n", flatKey, val))
+}
+
+func (s *StatsdSink) IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|c\n", flatKey, val))
+}
+
+func (s *StatsdSink) AddSample(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|ms\n", flatKey, val))
+}
+
+func (s *StatsdSink) AddSampleWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|ms\n", flatKey, val))
+}
+
+// Flattens the key for formatting, removes spaces
+func (s *StatsdSink) flattenKey(parts []string) string {
+	joined := strings.Join(parts, ".")
+	return strings.Map(func(r rune) rune {
+		switch r {
+		case ':':
+			fallthrough
+		case ' ':
+			return '_'
+		default:
+			return r
+		}
+	}, joined)
+}
+
+// Flattens the key along with labels for formatting, removes spaces
+func (s *StatsdSink) flattenKeyLabels(parts []string, labels []Label) string {
+	for _, label := range labels {
+		parts = append(parts, label.Value)
+	}
+	return s.flattenKey(parts)
+}
+
+// Does a non-blocking push to the metrics queue
+func (s *StatsdSink) pushMetric(m string) {
+	select {
+	case s.metricQueue <- m:
+	default:
+	}
+}
+
+// Flushes metrics
+func (s *StatsdSink) flushMetrics() {
+	var sock net.Conn
+	var err error
+	var wait <-chan time.Time
+	ticker := time.NewTicker(flushInterval)
+	defer ticker.Stop()
+
+CONNECT:
+	// Create a buffer
+	buf := bytes.NewBuffer(nil)
+
+	// Attempt to connect
+	sock, err = net.Dial("udp", s.addr)
+	if err != nil {
+		log.Printf("[ERR] Error connecting to statsd! Err: %s", err)
+		goto WAIT
+	}
+
+	for {
+		select {
+		case metric, ok := <-s.metricQueue:
+			// Get a metric from the queue
+			if !ok {
+				goto QUIT
+			}
+
+			// Check if this would overflow the packet size
+			if len(metric)+buf.Len() > statsdMaxLen {
+				_, err := sock.Write(buf.Bytes())
+				buf.Reset()
+				if err != nil {
+					log.Printf("[ERR] Error writing to statsd! Err: %s", err)
+					goto WAIT
+				}
+			}
+
+			// Append to the buffer
+			buf.WriteString(metric)
+
+		case <-ticker.C:
+			if buf.Len() == 0 {
+				continue
+			}
+
+			_, err := sock.Write(buf.Bytes())
+			buf.Reset()
+			if err != nil {
+				log.Printf("[ERR] Error flushing to statsd! Err: %s", err)
+				goto WAIT
+			}
+		}
+	}
+
+WAIT:
+	// Wait for a while
+	wait = time.After(time.Duration(5) * time.Second)
+	for {
+		select {
+		// Dequeue the messages to avoid backlog
+		case _, ok := <-s.metricQueue:
+			if !ok {
+				goto QUIT
+			}
+		case <-wait:
+			goto CONNECT
+		}
+	}
+QUIT:
+	s.metricQueue = nil
+}
diff --git a/vendor/github.com/armon/go-metrics/statsite.go b/vendor/github.com/armon/go-metrics/statsite.go
new file mode 100644
index 0000000..6c0d284
--- /dev/null
+++ b/vendor/github.com/armon/go-metrics/statsite.go
@@ -0,0 +1,172 @@
+package metrics
+
+import (
+	"bufio"
+	"fmt"
+	"log"
+	"net"
+	"net/url"
+	"strings"
+	"time"
+)
+
+const (
+	// We force flush the statsite metrics after this period of
+	// inactivity. Prevents stats from getting stuck in a buffer
+	// forever.
+	flushInterval = 100 * time.Millisecond
+)
+
+// NewStatsiteSinkFromURL creates an StatsiteSink from a URL. It is used
+// (and tested) from NewMetricSinkFromURL.
+func NewStatsiteSinkFromURL(u *url.URL) (MetricSink, error) {
+	return NewStatsiteSink(u.Host)
+}
+
+// StatsiteSink provides a MetricSink that can be used with a
+// statsite metrics server
+type StatsiteSink struct {
+	addr        string
+	metricQueue chan string
+}
+
+// NewStatsiteSink is used to create a new StatsiteSink
+func NewStatsiteSink(addr string) (*StatsiteSink, error) {
+	s := &StatsiteSink{
+		addr:        addr,
+		metricQueue: make(chan string, 4096),
+	}
+	go s.flushMetrics()
+	return s, nil
+}
+
+// Close is used to stop flushing to statsite
+func (s *StatsiteSink) Shutdown() {
+	close(s.metricQueue)
+}
+
+func (s *StatsiteSink) SetGauge(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|g\n", flatKey, val))
+}
+
+func (s *StatsiteSink) SetGaugeWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|g\n", flatKey, val))
+}
+
+func (s *StatsiteSink) EmitKey(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|kv\n", flatKey, val))
+}
+
+func (s *StatsiteSink) IncrCounter(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|c\n", flatKey, val))
+}
+
+func (s *StatsiteSink) IncrCounterWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|c\n", flatKey, val))
+}
+
+func (s *StatsiteSink) AddSample(key []string, val float32) {
+	flatKey := s.flattenKey(key)
+	s.pushMetric(fmt.Sprintf("%s:%f|ms\n", flatKey, val))
+}
+
+func (s *StatsiteSink) AddSampleWithLabels(key []string, val float32, labels []Label) {
+	flatKey := s.flattenKeyLabels(key, labels)
+	s.pushMetric(fmt.Sprintf("%s:%f|ms\n", flatKey, val))
+}
+
+// Flattens the key for formatting, removes spaces
+func (s *StatsiteSink) flattenKey(parts []string) string {
+	joined := strings.Join(parts, ".")
+	return strings.Map(func(r rune) rune {
+		switch r {
+		case ':':
+			fallthrough
+		case ' ':
+			return '_'
+		default:
+			return r
+		}
+	}, joined)
+}
+
+// Flattens the key along with labels for formatting, removes spaces
+func (s *StatsiteSink) flattenKeyLabels(parts []string, labels []Label) string {
+	for _, label := range labels {
+		parts = append(parts, label.Value)
+	}
+	return s.flattenKey(parts)
+}
+
+// Does a non-blocking push to the metrics queue
+func (s *StatsiteSink) pushMetric(m string) {
+	select {
+	case s.metricQueue <- m:
+	default:
+	}
+}
+
+// Flushes metrics
+func (s *StatsiteSink) flushMetrics() {
+	var sock net.Conn
+	var err error
+	var wait <-chan time.Time
+	var buffered *bufio.Writer
+	ticker := time.NewTicker(flushInterval)
+	defer ticker.Stop()
+
+CONNECT:
+	// Attempt to connect
+	sock, err = net.Dial("tcp", s.addr)
+	if err != nil {
+		log.Printf("[ERR] Error connecting to statsite! Err: %s", err)
+		goto WAIT
+	}
+
+	// Create a buffered writer
+	buffered = bufio.NewWriter(sock)
+
+	for {
+		select {
+		case metric, ok := <-s.metricQueue:
+			// Get a metric from the queue
+			if !ok {
+				goto QUIT
+			}
+
+			// Try to send to statsite
+			_, err := buffered.Write([]byte(metric))
+			if err != nil {
+				log.Printf("[ERR] Error writing to statsite! Err: %s", err)
+				goto WAIT
+			}
+		case <-ticker.C:
+			if err := buffered.Flush(); err != nil {
+				log.Printf("[ERR] Error flushing to statsite! Err: %s", err)
+				goto WAIT
+			}
+		}
+	}
+
+WAIT:
+	// Wait for a while
+	wait = time.After(time.Duration(5) * time.Second)
+	for {
+		select {
+		// Dequeue the messages to avoid backlog
+		case _, ok := <-s.metricQueue:
+			if !ok {
+				goto QUIT
+			}
+		case <-wait:
+			goto CONNECT
+		}
+	}
+QUIT:
+	s.metricQueue = nil
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/.gitignore b/vendor/github.com/bsm/sarama-cluster/.gitignore
new file mode 100644
index 0000000..88113c5
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/.gitignore
@@ -0,0 +1,4 @@
+*.log
+*.pid
+kafka*/
+vendor/
diff --git a/vendor/github.com/bsm/sarama-cluster/.travis.yml b/vendor/github.com/bsm/sarama-cluster/.travis.yml
new file mode 100644
index 0000000..07c7c10
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/.travis.yml
@@ -0,0 +1,18 @@
+sudo: false
+language: go
+go:
+  - 1.10.x
+  - 1.9.x
+install:
+  - go get -u github.com/golang/dep/cmd/dep
+  - dep ensure
+env:
+  - SCALA_VERSION=2.12 KAFKA_VERSION=0.11.0.1
+  - SCALA_VERSION=2.12 KAFKA_VERSION=1.0.1
+  - SCALA_VERSION=2.12 KAFKA_VERSION=1.1.0
+script:
+  - make default test-race
+addons:
+  apt:
+    packages:
+      - oracle-java8-set-default
diff --git a/vendor/github.com/bsm/sarama-cluster/Gopkg.lock b/vendor/github.com/bsm/sarama-cluster/Gopkg.lock
new file mode 100644
index 0000000..e1bc110
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/Gopkg.lock
@@ -0,0 +1,151 @@
+# This file is autogenerated, do not edit; changes may be undone by the next 'dep ensure'.
+
+
+[[projects]]
+  name = "github.com/Shopify/sarama"
+  packages = ["."]
+  revision = "35324cf48e33d8260e1c7c18854465a904ade249"
+  version = "v1.17.0"
+
+[[projects]]
+  name = "github.com/davecgh/go-spew"
+  packages = ["spew"]
+  revision = "346938d642f2ec3594ed81d874461961cd0faa76"
+  version = "v1.1.0"
+
+[[projects]]
+  name = "github.com/eapache/go-resiliency"
+  packages = ["breaker"]
+  revision = "ea41b0fad31007accc7f806884dcdf3da98b79ce"
+  version = "v1.1.0"
+
+[[projects]]
+  branch = "master"
+  name = "github.com/eapache/go-xerial-snappy"
+  packages = ["."]
+  revision = "bb955e01b9346ac19dc29eb16586c90ded99a98c"
+
+[[projects]]
+  name = "github.com/eapache/queue"
+  packages = ["."]
+  revision = "44cc805cf13205b55f69e14bcb69867d1ae92f98"
+  version = "v1.1.0"
+
+[[projects]]
+  branch = "master"
+  name = "github.com/golang/snappy"
+  packages = ["."]
+  revision = "2e65f85255dbc3072edf28d6b5b8efc472979f5a"
+
+[[projects]]
+  name = "github.com/onsi/ginkgo"
+  packages = [
+    ".",
+    "config",
+    "extensions/table",
+    "internal/codelocation",
+    "internal/containernode",
+    "internal/failer",
+    "internal/leafnodes",
+    "internal/remote",
+    "internal/spec",
+    "internal/spec_iterator",
+    "internal/specrunner",
+    "internal/suite",
+    "internal/testingtproxy",
+    "internal/writer",
+    "reporters",
+    "reporters/stenographer",
+    "reporters/stenographer/support/go-colorable",
+    "reporters/stenographer/support/go-isatty",
+    "types"
+  ]
+  revision = "fa5fabab2a1bfbd924faf4c067d07ae414e2aedf"
+  version = "v1.5.0"
+
+[[projects]]
+  name = "github.com/onsi/gomega"
+  packages = [
+    ".",
+    "format",
+    "internal/assertion",
+    "internal/asyncassertion",
+    "internal/oraclematcher",
+    "internal/testingtsupport",
+    "matchers",
+    "matchers/support/goraph/bipartitegraph",
+    "matchers/support/goraph/edge",
+    "matchers/support/goraph/node",
+    "matchers/support/goraph/util",
+    "types"
+  ]
+  revision = "62bff4df71bdbc266561a0caee19f0594b17c240"
+  version = "v1.4.0"
+
+[[projects]]
+  name = "github.com/pierrec/lz4"
+  packages = [
+    ".",
+    "internal/xxh32"
+  ]
+  revision = "6b9367c9ff401dbc54fabce3fb8d972e799b702d"
+  version = "v2.0.2"
+
+[[projects]]
+  branch = "master"
+  name = "github.com/rcrowley/go-metrics"
+  packages = ["."]
+  revision = "e2704e165165ec55d062f5919b4b29494e9fa790"
+
+[[projects]]
+  branch = "master"
+  name = "golang.org/x/net"
+  packages = [
+    "html",
+    "html/atom",
+    "html/charset"
+  ]
+  revision = "afe8f62b1d6bbd81f31868121a50b06d8188e1f9"
+
+[[projects]]
+  branch = "master"
+  name = "golang.org/x/sys"
+  packages = ["unix"]
+  revision = "63fc586f45fe72d95d5240a5d5eb95e6503907d3"
+
+[[projects]]
+  name = "golang.org/x/text"
+  packages = [
+    "encoding",
+    "encoding/charmap",
+    "encoding/htmlindex",
+    "encoding/internal",
+    "encoding/internal/identifier",
+    "encoding/japanese",
+    "encoding/korean",
+    "encoding/simplifiedchinese",
+    "encoding/traditionalchinese",
+    "encoding/unicode",
+    "internal/gen",
+    "internal/tag",
+    "internal/utf8internal",
+    "language",
+    "runes",
+    "transform",
+    "unicode/cldr"
+  ]
+  revision = "f21a4dfb5e38f5895301dc265a8def02365cc3d0"
+  version = "v0.3.0"
+
+[[projects]]
+  name = "gopkg.in/yaml.v2"
+  packages = ["."]
+  revision = "5420a8b6744d3b0345ab293f6fcba19c978f1183"
+  version = "v2.2.1"
+
+[solve-meta]
+  analyzer-name = "dep"
+  analyzer-version = 1
+  inputs-digest = "2fa33a2d1ae87e0905ef09332bb4b3fda29179f6bcd48fd3b94070774b9e458b"
+  solver-name = "gps-cdcl"
+  solver-version = 1
diff --git a/vendor/github.com/bsm/sarama-cluster/Gopkg.toml b/vendor/github.com/bsm/sarama-cluster/Gopkg.toml
new file mode 100644
index 0000000..1eecfef
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/Gopkg.toml
@@ -0,0 +1,26 @@
+
+# Gopkg.toml example
+#
+# Refer to https://github.com/golang/dep/blob/master/docs/Gopkg.toml.md
+# for detailed Gopkg.toml documentation.
+#
+# required = ["github.com/user/thing/cmd/thing"]
+# ignored = ["github.com/user/project/pkgX", "bitbucket.org/user/project/pkgA/pkgY"]
+#
+# [[constraint]]
+#   name = "github.com/user/project"
+#   version = "1.0.0"
+#
+# [[constraint]]
+#   name = "github.com/user/project2"
+#   branch = "dev"
+#   source = "github.com/myfork/project2"
+#
+# [[override]]
+#  name = "github.com/x/y"
+#  version = "2.4.0"
+
+
+[[constraint]]
+  name = "github.com/Shopify/sarama"
+  version = "^1.14.0"
diff --git a/vendor/github.com/bsm/sarama-cluster/LICENSE b/vendor/github.com/bsm/sarama-cluster/LICENSE
new file mode 100644
index 0000000..127751c
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/LICENSE
@@ -0,0 +1,22 @@
+(The MIT License)
+
+Copyright (c) 2017 Black Square Media Ltd
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+'Software'), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED 'AS IS', WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/vendor/github.com/bsm/sarama-cluster/Makefile b/vendor/github.com/bsm/sarama-cluster/Makefile
new file mode 100644
index 0000000..25c5bc2
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/Makefile
@@ -0,0 +1,35 @@
+SCALA_VERSION?= 2.12
+KAFKA_VERSION?= 1.1.0
+KAFKA_DIR= kafka_$(SCALA_VERSION)-$(KAFKA_VERSION)
+KAFKA_SRC= https://archive.apache.org/dist/kafka/$(KAFKA_VERSION)/$(KAFKA_DIR).tgz
+KAFKA_ROOT= testdata/$(KAFKA_DIR)
+PKG=$(shell go list ./... | grep -v vendor)
+
+default: vet test
+
+vet:
+	go vet $(PKG)
+
+test: testdeps
+	KAFKA_DIR=$(KAFKA_DIR) go test $(PKG) -ginkgo.slowSpecThreshold=60
+
+test-verbose: testdeps
+	KAFKA_DIR=$(KAFKA_DIR) go test $(PKG) -ginkgo.slowSpecThreshold=60 -v
+
+test-race: testdeps
+	KAFKA_DIR=$(KAFKA_DIR) go test $(PKG) -ginkgo.slowSpecThreshold=60 -v -race
+
+testdeps: $(KAFKA_ROOT)
+
+doc: README.md
+
+.PHONY: test testdeps vet doc
+
+# ---------------------------------------------------------------------
+
+$(KAFKA_ROOT):
+	@mkdir -p $(dir $@)
+	cd $(dir $@) && curl -sSL $(KAFKA_SRC) | tar xz
+
+README.md: README.md.tpl $(wildcard *.go)
+	becca -package $(subst $(GOPATH)/src/,,$(PWD))
diff --git a/vendor/github.com/bsm/sarama-cluster/README.md b/vendor/github.com/bsm/sarama-cluster/README.md
new file mode 100644
index 0000000..ebcd755
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/README.md
@@ -0,0 +1,151 @@
+# Sarama Cluster
+
+[![GoDoc](https://godoc.org/github.com/bsm/sarama-cluster?status.svg)](https://godoc.org/github.com/bsm/sarama-cluster)
+[![Build Status](https://travis-ci.org/bsm/sarama-cluster.svg?branch=master)](https://travis-ci.org/bsm/sarama-cluster)
+[![Go Report Card](https://goreportcard.com/badge/github.com/bsm/sarama-cluster)](https://goreportcard.com/report/github.com/bsm/sarama-cluster)
+[![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+
+Cluster extensions for [Sarama](https://github.com/Shopify/sarama), the Go client library for Apache Kafka 0.9 (and later).
+
+## Documentation
+
+Documentation and example are available via godoc at http://godoc.org/github.com/bsm/sarama-cluster
+
+## Examples
+
+Consumers have two modes of operation. In the default multiplexed mode messages (and errors) of multiple
+topics and partitions are all passed to the single channel:
+
+```go
+package main
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"os/signal"
+
+	cluster "github.com/bsm/sarama-cluster"
+)
+
+func main() {
+
+	// init (custom) config, enable errors and notifications
+	config := cluster.NewConfig()
+	config.Consumer.Return.Errors = true
+	config.Group.Return.Notifications = true
+
+	// init consumer
+	brokers := []string{"127.0.0.1:9092"}
+	topics := []string{"my_topic", "other_topic"}
+	consumer, err := cluster.NewConsumer(brokers, "my-consumer-group", topics, config)
+	if err != nil {
+		panic(err)
+	}
+	defer consumer.Close()
+
+	// trap SIGINT to trigger a shutdown.
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, os.Interrupt)
+
+	// consume errors
+	go func() {
+		for err := range consumer.Errors() {
+			log.Printf("Error: %s\n", err.Error())
+		}
+	}()
+
+	// consume notifications
+	go func() {
+		for ntf := range consumer.Notifications() {
+			log.Printf("Rebalanced: %+v\n", ntf)
+		}
+	}()
+
+	// consume messages, watch signals
+	for {
+		select {
+		case msg, ok := <-consumer.Messages():
+			if ok {
+				fmt.Fprintf(os.Stdout, "%s/%d/%d\t%s\t%s\n", msg.Topic, msg.Partition, msg.Offset, msg.Key, msg.Value)
+				consumer.MarkOffset(msg, "")	// mark message as processed
+			}
+		case <-signals:
+			return
+		}
+	}
+}
+```
+
+Users who require access to individual partitions can use the partitioned mode which exposes access to partition-level
+consumers:
+
+```go
+package main
+
+import (
+  "fmt"
+  "log"
+  "os"
+  "os/signal"
+
+  cluster "github.com/bsm/sarama-cluster"
+)
+
+func main() {
+
+	// init (custom) config, set mode to ConsumerModePartitions
+	config := cluster.NewConfig()
+	config.Group.Mode = cluster.ConsumerModePartitions
+
+	// init consumer
+	brokers := []string{"127.0.0.1:9092"}
+	topics := []string{"my_topic", "other_topic"}
+	consumer, err := cluster.NewConsumer(brokers, "my-consumer-group", topics, config)
+	if err != nil {
+		panic(err)
+	}
+	defer consumer.Close()
+
+	// trap SIGINT to trigger a shutdown.
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, os.Interrupt)
+
+	// consume partitions
+	for {
+		select {
+		case part, ok := <-consumer.Partitions():
+			if !ok {
+				return
+			}
+
+			// start a separate goroutine to consume messages
+			go func(pc cluster.PartitionConsumer) {
+				for msg := range pc.Messages() {
+					fmt.Fprintf(os.Stdout, "%s/%d/%d\t%s\t%s\n", msg.Topic, msg.Partition, msg.Offset, msg.Key, msg.Value)
+					consumer.MarkOffset(msg, "")	// mark message as processed
+				}
+			}(part)
+		case <-signals:
+			return
+		}
+	}
+}
+```
+
+## Running tests
+
+You need to install Ginkgo & Gomega to run tests. Please see
+http://onsi.github.io/ginkgo for more details.
+
+To run tests, call:
+
+	$ make test
+
+## Troubleshooting
+
+### Consumer not receiving any messages?
+
+By default, sarama's `Config.Consumer.Offsets.Initial` is set to `sarama.OffsetNewest`. This means that in the event that a brand new consumer is created, and it has never committed any offsets to kafka, it will only receive messages starting from the message after the current one that was written.
+
+If you wish to receive all messages (from the start of all messages in the topic) in the event that a consumer does not have any offsets committed to kafka, you need to set `Config.Consumer.Offsets.Initial` to `sarama.OffsetOldest`.
diff --git a/vendor/github.com/bsm/sarama-cluster/README.md.tpl b/vendor/github.com/bsm/sarama-cluster/README.md.tpl
new file mode 100644
index 0000000..5f63a69
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/README.md.tpl
@@ -0,0 +1,67 @@
+# Sarama Cluster
+
+[![GoDoc](https://godoc.org/github.com/bsm/sarama-cluster?status.svg)](https://godoc.org/github.com/bsm/sarama-cluster)
+[![Build Status](https://travis-ci.org/bsm/sarama-cluster.svg?branch=master)](https://travis-ci.org/bsm/sarama-cluster)
+[![Go Report Card](https://goreportcard.com/badge/github.com/bsm/sarama-cluster)](https://goreportcard.com/report/github.com/bsm/sarama-cluster)
+[![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+
+Cluster extensions for [Sarama](https://github.com/Shopify/sarama), the Go client library for Apache Kafka 0.9 (and later).
+
+## Documentation
+
+Documentation and example are available via godoc at http://godoc.org/github.com/bsm/sarama-cluster
+
+## Examples
+
+Consumers have two modes of operation. In the default multiplexed mode messages (and errors) of multiple
+topics and partitions are all passed to the single channel:
+
+```go
+package main
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"os/signal"
+
+	cluster "github.com/bsm/sarama-cluster"
+)
+
+func main() {{ "ExampleConsumer" | code }}
+```
+
+Users who require access to individual partitions can use the partitioned mode which exposes access to partition-level
+consumers:
+
+```go
+package main
+
+import (
+  "fmt"
+  "log"
+  "os"
+  "os/signal"
+
+  cluster "github.com/bsm/sarama-cluster"
+)
+
+func main() {{ "ExampleConsumer_Partitions" | code }}
+```
+
+## Running tests
+
+You need to install Ginkgo & Gomega to run tests. Please see
+http://onsi.github.io/ginkgo for more details.
+
+To run tests, call:
+
+	$ make test
+
+## Troubleshooting
+
+### Consumer not receiving any messages?
+
+By default, sarama's `Config.Consumer.Offsets.Initial` is set to `sarama.OffsetNewest`. This means that in the event that a brand new consumer is created, and it has never committed any offsets to kafka, it will only receive messages starting from the message after the current one that was written.
+
+If you wish to receive all messages (from the start of all messages in the topic) in the event that a consumer does not have any offsets committed to kafka, you need to set `Config.Consumer.Offsets.Initial` to `sarama.OffsetOldest`.
diff --git a/vendor/github.com/bsm/sarama-cluster/balancer.go b/vendor/github.com/bsm/sarama-cluster/balancer.go
new file mode 100644
index 0000000..3aeaece
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/balancer.go
@@ -0,0 +1,170 @@
+package cluster
+
+import (
+	"math"
+	"sort"
+
+	"github.com/Shopify/sarama"
+)
+
+// NotificationType defines the type of notification
+type NotificationType uint8
+
+// String describes the notification type
+func (t NotificationType) String() string {
+	switch t {
+	case RebalanceStart:
+		return "rebalance start"
+	case RebalanceOK:
+		return "rebalance OK"
+	case RebalanceError:
+		return "rebalance error"
+	}
+	return "unknown"
+}
+
+const (
+	UnknownNotification NotificationType = iota
+	RebalanceStart
+	RebalanceOK
+	RebalanceError
+)
+
+// Notification are state events emitted by the consumers on rebalance
+type Notification struct {
+	// Type exposes the notification type
+	Type NotificationType
+
+	// Claimed contains topic/partitions that were claimed by this rebalance cycle
+	Claimed map[string][]int32
+
+	// Released contains topic/partitions that were released as part of this rebalance cycle
+	Released map[string][]int32
+
+	// Current are topic/partitions that are currently claimed to the consumer
+	Current map[string][]int32
+}
+
+func newNotification(current map[string][]int32) *Notification {
+	return &Notification{
+		Type:    RebalanceStart,
+		Current: current,
+	}
+}
+
+func (n *Notification) success(current map[string][]int32) *Notification {
+	o := &Notification{
+		Type:     RebalanceOK,
+		Claimed:  make(map[string][]int32),
+		Released: make(map[string][]int32),
+		Current:  current,
+	}
+	for topic, partitions := range current {
+		o.Claimed[topic] = int32Slice(partitions).Diff(int32Slice(n.Current[topic]))
+	}
+	for topic, partitions := range n.Current {
+		o.Released[topic] = int32Slice(partitions).Diff(int32Slice(current[topic]))
+	}
+	return o
+}
+
+// --------------------------------------------------------------------
+
+type topicInfo struct {
+	Partitions []int32
+	MemberIDs  []string
+}
+
+func (info topicInfo) Perform(s Strategy) map[string][]int32 {
+	if s == StrategyRoundRobin {
+		return info.RoundRobin()
+	}
+	return info.Ranges()
+}
+
+func (info topicInfo) Ranges() map[string][]int32 {
+	sort.Strings(info.MemberIDs)
+
+	mlen := len(info.MemberIDs)
+	plen := len(info.Partitions)
+	res := make(map[string][]int32, mlen)
+
+	for pos, memberID := range info.MemberIDs {
+		n, i := float64(plen)/float64(mlen), float64(pos)
+		min := int(math.Floor(i*n + 0.5))
+		max := int(math.Floor((i+1)*n + 0.5))
+		sub := info.Partitions[min:max]
+		if len(sub) > 0 {
+			res[memberID] = sub
+		}
+	}
+	return res
+}
+
+func (info topicInfo) RoundRobin() map[string][]int32 {
+	sort.Strings(info.MemberIDs)
+
+	mlen := len(info.MemberIDs)
+	res := make(map[string][]int32, mlen)
+	for i, pnum := range info.Partitions {
+		memberID := info.MemberIDs[i%mlen]
+		res[memberID] = append(res[memberID], pnum)
+	}
+	return res
+}
+
+// --------------------------------------------------------------------
+
+type balancer struct {
+	client sarama.Client
+	topics map[string]topicInfo
+}
+
+func newBalancerFromMeta(client sarama.Client, members map[string]sarama.ConsumerGroupMemberMetadata) (*balancer, error) {
+	balancer := newBalancer(client)
+	for memberID, meta := range members {
+		for _, topic := range meta.Topics {
+			if err := balancer.Topic(topic, memberID); err != nil {
+				return nil, err
+			}
+		}
+	}
+	return balancer, nil
+}
+
+func newBalancer(client sarama.Client) *balancer {
+	return &balancer{
+		client: client,
+		topics: make(map[string]topicInfo),
+	}
+}
+
+func (r *balancer) Topic(name string, memberID string) error {
+	topic, ok := r.topics[name]
+	if !ok {
+		nums, err := r.client.Partitions(name)
+		if err != nil {
+			return err
+		}
+		topic = topicInfo{
+			Partitions: nums,
+			MemberIDs:  make([]string, 0, 1),
+		}
+	}
+	topic.MemberIDs = append(topic.MemberIDs, memberID)
+	r.topics[name] = topic
+	return nil
+}
+
+func (r *balancer) Perform(s Strategy) map[string]map[string][]int32 {
+	res := make(map[string]map[string][]int32, 1)
+	for topic, info := range r.topics {
+		for memberID, partitions := range info.Perform(s) {
+			if _, ok := res[memberID]; !ok {
+				res[memberID] = make(map[string][]int32, 1)
+			}
+			res[memberID][topic] = partitions
+		}
+	}
+	return res
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/client.go b/vendor/github.com/bsm/sarama-cluster/client.go
new file mode 100644
index 0000000..42ffb30
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/client.go
@@ -0,0 +1,50 @@
+package cluster
+
+import (
+	"errors"
+	"sync/atomic"
+
+	"github.com/Shopify/sarama"
+)
+
+var errClientInUse = errors.New("cluster: client is already used by another consumer")
+
+// Client is a group client
+type Client struct {
+	sarama.Client
+	config Config
+
+	inUse uint32
+}
+
+// NewClient creates a new client instance
+func NewClient(addrs []string, config *Config) (*Client, error) {
+	if config == nil {
+		config = NewConfig()
+	}
+
+	if err := config.Validate(); err != nil {
+		return nil, err
+	}
+
+	client, err := sarama.NewClient(addrs, &config.Config)
+	if err != nil {
+		return nil, err
+	}
+
+	return &Client{Client: client, config: *config}, nil
+}
+
+// ClusterConfig returns the cluster configuration.
+func (c *Client) ClusterConfig() *Config {
+	cfg := c.config
+	return &cfg
+}
+
+func (c *Client) claim() bool {
+	return atomic.CompareAndSwapUint32(&c.inUse, 0, 1)
+}
+
+func (c *Client) release() {
+	atomic.CompareAndSwapUint32(&c.inUse, 1, 0)
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/cluster.go b/vendor/github.com/bsm/sarama-cluster/cluster.go
new file mode 100644
index 0000000..adcf0e9
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/cluster.go
@@ -0,0 +1,25 @@
+package cluster
+
+// Strategy for partition to consumer assignement
+type Strategy string
+
+const (
+	// StrategyRange is the default and assigns partition ranges to consumers.
+	// Example with six partitions and two consumers:
+	//   C1: [0, 1, 2]
+	//   C2: [3, 4, 5]
+	StrategyRange Strategy = "range"
+
+	// StrategyRoundRobin assigns partitions by alternating over consumers.
+	// Example with six partitions and two consumers:
+	//   C1: [0, 2, 4]
+	//   C2: [1, 3, 5]
+	StrategyRoundRobin Strategy = "roundrobin"
+)
+
+// Error instances are wrappers for internal errors with a context and
+// may be returned through the consumer's Errors() channel
+type Error struct {
+	Ctx string
+	error
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/config.go b/vendor/github.com/bsm/sarama-cluster/config.go
new file mode 100644
index 0000000..084b835
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/config.go
@@ -0,0 +1,146 @@
+package cluster
+
+import (
+	"regexp"
+	"time"
+
+	"github.com/Shopify/sarama"
+)
+
+var minVersion = sarama.V0_9_0_0
+
+type ConsumerMode uint8
+
+const (
+	ConsumerModeMultiplex ConsumerMode = iota
+	ConsumerModePartitions
+)
+
+// Config extends sarama.Config with Group specific namespace
+type Config struct {
+	sarama.Config
+
+	// Group is the namespace for group management properties
+	Group struct {
+
+		// The strategy to use for the allocation of partitions to consumers (defaults to StrategyRange)
+		PartitionStrategy Strategy
+
+		// By default, messages and errors from the subscribed topics and partitions are all multiplexed and
+		// made available through the consumer's Messages() and Errors() channels.
+		//
+		// Users who require low-level access can enable ConsumerModePartitions where individual partitions
+		// are exposed on the Partitions() channel. Messages and errors must then be consumed on the partitions
+		// themselves.
+		Mode ConsumerMode
+
+		Offsets struct {
+			Retry struct {
+				// The numer retries when committing offsets (defaults to 3).
+				Max int
+			}
+			Synchronization struct {
+				// The duration allowed for other clients to commit their offsets before resumption in this client, e.g. during a rebalance
+				// NewConfig sets this to the Consumer.MaxProcessingTime duration of the Sarama configuration
+				DwellTime time.Duration
+			}
+		}
+
+		Session struct {
+			// The allowed session timeout for registered consumers (defaults to 30s).
+			// Must be within the allowed server range.
+			Timeout time.Duration
+		}
+
+		Heartbeat struct {
+			// Interval between each heartbeat (defaults to 3s). It should be no more
+			// than 1/3rd of the Group.Session.Timout setting
+			Interval time.Duration
+		}
+
+		// Return specifies which group channels will be populated. If they are set to true,
+		// you must read from the respective channels to prevent deadlock.
+		Return struct {
+			// If enabled, rebalance notification will be returned on the
+			// Notifications channel (default disabled).
+			Notifications bool
+		}
+
+		Topics struct {
+			// An additional whitelist of topics to subscribe to.
+			Whitelist *regexp.Regexp
+			// An additional blacklist of topics to avoid. If set, this will precede over
+			// the Whitelist setting.
+			Blacklist *regexp.Regexp
+		}
+
+		Member struct {
+			// Custom metadata to include when joining the group. The user data for all joined members
+			// can be retrieved by sending a DescribeGroupRequest to the broker that is the
+			// coordinator for the group.
+			UserData []byte
+		}
+	}
+}
+
+// NewConfig returns a new configuration instance with sane defaults.
+func NewConfig() *Config {
+	c := &Config{
+		Config: *sarama.NewConfig(),
+	}
+	c.Group.PartitionStrategy = StrategyRange
+	c.Group.Offsets.Retry.Max = 3
+	c.Group.Offsets.Synchronization.DwellTime = c.Consumer.MaxProcessingTime
+	c.Group.Session.Timeout = 30 * time.Second
+	c.Group.Heartbeat.Interval = 3 * time.Second
+	c.Config.Version = minVersion
+	return c
+}
+
+// Validate checks a Config instance. It will return a
+// sarama.ConfigurationError if the specified values don't make sense.
+func (c *Config) Validate() error {
+	if c.Group.Heartbeat.Interval%time.Millisecond != 0 {
+		sarama.Logger.Println("Group.Heartbeat.Interval only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Group.Session.Timeout%time.Millisecond != 0 {
+		sarama.Logger.Println("Group.Session.Timeout only supports millisecond precision; nanoseconds will be truncated.")
+	}
+	if c.Group.PartitionStrategy != StrategyRange && c.Group.PartitionStrategy != StrategyRoundRobin {
+		sarama.Logger.Println("Group.PartitionStrategy is not supported; range will be assumed.")
+	}
+	if !c.Version.IsAtLeast(minVersion) {
+		sarama.Logger.Println("Version is not supported; 0.9. will be assumed.")
+		c.Version = minVersion
+	}
+	if err := c.Config.Validate(); err != nil {
+		return err
+	}
+
+	// validate the Group values
+	switch {
+	case c.Group.Offsets.Retry.Max < 0:
+		return sarama.ConfigurationError("Group.Offsets.Retry.Max must be >= 0")
+	case c.Group.Offsets.Synchronization.DwellTime <= 0:
+		return sarama.ConfigurationError("Group.Offsets.Synchronization.DwellTime must be > 0")
+	case c.Group.Offsets.Synchronization.DwellTime > 10*time.Minute:
+		return sarama.ConfigurationError("Group.Offsets.Synchronization.DwellTime must be <= 10m")
+	case c.Group.Heartbeat.Interval <= 0:
+		return sarama.ConfigurationError("Group.Heartbeat.Interval must be > 0")
+	case c.Group.Session.Timeout <= 0:
+		return sarama.ConfigurationError("Group.Session.Timeout must be > 0")
+	case !c.Metadata.Full && c.Group.Topics.Whitelist != nil:
+		return sarama.ConfigurationError("Metadata.Full must be enabled when Group.Topics.Whitelist is used")
+	case !c.Metadata.Full && c.Group.Topics.Blacklist != nil:
+		return sarama.ConfigurationError("Metadata.Full must be enabled when Group.Topics.Blacklist is used")
+	}
+
+	// ensure offset is correct
+	switch c.Consumer.Offsets.Initial {
+	case sarama.OffsetOldest, sarama.OffsetNewest:
+	default:
+		return sarama.ConfigurationError("Consumer.Offsets.Initial must be either OffsetOldest or OffsetNewest")
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/consumer.go b/vendor/github.com/bsm/sarama-cluster/consumer.go
new file mode 100644
index 0000000..e7a67da
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/consumer.go
@@ -0,0 +1,919 @@
+package cluster
+
+import (
+	"sort"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/Shopify/sarama"
+)
+
+// Consumer is a cluster group consumer
+type Consumer struct {
+	client    *Client
+	ownClient bool
+
+	consumer sarama.Consumer
+	subs     *partitionMap
+
+	consumerID string
+	groupID    string
+
+	memberID     string
+	generationID int32
+	membershipMu sync.RWMutex
+
+	coreTopics  []string
+	extraTopics []string
+
+	dying, dead chan none
+	closeOnce   sync.Once
+
+	consuming     int32
+	messages      chan *sarama.ConsumerMessage
+	errors        chan error
+	partitions    chan PartitionConsumer
+	notifications chan *Notification
+
+	commitMu sync.Mutex
+}
+
+// NewConsumer initializes a new consumer
+func NewConsumer(addrs []string, groupID string, topics []string, config *Config) (*Consumer, error) {
+	client, err := NewClient(addrs, config)
+	if err != nil {
+		return nil, err
+	}
+
+	consumer, err := NewConsumerFromClient(client, groupID, topics)
+	if err != nil {
+		return nil, err
+	}
+	consumer.ownClient = true
+	return consumer, nil
+}
+
+// NewConsumerFromClient initializes a new consumer from an existing client.
+//
+// Please note that clients cannot be shared between consumers (due to Kafka internals),
+// they can only be re-used which requires the user to call Close() on the first consumer
+// before using this method again to initialize another one. Attempts to use a client with
+// more than one consumer at a time will return errors.
+func NewConsumerFromClient(client *Client, groupID string, topics []string) (*Consumer, error) {
+	if !client.claim() {
+		return nil, errClientInUse
+	}
+
+	consumer, err := sarama.NewConsumerFromClient(client.Client)
+	if err != nil {
+		client.release()
+		return nil, err
+	}
+
+	sort.Strings(topics)
+	c := &Consumer{
+		client:   client,
+		consumer: consumer,
+		subs:     newPartitionMap(),
+		groupID:  groupID,
+
+		coreTopics: topics,
+
+		dying: make(chan none),
+		dead:  make(chan none),
+
+		messages:      make(chan *sarama.ConsumerMessage),
+		errors:        make(chan error, client.config.ChannelBufferSize),
+		partitions:    make(chan PartitionConsumer, 1),
+		notifications: make(chan *Notification),
+	}
+	if err := c.client.RefreshCoordinator(groupID); err != nil {
+		client.release()
+		return nil, err
+	}
+
+	go c.mainLoop()
+	return c, nil
+}
+
+// Messages returns the read channel for the messages that are returned by
+// the broker.
+//
+// This channel will only return if Config.Group.Mode option is set to
+// ConsumerModeMultiplex (default).
+func (c *Consumer) Messages() <-chan *sarama.ConsumerMessage { return c.messages }
+
+// Partitions returns the read channels for individual partitions of this broker.
+//
+// This will channel will only return if Config.Group.Mode option is set to
+// ConsumerModePartitions.
+//
+// The Partitions() channel must be listened to for the life of this consumer;
+// when a rebalance happens old partitions will be closed (naturally come to
+// completion) and new ones will be emitted. The returned channel will only close
+// when the consumer is completely shut down.
+func (c *Consumer) Partitions() <-chan PartitionConsumer { return c.partitions }
+
+// Errors returns a read channel of errors that occur during offset management, if
+// enabled. By default, errors are logged and not returned over this channel. If
+// you want to implement any custom error handling, set your config's
+// Consumer.Return.Errors setting to true, and read from this channel.
+func (c *Consumer) Errors() <-chan error { return c.errors }
+
+// Notifications returns a channel of Notifications that occur during consumer
+// rebalancing. Notifications will only be emitted over this channel, if your config's
+// Group.Return.Notifications setting to true.
+func (c *Consumer) Notifications() <-chan *Notification { return c.notifications }
+
+// HighWaterMarks returns the current high water marks for each topic and partition
+// Consistency between partitions is not guaranteed since high water marks are updated separately.
+func (c *Consumer) HighWaterMarks() map[string]map[int32]int64 { return c.consumer.HighWaterMarks() }
+
+// MarkOffset marks the provided message as processed, alongside a metadata string
+// that represents the state of the partition consumer at that point in time. The
+// metadata string can be used by another consumer to restore that state, so it
+// can resume consumption.
+//
+// Note: calling MarkOffset does not necessarily commit the offset to the backend
+// store immediately for efficiency reasons, and it may never be committed if
+// your application crashes. This means that you may end up processing the same
+// message twice, and your processing should ideally be idempotent.
+func (c *Consumer) MarkOffset(msg *sarama.ConsumerMessage, metadata string) {
+	if sub := c.subs.Fetch(msg.Topic, msg.Partition); sub != nil {
+		sub.MarkOffset(msg.Offset, metadata)
+	}
+}
+
+// MarkPartitionOffset marks an offset of the provided topic/partition as processed.
+// See MarkOffset for additional explanation.
+func (c *Consumer) MarkPartitionOffset(topic string, partition int32, offset int64, metadata string) {
+	if sub := c.subs.Fetch(topic, partition); sub != nil {
+		sub.MarkOffset(offset, metadata)
+	}
+}
+
+// MarkOffsets marks stashed offsets as processed.
+// See MarkOffset for additional explanation.
+func (c *Consumer) MarkOffsets(s *OffsetStash) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	for tp, info := range s.offsets {
+		if sub := c.subs.Fetch(tp.Topic, tp.Partition); sub != nil {
+			sub.MarkOffset(info.Offset, info.Metadata)
+		}
+		delete(s.offsets, tp)
+	}
+}
+
+// ResetOffsets marks the provided message as processed, alongside a metadata string
+// that represents the state of the partition consumer at that point in time. The
+// metadata string can be used by another consumer to restore that state, so it
+// can resume consumption.
+//
+// Difference between ResetOffset and MarkOffset is that it allows to rewind to an earlier offset
+func (c *Consumer) ResetOffset(msg *sarama.ConsumerMessage, metadata string) {
+	if sub := c.subs.Fetch(msg.Topic, msg.Partition); sub != nil {
+		sub.ResetOffset(msg.Offset, metadata)
+	}
+}
+
+// ResetPartitionOffset marks an offset of the provided topic/partition as processed.
+// See ResetOffset for additional explanation.
+func (c *Consumer) ResetPartitionOffset(topic string, partition int32, offset int64, metadata string) {
+	sub := c.subs.Fetch(topic, partition)
+	if sub != nil {
+		sub.ResetOffset(offset, metadata)
+	}
+}
+
+// ResetOffsets marks stashed offsets as processed.
+// See ResetOffset for additional explanation.
+func (c *Consumer) ResetOffsets(s *OffsetStash) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	for tp, info := range s.offsets {
+		if sub := c.subs.Fetch(tp.Topic, tp.Partition); sub != nil {
+			sub.ResetOffset(info.Offset, info.Metadata)
+		}
+		delete(s.offsets, tp)
+	}
+}
+
+// Subscriptions returns the consumed topics and partitions
+func (c *Consumer) Subscriptions() map[string][]int32 {
+	return c.subs.Info()
+}
+
+// CommitOffsets allows to manually commit previously marked offsets. By default there is no
+// need to call this function as the consumer will commit offsets automatically
+// using the Config.Consumer.Offsets.CommitInterval setting.
+//
+// Please be aware that calling this function during an internal rebalance cycle may return
+// broker errors (e.g. sarama.ErrUnknownMemberId or sarama.ErrIllegalGeneration).
+func (c *Consumer) CommitOffsets() error {
+	c.commitMu.Lock()
+	defer c.commitMu.Unlock()
+
+	memberID, generationID := c.membership()
+	req := &sarama.OffsetCommitRequest{
+		Version:                 2,
+		ConsumerGroup:           c.groupID,
+		ConsumerGroupGeneration: generationID,
+		ConsumerID:              memberID,
+		RetentionTime:           -1,
+	}
+
+	if ns := c.client.config.Consumer.Offsets.Retention; ns != 0 {
+		req.RetentionTime = int64(ns / time.Millisecond)
+	}
+
+	snap := c.subs.Snapshot()
+	dirty := false
+	for tp, state := range snap {
+		if state.Dirty {
+			dirty = true
+			req.AddBlock(tp.Topic, tp.Partition, state.Info.Offset, 0, state.Info.Metadata)
+		}
+	}
+	if !dirty {
+		return nil
+	}
+
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return err
+	}
+
+	resp, err := broker.CommitOffset(req)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return err
+	}
+
+	for topic, errs := range resp.Errors {
+		for partition, kerr := range errs {
+			if kerr != sarama.ErrNoError {
+				err = kerr
+			} else if state, ok := snap[topicPartition{topic, partition}]; ok {
+				if sub := c.subs.Fetch(topic, partition); sub != nil {
+					sub.markCommitted(state.Info.Offset)
+				}
+			}
+		}
+	}
+	return err
+}
+
+// Close safely closes the consumer and releases all resources
+func (c *Consumer) Close() (err error) {
+	c.closeOnce.Do(func() {
+		close(c.dying)
+		<-c.dead
+
+		if e := c.release(); e != nil {
+			err = e
+		}
+		if e := c.consumer.Close(); e != nil {
+			err = e
+		}
+		close(c.messages)
+		close(c.errors)
+
+		if e := c.leaveGroup(); e != nil {
+			err = e
+		}
+		close(c.partitions)
+		close(c.notifications)
+
+		// drain
+		for range c.messages {
+		}
+		for range c.errors {
+		}
+		for p := range c.partitions {
+			_ = p.Close()
+		}
+		for range c.notifications {
+		}
+
+		c.client.release()
+		if c.ownClient {
+			if e := c.client.Close(); e != nil {
+				err = e
+			}
+		}
+	})
+	return
+}
+
+func (c *Consumer) mainLoop() {
+	defer close(c.dead)
+	defer atomic.StoreInt32(&c.consuming, 0)
+
+	for {
+		atomic.StoreInt32(&c.consuming, 0)
+
+		// Check if close was requested
+		select {
+		case <-c.dying:
+			return
+		default:
+		}
+
+		// Start next consume cycle
+		c.nextTick()
+	}
+}
+
+func (c *Consumer) nextTick() {
+	// Remember previous subscriptions
+	var notification *Notification
+	if c.client.config.Group.Return.Notifications {
+		notification = newNotification(c.subs.Info())
+	}
+
+	// Refresh coordinator
+	if err := c.refreshCoordinator(); err != nil {
+		c.rebalanceError(err, nil)
+		return
+	}
+
+	// Release subscriptions
+	if err := c.release(); err != nil {
+		c.rebalanceError(err, nil)
+		return
+	}
+
+	// Issue rebalance start notification
+	if c.client.config.Group.Return.Notifications {
+		c.handleNotification(notification)
+	}
+
+	// Rebalance, fetch new subscriptions
+	subs, err := c.rebalance()
+	if err != nil {
+		c.rebalanceError(err, notification)
+		return
+	}
+
+	// Coordinate loops, make sure everything is
+	// stopped on exit
+	tomb := newLoopTomb()
+	defer tomb.Close()
+
+	// Start the heartbeat
+	tomb.Go(c.hbLoop)
+
+	// Subscribe to topic/partitions
+	if err := c.subscribe(tomb, subs); err != nil {
+		c.rebalanceError(err, notification)
+		return
+	}
+
+	// Update/issue notification with new claims
+	if c.client.config.Group.Return.Notifications {
+		notification = notification.success(subs)
+		c.handleNotification(notification)
+	}
+
+	// Start topic watcher loop
+	tomb.Go(c.twLoop)
+
+	// Start consuming and committing offsets
+	tomb.Go(c.cmLoop)
+	atomic.StoreInt32(&c.consuming, 1)
+
+	// Wait for signals
+	select {
+	case <-tomb.Dying():
+	case <-c.dying:
+	}
+}
+
+// heartbeat loop, triggered by the mainLoop
+func (c *Consumer) hbLoop(stopped <-chan none) {
+	ticker := time.NewTicker(c.client.config.Group.Heartbeat.Interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ticker.C:
+			switch err := c.heartbeat(); err {
+			case nil, sarama.ErrNoError:
+			case sarama.ErrNotCoordinatorForConsumer, sarama.ErrRebalanceInProgress:
+				return
+			default:
+				c.handleError(&Error{Ctx: "heartbeat", error: err})
+				return
+			}
+		case <-stopped:
+			return
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+// topic watcher loop, triggered by the mainLoop
+func (c *Consumer) twLoop(stopped <-chan none) {
+	ticker := time.NewTicker(c.client.config.Metadata.RefreshFrequency / 2)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ticker.C:
+			topics, err := c.client.Topics()
+			if err != nil {
+				c.handleError(&Error{Ctx: "topics", error: err})
+				return
+			}
+
+			for _, topic := range topics {
+				if !c.isKnownCoreTopic(topic) &&
+					!c.isKnownExtraTopic(topic) &&
+					c.isPotentialExtraTopic(topic) {
+					return
+				}
+			}
+		case <-stopped:
+			return
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+// commit loop, triggered by the mainLoop
+func (c *Consumer) cmLoop(stopped <-chan none) {
+	ticker := time.NewTicker(c.client.config.Consumer.Offsets.CommitInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ticker.C:
+			if err := c.commitOffsetsWithRetry(c.client.config.Group.Offsets.Retry.Max); err != nil {
+				c.handleError(&Error{Ctx: "commit", error: err})
+				return
+			}
+		case <-stopped:
+			return
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+func (c *Consumer) rebalanceError(err error, n *Notification) {
+	if n != nil {
+		n.Type = RebalanceError
+		c.handleNotification(n)
+	}
+
+	switch err {
+	case sarama.ErrRebalanceInProgress:
+	default:
+		c.handleError(&Error{Ctx: "rebalance", error: err})
+	}
+
+	select {
+	case <-c.dying:
+	case <-time.After(c.client.config.Metadata.Retry.Backoff):
+	}
+}
+
+func (c *Consumer) handleNotification(n *Notification) {
+	if c.client.config.Group.Return.Notifications {
+		select {
+		case c.notifications <- n:
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+func (c *Consumer) handleError(e *Error) {
+	if c.client.config.Consumer.Return.Errors {
+		select {
+		case c.errors <- e:
+		case <-c.dying:
+			return
+		}
+	} else {
+		sarama.Logger.Printf("%s error: %s\n", e.Ctx, e.Error())
+	}
+}
+
+// Releases the consumer and commits offsets, called from rebalance() and Close()
+func (c *Consumer) release() (err error) {
+	// Stop all consumers
+	c.subs.Stop()
+
+	// Clear subscriptions on exit
+	defer c.subs.Clear()
+
+	// Wait for messages to be processed
+	timeout := time.NewTimer(c.client.config.Group.Offsets.Synchronization.DwellTime)
+	defer timeout.Stop()
+
+	select {
+	case <-c.dying:
+	case <-timeout.C:
+	}
+
+	// Commit offsets, continue on errors
+	if e := c.commitOffsetsWithRetry(c.client.config.Group.Offsets.Retry.Max); e != nil {
+		err = e
+	}
+
+	return
+}
+
+// --------------------------------------------------------------------
+
+// Performs a heartbeat, part of the mainLoop()
+func (c *Consumer) heartbeat() error {
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return err
+	}
+
+	memberID, generationID := c.membership()
+	resp, err := broker.Heartbeat(&sarama.HeartbeatRequest{
+		GroupId:      c.groupID,
+		MemberId:     memberID,
+		GenerationId: generationID,
+	})
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return err
+	}
+	return resp.Err
+}
+
+// Performs a rebalance, part of the mainLoop()
+func (c *Consumer) rebalance() (map[string][]int32, error) {
+	memberID, _ := c.membership()
+	sarama.Logger.Printf("cluster/consumer %s rebalance\n", memberID)
+
+	allTopics, err := c.client.Topics()
+	if err != nil {
+		return nil, err
+	}
+	c.extraTopics = c.selectExtraTopics(allTopics)
+	sort.Strings(c.extraTopics)
+
+	// Re-join consumer group
+	strategy, err := c.joinGroup()
+	switch {
+	case err == sarama.ErrUnknownMemberId:
+		c.membershipMu.Lock()
+		c.memberID = ""
+		c.membershipMu.Unlock()
+		return nil, err
+	case err != nil:
+		return nil, err
+	}
+
+	// Sync consumer group state, fetch subscriptions
+	subs, err := c.syncGroup(strategy)
+	switch {
+	case err == sarama.ErrRebalanceInProgress:
+		return nil, err
+	case err != nil:
+		_ = c.leaveGroup()
+		return nil, err
+	}
+	return subs, nil
+}
+
+// Performs the subscription, part of the mainLoop()
+func (c *Consumer) subscribe(tomb *loopTomb, subs map[string][]int32) error {
+	// fetch offsets
+	offsets, err := c.fetchOffsets(subs)
+	if err != nil {
+		_ = c.leaveGroup()
+		return err
+	}
+
+	// create consumers in parallel
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+
+	for topic, partitions := range subs {
+		for _, partition := range partitions {
+			wg.Add(1)
+
+			info := offsets[topic][partition]
+			go func(topic string, partition int32) {
+				if e := c.createConsumer(tomb, topic, partition, info); e != nil {
+					mu.Lock()
+					err = e
+					mu.Unlock()
+				}
+				wg.Done()
+			}(topic, partition)
+		}
+	}
+	wg.Wait()
+
+	if err != nil {
+		_ = c.release()
+		_ = c.leaveGroup()
+	}
+	return err
+}
+
+// --------------------------------------------------------------------
+
+// Send a request to the broker to join group on rebalance()
+func (c *Consumer) joinGroup() (*balancer, error) {
+	memberID, _ := c.membership()
+	req := &sarama.JoinGroupRequest{
+		GroupId:        c.groupID,
+		MemberId:       memberID,
+		SessionTimeout: int32(c.client.config.Group.Session.Timeout / time.Millisecond),
+		ProtocolType:   "consumer",
+	}
+
+	meta := &sarama.ConsumerGroupMemberMetadata{
+		Version:  1,
+		Topics:   append(c.coreTopics, c.extraTopics...),
+		UserData: c.client.config.Group.Member.UserData,
+	}
+	err := req.AddGroupProtocolMetadata(string(StrategyRange), meta)
+	if err != nil {
+		return nil, err
+	}
+	err = req.AddGroupProtocolMetadata(string(StrategyRoundRobin), meta)
+	if err != nil {
+		return nil, err
+	}
+
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	}
+
+	resp, err := broker.JoinGroup(req)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	} else if resp.Err != sarama.ErrNoError {
+		c.closeCoordinator(broker, resp.Err)
+		return nil, resp.Err
+	}
+
+	var strategy *balancer
+	if resp.LeaderId == resp.MemberId {
+		members, err := resp.GetMembers()
+		if err != nil {
+			return nil, err
+		}
+
+		strategy, err = newBalancerFromMeta(c.client, members)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	c.membershipMu.Lock()
+	c.memberID = resp.MemberId
+	c.generationID = resp.GenerationId
+	c.membershipMu.Unlock()
+
+	return strategy, nil
+}
+
+// Send a request to the broker to sync the group on rebalance().
+// Returns a list of topics and partitions to consume.
+func (c *Consumer) syncGroup(strategy *balancer) (map[string][]int32, error) {
+	memberID, generationID := c.membership()
+	req := &sarama.SyncGroupRequest{
+		GroupId:      c.groupID,
+		MemberId:     memberID,
+		GenerationId: generationID,
+	}
+
+	if strategy != nil {
+		for memberID, topics := range strategy.Perform(c.client.config.Group.PartitionStrategy) {
+			if err := req.AddGroupAssignmentMember(memberID, &sarama.ConsumerGroupMemberAssignment{
+				Topics: topics,
+			}); err != nil {
+				return nil, err
+			}
+		}
+	}
+
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	}
+
+	resp, err := broker.SyncGroup(req)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	} else if resp.Err != sarama.ErrNoError {
+		c.closeCoordinator(broker, resp.Err)
+		return nil, resp.Err
+	}
+
+	// Return if there is nothing to subscribe to
+	if len(resp.MemberAssignment) == 0 {
+		return nil, nil
+	}
+
+	// Get assigned subscriptions
+	members, err := resp.GetMemberAssignment()
+	if err != nil {
+		return nil, err
+	}
+
+	// Sort partitions, for each topic
+	for topic := range members.Topics {
+		sort.Sort(int32Slice(members.Topics[topic]))
+	}
+	return members.Topics, nil
+}
+
+// Fetches latest committed offsets for all subscriptions
+func (c *Consumer) fetchOffsets(subs map[string][]int32) (map[string]map[int32]offsetInfo, error) {
+	offsets := make(map[string]map[int32]offsetInfo, len(subs))
+	req := &sarama.OffsetFetchRequest{
+		Version:       1,
+		ConsumerGroup: c.groupID,
+	}
+
+	for topic, partitions := range subs {
+		offsets[topic] = make(map[int32]offsetInfo, len(partitions))
+		for _, partition := range partitions {
+			offsets[topic][partition] = offsetInfo{Offset: -1}
+			req.AddPartition(topic, partition)
+		}
+	}
+
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	}
+
+	resp, err := broker.FetchOffset(req)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return nil, err
+	}
+
+	for topic, partitions := range subs {
+		for _, partition := range partitions {
+			block := resp.GetBlock(topic, partition)
+			if block == nil {
+				return nil, sarama.ErrIncompleteResponse
+			}
+
+			if block.Err == sarama.ErrNoError {
+				offsets[topic][partition] = offsetInfo{Offset: block.Offset, Metadata: block.Metadata}
+			} else {
+				return nil, block.Err
+			}
+		}
+	}
+	return offsets, nil
+}
+
+// Send a request to the broker to leave the group on failes rebalance() and on Close()
+func (c *Consumer) leaveGroup() error {
+	broker, err := c.client.Coordinator(c.groupID)
+	if err != nil {
+		c.closeCoordinator(broker, err)
+		return err
+	}
+
+	memberID, _ := c.membership()
+	if _, err = broker.LeaveGroup(&sarama.LeaveGroupRequest{
+		GroupId:  c.groupID,
+		MemberId: memberID,
+	}); err != nil {
+		c.closeCoordinator(broker, err)
+	}
+	return err
+}
+
+// --------------------------------------------------------------------
+
+func (c *Consumer) createConsumer(tomb *loopTomb, topic string, partition int32, info offsetInfo) error {
+	memberID, _ := c.membership()
+	sarama.Logger.Printf("cluster/consumer %s consume %s/%d from %d\n", memberID, topic, partition, info.NextOffset(c.client.config.Consumer.Offsets.Initial))
+
+	// Create partitionConsumer
+	pc, err := newPartitionConsumer(c.consumer, topic, partition, info, c.client.config.Consumer.Offsets.Initial)
+	if err != nil {
+		return err
+	}
+
+	// Store in subscriptions
+	c.subs.Store(topic, partition, pc)
+
+	// Start partition consumer goroutine
+	tomb.Go(func(stopper <-chan none) {
+		if c.client.config.Group.Mode == ConsumerModePartitions {
+			pc.waitFor(stopper, c.errors)
+		} else {
+			pc.multiplex(stopper, c.messages, c.errors)
+		}
+	})
+
+	if c.client.config.Group.Mode == ConsumerModePartitions {
+		c.partitions <- pc
+	}
+	return nil
+}
+
+func (c *Consumer) commitOffsetsWithRetry(retries int) error {
+	err := c.CommitOffsets()
+	if err != nil && retries > 0 {
+		return c.commitOffsetsWithRetry(retries - 1)
+	}
+	return err
+}
+
+func (c *Consumer) closeCoordinator(broker *sarama.Broker, err error) {
+	if broker != nil {
+		_ = broker.Close()
+	}
+
+	switch err {
+	case sarama.ErrConsumerCoordinatorNotAvailable, sarama.ErrNotCoordinatorForConsumer:
+		_ = c.client.RefreshCoordinator(c.groupID)
+	}
+}
+
+func (c *Consumer) selectExtraTopics(allTopics []string) []string {
+	extra := allTopics[:0]
+	for _, topic := range allTopics {
+		if !c.isKnownCoreTopic(topic) && c.isPotentialExtraTopic(topic) {
+			extra = append(extra, topic)
+		}
+	}
+	return extra
+}
+
+func (c *Consumer) isKnownCoreTopic(topic string) bool {
+	pos := sort.SearchStrings(c.coreTopics, topic)
+	return pos < len(c.coreTopics) && c.coreTopics[pos] == topic
+}
+
+func (c *Consumer) isKnownExtraTopic(topic string) bool {
+	pos := sort.SearchStrings(c.extraTopics, topic)
+	return pos < len(c.extraTopics) && c.extraTopics[pos] == topic
+}
+
+func (c *Consumer) isPotentialExtraTopic(topic string) bool {
+	rx := c.client.config.Group.Topics
+	if rx.Blacklist != nil && rx.Blacklist.MatchString(topic) {
+		return false
+	}
+	if rx.Whitelist != nil && rx.Whitelist.MatchString(topic) {
+		return true
+	}
+	return false
+}
+
+func (c *Consumer) refreshCoordinator() error {
+	if err := c.refreshMetadata(); err != nil {
+		return err
+	}
+	return c.client.RefreshCoordinator(c.groupID)
+}
+
+func (c *Consumer) refreshMetadata() (err error) {
+	if c.client.config.Metadata.Full {
+		err = c.client.RefreshMetadata()
+	} else {
+		var topics []string
+		if topics, err = c.client.Topics(); err == nil && len(topics) != 0 {
+			err = c.client.RefreshMetadata(topics...)
+		}
+	}
+
+	// maybe we didn't have authorization to describe all topics
+	switch err {
+	case sarama.ErrTopicAuthorizationFailed:
+		err = c.client.RefreshMetadata(c.coreTopics...)
+	}
+	return
+}
+
+func (c *Consumer) membership() (memberID string, generationID int32) {
+	c.membershipMu.RLock()
+	memberID, generationID = c.memberID, c.generationID
+	c.membershipMu.RUnlock()
+	return
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/doc.go b/vendor/github.com/bsm/sarama-cluster/doc.go
new file mode 100644
index 0000000..9c8ff16
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/doc.go
@@ -0,0 +1,8 @@
+/*
+Package cluster provides cluster extensions for Sarama, enabing users
+to consume topics across from multiple, balanced nodes.
+
+It requires Kafka v0.9+ and follows the steps guide, described in:
+https://cwiki.apache.org/confluence/display/KAFKA/Kafka+0.9+Consumer+Rewrite+Design
+*/
+package cluster
diff --git a/vendor/github.com/bsm/sarama-cluster/offsets.go b/vendor/github.com/bsm/sarama-cluster/offsets.go
new file mode 100644
index 0000000..4223ac5
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/offsets.go
@@ -0,0 +1,69 @@
+package cluster
+
+import (
+	"sync"
+
+	"github.com/Shopify/sarama"
+)
+
+// OffsetStash allows to accumulate offsets and
+// mark them as processed in a bulk
+type OffsetStash struct {
+	offsets map[topicPartition]offsetInfo
+	mu      sync.Mutex
+}
+
+// NewOffsetStash inits a blank stash
+func NewOffsetStash() *OffsetStash {
+	return &OffsetStash{offsets: make(map[topicPartition]offsetInfo)}
+}
+
+// MarkOffset stashes the provided message offset
+func (s *OffsetStash) MarkOffset(msg *sarama.ConsumerMessage, metadata string) {
+	s.MarkPartitionOffset(msg.Topic, msg.Partition, msg.Offset, metadata)
+}
+
+// MarkPartitionOffset stashes the offset for the provided topic/partition combination
+func (s *OffsetStash) MarkPartitionOffset(topic string, partition int32, offset int64, metadata string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	key := topicPartition{Topic: topic, Partition: partition}
+	if info := s.offsets[key]; offset >= info.Offset {
+		info.Offset = offset
+		info.Metadata = metadata
+		s.offsets[key] = info
+	}
+}
+
+// ResetPartitionOffset stashes the offset for the provided topic/partition combination.
+// Difference between ResetPartitionOffset and MarkPartitionOffset is that, ResetPartitionOffset supports earlier offsets
+func (s *OffsetStash) ResetPartitionOffset(topic string, partition int32, offset int64, metadata string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	key := topicPartition{Topic: topic, Partition: partition}
+	if info := s.offsets[key]; offset <= info.Offset {
+		info.Offset = offset
+		info.Metadata = metadata
+		s.offsets[key] = info
+	}
+}
+
+// ResetOffset stashes the provided message offset
+// See ResetPartitionOffset for explanation
+func (s *OffsetStash) ResetOffset(msg *sarama.ConsumerMessage, metadata string) {
+	s.ResetPartitionOffset(msg.Topic, msg.Partition, msg.Offset, metadata)
+}
+
+// Offsets returns the latest stashed offsets by topic-partition
+func (s *OffsetStash) Offsets() map[string]int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	res := make(map[string]int64, len(s.offsets))
+	for tp, info := range s.offsets {
+		res[tp.String()] = info.Offset
+	}
+	return res
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/partitions.go b/vendor/github.com/bsm/sarama-cluster/partitions.go
new file mode 100644
index 0000000..bfaa587
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/partitions.go
@@ -0,0 +1,290 @@
+package cluster
+
+import (
+	"sort"
+	"sync"
+	"time"
+
+	"github.com/Shopify/sarama"
+)
+
+// PartitionConsumer allows code to consume individual partitions from the cluster.
+//
+// See docs for Consumer.Partitions() for more on how to implement this.
+type PartitionConsumer interface {
+	sarama.PartitionConsumer
+
+	// Topic returns the consumed topic name
+	Topic() string
+
+	// Partition returns the consumed partition
+	Partition() int32
+
+	// InitialOffset returns the offset used for creating the PartitionConsumer instance.
+	// The returned offset can be a literal offset, or OffsetNewest, or OffsetOldest
+	InitialOffset() int64
+  
+	// MarkOffset marks the offset of a message as preocessed.
+	MarkOffset(offset int64, metadata string)
+
+	// ResetOffset resets the offset to a previously processed message.
+	ResetOffset(offset int64, metadata string)
+}
+
+type partitionConsumer struct {
+	sarama.PartitionConsumer
+
+	state partitionState
+	mu    sync.Mutex
+
+	topic         string
+	partition     int32
+	initialOffset int64
+
+	closeOnce sync.Once
+	closeErr  error
+
+	dying, dead chan none
+}
+
+func newPartitionConsumer(manager sarama.Consumer, topic string, partition int32, info offsetInfo, defaultOffset int64) (*partitionConsumer, error) {
+	offset := info.NextOffset(defaultOffset)
+	pcm, err := manager.ConsumePartition(topic, partition, offset)
+
+	// Resume from default offset, if requested offset is out-of-range
+	if err == sarama.ErrOffsetOutOfRange {
+		info.Offset = -1
+		offset = defaultOffset
+		pcm, err = manager.ConsumePartition(topic, partition, offset)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	return &partitionConsumer{
+		PartitionConsumer: pcm,
+		state:             partitionState{Info: info},
+
+		topic:         topic,
+		partition:     partition,
+		initialOffset: offset,
+
+		dying: make(chan none),
+		dead:  make(chan none),
+	}, nil
+}
+
+// Topic implements PartitionConsumer
+func (c *partitionConsumer) Topic() string { return c.topic }
+
+// Partition implements PartitionConsumer
+func (c *partitionConsumer) Partition() int32 { return c.partition }
+
+// InitialOffset implements PartitionConsumer
+func (c *partitionConsumer) InitialOffset() int64 { return c.initialOffset }
+
+// AsyncClose implements PartitionConsumer
+func (c *partitionConsumer) AsyncClose() {
+	c.closeOnce.Do(func() {
+		c.closeErr = c.PartitionConsumer.Close()
+		close(c.dying)
+	})
+}
+
+// Close implements PartitionConsumer
+func (c *partitionConsumer) Close() error {
+	c.AsyncClose()
+	<-c.dead
+	return c.closeErr
+}
+
+func (c *partitionConsumer) waitFor(stopper <-chan none, errors chan<- error) {
+	defer close(c.dead)
+
+	for {
+		select {
+		case err, ok := <-c.Errors():
+			if !ok {
+				return
+			}
+			select {
+			case errors <- err:
+			case <-stopper:
+				return
+			case <-c.dying:
+				return
+			}
+		case <-stopper:
+			return
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+func (c *partitionConsumer) multiplex(stopper <-chan none, messages chan<- *sarama.ConsumerMessage, errors chan<- error) {
+	defer close(c.dead)
+
+	for {
+		select {
+		case msg, ok := <-c.Messages():
+			if !ok {
+				return
+			}
+			select {
+			case messages <- msg:
+			case <-stopper:
+				return
+			case <-c.dying:
+				return
+			}
+		case err, ok := <-c.Errors():
+			if !ok {
+				return
+			}
+			select {
+			case errors <- err:
+			case <-stopper:
+				return
+			case <-c.dying:
+				return
+			}
+		case <-stopper:
+			return
+		case <-c.dying:
+			return
+		}
+	}
+}
+
+func (c *partitionConsumer) getState() partitionState {
+	c.mu.Lock()
+	state := c.state
+	c.mu.Unlock()
+
+	return state
+}
+
+func (c *partitionConsumer) markCommitted(offset int64) {
+	c.mu.Lock()
+	if offset == c.state.Info.Offset {
+		c.state.Dirty = false
+	}
+	c.mu.Unlock()
+}
+
+// MarkOffset implements PartitionConsumer
+func (c *partitionConsumer) MarkOffset(offset int64, metadata string) {
+	c.mu.Lock()
+	if next := offset + 1; next > c.state.Info.Offset {
+		c.state.Info.Offset = next
+		c.state.Info.Metadata = metadata
+		c.state.Dirty = true
+	}
+	c.mu.Unlock()
+}
+
+// ResetOffset implements PartitionConsumer
+func (c *partitionConsumer) ResetOffset(offset int64, metadata string) {
+	c.mu.Lock()
+	if next := offset + 1; next <= c.state.Info.Offset {
+		c.state.Info.Offset = next
+		c.state.Info.Metadata = metadata
+		c.state.Dirty = true
+	}
+	c.mu.Unlock()
+}
+
+// --------------------------------------------------------------------
+
+type partitionState struct {
+	Info       offsetInfo
+	Dirty      bool
+	LastCommit time.Time
+}
+
+// --------------------------------------------------------------------
+
+type partitionMap struct {
+	data map[topicPartition]*partitionConsumer
+	mu   sync.RWMutex
+}
+
+func newPartitionMap() *partitionMap {
+	return &partitionMap{
+		data: make(map[topicPartition]*partitionConsumer),
+	}
+}
+
+func (m *partitionMap) IsSubscribedTo(topic string) bool {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+
+	for tp := range m.data {
+		if tp.Topic == topic {
+			return true
+		}
+	}
+	return false
+}
+
+func (m *partitionMap) Fetch(topic string, partition int32) *partitionConsumer {
+	m.mu.RLock()
+	pc, _ := m.data[topicPartition{topic, partition}]
+	m.mu.RUnlock()
+	return pc
+}
+
+func (m *partitionMap) Store(topic string, partition int32, pc *partitionConsumer) {
+	m.mu.Lock()
+	m.data[topicPartition{topic, partition}] = pc
+	m.mu.Unlock()
+}
+
+func (m *partitionMap) Snapshot() map[topicPartition]partitionState {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+
+	snap := make(map[topicPartition]partitionState, len(m.data))
+	for tp, pc := range m.data {
+		snap[tp] = pc.getState()
+	}
+	return snap
+}
+
+func (m *partitionMap) Stop() {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+
+	var wg sync.WaitGroup
+	for tp := range m.data {
+		wg.Add(1)
+		go func(p *partitionConsumer) {
+			_ = p.Close()
+			wg.Done()
+		}(m.data[tp])
+	}
+	wg.Wait()
+}
+
+func (m *partitionMap) Clear() {
+	m.mu.Lock()
+	for tp := range m.data {
+		delete(m.data, tp)
+	}
+	m.mu.Unlock()
+}
+
+func (m *partitionMap) Info() map[string][]int32 {
+	info := make(map[string][]int32)
+	m.mu.RLock()
+	for tp := range m.data {
+		info[tp.Topic] = append(info[tp.Topic], tp.Partition)
+	}
+	m.mu.RUnlock()
+
+	for topic := range info {
+		sort.Sort(int32Slice(info[topic]))
+	}
+	return info
+}
diff --git a/vendor/github.com/bsm/sarama-cluster/util.go b/vendor/github.com/bsm/sarama-cluster/util.go
new file mode 100644
index 0000000..e7cb5dd
--- /dev/null
+++ b/vendor/github.com/bsm/sarama-cluster/util.go
@@ -0,0 +1,75 @@
+package cluster
+
+import (
+	"fmt"
+	"sort"
+	"sync"
+)
+
+type none struct{}
+
+type topicPartition struct {
+	Topic     string
+	Partition int32
+}
+
+func (tp *topicPartition) String() string {
+	return fmt.Sprintf("%s-%d", tp.Topic, tp.Partition)
+}
+
+type offsetInfo struct {
+	Offset   int64
+	Metadata string
+}
+
+func (i offsetInfo) NextOffset(fallback int64) int64 {
+	if i.Offset > -1 {
+		return i.Offset
+	}
+	return fallback
+}
+
+type int32Slice []int32
+
+func (p int32Slice) Len() int           { return len(p) }
+func (p int32Slice) Less(i, j int) bool { return p[i] < p[j] }
+func (p int32Slice) Swap(i, j int)      { p[i], p[j] = p[j], p[i] }
+
+func (p int32Slice) Diff(o int32Slice) (res []int32) {
+	on := len(o)
+	for _, x := range p {
+		n := sort.Search(on, func(i int) bool { return o[i] >= x })
+		if n < on && o[n] == x {
+			continue
+		}
+		res = append(res, x)
+	}
+	return
+}
+
+// --------------------------------------------------------------------
+
+type loopTomb struct {
+	c chan none
+	o sync.Once
+	w sync.WaitGroup
+}
+
+func newLoopTomb() *loopTomb {
+	return &loopTomb{c: make(chan none)}
+}
+
+func (t *loopTomb) stop()  { t.o.Do(func() { close(t.c) }) }
+func (t *loopTomb) Close() { t.stop(); t.w.Wait() }
+
+func (t *loopTomb) Dying() <-chan none { return t.c }
+func (t *loopTomb) Go(f func(<-chan none)) {
+	t.w.Add(1)
+
+	go func() {
+		defer t.stop()
+		defer t.w.Done()
+
+		f(t.c)
+	}()
+}
diff --git a/vendor/github.com/coreos/go-systemd/LICENSE b/vendor/github.com/coreos/go-systemd/LICENSE
new file mode 100644
index 0000000..37ec93a
--- /dev/null
+++ b/vendor/github.com/coreos/go-systemd/LICENSE
@@ -0,0 +1,191 @@
+Apache License
+Version 2.0, January 2004
+http://www.apache.org/licenses/
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+1. Definitions.
+
+"License" shall mean the terms and conditions for use, reproduction, and
+distribution as defined by Sections 1 through 9 of this document.
+
+"Licensor" shall mean the copyright owner or entity authorized by the copyright
+owner that is granting the License.
+
+"Legal Entity" shall mean the union of the acting entity and all other entities
+that control, are controlled by, or are under common control with that entity.
+For the purposes of this definition, "control" means (i) the power, direct or
+indirect, to cause the direction or management of such entity, whether by
+contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the
+outstanding shares, or (iii) beneficial ownership of such entity.
+
+"You" (or "Your") shall mean an individual or Legal Entity exercising
+permissions granted by this License.
+
+"Source" form shall mean the preferred form for making modifications, including
+but not limited to software source code, documentation source, and configuration
+files.
+
+"Object" form shall mean any form resulting from mechanical transformation or
+translation of a Source form, including but not limited to compiled object code,
+generated documentation, and conversions to other media types.
+
+"Work" shall mean the work of authorship, whether in Source or Object form, made
+available under the License, as indicated by a copyright notice that is included
+in or attached to the work (an example is provided in the Appendix below).
+
+"Derivative Works" shall mean any work, whether in Source or Object form, that
+is based on (or derived from) the Work and for which the editorial revisions,
+annotations, elaborations, or other modifications represent, as a whole, an
+original work of authorship. For the purposes of this License, Derivative Works
+shall not include works that remain separable from, or merely link (or bind by
+name) to the interfaces of, the Work and Derivative Works thereof.
+
+"Contribution" shall mean any work of authorship, including the original version
+of the Work and any modifications or additions to that Work or Derivative Works
+thereof, that is intentionally submitted to Licensor for inclusion in the Work
+by the copyright owner or by an individual or Legal Entity authorized to submit
+on behalf of the copyright owner. For the purposes of this definition,
+"submitted" means any form of electronic, verbal, or written communication sent
+to the Licensor or its representatives, including but not limited to
+communication on electronic mailing lists, source code control systems, and
+issue tracking systems that are managed by, or on behalf of, the Licensor for
+the purpose of discussing and improving the Work, but excluding communication
+that is conspicuously marked or otherwise designated in writing by the copyright
+owner as "Not a Contribution."
+
+"Contributor" shall mean Licensor and any individual or Legal Entity on behalf
+of whom a Contribution has been received by Licensor and subsequently
+incorporated within the Work.
+
+2. Grant of Copyright License.
+
+Subject to the terms and conditions of this License, each Contributor hereby
+grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free,
+irrevocable copyright license to reproduce, prepare Derivative Works of,
+publicly display, publicly perform, sublicense, and distribute the Work and such
+Derivative Works in Source or Object form.
+
+3. Grant of Patent License.
+
+Subject to the terms and conditions of this License, each Contributor hereby
+grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free,
+irrevocable (except as stated in this section) patent license to make, have
+made, use, offer to sell, sell, import, and otherwise transfer the Work, where
+such license applies only to those patent claims licensable by such Contributor
+that are necessarily infringed by their Contribution(s) alone or by combination
+of their Contribution(s) with the Work to which such Contribution(s) was
+submitted. If You institute patent litigation against any entity (including a
+cross-claim or counterclaim in a lawsuit) alleging that the Work or a
+Contribution incorporated within the Work constitutes direct or contributory
+patent infringement, then any patent licenses granted to You under this License
+for that Work shall terminate as of the date such litigation is filed.
+
+4. Redistribution.
+
+You may reproduce and distribute copies of the Work or Derivative Works thereof
+in any medium, with or without modifications, and in Source or Object form,
+provided that You meet the following conditions:
+
+You must give any other recipients of the Work or Derivative Works a copy of
+this License; and
+You must cause any modified files to carry prominent notices stating that You
+changed the files; and
+You must retain, in the Source form of any Derivative Works that You distribute,
+all copyright, patent, trademark, and attribution notices from the Source form
+of the Work, excluding those notices that do not pertain to any part of the
+Derivative Works; and
+If the Work includes a "NOTICE" text file as part of its distribution, then any
+Derivative Works that You distribute must include a readable copy of the
+attribution notices contained within such NOTICE file, excluding those notices
+that do not pertain to any part of the Derivative Works, in at least one of the
+following places: within a NOTICE text file distributed as part of the
+Derivative Works; within the Source form or documentation, if provided along
+with the Derivative Works; or, within a display generated by the Derivative
+Works, if and wherever such third-party notices normally appear. The contents of
+the NOTICE file are for informational purposes only and do not modify the
+License. You may add Your own attribution notices within Derivative Works that
+You distribute, alongside or as an addendum to the NOTICE text from the Work,
+provided that such additional attribution notices cannot be construed as
+modifying the License.
+You may add Your own copyright statement to Your modifications and may provide
+additional or different license terms and conditions for use, reproduction, or
+distribution of Your modifications, or for any such Derivative Works as a whole,
+provided Your use, reproduction, and distribution of the Work otherwise complies
+with the conditions stated in this License.
+
+5. Submission of Contributions.
+
+Unless You explicitly state otherwise, any Contribution intentionally submitted
+for inclusion in the Work by You to the Licensor shall be under the terms and
+conditions of this License, without any additional terms or conditions.
+Notwithstanding the above, nothing herein shall supersede or modify the terms of
+any separate license agreement you may have executed with Licensor regarding
+such Contributions.
+
+6. Trademarks.
+
+This License does not grant permission to use the trade names, trademarks,
+service marks, or product names of the Licensor, except as required for
+reasonable and customary use in describing the origin of the Work and
+reproducing the content of the NOTICE file.
+
+7. Disclaimer of Warranty.
+
+Unless required by applicable law or agreed to in writing, Licensor provides the
+Work (and each Contributor provides its Contributions) on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied,
+including, without limitation, any warranties or conditions of TITLE,
+NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are
+solely responsible for determining the appropriateness of using or
+redistributing the Work and assume any risks associated with Your exercise of
+permissions under this License.
+
+8. Limitation of Liability.
+
+In no event and under no legal theory, whether in tort (including negligence),
+contract, or otherwise, unless required by applicable law (such as deliberate
+and grossly negligent acts) or agreed to in writing, shall any Contributor be
+liable to You for damages, including any direct, indirect, special, incidental,
+or consequential damages of any character arising as a result of this License or
+out of the use or inability to use the Work (including but not limited to
+damages for loss of goodwill, work stoppage, computer failure or malfunction, or
+any and all other commercial damages or losses), even if such Contributor has
+been advised of the possibility of such damages.
+
+9. Accepting Warranty or Additional Liability.
+
+While redistributing the Work or Derivative Works thereof, You may choose to
+offer, and charge a fee for, acceptance of support, warranty, indemnity, or
+other liability obligations and/or rights consistent with this License. However,
+in accepting such obligations, You may act only on Your own behalf and on Your
+sole responsibility, not on behalf of any other Contributor, and only if You
+agree to indemnify, defend, and hold each Contributor harmless for any liability
+incurred by, or claims asserted against, such Contributor by reason of your
+accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS
+
+APPENDIX: How to apply the Apache License to your work
+
+To apply the Apache License to your work, attach the following boilerplate
+notice, with the fields enclosed by brackets "[]" replaced with your own
+identifying information. (Don't include the brackets!) The text should be
+enclosed in the appropriate comment syntax for the file format. We also
+recommend that a file or class name and description of purpose be included on
+the same "printed page" as the copyright notice for easier identification within
+third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+     http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
diff --git a/vendor/github.com/coreos/go-systemd/NOTICE b/vendor/github.com/coreos/go-systemd/NOTICE
new file mode 100644
index 0000000..23a0ada
--- /dev/null
+++ b/vendor/github.com/coreos/go-systemd/NOTICE
@@ -0,0 +1,5 @@
+CoreOS Project
+Copyright 2018 CoreOS, Inc
+
+This product includes software developed at CoreOS, Inc.
+(http://www.coreos.com/).
diff --git a/vendor/github.com/coreos/go-systemd/journal/journal.go b/vendor/github.com/coreos/go-systemd/journal/journal.go
new file mode 100644
index 0000000..a0f4837
--- /dev/null
+++ b/vendor/github.com/coreos/go-systemd/journal/journal.go
@@ -0,0 +1,225 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package journal provides write bindings to the local systemd journal.
+// It is implemented in pure Go and connects to the journal directly over its
+// unix socket.
+//
+// To read from the journal, see the "sdjournal" package, which wraps the
+// sd-journal a C API.
+//
+// http://www.freedesktop.org/software/systemd/man/systemd-journald.service.html
+package journal
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"net"
+	"os"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+)
+
+// Priority of a journal message
+type Priority int
+
+const (
+	PriEmerg Priority = iota
+	PriAlert
+	PriCrit
+	PriErr
+	PriWarning
+	PriNotice
+	PriInfo
+	PriDebug
+)
+
+var (
+	// This can be overridden at build-time:
+	// https://github.com/golang/go/wiki/GcToolchainTricks#including-build-information-in-the-executable
+	journalSocket = "/run/systemd/journal/socket"
+
+	// unixConnPtr atomically holds the local unconnected Unix-domain socket.
+	// Concrete safe pointer type: *net.UnixConn
+	unixConnPtr unsafe.Pointer
+	// onceConn ensures that unixConnPtr is initialized exactly once.
+	onceConn sync.Once
+)
+
+func init() {
+	onceConn.Do(initConn)
+}
+
+// Enabled checks whether the local systemd journal is available for logging.
+func Enabled() bool {
+	onceConn.Do(initConn)
+
+	if (*net.UnixConn)(atomic.LoadPointer(&unixConnPtr)) == nil {
+		return false
+	}
+
+	if _, err := net.Dial("unixgram", journalSocket); err != nil {
+		return false
+	}
+
+	return true
+}
+
+// Send a message to the local systemd journal. vars is a map of journald
+// fields to values.  Fields must be composed of uppercase letters, numbers,
+// and underscores, but must not start with an underscore. Within these
+// restrictions, any arbitrary field name may be used.  Some names have special
+// significance: see the journalctl documentation
+// (http://www.freedesktop.org/software/systemd/man/systemd.journal-fields.html)
+// for more details.  vars may be nil.
+func Send(message string, priority Priority, vars map[string]string) error {
+	conn := (*net.UnixConn)(atomic.LoadPointer(&unixConnPtr))
+	if conn == nil {
+		return errors.New("could not initialize socket to journald")
+	}
+
+	socketAddr := &net.UnixAddr{
+		Name: journalSocket,
+		Net:  "unixgram",
+	}
+
+	data := new(bytes.Buffer)
+	appendVariable(data, "PRIORITY", strconv.Itoa(int(priority)))
+	appendVariable(data, "MESSAGE", message)
+	for k, v := range vars {
+		appendVariable(data, k, v)
+	}
+
+	_, _, err := conn.WriteMsgUnix(data.Bytes(), nil, socketAddr)
+	if err == nil {
+		return nil
+	}
+	if !isSocketSpaceError(err) {
+		return err
+	}
+
+	// Large log entry, send it via tempfile and ancillary-fd.
+	file, err := tempFd()
+	if err != nil {
+		return err
+	}
+	defer file.Close()
+	_, err = io.Copy(file, data)
+	if err != nil {
+		return err
+	}
+	rights := syscall.UnixRights(int(file.Fd()))
+	_, _, err = conn.WriteMsgUnix([]byte{}, rights, socketAddr)
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+// Print prints a message to the local systemd journal using Send().
+func Print(priority Priority, format string, a ...interface{}) error {
+	return Send(fmt.Sprintf(format, a...), priority, nil)
+}
+
+func appendVariable(w io.Writer, name, value string) {
+	if err := validVarName(name); err != nil {
+		fmt.Fprintf(os.Stderr, "variable name %s contains invalid character, ignoring\n", name)
+	}
+	if strings.ContainsRune(value, '\n') {
+		/* When the value contains a newline, we write:
+		 * - the variable name, followed by a newline
+		 * - the size (in 64bit little endian format)
+		 * - the data, followed by a newline
+		 */
+		fmt.Fprintln(w, name)
+		binary.Write(w, binary.LittleEndian, uint64(len(value)))
+		fmt.Fprintln(w, value)
+	} else {
+		/* just write the variable and value all on one line */
+		fmt.Fprintf(w, "%s=%s\n", name, value)
+	}
+}
+
+// validVarName validates a variable name to make sure journald will accept it.
+// The variable name must be in uppercase and consist only of characters,
+// numbers and underscores, and may not begin with an underscore:
+// https://www.freedesktop.org/software/systemd/man/sd_journal_print.html
+func validVarName(name string) error {
+	if name == "" {
+		return errors.New("Empty variable name")
+	} else if name[0] == '_' {
+		return errors.New("Variable name begins with an underscore")
+	}
+
+	for _, c := range name {
+		if !(('A' <= c && c <= 'Z') || ('0' <= c && c <= '9') || c == '_') {
+			return errors.New("Variable name contains invalid characters")
+		}
+	}
+	return nil
+}
+
+// isSocketSpaceError checks whether the error is signaling
+// an "overlarge message" condition.
+func isSocketSpaceError(err error) bool {
+	opErr, ok := err.(*net.OpError)
+	if !ok || opErr == nil {
+		return false
+	}
+
+	sysErr, ok := opErr.Err.(*os.SyscallError)
+	if !ok || sysErr == nil {
+		return false
+	}
+
+	return sysErr.Err == syscall.EMSGSIZE || sysErr.Err == syscall.ENOBUFS
+}
+
+// tempFd creates a temporary, unlinked file under `/dev/shm`.
+func tempFd() (*os.File, error) {
+	file, err := ioutil.TempFile("/dev/shm/", "journal.XXXXX")
+	if err != nil {
+		return nil, err
+	}
+	err = syscall.Unlink(file.Name())
+	if err != nil {
+		return nil, err
+	}
+	return file, nil
+}
+
+// initConn initializes the global `unixConnPtr` socket.
+// It is meant to be called exactly once, at program startup.
+func initConn() {
+	autobind, err := net.ResolveUnixAddr("unixgram", "")
+	if err != nil {
+		return
+	}
+
+	sock, err := net.ListenUnixgram("unixgram", autobind)
+	if err != nil {
+		return
+	}
+
+	atomic.StorePointer(&unixConnPtr, unsafe.Pointer(sock))
+}
diff --git a/vendor/github.com/coreos/pkg/LICENSE b/vendor/github.com/coreos/pkg/LICENSE
new file mode 100644
index 0000000..e06d208
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/LICENSE
@@ -0,0 +1,202 @@
+Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright {yyyy} {name of copyright owner}
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
diff --git a/vendor/github.com/coreos/pkg/NOTICE b/vendor/github.com/coreos/pkg/NOTICE
new file mode 100644
index 0000000..b39ddfa
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/NOTICE
@@ -0,0 +1,5 @@
+CoreOS Project
+Copyright 2014 CoreOS, Inc
+
+This product includes software developed at CoreOS, Inc.
+(http://www.coreos.com/).
diff --git a/vendor/github.com/coreos/pkg/capnslog/README.md b/vendor/github.com/coreos/pkg/capnslog/README.md
new file mode 100644
index 0000000..f79dbfc
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/README.md
@@ -0,0 +1,39 @@
+# capnslog, the CoreOS logging package
+
+There are far too many logging packages out there, with varying degrees of licenses, far too many features (colorization, all sorts of log frameworks) or are just a pain to use (lack of `Fatalln()`?).
+capnslog provides a simple but consistent logging interface suitable for all kinds of projects.
+
+### Design Principles
+
+##### `package main` is the place where logging gets turned on and routed
+
+A library should not touch log options, only generate log entries. Libraries are silent until main lets them speak.
+
+##### All log options are runtime-configurable. 
+
+Still the job of `main` to expose these configurations. `main` may delegate this to, say, a configuration webhook, but does so explicitly. 
+
+##### There is one log object per package. It is registered under its repository and package name.
+
+`main` activates logging for its repository and any dependency repositories it would also like to have output in its logstream. `main` also dictates at which level each subpackage logs.
+
+##### There is *one* output stream, and it is an `io.Writer` composed with a formatter.
+
+Splitting streams is probably not the job of your program, but rather, your log aggregation framework. If you must split output streams, again, `main` configures this and you can write a very simple two-output struct that satisfies io.Writer.
+
+Fancy colorful formatting and JSON output are beyond the scope of a basic logging framework -- they're application/log-collector dependent. These are, at best, provided as options, but more likely, provided by your application.
+
+##### Log objects are an interface
+
+An object knows best how to print itself. Log objects can collect more interesting metadata if they wish, however, because text isn't going away anytime soon, they must all be marshalable to text. The simplest log object is a string, which returns itself. If you wish to do more fancy tricks for printing your log objects, see also JSON output -- introspect and write a formatter which can handle your advanced log interface. Making strings is the only thing guaranteed.
+
+##### Log levels have specific meanings:
+
+  * Critical: Unrecoverable. Must fail.
+  * Error: Data has been lost, a request has failed for a bad reason, or a required resource has been lost
+  * Warning: (Hopefully) Temporary conditions that may cause errors, but may work fine. A replica disappearing (that may reconnect) is a warning.
+  * Notice: Normal, but important (uncommon) log information.
+  * Info: Normal, working log information, everything is fine, but helpful notices for auditing or common operations.
+  * Debug: Everything is still fine, but even common operations may be logged, and less helpful but more quantity of notices.
+  * Trace: Anything goes, from logging every function call as part of a common operation, to tracing execution of a query.
+
diff --git a/vendor/github.com/coreos/pkg/capnslog/formatters.go b/vendor/github.com/coreos/pkg/capnslog/formatters.go
new file mode 100644
index 0000000..b305a84
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/formatters.go
@@ -0,0 +1,157 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"log"
+	"runtime"
+	"strings"
+	"time"
+)
+
+type Formatter interface {
+	Format(pkg string, level LogLevel, depth int, entries ...interface{})
+	Flush()
+}
+
+func NewStringFormatter(w io.Writer) Formatter {
+	return &StringFormatter{
+		w: bufio.NewWriter(w),
+	}
+}
+
+type StringFormatter struct {
+	w *bufio.Writer
+}
+
+func (s *StringFormatter) Format(pkg string, l LogLevel, i int, entries ...interface{}) {
+	now := time.Now().UTC()
+	s.w.WriteString(now.Format(time.RFC3339))
+	s.w.WriteByte(' ')
+	writeEntries(s.w, pkg, l, i, entries...)
+	s.Flush()
+}
+
+func writeEntries(w *bufio.Writer, pkg string, _ LogLevel, _ int, entries ...interface{}) {
+	if pkg != "" {
+		w.WriteString(pkg + ": ")
+	}
+	str := fmt.Sprint(entries...)
+	endsInNL := strings.HasSuffix(str, "\n")
+	w.WriteString(str)
+	if !endsInNL {
+		w.WriteString("\n")
+	}
+}
+
+func (s *StringFormatter) Flush() {
+	s.w.Flush()
+}
+
+func NewPrettyFormatter(w io.Writer, debug bool) Formatter {
+	return &PrettyFormatter{
+		w:     bufio.NewWriter(w),
+		debug: debug,
+	}
+}
+
+type PrettyFormatter struct {
+	w     *bufio.Writer
+	debug bool
+}
+
+func (c *PrettyFormatter) Format(pkg string, l LogLevel, depth int, entries ...interface{}) {
+	now := time.Now()
+	ts := now.Format("2006-01-02 15:04:05")
+	c.w.WriteString(ts)
+	ms := now.Nanosecond() / 1000
+	c.w.WriteString(fmt.Sprintf(".%06d", ms))
+	if c.debug {
+		_, file, line, ok := runtime.Caller(depth) // It's always the same number of frames to the user's call.
+		if !ok {
+			file = "???"
+			line = 1
+		} else {
+			slash := strings.LastIndex(file, "/")
+			if slash >= 0 {
+				file = file[slash+1:]
+			}
+		}
+		if line < 0 {
+			line = 0 // not a real line number
+		}
+		c.w.WriteString(fmt.Sprintf(" [%s:%d]", file, line))
+	}
+	c.w.WriteString(fmt.Sprint(" ", l.Char(), " | "))
+	writeEntries(c.w, pkg, l, depth, entries...)
+	c.Flush()
+}
+
+func (c *PrettyFormatter) Flush() {
+	c.w.Flush()
+}
+
+// LogFormatter emulates the form of the traditional built-in logger.
+type LogFormatter struct {
+	logger *log.Logger
+	prefix string
+}
+
+// NewLogFormatter is a helper to produce a new LogFormatter struct. It uses the
+// golang log package to actually do the logging work so that logs look similar.
+func NewLogFormatter(w io.Writer, prefix string, flag int) Formatter {
+	return &LogFormatter{
+		logger: log.New(w, "", flag), // don't use prefix here
+		prefix: prefix,               // save it instead
+	}
+}
+
+// Format builds a log message for the LogFormatter. The LogLevel is ignored.
+func (lf *LogFormatter) Format(pkg string, _ LogLevel, _ int, entries ...interface{}) {
+	str := fmt.Sprint(entries...)
+	prefix := lf.prefix
+	if pkg != "" {
+		prefix = fmt.Sprintf("%s%s: ", prefix, pkg)
+	}
+	lf.logger.Output(5, fmt.Sprintf("%s%v", prefix, str)) // call depth is 5
+}
+
+// Flush is included so that the interface is complete, but is a no-op.
+func (lf *LogFormatter) Flush() {
+	// noop
+}
+
+// NilFormatter is a no-op log formatter that does nothing.
+type NilFormatter struct {
+}
+
+// NewNilFormatter is a helper to produce a new LogFormatter struct. It logs no
+// messages so that you can cause part of your logging to be silent.
+func NewNilFormatter() Formatter {
+	return &NilFormatter{}
+}
+
+// Format does nothing.
+func (_ *NilFormatter) Format(_ string, _ LogLevel, _ int, _ ...interface{}) {
+	// noop
+}
+
+// Flush is included so that the interface is complete, but is a no-op.
+func (_ *NilFormatter) Flush() {
+	// noop
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/glog_formatter.go b/vendor/github.com/coreos/pkg/capnslog/glog_formatter.go
new file mode 100644
index 0000000..426603e
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/glog_formatter.go
@@ -0,0 +1,96 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import (
+	"bufio"
+	"bytes"
+	"io"
+	"os"
+	"runtime"
+	"strconv"
+	"strings"
+	"time"
+)
+
+var pid = os.Getpid()
+
+type GlogFormatter struct {
+	StringFormatter
+}
+
+func NewGlogFormatter(w io.Writer) *GlogFormatter {
+	g := &GlogFormatter{}
+	g.w = bufio.NewWriter(w)
+	return g
+}
+
+func (g GlogFormatter) Format(pkg string, level LogLevel, depth int, entries ...interface{}) {
+	g.w.Write(GlogHeader(level, depth+1))
+	g.StringFormatter.Format(pkg, level, depth+1, entries...)
+}
+
+func GlogHeader(level LogLevel, depth int) []byte {
+	// Lmmdd hh:mm:ss.uuuuuu threadid file:line]
+	now := time.Now().UTC()
+	_, file, line, ok := runtime.Caller(depth) // It's always the same number of frames to the user's call.
+	if !ok {
+		file = "???"
+		line = 1
+	} else {
+		slash := strings.LastIndex(file, "/")
+		if slash >= 0 {
+			file = file[slash+1:]
+		}
+	}
+	if line < 0 {
+		line = 0 // not a real line number
+	}
+	buf := &bytes.Buffer{}
+	buf.Grow(30)
+	_, month, day := now.Date()
+	hour, minute, second := now.Clock()
+	buf.WriteString(level.Char())
+	twoDigits(buf, int(month))
+	twoDigits(buf, day)
+	buf.WriteByte(' ')
+	twoDigits(buf, hour)
+	buf.WriteByte(':')
+	twoDigits(buf, minute)
+	buf.WriteByte(':')
+	twoDigits(buf, second)
+	buf.WriteByte('.')
+	buf.WriteString(strconv.Itoa(now.Nanosecond() / 1000))
+	buf.WriteByte('Z')
+	buf.WriteByte(' ')
+	buf.WriteString(strconv.Itoa(pid))
+	buf.WriteByte(' ')
+	buf.WriteString(file)
+	buf.WriteByte(':')
+	buf.WriteString(strconv.Itoa(line))
+	buf.WriteByte(']')
+	buf.WriteByte(' ')
+	return buf.Bytes()
+}
+
+const digits = "0123456789"
+
+func twoDigits(b *bytes.Buffer, d int) {
+	c2 := digits[d%10]
+	d /= 10
+	c1 := digits[d%10]
+	b.WriteByte(c1)
+	b.WriteByte(c2)
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/init.go b/vendor/github.com/coreos/pkg/capnslog/init.go
new file mode 100644
index 0000000..44b8cd3
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/init.go
@@ -0,0 +1,49 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// +build !windows
+
+package capnslog
+
+import (
+	"io"
+	"os"
+	"syscall"
+)
+
+// Here's where the opinionation comes in. We need some sensible defaults,
+// especially after taking over the log package. Your project (whatever it may
+// be) may see things differently. That's okay; there should be no defaults in
+// the main package that cannot be controlled or overridden programatically,
+// otherwise it's a bug. Doing so is creating your own init_log.go file much
+// like this one.
+
+func init() {
+	initHijack()
+
+	// Go `log` pacakge uses os.Stderr.
+	SetFormatter(NewDefaultFormatter(os.Stderr))
+	SetGlobalLogLevel(INFO)
+}
+
+func NewDefaultFormatter(out io.Writer) Formatter {
+	if syscall.Getppid() == 1 {
+		// We're running under init, which may be systemd.
+		f, err := NewJournaldFormatter()
+		if err == nil {
+			return f
+		}
+	}
+	return NewPrettyFormatter(out, false)
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/init_windows.go b/vendor/github.com/coreos/pkg/capnslog/init_windows.go
new file mode 100644
index 0000000..4553050
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/init_windows.go
@@ -0,0 +1,25 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import "os"
+
+func init() {
+	initHijack()
+
+	// Go `log` package uses os.Stderr.
+	SetFormatter(NewPrettyFormatter(os.Stderr, false))
+	SetGlobalLogLevel(INFO)
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/journald_formatter.go b/vendor/github.com/coreos/pkg/capnslog/journald_formatter.go
new file mode 100644
index 0000000..72e0520
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/journald_formatter.go
@@ -0,0 +1,68 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// +build !windows
+
+package capnslog
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	"github.com/coreos/go-systemd/journal"
+)
+
+func NewJournaldFormatter() (Formatter, error) {
+	if !journal.Enabled() {
+		return nil, errors.New("No systemd detected")
+	}
+	return &journaldFormatter{}, nil
+}
+
+type journaldFormatter struct{}
+
+func (j *journaldFormatter) Format(pkg string, l LogLevel, _ int, entries ...interface{}) {
+	var pri journal.Priority
+	switch l {
+	case CRITICAL:
+		pri = journal.PriCrit
+	case ERROR:
+		pri = journal.PriErr
+	case WARNING:
+		pri = journal.PriWarning
+	case NOTICE:
+		pri = journal.PriNotice
+	case INFO:
+		pri = journal.PriInfo
+	case DEBUG:
+		pri = journal.PriDebug
+	case TRACE:
+		pri = journal.PriDebug
+	default:
+		panic("Unhandled loglevel")
+	}
+	msg := fmt.Sprint(entries...)
+	tags := map[string]string{
+		"PACKAGE":           pkg,
+		"SYSLOG_IDENTIFIER": filepath.Base(os.Args[0]),
+	}
+	err := journal.Send(msg, pri, tags)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+	}
+}
+
+func (j *journaldFormatter) Flush() {}
diff --git a/vendor/github.com/coreos/pkg/capnslog/log_hijack.go b/vendor/github.com/coreos/pkg/capnslog/log_hijack.go
new file mode 100644
index 0000000..970086b
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/log_hijack.go
@@ -0,0 +1,39 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import (
+	"log"
+)
+
+func initHijack() {
+	pkg := NewPackageLogger("log", "")
+	w := packageWriter{pkg}
+	log.SetFlags(0)
+	log.SetPrefix("")
+	log.SetOutput(w)
+}
+
+type packageWriter struct {
+	pl *PackageLogger
+}
+
+func (p packageWriter) Write(b []byte) (int, error) {
+	if p.pl.level < INFO {
+		return 0, nil
+	}
+	p.pl.internalLog(calldepth+2, INFO, string(b))
+	return len(b), nil
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/logmap.go b/vendor/github.com/coreos/pkg/capnslog/logmap.go
new file mode 100644
index 0000000..226b60c
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/logmap.go
@@ -0,0 +1,245 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import (
+	"errors"
+	"strings"
+	"sync"
+)
+
+// LogLevel is the set of all log levels.
+type LogLevel int8
+
+const (
+	// CRITICAL is the lowest log level; only errors which will end the program will be propagated.
+	CRITICAL LogLevel = iota - 1
+	// ERROR is for errors that are not fatal but lead to troubling behavior.
+	ERROR
+	// WARNING is for errors which are not fatal and not errors, but are unusual. Often sourced from misconfigurations.
+	WARNING
+	// NOTICE is for normal but significant conditions.
+	NOTICE
+	// INFO is a log level for common, everyday log updates.
+	INFO
+	// DEBUG is the default hidden level for more verbose updates about internal processes.
+	DEBUG
+	// TRACE is for (potentially) call by call tracing of programs.
+	TRACE
+)
+
+// Char returns a single-character representation of the log level.
+func (l LogLevel) Char() string {
+	switch l {
+	case CRITICAL:
+		return "C"
+	case ERROR:
+		return "E"
+	case WARNING:
+		return "W"
+	case NOTICE:
+		return "N"
+	case INFO:
+		return "I"
+	case DEBUG:
+		return "D"
+	case TRACE:
+		return "T"
+	default:
+		panic("Unhandled loglevel")
+	}
+}
+
+// String returns a multi-character representation of the log level.
+func (l LogLevel) String() string {
+	switch l {
+	case CRITICAL:
+		return "CRITICAL"
+	case ERROR:
+		return "ERROR"
+	case WARNING:
+		return "WARNING"
+	case NOTICE:
+		return "NOTICE"
+	case INFO:
+		return "INFO"
+	case DEBUG:
+		return "DEBUG"
+	case TRACE:
+		return "TRACE"
+	default:
+		panic("Unhandled loglevel")
+	}
+}
+
+// Update using the given string value. Fulfills the flag.Value interface.
+func (l *LogLevel) Set(s string) error {
+	value, err := ParseLevel(s)
+	if err != nil {
+		return err
+	}
+
+	*l = value
+	return nil
+}
+
+// Returns an empty string, only here to fulfill the pflag.Value interface.
+func (l *LogLevel) Type() string {
+	return ""
+}
+
+// ParseLevel translates some potential loglevel strings into their corresponding levels.
+func ParseLevel(s string) (LogLevel, error) {
+	switch s {
+	case "CRITICAL", "C":
+		return CRITICAL, nil
+	case "ERROR", "0", "E":
+		return ERROR, nil
+	case "WARNING", "1", "W":
+		return WARNING, nil
+	case "NOTICE", "2", "N":
+		return NOTICE, nil
+	case "INFO", "3", "I":
+		return INFO, nil
+	case "DEBUG", "4", "D":
+		return DEBUG, nil
+	case "TRACE", "5", "T":
+		return TRACE, nil
+	}
+	return CRITICAL, errors.New("couldn't parse log level " + s)
+}
+
+type RepoLogger map[string]*PackageLogger
+
+type loggerStruct struct {
+	sync.Mutex
+	repoMap   map[string]RepoLogger
+	formatter Formatter
+}
+
+// logger is the global logger
+var logger = new(loggerStruct)
+
+// SetGlobalLogLevel sets the log level for all packages in all repositories
+// registered with capnslog.
+func SetGlobalLogLevel(l LogLevel) {
+	logger.Lock()
+	defer logger.Unlock()
+	for _, r := range logger.repoMap {
+		r.setRepoLogLevelInternal(l)
+	}
+}
+
+// GetRepoLogger may return the handle to the repository's set of packages' loggers.
+func GetRepoLogger(repo string) (RepoLogger, error) {
+	logger.Lock()
+	defer logger.Unlock()
+	r, ok := logger.repoMap[repo]
+	if !ok {
+		return nil, errors.New("no packages registered for repo " + repo)
+	}
+	return r, nil
+}
+
+// MustRepoLogger returns the handle to the repository's packages' loggers.
+func MustRepoLogger(repo string) RepoLogger {
+	r, err := GetRepoLogger(repo)
+	if err != nil {
+		panic(err)
+	}
+	return r
+}
+
+// SetRepoLogLevel sets the log level for all packages in the repository.
+func (r RepoLogger) SetRepoLogLevel(l LogLevel) {
+	logger.Lock()
+	defer logger.Unlock()
+	r.setRepoLogLevelInternal(l)
+}
+
+func (r RepoLogger) setRepoLogLevelInternal(l LogLevel) {
+	for _, v := range r {
+		v.level = l
+	}
+}
+
+// ParseLogLevelConfig parses a comma-separated string of "package=loglevel", in
+// order, and returns a map of the results, for use in SetLogLevel.
+func (r RepoLogger) ParseLogLevelConfig(conf string) (map[string]LogLevel, error) {
+	setlist := strings.Split(conf, ",")
+	out := make(map[string]LogLevel)
+	for _, setstring := range setlist {
+		setting := strings.Split(setstring, "=")
+		if len(setting) != 2 {
+			return nil, errors.New("oddly structured `pkg=level` option: " + setstring)
+		}
+		l, err := ParseLevel(setting[1])
+		if err != nil {
+			return nil, err
+		}
+		out[setting[0]] = l
+	}
+	return out, nil
+}
+
+// SetLogLevel takes a map of package names within a repository to their desired
+// loglevel, and sets the levels appropriately. Unknown packages are ignored.
+// "*" is a special package name that corresponds to all packages, and will be
+// processed first.
+func (r RepoLogger) SetLogLevel(m map[string]LogLevel) {
+	logger.Lock()
+	defer logger.Unlock()
+	if l, ok := m["*"]; ok {
+		r.setRepoLogLevelInternal(l)
+	}
+	for k, v := range m {
+		l, ok := r[k]
+		if !ok {
+			continue
+		}
+		l.level = v
+	}
+}
+
+// SetFormatter sets the formatting function for all logs.
+func SetFormatter(f Formatter) {
+	logger.Lock()
+	defer logger.Unlock()
+	logger.formatter = f
+}
+
+// NewPackageLogger creates a package logger object.
+// This should be defined as a global var in your package, referencing your repo.
+func NewPackageLogger(repo string, pkg string) (p *PackageLogger) {
+	logger.Lock()
+	defer logger.Unlock()
+	if logger.repoMap == nil {
+		logger.repoMap = make(map[string]RepoLogger)
+	}
+	r, rok := logger.repoMap[repo]
+	if !rok {
+		logger.repoMap[repo] = make(RepoLogger)
+		r = logger.repoMap[repo]
+	}
+	p, pok := r[pkg]
+	if !pok {
+		r[pkg] = &PackageLogger{
+			pkg:   pkg,
+			level: INFO,
+		}
+		p = r[pkg]
+	}
+	return
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/pkg_logger.go b/vendor/github.com/coreos/pkg/capnslog/pkg_logger.go
new file mode 100644
index 0000000..00ff371
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/pkg_logger.go
@@ -0,0 +1,191 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package capnslog
+
+import (
+	"fmt"
+	"os"
+)
+
+type PackageLogger struct {
+	pkg   string
+	level LogLevel
+}
+
+const calldepth = 2
+
+func (p *PackageLogger) internalLog(depth int, inLevel LogLevel, entries ...interface{}) {
+	logger.Lock()
+	defer logger.Unlock()
+	if inLevel != CRITICAL && p.level < inLevel {
+		return
+	}
+	if logger.formatter != nil {
+		logger.formatter.Format(p.pkg, inLevel, depth+1, entries...)
+	}
+}
+
+// SetLevel allows users to change the current logging level.
+func (p *PackageLogger) SetLevel(l LogLevel) {
+	logger.Lock()
+	defer logger.Unlock()
+	p.level = l
+}
+
+// LevelAt checks if the given log level will be outputted under current setting.
+func (p *PackageLogger) LevelAt(l LogLevel) bool {
+	logger.Lock()
+	defer logger.Unlock()
+	return p.level >= l
+}
+
+// Log a formatted string at any level between ERROR and TRACE
+func (p *PackageLogger) Logf(l LogLevel, format string, args ...interface{}) {
+	p.internalLog(calldepth, l, fmt.Sprintf(format, args...))
+}
+
+// Log a message at any level between ERROR and TRACE
+func (p *PackageLogger) Log(l LogLevel, args ...interface{}) {
+	p.internalLog(calldepth, l, fmt.Sprint(args...))
+}
+
+// log stdlib compatibility
+
+func (p *PackageLogger) Println(args ...interface{}) {
+	p.internalLog(calldepth, INFO, fmt.Sprintln(args...))
+}
+
+func (p *PackageLogger) Printf(format string, args ...interface{}) {
+	p.Logf(INFO, format, args...)
+}
+
+func (p *PackageLogger) Print(args ...interface{}) {
+	p.internalLog(calldepth, INFO, fmt.Sprint(args...))
+}
+
+// Panic and fatal
+
+func (p *PackageLogger) Panicf(format string, args ...interface{}) {
+	s := fmt.Sprintf(format, args...)
+	p.internalLog(calldepth, CRITICAL, s)
+	panic(s)
+}
+
+func (p *PackageLogger) Panic(args ...interface{}) {
+	s := fmt.Sprint(args...)
+	p.internalLog(calldepth, CRITICAL, s)
+	panic(s)
+}
+
+func (p *PackageLogger) Panicln(args ...interface{}) {
+	s := fmt.Sprintln(args...)
+	p.internalLog(calldepth, CRITICAL, s)
+	panic(s)
+}
+
+func (p *PackageLogger) Fatalf(format string, args ...interface{}) {
+	p.Logf(CRITICAL, format, args...)
+	os.Exit(1)
+}
+
+func (p *PackageLogger) Fatal(args ...interface{}) {
+	s := fmt.Sprint(args...)
+	p.internalLog(calldepth, CRITICAL, s)
+	os.Exit(1)
+}
+
+func (p *PackageLogger) Fatalln(args ...interface{}) {
+	s := fmt.Sprintln(args...)
+	p.internalLog(calldepth, CRITICAL, s)
+	os.Exit(1)
+}
+
+// Error Functions
+
+func (p *PackageLogger) Errorf(format string, args ...interface{}) {
+	p.Logf(ERROR, format, args...)
+}
+
+func (p *PackageLogger) Error(entries ...interface{}) {
+	p.internalLog(calldepth, ERROR, entries...)
+}
+
+// Warning Functions
+
+func (p *PackageLogger) Warningf(format string, args ...interface{}) {
+	p.Logf(WARNING, format, args...)
+}
+
+func (p *PackageLogger) Warning(entries ...interface{}) {
+	p.internalLog(calldepth, WARNING, entries...)
+}
+
+// Notice Functions
+
+func (p *PackageLogger) Noticef(format string, args ...interface{}) {
+	p.Logf(NOTICE, format, args...)
+}
+
+func (p *PackageLogger) Notice(entries ...interface{}) {
+	p.internalLog(calldepth, NOTICE, entries...)
+}
+
+// Info Functions
+
+func (p *PackageLogger) Infof(format string, args ...interface{}) {
+	p.Logf(INFO, format, args...)
+}
+
+func (p *PackageLogger) Info(entries ...interface{}) {
+	p.internalLog(calldepth, INFO, entries...)
+}
+
+// Debug Functions
+
+func (p *PackageLogger) Debugf(format string, args ...interface{}) {
+	if p.level < DEBUG {
+		return
+	}
+	p.Logf(DEBUG, format, args...)
+}
+
+func (p *PackageLogger) Debug(entries ...interface{}) {
+	if p.level < DEBUG {
+		return
+	}
+	p.internalLog(calldepth, DEBUG, entries...)
+}
+
+// Trace Functions
+
+func (p *PackageLogger) Tracef(format string, args ...interface{}) {
+	if p.level < TRACE {
+		return
+	}
+	p.Logf(TRACE, format, args...)
+}
+
+func (p *PackageLogger) Trace(entries ...interface{}) {
+	if p.level < TRACE {
+		return
+	}
+	p.internalLog(calldepth, TRACE, entries...)
+}
+
+func (p *PackageLogger) Flush() {
+	logger.Lock()
+	defer logger.Unlock()
+	logger.formatter.Flush()
+}
diff --git a/vendor/github.com/coreos/pkg/capnslog/syslog_formatter.go b/vendor/github.com/coreos/pkg/capnslog/syslog_formatter.go
new file mode 100644
index 0000000..4be5a1f
--- /dev/null
+++ b/vendor/github.com/coreos/pkg/capnslog/syslog_formatter.go
@@ -0,0 +1,65 @@
+// Copyright 2015 CoreOS, Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// +build !windows
+
+package capnslog
+
+import (
+	"fmt"
+	"log/syslog"
+)
+
+func NewSyslogFormatter(w *syslog.Writer) Formatter {
+	return &syslogFormatter{w}
+}
+
+func NewDefaultSyslogFormatter(tag string) (Formatter, error) {
+	w, err := syslog.New(syslog.LOG_DEBUG, tag)
+	if err != nil {
+		return nil, err
+	}
+	return NewSyslogFormatter(w), nil
+}
+
+type syslogFormatter struct {
+	w *syslog.Writer
+}
+
+func (s *syslogFormatter) Format(pkg string, l LogLevel, _ int, entries ...interface{}) {
+	for _, entry := range entries {
+		str := fmt.Sprint(entry)
+		switch l {
+		case CRITICAL:
+			s.w.Crit(str)
+		case ERROR:
+			s.w.Err(str)
+		case WARNING:
+			s.w.Warning(str)
+		case NOTICE:
+			s.w.Notice(str)
+		case INFO:
+			s.w.Info(str)
+		case DEBUG:
+			s.w.Debug(str)
+		case TRACE:
+			s.w.Debug(str)
+		default:
+			panic("Unhandled loglevel")
+		}
+	}
+}
+
+func (s *syslogFormatter) Flush() {
+}
diff --git a/vendor/github.com/davecgh/go-spew/LICENSE b/vendor/github.com/davecgh/go-spew/LICENSE
new file mode 100644
index 0000000..bc52e96
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/LICENSE
@@ -0,0 +1,15 @@
+ISC License
+
+Copyright (c) 2012-2016 Dave Collins <dave@davec.name>
+
+Permission to use, copy, modify, and/or distribute this software for any
+purpose with or without fee is hereby granted, provided that the above
+copyright notice and this permission notice appear in all copies.
+
+THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
diff --git a/vendor/github.com/davecgh/go-spew/spew/bypass.go b/vendor/github.com/davecgh/go-spew/spew/bypass.go
new file mode 100644
index 0000000..7929947
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/bypass.go
@@ -0,0 +1,145 @@
+// Copyright (c) 2015-2016 Dave Collins <dave@davec.name>
+//
+// Permission to use, copy, modify, and distribute this software for any
+// purpose with or without fee is hereby granted, provided that the above
+// copyright notice and this permission notice appear in all copies.
+//
+// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+// WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+// MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+// ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+// WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+// ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+// OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+
+// NOTE: Due to the following build constraints, this file will only be compiled
+// when the code is not running on Google App Engine, compiled by GopherJS, and
+// "-tags safe" is not added to the go build command line.  The "disableunsafe"
+// tag is deprecated and thus should not be used.
+// Go versions prior to 1.4 are disabled because they use a different layout
+// for interfaces which make the implementation of unsafeReflectValue more complex.
+// +build !js,!appengine,!safe,!disableunsafe,go1.4
+
+package spew
+
+import (
+	"reflect"
+	"unsafe"
+)
+
+const (
+	// UnsafeDisabled is a build-time constant which specifies whether or
+	// not access to the unsafe package is available.
+	UnsafeDisabled = false
+
+	// ptrSize is the size of a pointer on the current arch.
+	ptrSize = unsafe.Sizeof((*byte)(nil))
+)
+
+type flag uintptr
+
+var (
+	// flagRO indicates whether the value field of a reflect.Value
+	// is read-only.
+	flagRO flag
+
+	// flagAddr indicates whether the address of the reflect.Value's
+	// value may be taken.
+	flagAddr flag
+)
+
+// flagKindMask holds the bits that make up the kind
+// part of the flags field. In all the supported versions,
+// it is in the lower 5 bits.
+const flagKindMask = flag(0x1f)
+
+// Different versions of Go have used different
+// bit layouts for the flags type. This table
+// records the known combinations.
+var okFlags = []struct {
+	ro, addr flag
+}{{
+	// From Go 1.4 to 1.5
+	ro:   1 << 5,
+	addr: 1 << 7,
+}, {
+	// Up to Go tip.
+	ro:   1<<5 | 1<<6,
+	addr: 1 << 8,
+}}
+
+var flagValOffset = func() uintptr {
+	field, ok := reflect.TypeOf(reflect.Value{}).FieldByName("flag")
+	if !ok {
+		panic("reflect.Value has no flag field")
+	}
+	return field.Offset
+}()
+
+// flagField returns a pointer to the flag field of a reflect.Value.
+func flagField(v *reflect.Value) *flag {
+	return (*flag)(unsafe.Pointer(uintptr(unsafe.Pointer(v)) + flagValOffset))
+}
+
+// unsafeReflectValue converts the passed reflect.Value into a one that bypasses
+// the typical safety restrictions preventing access to unaddressable and
+// unexported data.  It works by digging the raw pointer to the underlying
+// value out of the protected value and generating a new unprotected (unsafe)
+// reflect.Value to it.
+//
+// This allows us to check for implementations of the Stringer and error
+// interfaces to be used for pretty printing ordinarily unaddressable and
+// inaccessible values such as unexported struct fields.
+func unsafeReflectValue(v reflect.Value) reflect.Value {
+	if !v.IsValid() || (v.CanInterface() && v.CanAddr()) {
+		return v
+	}
+	flagFieldPtr := flagField(&v)
+	*flagFieldPtr &^= flagRO
+	*flagFieldPtr |= flagAddr
+	return v
+}
+
+// Sanity checks against future reflect package changes
+// to the type or semantics of the Value.flag field.
+func init() {
+	field, ok := reflect.TypeOf(reflect.Value{}).FieldByName("flag")
+	if !ok {
+		panic("reflect.Value has no flag field")
+	}
+	if field.Type.Kind() != reflect.TypeOf(flag(0)).Kind() {
+		panic("reflect.Value flag field has changed kind")
+	}
+	type t0 int
+	var t struct {
+		A t0
+		// t0 will have flagEmbedRO set.
+		t0
+		// a will have flagStickyRO set
+		a t0
+	}
+	vA := reflect.ValueOf(t).FieldByName("A")
+	va := reflect.ValueOf(t).FieldByName("a")
+	vt0 := reflect.ValueOf(t).FieldByName("t0")
+
+	// Infer flagRO from the difference between the flags
+	// for the (otherwise identical) fields in t.
+	flagPublic := *flagField(&vA)
+	flagWithRO := *flagField(&va) | *flagField(&vt0)
+	flagRO = flagPublic ^ flagWithRO
+
+	// Infer flagAddr from the difference between a value
+	// taken from a pointer and not.
+	vPtrA := reflect.ValueOf(&t).Elem().FieldByName("A")
+	flagNoPtr := *flagField(&vA)
+	flagPtr := *flagField(&vPtrA)
+	flagAddr = flagNoPtr ^ flagPtr
+
+	// Check that the inferred flags tally with one of the known versions.
+	for _, f := range okFlags {
+		if flagRO == f.ro && flagAddr == f.addr {
+			return
+		}
+	}
+	panic("reflect.Value read-only flag has changed semantics")
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/bypasssafe.go b/vendor/github.com/davecgh/go-spew/spew/bypasssafe.go
new file mode 100644
index 0000000..205c28d
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/bypasssafe.go
@@ -0,0 +1,38 @@
+// Copyright (c) 2015-2016 Dave Collins <dave@davec.name>
+//
+// Permission to use, copy, modify, and distribute this software for any
+// purpose with or without fee is hereby granted, provided that the above
+// copyright notice and this permission notice appear in all copies.
+//
+// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+// WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+// MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+// ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+// WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+// ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+// OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+
+// NOTE: Due to the following build constraints, this file will only be compiled
+// when the code is running on Google App Engine, compiled by GopherJS, or
+// "-tags safe" is added to the go build command line.  The "disableunsafe"
+// tag is deprecated and thus should not be used.
+// +build js appengine safe disableunsafe !go1.4
+
+package spew
+
+import "reflect"
+
+const (
+	// UnsafeDisabled is a build-time constant which specifies whether or
+	// not access to the unsafe package is available.
+	UnsafeDisabled = true
+)
+
+// unsafeReflectValue typically converts the passed reflect.Value into a one
+// that bypasses the typical safety restrictions preventing access to
+// unaddressable and unexported data.  However, doing this relies on access to
+// the unsafe package.  This is a stub version which simply returns the passed
+// reflect.Value when the unsafe package is not available.
+func unsafeReflectValue(v reflect.Value) reflect.Value {
+	return v
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/common.go b/vendor/github.com/davecgh/go-spew/spew/common.go
new file mode 100644
index 0000000..1be8ce9
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/common.go
@@ -0,0 +1,341 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+package spew
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"reflect"
+	"sort"
+	"strconv"
+)
+
+// Some constants in the form of bytes to avoid string overhead.  This mirrors
+// the technique used in the fmt package.
+var (
+	panicBytes            = []byte("(PANIC=")
+	plusBytes             = []byte("+")
+	iBytes                = []byte("i")
+	trueBytes             = []byte("true")
+	falseBytes            = []byte("false")
+	interfaceBytes        = []byte("(interface {})")
+	commaNewlineBytes     = []byte(",\n")
+	newlineBytes          = []byte("\n")
+	openBraceBytes        = []byte("{")
+	openBraceNewlineBytes = []byte("{\n")
+	closeBraceBytes       = []byte("}")
+	asteriskBytes         = []byte("*")
+	colonBytes            = []byte(":")
+	colonSpaceBytes       = []byte(": ")
+	openParenBytes        = []byte("(")
+	closeParenBytes       = []byte(")")
+	spaceBytes            = []byte(" ")
+	pointerChainBytes     = []byte("->")
+	nilAngleBytes         = []byte("<nil>")
+	maxNewlineBytes       = []byte("<max depth reached>\n")
+	maxShortBytes         = []byte("<max>")
+	circularBytes         = []byte("<already shown>")
+	circularShortBytes    = []byte("<shown>")
+	invalidAngleBytes     = []byte("<invalid>")
+	openBracketBytes      = []byte("[")
+	closeBracketBytes     = []byte("]")
+	percentBytes          = []byte("%")
+	precisionBytes        = []byte(".")
+	openAngleBytes        = []byte("<")
+	closeAngleBytes       = []byte(">")
+	openMapBytes          = []byte("map[")
+	closeMapBytes         = []byte("]")
+	lenEqualsBytes        = []byte("len=")
+	capEqualsBytes        = []byte("cap=")
+)
+
+// hexDigits is used to map a decimal value to a hex digit.
+var hexDigits = "0123456789abcdef"
+
+// catchPanic handles any panics that might occur during the handleMethods
+// calls.
+func catchPanic(w io.Writer, v reflect.Value) {
+	if err := recover(); err != nil {
+		w.Write(panicBytes)
+		fmt.Fprintf(w, "%v", err)
+		w.Write(closeParenBytes)
+	}
+}
+
+// handleMethods attempts to call the Error and String methods on the underlying
+// type the passed reflect.Value represents and outputes the result to Writer w.
+//
+// It handles panics in any called methods by catching and displaying the error
+// as the formatted value.
+func handleMethods(cs *ConfigState, w io.Writer, v reflect.Value) (handled bool) {
+	// We need an interface to check if the type implements the error or
+	// Stringer interface.  However, the reflect package won't give us an
+	// interface on certain things like unexported struct fields in order
+	// to enforce visibility rules.  We use unsafe, when it's available,
+	// to bypass these restrictions since this package does not mutate the
+	// values.
+	if !v.CanInterface() {
+		if UnsafeDisabled {
+			return false
+		}
+
+		v = unsafeReflectValue(v)
+	}
+
+	// Choose whether or not to do error and Stringer interface lookups against
+	// the base type or a pointer to the base type depending on settings.
+	// Technically calling one of these methods with a pointer receiver can
+	// mutate the value, however, types which choose to satisify an error or
+	// Stringer interface with a pointer receiver should not be mutating their
+	// state inside these interface methods.
+	if !cs.DisablePointerMethods && !UnsafeDisabled && !v.CanAddr() {
+		v = unsafeReflectValue(v)
+	}
+	if v.CanAddr() {
+		v = v.Addr()
+	}
+
+	// Is it an error or Stringer?
+	switch iface := v.Interface().(type) {
+	case error:
+		defer catchPanic(w, v)
+		if cs.ContinueOnMethod {
+			w.Write(openParenBytes)
+			w.Write([]byte(iface.Error()))
+			w.Write(closeParenBytes)
+			w.Write(spaceBytes)
+			return false
+		}
+
+		w.Write([]byte(iface.Error()))
+		return true
+
+	case fmt.Stringer:
+		defer catchPanic(w, v)
+		if cs.ContinueOnMethod {
+			w.Write(openParenBytes)
+			w.Write([]byte(iface.String()))
+			w.Write(closeParenBytes)
+			w.Write(spaceBytes)
+			return false
+		}
+		w.Write([]byte(iface.String()))
+		return true
+	}
+	return false
+}
+
+// printBool outputs a boolean value as true or false to Writer w.
+func printBool(w io.Writer, val bool) {
+	if val {
+		w.Write(trueBytes)
+	} else {
+		w.Write(falseBytes)
+	}
+}
+
+// printInt outputs a signed integer value to Writer w.
+func printInt(w io.Writer, val int64, base int) {
+	w.Write([]byte(strconv.FormatInt(val, base)))
+}
+
+// printUint outputs an unsigned integer value to Writer w.
+func printUint(w io.Writer, val uint64, base int) {
+	w.Write([]byte(strconv.FormatUint(val, base)))
+}
+
+// printFloat outputs a floating point value using the specified precision,
+// which is expected to be 32 or 64bit, to Writer w.
+func printFloat(w io.Writer, val float64, precision int) {
+	w.Write([]byte(strconv.FormatFloat(val, 'g', -1, precision)))
+}
+
+// printComplex outputs a complex value using the specified float precision
+// for the real and imaginary parts to Writer w.
+func printComplex(w io.Writer, c complex128, floatPrecision int) {
+	r := real(c)
+	w.Write(openParenBytes)
+	w.Write([]byte(strconv.FormatFloat(r, 'g', -1, floatPrecision)))
+	i := imag(c)
+	if i >= 0 {
+		w.Write(plusBytes)
+	}
+	w.Write([]byte(strconv.FormatFloat(i, 'g', -1, floatPrecision)))
+	w.Write(iBytes)
+	w.Write(closeParenBytes)
+}
+
+// printHexPtr outputs a uintptr formatted as hexadecimal with a leading '0x'
+// prefix to Writer w.
+func printHexPtr(w io.Writer, p uintptr) {
+	// Null pointer.
+	num := uint64(p)
+	if num == 0 {
+		w.Write(nilAngleBytes)
+		return
+	}
+
+	// Max uint64 is 16 bytes in hex + 2 bytes for '0x' prefix
+	buf := make([]byte, 18)
+
+	// It's simpler to construct the hex string right to left.
+	base := uint64(16)
+	i := len(buf) - 1
+	for num >= base {
+		buf[i] = hexDigits[num%base]
+		num /= base
+		i--
+	}
+	buf[i] = hexDigits[num]
+
+	// Add '0x' prefix.
+	i--
+	buf[i] = 'x'
+	i--
+	buf[i] = '0'
+
+	// Strip unused leading bytes.
+	buf = buf[i:]
+	w.Write(buf)
+}
+
+// valuesSorter implements sort.Interface to allow a slice of reflect.Value
+// elements to be sorted.
+type valuesSorter struct {
+	values  []reflect.Value
+	strings []string // either nil or same len and values
+	cs      *ConfigState
+}
+
+// newValuesSorter initializes a valuesSorter instance, which holds a set of
+// surrogate keys on which the data should be sorted.  It uses flags in
+// ConfigState to decide if and how to populate those surrogate keys.
+func newValuesSorter(values []reflect.Value, cs *ConfigState) sort.Interface {
+	vs := &valuesSorter{values: values, cs: cs}
+	if canSortSimply(vs.values[0].Kind()) {
+		return vs
+	}
+	if !cs.DisableMethods {
+		vs.strings = make([]string, len(values))
+		for i := range vs.values {
+			b := bytes.Buffer{}
+			if !handleMethods(cs, &b, vs.values[i]) {
+				vs.strings = nil
+				break
+			}
+			vs.strings[i] = b.String()
+		}
+	}
+	if vs.strings == nil && cs.SpewKeys {
+		vs.strings = make([]string, len(values))
+		for i := range vs.values {
+			vs.strings[i] = Sprintf("%#v", vs.values[i].Interface())
+		}
+	}
+	return vs
+}
+
+// canSortSimply tests whether a reflect.Kind is a primitive that can be sorted
+// directly, or whether it should be considered for sorting by surrogate keys
+// (if the ConfigState allows it).
+func canSortSimply(kind reflect.Kind) bool {
+	// This switch parallels valueSortLess, except for the default case.
+	switch kind {
+	case reflect.Bool:
+		return true
+	case reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Int:
+		return true
+	case reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uint:
+		return true
+	case reflect.Float32, reflect.Float64:
+		return true
+	case reflect.String:
+		return true
+	case reflect.Uintptr:
+		return true
+	case reflect.Array:
+		return true
+	}
+	return false
+}
+
+// Len returns the number of values in the slice.  It is part of the
+// sort.Interface implementation.
+func (s *valuesSorter) Len() int {
+	return len(s.values)
+}
+
+// Swap swaps the values at the passed indices.  It is part of the
+// sort.Interface implementation.
+func (s *valuesSorter) Swap(i, j int) {
+	s.values[i], s.values[j] = s.values[j], s.values[i]
+	if s.strings != nil {
+		s.strings[i], s.strings[j] = s.strings[j], s.strings[i]
+	}
+}
+
+// valueSortLess returns whether the first value should sort before the second
+// value.  It is used by valueSorter.Less as part of the sort.Interface
+// implementation.
+func valueSortLess(a, b reflect.Value) bool {
+	switch a.Kind() {
+	case reflect.Bool:
+		return !a.Bool() && b.Bool()
+	case reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Int:
+		return a.Int() < b.Int()
+	case reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uint:
+		return a.Uint() < b.Uint()
+	case reflect.Float32, reflect.Float64:
+		return a.Float() < b.Float()
+	case reflect.String:
+		return a.String() < b.String()
+	case reflect.Uintptr:
+		return a.Uint() < b.Uint()
+	case reflect.Array:
+		// Compare the contents of both arrays.
+		l := a.Len()
+		for i := 0; i < l; i++ {
+			av := a.Index(i)
+			bv := b.Index(i)
+			if av.Interface() == bv.Interface() {
+				continue
+			}
+			return valueSortLess(av, bv)
+		}
+	}
+	return a.String() < b.String()
+}
+
+// Less returns whether the value at index i should sort before the
+// value at index j.  It is part of the sort.Interface implementation.
+func (s *valuesSorter) Less(i, j int) bool {
+	if s.strings == nil {
+		return valueSortLess(s.values[i], s.values[j])
+	}
+	return s.strings[i] < s.strings[j]
+}
+
+// sortValues is a sort function that handles both native types and any type that
+// can be converted to error or Stringer.  Other inputs are sorted according to
+// their Value.String() value to ensure display stability.
+func sortValues(values []reflect.Value, cs *ConfigState) {
+	if len(values) == 0 {
+		return
+	}
+	sort.Sort(newValuesSorter(values, cs))
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/config.go b/vendor/github.com/davecgh/go-spew/spew/config.go
new file mode 100644
index 0000000..2e3d22f
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/config.go
@@ -0,0 +1,306 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+package spew
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+)
+
+// ConfigState houses the configuration options used by spew to format and
+// display values.  There is a global instance, Config, that is used to control
+// all top-level Formatter and Dump functionality.  Each ConfigState instance
+// provides methods equivalent to the top-level functions.
+//
+// The zero value for ConfigState provides no indentation.  You would typically
+// want to set it to a space or a tab.
+//
+// Alternatively, you can use NewDefaultConfig to get a ConfigState instance
+// with default settings.  See the documentation of NewDefaultConfig for default
+// values.
+type ConfigState struct {
+	// Indent specifies the string to use for each indentation level.  The
+	// global config instance that all top-level functions use set this to a
+	// single space by default.  If you would like more indentation, you might
+	// set this to a tab with "\t" or perhaps two spaces with "  ".
+	Indent string
+
+	// MaxDepth controls the maximum number of levels to descend into nested
+	// data structures.  The default, 0, means there is no limit.
+	//
+	// NOTE: Circular data structures are properly detected, so it is not
+	// necessary to set this value unless you specifically want to limit deeply
+	// nested data structures.
+	MaxDepth int
+
+	// DisableMethods specifies whether or not error and Stringer interfaces are
+	// invoked for types that implement them.
+	DisableMethods bool
+
+	// DisablePointerMethods specifies whether or not to check for and invoke
+	// error and Stringer interfaces on types which only accept a pointer
+	// receiver when the current type is not a pointer.
+	//
+	// NOTE: This might be an unsafe action since calling one of these methods
+	// with a pointer receiver could technically mutate the value, however,
+	// in practice, types which choose to satisify an error or Stringer
+	// interface with a pointer receiver should not be mutating their state
+	// inside these interface methods.  As a result, this option relies on
+	// access to the unsafe package, so it will not have any effect when
+	// running in environments without access to the unsafe package such as
+	// Google App Engine or with the "safe" build tag specified.
+	DisablePointerMethods bool
+
+	// DisablePointerAddresses specifies whether to disable the printing of
+	// pointer addresses. This is useful when diffing data structures in tests.
+	DisablePointerAddresses bool
+
+	// DisableCapacities specifies whether to disable the printing of capacities
+	// for arrays, slices, maps and channels. This is useful when diffing
+	// data structures in tests.
+	DisableCapacities bool
+
+	// ContinueOnMethod specifies whether or not recursion should continue once
+	// a custom error or Stringer interface is invoked.  The default, false,
+	// means it will print the results of invoking the custom error or Stringer
+	// interface and return immediately instead of continuing to recurse into
+	// the internals of the data type.
+	//
+	// NOTE: This flag does not have any effect if method invocation is disabled
+	// via the DisableMethods or DisablePointerMethods options.
+	ContinueOnMethod bool
+
+	// SortKeys specifies map keys should be sorted before being printed. Use
+	// this to have a more deterministic, diffable output.  Note that only
+	// native types (bool, int, uint, floats, uintptr and string) and types
+	// that support the error or Stringer interfaces (if methods are
+	// enabled) are supported, with other types sorted according to the
+	// reflect.Value.String() output which guarantees display stability.
+	SortKeys bool
+
+	// SpewKeys specifies that, as a last resort attempt, map keys should
+	// be spewed to strings and sorted by those strings.  This is only
+	// considered if SortKeys is true.
+	SpewKeys bool
+}
+
+// Config is the active configuration of the top-level functions.
+// The configuration can be changed by modifying the contents of spew.Config.
+var Config = ConfigState{Indent: " "}
+
+// Errorf is a wrapper for fmt.Errorf that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the formatted string as a value that satisfies error.  See NewFormatter
+// for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Errorf(format, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Errorf(format string, a ...interface{}) (err error) {
+	return fmt.Errorf(format, c.convertArgs(a)...)
+}
+
+// Fprint is a wrapper for fmt.Fprint that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprint(w, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Fprint(w io.Writer, a ...interface{}) (n int, err error) {
+	return fmt.Fprint(w, c.convertArgs(a)...)
+}
+
+// Fprintf is a wrapper for fmt.Fprintf that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprintf(w, format, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Fprintf(w io.Writer, format string, a ...interface{}) (n int, err error) {
+	return fmt.Fprintf(w, format, c.convertArgs(a)...)
+}
+
+// Fprintln is a wrapper for fmt.Fprintln that treats each argument as if it
+// passed with a Formatter interface returned by c.NewFormatter.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprintln(w, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Fprintln(w io.Writer, a ...interface{}) (n int, err error) {
+	return fmt.Fprintln(w, c.convertArgs(a)...)
+}
+
+// Print is a wrapper for fmt.Print that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Print(c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Print(a ...interface{}) (n int, err error) {
+	return fmt.Print(c.convertArgs(a)...)
+}
+
+// Printf is a wrapper for fmt.Printf that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Printf(format, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Printf(format string, a ...interface{}) (n int, err error) {
+	return fmt.Printf(format, c.convertArgs(a)...)
+}
+
+// Println is a wrapper for fmt.Println that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Println(c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Println(a ...interface{}) (n int, err error) {
+	return fmt.Println(c.convertArgs(a)...)
+}
+
+// Sprint is a wrapper for fmt.Sprint that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprint(c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Sprint(a ...interface{}) string {
+	return fmt.Sprint(c.convertArgs(a)...)
+}
+
+// Sprintf is a wrapper for fmt.Sprintf that treats each argument as if it were
+// passed with a Formatter interface returned by c.NewFormatter.  It returns
+// the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprintf(format, c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Sprintf(format string, a ...interface{}) string {
+	return fmt.Sprintf(format, c.convertArgs(a)...)
+}
+
+// Sprintln is a wrapper for fmt.Sprintln that treats each argument as if it
+// were passed with a Formatter interface returned by c.NewFormatter.  It
+// returns the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprintln(c.NewFormatter(a), c.NewFormatter(b))
+func (c *ConfigState) Sprintln(a ...interface{}) string {
+	return fmt.Sprintln(c.convertArgs(a)...)
+}
+
+/*
+NewFormatter returns a custom formatter that satisfies the fmt.Formatter
+interface.  As a result, it integrates cleanly with standard fmt package
+printing functions.  The formatter is useful for inline printing of smaller data
+types similar to the standard %v format specifier.
+
+The custom formatter only responds to the %v (most compact), %+v (adds pointer
+addresses), %#v (adds types), and %#+v (adds types and pointer addresses) verb
+combinations.  Any other verbs such as %x and %q will be sent to the the
+standard fmt package for formatting.  In addition, the custom formatter ignores
+the width and precision arguments (however they will still work on the format
+specifiers not handled by the custom formatter).
+
+Typically this function shouldn't be called directly.  It is much easier to make
+use of the custom formatter by calling one of the convenience functions such as
+c.Printf, c.Println, or c.Printf.
+*/
+func (c *ConfigState) NewFormatter(v interface{}) fmt.Formatter {
+	return newFormatter(c, v)
+}
+
+// Fdump formats and displays the passed arguments to io.Writer w.  It formats
+// exactly the same as Dump.
+func (c *ConfigState) Fdump(w io.Writer, a ...interface{}) {
+	fdump(c, w, a...)
+}
+
+/*
+Dump displays the passed parameters to standard out with newlines, customizable
+indentation, and additional debug information such as complete types and all
+pointer addresses used to indirect to the final value.  It provides the
+following features over the built-in printing facilities provided by the fmt
+package:
+
+	* Pointers are dereferenced and followed
+	* Circular data structures are detected and handled properly
+	* Custom Stringer/error interfaces are optionally invoked, including
+	  on unexported types
+	* Custom types which only implement the Stringer/error interfaces via
+	  a pointer receiver are optionally invoked when passing non-pointer
+	  variables
+	* Byte arrays and slices are dumped like the hexdump -C command which
+	  includes offsets, byte values in hex, and ASCII output
+
+The configuration options are controlled by modifying the public members
+of c.  See ConfigState for options documentation.
+
+See Fdump if you would prefer dumping to an arbitrary io.Writer or Sdump to
+get the formatted result as a string.
+*/
+func (c *ConfigState) Dump(a ...interface{}) {
+	fdump(c, os.Stdout, a...)
+}
+
+// Sdump returns a string with the passed arguments formatted exactly the same
+// as Dump.
+func (c *ConfigState) Sdump(a ...interface{}) string {
+	var buf bytes.Buffer
+	fdump(c, &buf, a...)
+	return buf.String()
+}
+
+// convertArgs accepts a slice of arguments and returns a slice of the same
+// length with each argument converted to a spew Formatter interface using
+// the ConfigState associated with s.
+func (c *ConfigState) convertArgs(args []interface{}) (formatters []interface{}) {
+	formatters = make([]interface{}, len(args))
+	for index, arg := range args {
+		formatters[index] = newFormatter(c, arg)
+	}
+	return formatters
+}
+
+// NewDefaultConfig returns a ConfigState with the following default settings.
+//
+// 	Indent: " "
+// 	MaxDepth: 0
+// 	DisableMethods: false
+// 	DisablePointerMethods: false
+// 	ContinueOnMethod: false
+// 	SortKeys: false
+func NewDefaultConfig() *ConfigState {
+	return &ConfigState{Indent: " "}
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/doc.go b/vendor/github.com/davecgh/go-spew/spew/doc.go
new file mode 100644
index 0000000..aacaac6
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/doc.go
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+/*
+Package spew implements a deep pretty printer for Go data structures to aid in
+debugging.
+
+A quick overview of the additional features spew provides over the built-in
+printing facilities for Go data types are as follows:
+
+	* Pointers are dereferenced and followed
+	* Circular data structures are detected and handled properly
+	* Custom Stringer/error interfaces are optionally invoked, including
+	  on unexported types
+	* Custom types which only implement the Stringer/error interfaces via
+	  a pointer receiver are optionally invoked when passing non-pointer
+	  variables
+	* Byte arrays and slices are dumped like the hexdump -C command which
+	  includes offsets, byte values in hex, and ASCII output (only when using
+	  Dump style)
+
+There are two different approaches spew allows for dumping Go data structures:
+
+	* Dump style which prints with newlines, customizable indentation,
+	  and additional debug information such as types and all pointer addresses
+	  used to indirect to the final value
+	* A custom Formatter interface that integrates cleanly with the standard fmt
+	  package and replaces %v, %+v, %#v, and %#+v to provide inline printing
+	  similar to the default %v while providing the additional functionality
+	  outlined above and passing unsupported format verbs such as %x and %q
+	  along to fmt
+
+Quick Start
+
+This section demonstrates how to quickly get started with spew.  See the
+sections below for further details on formatting and configuration options.
+
+To dump a variable with full newlines, indentation, type, and pointer
+information use Dump, Fdump, or Sdump:
+	spew.Dump(myVar1, myVar2, ...)
+	spew.Fdump(someWriter, myVar1, myVar2, ...)
+	str := spew.Sdump(myVar1, myVar2, ...)
+
+Alternatively, if you would prefer to use format strings with a compacted inline
+printing style, use the convenience wrappers Printf, Fprintf, etc with
+%v (most compact), %+v (adds pointer addresses), %#v (adds types), or
+%#+v (adds types and pointer addresses):
+	spew.Printf("myVar1: %v -- myVar2: %+v", myVar1, myVar2)
+	spew.Printf("myVar3: %#v -- myVar4: %#+v", myVar3, myVar4)
+	spew.Fprintf(someWriter, "myVar1: %v -- myVar2: %+v", myVar1, myVar2)
+	spew.Fprintf(someWriter, "myVar3: %#v -- myVar4: %#+v", myVar3, myVar4)
+
+Configuration Options
+
+Configuration of spew is handled by fields in the ConfigState type.  For
+convenience, all of the top-level functions use a global state available
+via the spew.Config global.
+
+It is also possible to create a ConfigState instance that provides methods
+equivalent to the top-level functions.  This allows concurrent configuration
+options.  See the ConfigState documentation for more details.
+
+The following configuration options are available:
+	* Indent
+		String to use for each indentation level for Dump functions.
+		It is a single space by default.  A popular alternative is "\t".
+
+	* MaxDepth
+		Maximum number of levels to descend into nested data structures.
+		There is no limit by default.
+
+	* DisableMethods
+		Disables invocation of error and Stringer interface methods.
+		Method invocation is enabled by default.
+
+	* DisablePointerMethods
+		Disables invocation of error and Stringer interface methods on types
+		which only accept pointer receivers from non-pointer variables.
+		Pointer method invocation is enabled by default.
+
+	* DisablePointerAddresses
+		DisablePointerAddresses specifies whether to disable the printing of
+		pointer addresses. This is useful when diffing data structures in tests.
+
+	* DisableCapacities
+		DisableCapacities specifies whether to disable the printing of
+		capacities for arrays, slices, maps and channels. This is useful when
+		diffing data structures in tests.
+
+	* ContinueOnMethod
+		Enables recursion into types after invoking error and Stringer interface
+		methods. Recursion after method invocation is disabled by default.
+
+	* SortKeys
+		Specifies map keys should be sorted before being printed. Use
+		this to have a more deterministic, diffable output.  Note that
+		only native types (bool, int, uint, floats, uintptr and string)
+		and types which implement error or Stringer interfaces are
+		supported with other types sorted according to the
+		reflect.Value.String() output which guarantees display
+		stability.  Natural map order is used by default.
+
+	* SpewKeys
+		Specifies that, as a last resort attempt, map keys should be
+		spewed to strings and sorted by those strings.  This is only
+		considered if SortKeys is true.
+
+Dump Usage
+
+Simply call spew.Dump with a list of variables you want to dump:
+
+	spew.Dump(myVar1, myVar2, ...)
+
+You may also call spew.Fdump if you would prefer to output to an arbitrary
+io.Writer.  For example, to dump to standard error:
+
+	spew.Fdump(os.Stderr, myVar1, myVar2, ...)
+
+A third option is to call spew.Sdump to get the formatted output as a string:
+
+	str := spew.Sdump(myVar1, myVar2, ...)
+
+Sample Dump Output
+
+See the Dump example for details on the setup of the types and variables being
+shown here.
+
+	(main.Foo) {
+	 unexportedField: (*main.Bar)(0xf84002e210)({
+	  flag: (main.Flag) flagTwo,
+	  data: (uintptr) <nil>
+	 }),
+	 ExportedField: (map[interface {}]interface {}) (len=1) {
+	  (string) (len=3) "one": (bool) true
+	 }
+	}
+
+Byte (and uint8) arrays and slices are displayed uniquely like the hexdump -C
+command as shown.
+	([]uint8) (len=32 cap=32) {
+	 00000000  11 12 13 14 15 16 17 18  19 1a 1b 1c 1d 1e 1f 20  |............... |
+	 00000010  21 22 23 24 25 26 27 28  29 2a 2b 2c 2d 2e 2f 30  |!"#$%&'()*+,-./0|
+	 00000020  31 32                                             |12|
+	}
+
+Custom Formatter
+
+Spew provides a custom formatter that implements the fmt.Formatter interface
+so that it integrates cleanly with standard fmt package printing functions. The
+formatter is useful for inline printing of smaller data types similar to the
+standard %v format specifier.
+
+The custom formatter only responds to the %v (most compact), %+v (adds pointer
+addresses), %#v (adds types), or %#+v (adds types and pointer addresses) verb
+combinations.  Any other verbs such as %x and %q will be sent to the the
+standard fmt package for formatting.  In addition, the custom formatter ignores
+the width and precision arguments (however they will still work on the format
+specifiers not handled by the custom formatter).
+
+Custom Formatter Usage
+
+The simplest way to make use of the spew custom formatter is to call one of the
+convenience functions such as spew.Printf, spew.Println, or spew.Printf.  The
+functions have syntax you are most likely already familiar with:
+
+	spew.Printf("myVar1: %v -- myVar2: %+v", myVar1, myVar2)
+	spew.Printf("myVar3: %#v -- myVar4: %#+v", myVar3, myVar4)
+	spew.Println(myVar, myVar2)
+	spew.Fprintf(os.Stderr, "myVar1: %v -- myVar2: %+v", myVar1, myVar2)
+	spew.Fprintf(os.Stderr, "myVar3: %#v -- myVar4: %#+v", myVar3, myVar4)
+
+See the Index for the full list convenience functions.
+
+Sample Formatter Output
+
+Double pointer to a uint8:
+	  %v: <**>5
+	 %+v: <**>(0xf8400420d0->0xf8400420c8)5
+	 %#v: (**uint8)5
+	%#+v: (**uint8)(0xf8400420d0->0xf8400420c8)5
+
+Pointer to circular struct with a uint8 field and a pointer to itself:
+	  %v: <*>{1 <*><shown>}
+	 %+v: <*>(0xf84003e260){ui8:1 c:<*>(0xf84003e260)<shown>}
+	 %#v: (*main.circular){ui8:(uint8)1 c:(*main.circular)<shown>}
+	%#+v: (*main.circular)(0xf84003e260){ui8:(uint8)1 c:(*main.circular)(0xf84003e260)<shown>}
+
+See the Printf example for details on the setup of variables being shown
+here.
+
+Errors
+
+Since it is possible for custom Stringer/error interfaces to panic, spew
+detects them and handles them internally by printing the panic information
+inline with the output.  Since spew is intended to provide deep pretty printing
+capabilities on structures, it intentionally does not return any errors.
+*/
+package spew
diff --git a/vendor/github.com/davecgh/go-spew/spew/dump.go b/vendor/github.com/davecgh/go-spew/spew/dump.go
new file mode 100644
index 0000000..f78d89f
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/dump.go
@@ -0,0 +1,509 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+package spew
+
+import (
+	"bytes"
+	"encoding/hex"
+	"fmt"
+	"io"
+	"os"
+	"reflect"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+var (
+	// uint8Type is a reflect.Type representing a uint8.  It is used to
+	// convert cgo types to uint8 slices for hexdumping.
+	uint8Type = reflect.TypeOf(uint8(0))
+
+	// cCharRE is a regular expression that matches a cgo char.
+	// It is used to detect character arrays to hexdump them.
+	cCharRE = regexp.MustCompile(`^.*\._Ctype_char$`)
+
+	// cUnsignedCharRE is a regular expression that matches a cgo unsigned
+	// char.  It is used to detect unsigned character arrays to hexdump
+	// them.
+	cUnsignedCharRE = regexp.MustCompile(`^.*\._Ctype_unsignedchar$`)
+
+	// cUint8tCharRE is a regular expression that matches a cgo uint8_t.
+	// It is used to detect uint8_t arrays to hexdump them.
+	cUint8tCharRE = regexp.MustCompile(`^.*\._Ctype_uint8_t$`)
+)
+
+// dumpState contains information about the state of a dump operation.
+type dumpState struct {
+	w                io.Writer
+	depth            int
+	pointers         map[uintptr]int
+	ignoreNextType   bool
+	ignoreNextIndent bool
+	cs               *ConfigState
+}
+
+// indent performs indentation according to the depth level and cs.Indent
+// option.
+func (d *dumpState) indent() {
+	if d.ignoreNextIndent {
+		d.ignoreNextIndent = false
+		return
+	}
+	d.w.Write(bytes.Repeat([]byte(d.cs.Indent), d.depth))
+}
+
+// unpackValue returns values inside of non-nil interfaces when possible.
+// This is useful for data types like structs, arrays, slices, and maps which
+// can contain varying types packed inside an interface.
+func (d *dumpState) unpackValue(v reflect.Value) reflect.Value {
+	if v.Kind() == reflect.Interface && !v.IsNil() {
+		v = v.Elem()
+	}
+	return v
+}
+
+// dumpPtr handles formatting of pointers by indirecting them as necessary.
+func (d *dumpState) dumpPtr(v reflect.Value) {
+	// Remove pointers at or below the current depth from map used to detect
+	// circular refs.
+	for k, depth := range d.pointers {
+		if depth >= d.depth {
+			delete(d.pointers, k)
+		}
+	}
+
+	// Keep list of all dereferenced pointers to show later.
+	pointerChain := make([]uintptr, 0)
+
+	// Figure out how many levels of indirection there are by dereferencing
+	// pointers and unpacking interfaces down the chain while detecting circular
+	// references.
+	nilFound := false
+	cycleFound := false
+	indirects := 0
+	ve := v
+	for ve.Kind() == reflect.Ptr {
+		if ve.IsNil() {
+			nilFound = true
+			break
+		}
+		indirects++
+		addr := ve.Pointer()
+		pointerChain = append(pointerChain, addr)
+		if pd, ok := d.pointers[addr]; ok && pd < d.depth {
+			cycleFound = true
+			indirects--
+			break
+		}
+		d.pointers[addr] = d.depth
+
+		ve = ve.Elem()
+		if ve.Kind() == reflect.Interface {
+			if ve.IsNil() {
+				nilFound = true
+				break
+			}
+			ve = ve.Elem()
+		}
+	}
+
+	// Display type information.
+	d.w.Write(openParenBytes)
+	d.w.Write(bytes.Repeat(asteriskBytes, indirects))
+	d.w.Write([]byte(ve.Type().String()))
+	d.w.Write(closeParenBytes)
+
+	// Display pointer information.
+	if !d.cs.DisablePointerAddresses && len(pointerChain) > 0 {
+		d.w.Write(openParenBytes)
+		for i, addr := range pointerChain {
+			if i > 0 {
+				d.w.Write(pointerChainBytes)
+			}
+			printHexPtr(d.w, addr)
+		}
+		d.w.Write(closeParenBytes)
+	}
+
+	// Display dereferenced value.
+	d.w.Write(openParenBytes)
+	switch {
+	case nilFound:
+		d.w.Write(nilAngleBytes)
+
+	case cycleFound:
+		d.w.Write(circularBytes)
+
+	default:
+		d.ignoreNextType = true
+		d.dump(ve)
+	}
+	d.w.Write(closeParenBytes)
+}
+
+// dumpSlice handles formatting of arrays and slices.  Byte (uint8 under
+// reflection) arrays and slices are dumped in hexdump -C fashion.
+func (d *dumpState) dumpSlice(v reflect.Value) {
+	// Determine whether this type should be hex dumped or not.  Also,
+	// for types which should be hexdumped, try to use the underlying data
+	// first, then fall back to trying to convert them to a uint8 slice.
+	var buf []uint8
+	doConvert := false
+	doHexDump := false
+	numEntries := v.Len()
+	if numEntries > 0 {
+		vt := v.Index(0).Type()
+		vts := vt.String()
+		switch {
+		// C types that need to be converted.
+		case cCharRE.MatchString(vts):
+			fallthrough
+		case cUnsignedCharRE.MatchString(vts):
+			fallthrough
+		case cUint8tCharRE.MatchString(vts):
+			doConvert = true
+
+		// Try to use existing uint8 slices and fall back to converting
+		// and copying if that fails.
+		case vt.Kind() == reflect.Uint8:
+			// We need an addressable interface to convert the type
+			// to a byte slice.  However, the reflect package won't
+			// give us an interface on certain things like
+			// unexported struct fields in order to enforce
+			// visibility rules.  We use unsafe, when available, to
+			// bypass these restrictions since this package does not
+			// mutate the values.
+			vs := v
+			if !vs.CanInterface() || !vs.CanAddr() {
+				vs = unsafeReflectValue(vs)
+			}
+			if !UnsafeDisabled {
+				vs = vs.Slice(0, numEntries)
+
+				// Use the existing uint8 slice if it can be
+				// type asserted.
+				iface := vs.Interface()
+				if slice, ok := iface.([]uint8); ok {
+					buf = slice
+					doHexDump = true
+					break
+				}
+			}
+
+			// The underlying data needs to be converted if it can't
+			// be type asserted to a uint8 slice.
+			doConvert = true
+		}
+
+		// Copy and convert the underlying type if needed.
+		if doConvert && vt.ConvertibleTo(uint8Type) {
+			// Convert and copy each element into a uint8 byte
+			// slice.
+			buf = make([]uint8, numEntries)
+			for i := 0; i < numEntries; i++ {
+				vv := v.Index(i)
+				buf[i] = uint8(vv.Convert(uint8Type).Uint())
+			}
+			doHexDump = true
+		}
+	}
+
+	// Hexdump the entire slice as needed.
+	if doHexDump {
+		indent := strings.Repeat(d.cs.Indent, d.depth)
+		str := indent + hex.Dump(buf)
+		str = strings.Replace(str, "\n", "\n"+indent, -1)
+		str = strings.TrimRight(str, d.cs.Indent)
+		d.w.Write([]byte(str))
+		return
+	}
+
+	// Recursively call dump for each item.
+	for i := 0; i < numEntries; i++ {
+		d.dump(d.unpackValue(v.Index(i)))
+		if i < (numEntries - 1) {
+			d.w.Write(commaNewlineBytes)
+		} else {
+			d.w.Write(newlineBytes)
+		}
+	}
+}
+
+// dump is the main workhorse for dumping a value.  It uses the passed reflect
+// value to figure out what kind of object we are dealing with and formats it
+// appropriately.  It is a recursive function, however circular data structures
+// are detected and handled properly.
+func (d *dumpState) dump(v reflect.Value) {
+	// Handle invalid reflect values immediately.
+	kind := v.Kind()
+	if kind == reflect.Invalid {
+		d.w.Write(invalidAngleBytes)
+		return
+	}
+
+	// Handle pointers specially.
+	if kind == reflect.Ptr {
+		d.indent()
+		d.dumpPtr(v)
+		return
+	}
+
+	// Print type information unless already handled elsewhere.
+	if !d.ignoreNextType {
+		d.indent()
+		d.w.Write(openParenBytes)
+		d.w.Write([]byte(v.Type().String()))
+		d.w.Write(closeParenBytes)
+		d.w.Write(spaceBytes)
+	}
+	d.ignoreNextType = false
+
+	// Display length and capacity if the built-in len and cap functions
+	// work with the value's kind and the len/cap itself is non-zero.
+	valueLen, valueCap := 0, 0
+	switch v.Kind() {
+	case reflect.Array, reflect.Slice, reflect.Chan:
+		valueLen, valueCap = v.Len(), v.Cap()
+	case reflect.Map, reflect.String:
+		valueLen = v.Len()
+	}
+	if valueLen != 0 || !d.cs.DisableCapacities && valueCap != 0 {
+		d.w.Write(openParenBytes)
+		if valueLen != 0 {
+			d.w.Write(lenEqualsBytes)
+			printInt(d.w, int64(valueLen), 10)
+		}
+		if !d.cs.DisableCapacities && valueCap != 0 {
+			if valueLen != 0 {
+				d.w.Write(spaceBytes)
+			}
+			d.w.Write(capEqualsBytes)
+			printInt(d.w, int64(valueCap), 10)
+		}
+		d.w.Write(closeParenBytes)
+		d.w.Write(spaceBytes)
+	}
+
+	// Call Stringer/error interfaces if they exist and the handle methods flag
+	// is enabled
+	if !d.cs.DisableMethods {
+		if (kind != reflect.Invalid) && (kind != reflect.Interface) {
+			if handled := handleMethods(d.cs, d.w, v); handled {
+				return
+			}
+		}
+	}
+
+	switch kind {
+	case reflect.Invalid:
+		// Do nothing.  We should never get here since invalid has already
+		// been handled above.
+
+	case reflect.Bool:
+		printBool(d.w, v.Bool())
+
+	case reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Int:
+		printInt(d.w, v.Int(), 10)
+
+	case reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uint:
+		printUint(d.w, v.Uint(), 10)
+
+	case reflect.Float32:
+		printFloat(d.w, v.Float(), 32)
+
+	case reflect.Float64:
+		printFloat(d.w, v.Float(), 64)
+
+	case reflect.Complex64:
+		printComplex(d.w, v.Complex(), 32)
+
+	case reflect.Complex128:
+		printComplex(d.w, v.Complex(), 64)
+
+	case reflect.Slice:
+		if v.IsNil() {
+			d.w.Write(nilAngleBytes)
+			break
+		}
+		fallthrough
+
+	case reflect.Array:
+		d.w.Write(openBraceNewlineBytes)
+		d.depth++
+		if (d.cs.MaxDepth != 0) && (d.depth > d.cs.MaxDepth) {
+			d.indent()
+			d.w.Write(maxNewlineBytes)
+		} else {
+			d.dumpSlice(v)
+		}
+		d.depth--
+		d.indent()
+		d.w.Write(closeBraceBytes)
+
+	case reflect.String:
+		d.w.Write([]byte(strconv.Quote(v.String())))
+
+	case reflect.Interface:
+		// The only time we should get here is for nil interfaces due to
+		// unpackValue calls.
+		if v.IsNil() {
+			d.w.Write(nilAngleBytes)
+		}
+
+	case reflect.Ptr:
+		// Do nothing.  We should never get here since pointers have already
+		// been handled above.
+
+	case reflect.Map:
+		// nil maps should be indicated as different than empty maps
+		if v.IsNil() {
+			d.w.Write(nilAngleBytes)
+			break
+		}
+
+		d.w.Write(openBraceNewlineBytes)
+		d.depth++
+		if (d.cs.MaxDepth != 0) && (d.depth > d.cs.MaxDepth) {
+			d.indent()
+			d.w.Write(maxNewlineBytes)
+		} else {
+			numEntries := v.Len()
+			keys := v.MapKeys()
+			if d.cs.SortKeys {
+				sortValues(keys, d.cs)
+			}
+			for i, key := range keys {
+				d.dump(d.unpackValue(key))
+				d.w.Write(colonSpaceBytes)
+				d.ignoreNextIndent = true
+				d.dump(d.unpackValue(v.MapIndex(key)))
+				if i < (numEntries - 1) {
+					d.w.Write(commaNewlineBytes)
+				} else {
+					d.w.Write(newlineBytes)
+				}
+			}
+		}
+		d.depth--
+		d.indent()
+		d.w.Write(closeBraceBytes)
+
+	case reflect.Struct:
+		d.w.Write(openBraceNewlineBytes)
+		d.depth++
+		if (d.cs.MaxDepth != 0) && (d.depth > d.cs.MaxDepth) {
+			d.indent()
+			d.w.Write(maxNewlineBytes)
+		} else {
+			vt := v.Type()
+			numFields := v.NumField()
+			for i := 0; i < numFields; i++ {
+				d.indent()
+				vtf := vt.Field(i)
+				d.w.Write([]byte(vtf.Name))
+				d.w.Write(colonSpaceBytes)
+				d.ignoreNextIndent = true
+				d.dump(d.unpackValue(v.Field(i)))
+				if i < (numFields - 1) {
+					d.w.Write(commaNewlineBytes)
+				} else {
+					d.w.Write(newlineBytes)
+				}
+			}
+		}
+		d.depth--
+		d.indent()
+		d.w.Write(closeBraceBytes)
+
+	case reflect.Uintptr:
+		printHexPtr(d.w, uintptr(v.Uint()))
+
+	case reflect.UnsafePointer, reflect.Chan, reflect.Func:
+		printHexPtr(d.w, v.Pointer())
+
+	// There were not any other types at the time this code was written, but
+	// fall back to letting the default fmt package handle it in case any new
+	// types are added.
+	default:
+		if v.CanInterface() {
+			fmt.Fprintf(d.w, "%v", v.Interface())
+		} else {
+			fmt.Fprintf(d.w, "%v", v.String())
+		}
+	}
+}
+
+// fdump is a helper function to consolidate the logic from the various public
+// methods which take varying writers and config states.
+func fdump(cs *ConfigState, w io.Writer, a ...interface{}) {
+	for _, arg := range a {
+		if arg == nil {
+			w.Write(interfaceBytes)
+			w.Write(spaceBytes)
+			w.Write(nilAngleBytes)
+			w.Write(newlineBytes)
+			continue
+		}
+
+		d := dumpState{w: w, cs: cs}
+		d.pointers = make(map[uintptr]int)
+		d.dump(reflect.ValueOf(arg))
+		d.w.Write(newlineBytes)
+	}
+}
+
+// Fdump formats and displays the passed arguments to io.Writer w.  It formats
+// exactly the same as Dump.
+func Fdump(w io.Writer, a ...interface{}) {
+	fdump(&Config, w, a...)
+}
+
+// Sdump returns a string with the passed arguments formatted exactly the same
+// as Dump.
+func Sdump(a ...interface{}) string {
+	var buf bytes.Buffer
+	fdump(&Config, &buf, a...)
+	return buf.String()
+}
+
+/*
+Dump displays the passed parameters to standard out with newlines, customizable
+indentation, and additional debug information such as complete types and all
+pointer addresses used to indirect to the final value.  It provides the
+following features over the built-in printing facilities provided by the fmt
+package:
+
+	* Pointers are dereferenced and followed
+	* Circular data structures are detected and handled properly
+	* Custom Stringer/error interfaces are optionally invoked, including
+	  on unexported types
+	* Custom types which only implement the Stringer/error interfaces via
+	  a pointer receiver are optionally invoked when passing non-pointer
+	  variables
+	* Byte arrays and slices are dumped like the hexdump -C command which
+	  includes offsets, byte values in hex, and ASCII output
+
+The configuration options are controlled by an exported package global,
+spew.Config.  See ConfigState for options documentation.
+
+See Fdump if you would prefer dumping to an arbitrary io.Writer or Sdump to
+get the formatted result as a string.
+*/
+func Dump(a ...interface{}) {
+	fdump(&Config, os.Stdout, a...)
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/format.go b/vendor/github.com/davecgh/go-spew/spew/format.go
new file mode 100644
index 0000000..b04edb7
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/format.go
@@ -0,0 +1,419 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+package spew
+
+import (
+	"bytes"
+	"fmt"
+	"reflect"
+	"strconv"
+	"strings"
+)
+
+// supportedFlags is a list of all the character flags supported by fmt package.
+const supportedFlags = "0-+# "
+
+// formatState implements the fmt.Formatter interface and contains information
+// about the state of a formatting operation.  The NewFormatter function can
+// be used to get a new Formatter which can be used directly as arguments
+// in standard fmt package printing calls.
+type formatState struct {
+	value          interface{}
+	fs             fmt.State
+	depth          int
+	pointers       map[uintptr]int
+	ignoreNextType bool
+	cs             *ConfigState
+}
+
+// buildDefaultFormat recreates the original format string without precision
+// and width information to pass in to fmt.Sprintf in the case of an
+// unrecognized type.  Unless new types are added to the language, this
+// function won't ever be called.
+func (f *formatState) buildDefaultFormat() (format string) {
+	buf := bytes.NewBuffer(percentBytes)
+
+	for _, flag := range supportedFlags {
+		if f.fs.Flag(int(flag)) {
+			buf.WriteRune(flag)
+		}
+	}
+
+	buf.WriteRune('v')
+
+	format = buf.String()
+	return format
+}
+
+// constructOrigFormat recreates the original format string including precision
+// and width information to pass along to the standard fmt package.  This allows
+// automatic deferral of all format strings this package doesn't support.
+func (f *formatState) constructOrigFormat(verb rune) (format string) {
+	buf := bytes.NewBuffer(percentBytes)
+
+	for _, flag := range supportedFlags {
+		if f.fs.Flag(int(flag)) {
+			buf.WriteRune(flag)
+		}
+	}
+
+	if width, ok := f.fs.Width(); ok {
+		buf.WriteString(strconv.Itoa(width))
+	}
+
+	if precision, ok := f.fs.Precision(); ok {
+		buf.Write(precisionBytes)
+		buf.WriteString(strconv.Itoa(precision))
+	}
+
+	buf.WriteRune(verb)
+
+	format = buf.String()
+	return format
+}
+
+// unpackValue returns values inside of non-nil interfaces when possible and
+// ensures that types for values which have been unpacked from an interface
+// are displayed when the show types flag is also set.
+// This is useful for data types like structs, arrays, slices, and maps which
+// can contain varying types packed inside an interface.
+func (f *formatState) unpackValue(v reflect.Value) reflect.Value {
+	if v.Kind() == reflect.Interface {
+		f.ignoreNextType = false
+		if !v.IsNil() {
+			v = v.Elem()
+		}
+	}
+	return v
+}
+
+// formatPtr handles formatting of pointers by indirecting them as necessary.
+func (f *formatState) formatPtr(v reflect.Value) {
+	// Display nil if top level pointer is nil.
+	showTypes := f.fs.Flag('#')
+	if v.IsNil() && (!showTypes || f.ignoreNextType) {
+		f.fs.Write(nilAngleBytes)
+		return
+	}
+
+	// Remove pointers at or below the current depth from map used to detect
+	// circular refs.
+	for k, depth := range f.pointers {
+		if depth >= f.depth {
+			delete(f.pointers, k)
+		}
+	}
+
+	// Keep list of all dereferenced pointers to possibly show later.
+	pointerChain := make([]uintptr, 0)
+
+	// Figure out how many levels of indirection there are by derferencing
+	// pointers and unpacking interfaces down the chain while detecting circular
+	// references.
+	nilFound := false
+	cycleFound := false
+	indirects := 0
+	ve := v
+	for ve.Kind() == reflect.Ptr {
+		if ve.IsNil() {
+			nilFound = true
+			break
+		}
+		indirects++
+		addr := ve.Pointer()
+		pointerChain = append(pointerChain, addr)
+		if pd, ok := f.pointers[addr]; ok && pd < f.depth {
+			cycleFound = true
+			indirects--
+			break
+		}
+		f.pointers[addr] = f.depth
+
+		ve = ve.Elem()
+		if ve.Kind() == reflect.Interface {
+			if ve.IsNil() {
+				nilFound = true
+				break
+			}
+			ve = ve.Elem()
+		}
+	}
+
+	// Display type or indirection level depending on flags.
+	if showTypes && !f.ignoreNextType {
+		f.fs.Write(openParenBytes)
+		f.fs.Write(bytes.Repeat(asteriskBytes, indirects))
+		f.fs.Write([]byte(ve.Type().String()))
+		f.fs.Write(closeParenBytes)
+	} else {
+		if nilFound || cycleFound {
+			indirects += strings.Count(ve.Type().String(), "*")
+		}
+		f.fs.Write(openAngleBytes)
+		f.fs.Write([]byte(strings.Repeat("*", indirects)))
+		f.fs.Write(closeAngleBytes)
+	}
+
+	// Display pointer information depending on flags.
+	if f.fs.Flag('+') && (len(pointerChain) > 0) {
+		f.fs.Write(openParenBytes)
+		for i, addr := range pointerChain {
+			if i > 0 {
+				f.fs.Write(pointerChainBytes)
+			}
+			printHexPtr(f.fs, addr)
+		}
+		f.fs.Write(closeParenBytes)
+	}
+
+	// Display dereferenced value.
+	switch {
+	case nilFound:
+		f.fs.Write(nilAngleBytes)
+
+	case cycleFound:
+		f.fs.Write(circularShortBytes)
+
+	default:
+		f.ignoreNextType = true
+		f.format(ve)
+	}
+}
+
+// format is the main workhorse for providing the Formatter interface.  It
+// uses the passed reflect value to figure out what kind of object we are
+// dealing with and formats it appropriately.  It is a recursive function,
+// however circular data structures are detected and handled properly.
+func (f *formatState) format(v reflect.Value) {
+	// Handle invalid reflect values immediately.
+	kind := v.Kind()
+	if kind == reflect.Invalid {
+		f.fs.Write(invalidAngleBytes)
+		return
+	}
+
+	// Handle pointers specially.
+	if kind == reflect.Ptr {
+		f.formatPtr(v)
+		return
+	}
+
+	// Print type information unless already handled elsewhere.
+	if !f.ignoreNextType && f.fs.Flag('#') {
+		f.fs.Write(openParenBytes)
+		f.fs.Write([]byte(v.Type().String()))
+		f.fs.Write(closeParenBytes)
+	}
+	f.ignoreNextType = false
+
+	// Call Stringer/error interfaces if they exist and the handle methods
+	// flag is enabled.
+	if !f.cs.DisableMethods {
+		if (kind != reflect.Invalid) && (kind != reflect.Interface) {
+			if handled := handleMethods(f.cs, f.fs, v); handled {
+				return
+			}
+		}
+	}
+
+	switch kind {
+	case reflect.Invalid:
+		// Do nothing.  We should never get here since invalid has already
+		// been handled above.
+
+	case reflect.Bool:
+		printBool(f.fs, v.Bool())
+
+	case reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Int:
+		printInt(f.fs, v.Int(), 10)
+
+	case reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uint:
+		printUint(f.fs, v.Uint(), 10)
+
+	case reflect.Float32:
+		printFloat(f.fs, v.Float(), 32)
+
+	case reflect.Float64:
+		printFloat(f.fs, v.Float(), 64)
+
+	case reflect.Complex64:
+		printComplex(f.fs, v.Complex(), 32)
+
+	case reflect.Complex128:
+		printComplex(f.fs, v.Complex(), 64)
+
+	case reflect.Slice:
+		if v.IsNil() {
+			f.fs.Write(nilAngleBytes)
+			break
+		}
+		fallthrough
+
+	case reflect.Array:
+		f.fs.Write(openBracketBytes)
+		f.depth++
+		if (f.cs.MaxDepth != 0) && (f.depth > f.cs.MaxDepth) {
+			f.fs.Write(maxShortBytes)
+		} else {
+			numEntries := v.Len()
+			for i := 0; i < numEntries; i++ {
+				if i > 0 {
+					f.fs.Write(spaceBytes)
+				}
+				f.ignoreNextType = true
+				f.format(f.unpackValue(v.Index(i)))
+			}
+		}
+		f.depth--
+		f.fs.Write(closeBracketBytes)
+
+	case reflect.String:
+		f.fs.Write([]byte(v.String()))
+
+	case reflect.Interface:
+		// The only time we should get here is for nil interfaces due to
+		// unpackValue calls.
+		if v.IsNil() {
+			f.fs.Write(nilAngleBytes)
+		}
+
+	case reflect.Ptr:
+		// Do nothing.  We should never get here since pointers have already
+		// been handled above.
+
+	case reflect.Map:
+		// nil maps should be indicated as different than empty maps
+		if v.IsNil() {
+			f.fs.Write(nilAngleBytes)
+			break
+		}
+
+		f.fs.Write(openMapBytes)
+		f.depth++
+		if (f.cs.MaxDepth != 0) && (f.depth > f.cs.MaxDepth) {
+			f.fs.Write(maxShortBytes)
+		} else {
+			keys := v.MapKeys()
+			if f.cs.SortKeys {
+				sortValues(keys, f.cs)
+			}
+			for i, key := range keys {
+				if i > 0 {
+					f.fs.Write(spaceBytes)
+				}
+				f.ignoreNextType = true
+				f.format(f.unpackValue(key))
+				f.fs.Write(colonBytes)
+				f.ignoreNextType = true
+				f.format(f.unpackValue(v.MapIndex(key)))
+			}
+		}
+		f.depth--
+		f.fs.Write(closeMapBytes)
+
+	case reflect.Struct:
+		numFields := v.NumField()
+		f.fs.Write(openBraceBytes)
+		f.depth++
+		if (f.cs.MaxDepth != 0) && (f.depth > f.cs.MaxDepth) {
+			f.fs.Write(maxShortBytes)
+		} else {
+			vt := v.Type()
+			for i := 0; i < numFields; i++ {
+				if i > 0 {
+					f.fs.Write(spaceBytes)
+				}
+				vtf := vt.Field(i)
+				if f.fs.Flag('+') || f.fs.Flag('#') {
+					f.fs.Write([]byte(vtf.Name))
+					f.fs.Write(colonBytes)
+				}
+				f.format(f.unpackValue(v.Field(i)))
+			}
+		}
+		f.depth--
+		f.fs.Write(closeBraceBytes)
+
+	case reflect.Uintptr:
+		printHexPtr(f.fs, uintptr(v.Uint()))
+
+	case reflect.UnsafePointer, reflect.Chan, reflect.Func:
+		printHexPtr(f.fs, v.Pointer())
+
+	// There were not any other types at the time this code was written, but
+	// fall back to letting the default fmt package handle it if any get added.
+	default:
+		format := f.buildDefaultFormat()
+		if v.CanInterface() {
+			fmt.Fprintf(f.fs, format, v.Interface())
+		} else {
+			fmt.Fprintf(f.fs, format, v.String())
+		}
+	}
+}
+
+// Format satisfies the fmt.Formatter interface. See NewFormatter for usage
+// details.
+func (f *formatState) Format(fs fmt.State, verb rune) {
+	f.fs = fs
+
+	// Use standard formatting for verbs that are not v.
+	if verb != 'v' {
+		format := f.constructOrigFormat(verb)
+		fmt.Fprintf(fs, format, f.value)
+		return
+	}
+
+	if f.value == nil {
+		if fs.Flag('#') {
+			fs.Write(interfaceBytes)
+		}
+		fs.Write(nilAngleBytes)
+		return
+	}
+
+	f.format(reflect.ValueOf(f.value))
+}
+
+// newFormatter is a helper function to consolidate the logic from the various
+// public methods which take varying config states.
+func newFormatter(cs *ConfigState, v interface{}) fmt.Formatter {
+	fs := &formatState{value: v, cs: cs}
+	fs.pointers = make(map[uintptr]int)
+	return fs
+}
+
+/*
+NewFormatter returns a custom formatter that satisfies the fmt.Formatter
+interface.  As a result, it integrates cleanly with standard fmt package
+printing functions.  The formatter is useful for inline printing of smaller data
+types similar to the standard %v format specifier.
+
+The custom formatter only responds to the %v (most compact), %+v (adds pointer
+addresses), %#v (adds types), or %#+v (adds types and pointer addresses) verb
+combinations.  Any other verbs such as %x and %q will be sent to the the
+standard fmt package for formatting.  In addition, the custom formatter ignores
+the width and precision arguments (however they will still work on the format
+specifiers not handled by the custom formatter).
+
+Typically this function shouldn't be called directly.  It is much easier to make
+use of the custom formatter by calling one of the convenience functions such as
+Printf, Println, or Fprintf.
+*/
+func NewFormatter(v interface{}) fmt.Formatter {
+	return newFormatter(&Config, v)
+}
diff --git a/vendor/github.com/davecgh/go-spew/spew/spew.go b/vendor/github.com/davecgh/go-spew/spew/spew.go
new file mode 100644
index 0000000..32c0e33
--- /dev/null
+++ b/vendor/github.com/davecgh/go-spew/spew/spew.go
@@ -0,0 +1,148 @@
+/*
+ * Copyright (c) 2013-2016 Dave Collins <dave@davec.name>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+package spew
+
+import (
+	"fmt"
+	"io"
+)
+
+// Errorf is a wrapper for fmt.Errorf that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the formatted string as a value that satisfies error.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Errorf(format, spew.NewFormatter(a), spew.NewFormatter(b))
+func Errorf(format string, a ...interface{}) (err error) {
+	return fmt.Errorf(format, convertArgs(a)...)
+}
+
+// Fprint is a wrapper for fmt.Fprint that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprint(w, spew.NewFormatter(a), spew.NewFormatter(b))
+func Fprint(w io.Writer, a ...interface{}) (n int, err error) {
+	return fmt.Fprint(w, convertArgs(a)...)
+}
+
+// Fprintf is a wrapper for fmt.Fprintf that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprintf(w, format, spew.NewFormatter(a), spew.NewFormatter(b))
+func Fprintf(w io.Writer, format string, a ...interface{}) (n int, err error) {
+	return fmt.Fprintf(w, format, convertArgs(a)...)
+}
+
+// Fprintln is a wrapper for fmt.Fprintln that treats each argument as if it
+// passed with a default Formatter interface returned by NewFormatter.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Fprintln(w, spew.NewFormatter(a), spew.NewFormatter(b))
+func Fprintln(w io.Writer, a ...interface{}) (n int, err error) {
+	return fmt.Fprintln(w, convertArgs(a)...)
+}
+
+// Print is a wrapper for fmt.Print that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Print(spew.NewFormatter(a), spew.NewFormatter(b))
+func Print(a ...interface{}) (n int, err error) {
+	return fmt.Print(convertArgs(a)...)
+}
+
+// Printf is a wrapper for fmt.Printf that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Printf(format, spew.NewFormatter(a), spew.NewFormatter(b))
+func Printf(format string, a ...interface{}) (n int, err error) {
+	return fmt.Printf(format, convertArgs(a)...)
+}
+
+// Println is a wrapper for fmt.Println that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the number of bytes written and any write error encountered.  See
+// NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Println(spew.NewFormatter(a), spew.NewFormatter(b))
+func Println(a ...interface{}) (n int, err error) {
+	return fmt.Println(convertArgs(a)...)
+}
+
+// Sprint is a wrapper for fmt.Sprint that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprint(spew.NewFormatter(a), spew.NewFormatter(b))
+func Sprint(a ...interface{}) string {
+	return fmt.Sprint(convertArgs(a)...)
+}
+
+// Sprintf is a wrapper for fmt.Sprintf that treats each argument as if it were
+// passed with a default Formatter interface returned by NewFormatter.  It
+// returns the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprintf(format, spew.NewFormatter(a), spew.NewFormatter(b))
+func Sprintf(format string, a ...interface{}) string {
+	return fmt.Sprintf(format, convertArgs(a)...)
+}
+
+// Sprintln is a wrapper for fmt.Sprintln that treats each argument as if it
+// were passed with a default Formatter interface returned by NewFormatter.  It
+// returns the resulting string.  See NewFormatter for formatting details.
+//
+// This function is shorthand for the following syntax:
+//
+//	fmt.Sprintln(spew.NewFormatter(a), spew.NewFormatter(b))
+func Sprintln(a ...interface{}) string {
+	return fmt.Sprintln(convertArgs(a)...)
+}
+
+// convertArgs accepts a slice of arguments and returns a slice of the same
+// length with each argument converted to a default spew Formatter interface.
+func convertArgs(args []interface{}) (formatters []interface{}) {
+	formatters = make([]interface{}, len(args))
+	for index, arg := range args {
+		formatters[index] = NewFormatter(arg)
+	}
+	return formatters
+}
diff --git a/vendor/github.com/eapache/go-resiliency/LICENSE b/vendor/github.com/eapache/go-resiliency/LICENSE
new file mode 100644
index 0000000..698a3f5
--- /dev/null
+++ b/vendor/github.com/eapache/go-resiliency/LICENSE
@@ -0,0 +1,22 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 Evan Huus
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
diff --git a/vendor/github.com/eapache/go-resiliency/breaker/README.md b/vendor/github.com/eapache/go-resiliency/breaker/README.md
new file mode 100644
index 0000000..2d1b3d9
--- /dev/null
+++ b/vendor/github.com/eapache/go-resiliency/breaker/README.md
@@ -0,0 +1,34 @@
+circuit-breaker
+===============
+
+[![Build Status](https://travis-ci.org/eapache/go-resiliency.svg?branch=master)](https://travis-ci.org/eapache/go-resiliency)
+[![GoDoc](https://godoc.org/github.com/eapache/go-resiliency/breaker?status.svg)](https://godoc.org/github.com/eapache/go-resiliency/breaker)
+[![Code of Conduct](https://img.shields.io/badge/code%20of%20conduct-active-blue.svg)](https://eapache.github.io/conduct.html)
+
+The circuit-breaker resiliency pattern for golang.
+
+Creating a breaker takes three parameters:
+- error threshold (for opening the breaker)
+- success threshold (for closing the breaker)
+- timeout (how long to keep the breaker open)
+
+```go
+b := breaker.New(3, 1, 5*time.Second)
+
+for {
+	result := b.Run(func() error {
+		// communicate with some external service and
+		// return an error if the communication failed
+		return nil
+	})
+
+	switch result {
+	case nil:
+		// success!
+	case breaker.ErrBreakerOpen:
+		// our function wasn't run because the breaker was open
+	default:
+		// some other error
+	}
+}
+```
diff --git a/vendor/github.com/eapache/go-resiliency/breaker/breaker.go b/vendor/github.com/eapache/go-resiliency/breaker/breaker.go
new file mode 100644
index 0000000..f88ca72
--- /dev/null
+++ b/vendor/github.com/eapache/go-resiliency/breaker/breaker.go
@@ -0,0 +1,161 @@
+// Package breaker implements the circuit-breaker resiliency pattern for Go.
+package breaker
+
+import (
+	"errors"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// ErrBreakerOpen is the error returned from Run() when the function is not executed
+// because the breaker is currently open.
+var ErrBreakerOpen = errors.New("circuit breaker is open")
+
+const (
+	closed uint32 = iota
+	open
+	halfOpen
+)
+
+// Breaker implements the circuit-breaker resiliency pattern
+type Breaker struct {
+	errorThreshold, successThreshold int
+	timeout                          time.Duration
+
+	lock              sync.Mutex
+	state             uint32
+	errors, successes int
+	lastError         time.Time
+}
+
+// New constructs a new circuit-breaker that starts closed.
+// From closed, the breaker opens if "errorThreshold" errors are seen
+// without an error-free period of at least "timeout". From open, the
+// breaker half-closes after "timeout". From half-open, the breaker closes
+// after "successThreshold" consecutive successes, or opens on a single error.
+func New(errorThreshold, successThreshold int, timeout time.Duration) *Breaker {
+	return &Breaker{
+		errorThreshold:   errorThreshold,
+		successThreshold: successThreshold,
+		timeout:          timeout,
+	}
+}
+
+// Run will either return ErrBreakerOpen immediately if the circuit-breaker is
+// already open, or it will run the given function and pass along its return
+// value. It is safe to call Run concurrently on the same Breaker.
+func (b *Breaker) Run(work func() error) error {
+	state := atomic.LoadUint32(&b.state)
+
+	if state == open {
+		return ErrBreakerOpen
+	}
+
+	return b.doWork(state, work)
+}
+
+// Go will either return ErrBreakerOpen immediately if the circuit-breaker is
+// already open, or it will run the given function in a separate goroutine.
+// If the function is run, Go will return nil immediately, and will *not* return
+// the return value of the function. It is safe to call Go concurrently on the
+// same Breaker.
+func (b *Breaker) Go(work func() error) error {
+	state := atomic.LoadUint32(&b.state)
+
+	if state == open {
+		return ErrBreakerOpen
+	}
+
+	// errcheck complains about ignoring the error return value, but
+	// that's on purpose; if you want an error from a goroutine you have to
+	// get it over a channel or something
+	go b.doWork(state, work)
+
+	return nil
+}
+
+func (b *Breaker) doWork(state uint32, work func() error) error {
+	var panicValue interface{}
+
+	result := func() error {
+		defer func() {
+			panicValue = recover()
+		}()
+		return work()
+	}()
+
+	if result == nil && panicValue == nil && state == closed {
+		// short-circuit the normal, success path without contending
+		// on the lock
+		return nil
+	}
+
+	// oh well, I guess we have to contend on the lock
+	b.processResult(result, panicValue)
+
+	if panicValue != nil {
+		// as close as Go lets us come to a "rethrow" although unfortunately
+		// we lose the original panicing location
+		panic(panicValue)
+	}
+
+	return result
+}
+
+func (b *Breaker) processResult(result error, panicValue interface{}) {
+	b.lock.Lock()
+	defer b.lock.Unlock()
+
+	if result == nil && panicValue == nil {
+		if b.state == halfOpen {
+			b.successes++
+			if b.successes == b.successThreshold {
+				b.closeBreaker()
+			}
+		}
+	} else {
+		if b.errors > 0 {
+			expiry := b.lastError.Add(b.timeout)
+			if time.Now().After(expiry) {
+				b.errors = 0
+			}
+		}
+
+		switch b.state {
+		case closed:
+			b.errors++
+			if b.errors == b.errorThreshold {
+				b.openBreaker()
+			} else {
+				b.lastError = time.Now()
+			}
+		case halfOpen:
+			b.openBreaker()
+		}
+	}
+}
+
+func (b *Breaker) openBreaker() {
+	b.changeState(open)
+	go b.timer()
+}
+
+func (b *Breaker) closeBreaker() {
+	b.changeState(closed)
+}
+
+func (b *Breaker) timer() {
+	time.Sleep(b.timeout)
+
+	b.lock.Lock()
+	defer b.lock.Unlock()
+
+	b.changeState(halfOpen)
+}
+
+func (b *Breaker) changeState(newState uint32) {
+	b.errors = 0
+	b.successes = 0
+	atomic.StoreUint32(&b.state, newState)
+}
diff --git a/vendor/github.com/eapache/go-xerial-snappy/.gitignore b/vendor/github.com/eapache/go-xerial-snappy/.gitignore
new file mode 100644
index 0000000..daf913b
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/.gitignore
@@ -0,0 +1,24 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+
+# Folders
+_obj
+_test
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+*.test
+*.prof
diff --git a/vendor/github.com/eapache/go-xerial-snappy/.travis.yml b/vendor/github.com/eapache/go-xerial-snappy/.travis.yml
new file mode 100644
index 0000000..d6cf4f1
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/.travis.yml
@@ -0,0 +1,7 @@
+language: go
+
+go:
+- 1.5.4
+- 1.6.1
+
+sudo: false
diff --git a/vendor/github.com/eapache/go-xerial-snappy/LICENSE b/vendor/github.com/eapache/go-xerial-snappy/LICENSE
new file mode 100644
index 0000000..5bf3688
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/LICENSE
@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2016 Evan Huus
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/vendor/github.com/eapache/go-xerial-snappy/README.md b/vendor/github.com/eapache/go-xerial-snappy/README.md
new file mode 100644
index 0000000..3f2695c
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/README.md
@@ -0,0 +1,13 @@
+# go-xerial-snappy
+
+[![Build Status](https://travis-ci.org/eapache/go-xerial-snappy.svg?branch=master)](https://travis-ci.org/eapache/go-xerial-snappy)
+
+Xerial-compatible Snappy framing support for golang.
+
+Packages using Xerial for snappy encoding use a framing format incompatible with
+basically everything else in existence. This package wraps Go's built-in snappy
+package to support it.
+
+Apps that use this format include Apache Kafka (see
+https://github.com/dpkp/kafka-python/issues/126#issuecomment-35478921 for
+details).
diff --git a/vendor/github.com/eapache/go-xerial-snappy/fuzz.go b/vendor/github.com/eapache/go-xerial-snappy/fuzz.go
new file mode 100644
index 0000000..6a46f47
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/fuzz.go
@@ -0,0 +1,16 @@
+// +build gofuzz
+
+package snappy
+
+func Fuzz(data []byte) int {
+	decode, err := Decode(data)
+	if decode == nil && err == nil {
+		panic("nil error with nil result")
+	}
+
+	if err != nil {
+		return 0
+	}
+
+	return 1
+}
diff --git a/vendor/github.com/eapache/go-xerial-snappy/snappy.go b/vendor/github.com/eapache/go-xerial-snappy/snappy.go
new file mode 100644
index 0000000..ea8f7af
--- /dev/null
+++ b/vendor/github.com/eapache/go-xerial-snappy/snappy.go
@@ -0,0 +1,131 @@
+package snappy
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+
+	master "github.com/golang/snappy"
+)
+
+const (
+	sizeOffset = 16
+	sizeBytes  = 4
+)
+
+var (
+	xerialHeader = []byte{130, 83, 78, 65, 80, 80, 89, 0}
+
+	// This is xerial version 1 and minimally compatible with version 1
+	xerialVersionInfo = []byte{0, 0, 0, 1, 0, 0, 0, 1}
+
+	// ErrMalformed is returned by the decoder when the xerial framing
+	// is malformed
+	ErrMalformed = errors.New("malformed xerial framing")
+)
+
+func min(x, y int) int {
+    if x < y {
+        return x
+    }
+    return y
+}
+
+// Encode encodes data as snappy with no framing header.
+func Encode(src []byte) []byte {
+	return master.Encode(nil, src)
+}
+
+// EncodeStream *appends* to the specified 'dst' the compressed
+// 'src' in xerial framing format. If 'dst' does not have enough
+// capacity, then a new slice will be allocated. If 'dst' has
+// non-zero length, then if *must* have been built using this function.
+func EncodeStream(dst, src []byte) []byte {
+	if len(dst) == 0 {
+		dst = append(dst, xerialHeader...)
+		dst = append(dst, xerialVersionInfo...)
+	}
+
+	// Snappy encode in blocks of maximum 32KB
+	var (
+		max = len(src)
+		blockSize = 32 * 1024
+		pos   = 0
+		chunk []byte
+	)
+
+	for pos < max {
+		newPos := min(pos + blockSize, max)
+		chunk = master.Encode(chunk[:cap(chunk)], src[pos:newPos])
+
+		// First encode the compressed size (big-endian)
+		// Put* panics if the buffer is too small, so pad 4 bytes first
+		origLen := len(dst)
+		dst = append(dst, dst[0:4]...)
+		binary.BigEndian.PutUint32(dst[origLen:], uint32(len(chunk)))
+
+		// And now the compressed data
+		dst = append(dst, chunk...)
+		pos = newPos
+	}
+	return dst
+}
+
+// Decode decodes snappy data whether it is traditional unframed
+// or includes the xerial framing format.
+func Decode(src []byte) ([]byte, error) {
+	return DecodeInto(nil, src)
+}
+
+// DecodeInto decodes snappy data whether it is traditional unframed
+// or includes the xerial framing format into the specified `dst`.
+// It is assumed that the entirety of `dst` including all capacity is available
+// for use by this function. If `dst` is nil *or* insufficiently large to hold
+// the decoded `src`, new space will be allocated.
+func DecodeInto(dst, src []byte) ([]byte, error) {
+	var max = len(src)
+	if max < len(xerialHeader) {
+		return nil, ErrMalformed
+	}
+
+	if !bytes.Equal(src[:8], xerialHeader) {
+		return master.Decode(dst[:cap(dst)], src)
+	}
+
+	if max < sizeOffset+sizeBytes {
+		return nil, ErrMalformed
+	}
+
+	if dst == nil {
+		dst = make([]byte, 0, len(src))
+	}
+
+	dst = dst[:0]
+	var (
+		pos   = sizeOffset
+		chunk []byte
+		err       error
+	)
+
+	for pos+sizeBytes <= max {
+		size := int(binary.BigEndian.Uint32(src[pos : pos+sizeBytes]))
+		pos += sizeBytes
+
+		nextPos := pos + size
+		// On architectures where int is 32-bytes wide size + pos could
+		// overflow so we need to check the low bound as well as the
+		// high
+		if nextPos < pos || nextPos > max {
+			return nil, ErrMalformed
+		}
+
+		chunk, err = master.Decode(chunk[:cap(chunk)], src[pos:nextPos])
+
+		if err != nil {
+			return nil, err
+		}
+		pos = nextPos
+		dst = append(dst, chunk...)
+	}
+	return dst, nil
+}
diff --git a/vendor/github.com/eapache/queue/.gitignore b/vendor/github.com/eapache/queue/.gitignore
new file mode 100644
index 0000000..8365624
--- /dev/null
+++ b/vendor/github.com/eapache/queue/.gitignore
@@ -0,0 +1,23 @@
+# Compiled Object files, Static and Dynamic libs (Shared Objects)
+*.o
+*.a
+*.so
+
+# Folders
+_obj
+_test
+
+# Architecture specific extensions/prefixes
+*.[568vq]
+[568vq].out
+
+*.cgo1.go
+*.cgo2.c
+_cgo_defun.c
+_cgo_gotypes.go
+_cgo_export.*
+
+_testmain.go
+
+*.exe
+*.test
diff --git a/vendor/github.com/eapache/queue/.travis.yml b/vendor/github.com/eapache/queue/.travis.yml
new file mode 100644
index 0000000..235a40a
--- /dev/null
+++ b/vendor/github.com/eapache/queue/.travis.yml
@@ -0,0 +1,7 @@
+language: go
+sudo: false
+
+go:
+  - 1.2
+  - 1.3
+  - 1.4
diff --git a/vendor/github.com/eapache/queue/LICENSE b/vendor/github.com/eapache/queue/LICENSE
new file mode 100644
index 0000000..d5f36db
--- /dev/null
+++ b/vendor/github.com/eapache/queue/LICENSE
@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 Evan Huus
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
\ No newline at end of file
diff --git a/vendor/github.com/eapache/queue/README.md b/vendor/github.com/eapache/queue/README.md
new file mode 100644
index 0000000..8e78233
--- /dev/null
+++ b/vendor/github.com/eapache/queue/README.md
@@ -0,0 +1,16 @@
+Queue
+=====
+
+[![Build Status](https://travis-ci.org/eapache/queue.svg)](https://travis-ci.org/eapache/queue)
+[![GoDoc](https://godoc.org/github.com/eapache/queue?status.png)](https://godoc.org/github.com/eapache/queue)
+[![Code of Conduct](https://img.shields.io/badge/code%20of%20conduct-active-blue.svg)](https://eapache.github.io/conduct.html)
+
+A fast Golang queue using a ring-buffer, based on the version suggested by Dariusz Górecki.
+Using this instead of other, simpler, queue implementations (slice+append or linked list) provides
+substantial memory and time benefits, and fewer GC pauses.
+
+The queue implemented here is as fast as it is in part because it is *not* thread-safe.
+
+Follows semantic versioning using https://gopkg.in/ - import from
+[`gopkg.in/eapache/queue.v1`](https://gopkg.in/eapache/queue.v1)
+for guaranteed API stability.
diff --git a/vendor/github.com/eapache/queue/queue.go b/vendor/github.com/eapache/queue/queue.go
new file mode 100644
index 0000000..71d1acd
--- /dev/null
+++ b/vendor/github.com/eapache/queue/queue.go
@@ -0,0 +1,102 @@
+/*
+Package queue provides a fast, ring-buffer queue based on the version suggested by Dariusz Górecki.
+Using this instead of other, simpler, queue implementations (slice+append or linked list) provides
+substantial memory and time benefits, and fewer GC pauses.
+
+The queue implemented here is as fast as it is for an additional reason: it is *not* thread-safe.
+*/
+package queue
+
+// minQueueLen is smallest capacity that queue may have.
+// Must be power of 2 for bitwise modulus: x % n == x & (n - 1).
+const minQueueLen = 16
+
+// Queue represents a single instance of the queue data structure.
+type Queue struct {
+	buf               []interface{}
+	head, tail, count int
+}
+
+// New constructs and returns a new Queue.
+func New() *Queue {
+	return &Queue{
+		buf: make([]interface{}, minQueueLen),
+	}
+}
+
+// Length returns the number of elements currently stored in the queue.
+func (q *Queue) Length() int {
+	return q.count
+}
+
+// resizes the queue to fit exactly twice its current contents
+// this can result in shrinking if the queue is less than half-full
+func (q *Queue) resize() {
+	newBuf := make([]interface{}, q.count<<1)
+
+	if q.tail > q.head {
+		copy(newBuf, q.buf[q.head:q.tail])
+	} else {
+		n := copy(newBuf, q.buf[q.head:])
+		copy(newBuf[n:], q.buf[:q.tail])
+	}
+
+	q.head = 0
+	q.tail = q.count
+	q.buf = newBuf
+}
+
+// Add puts an element on the end of the queue.
+func (q *Queue) Add(elem interface{}) {
+	if q.count == len(q.buf) {
+		q.resize()
+	}
+
+	q.buf[q.tail] = elem
+	// bitwise modulus
+	q.tail = (q.tail + 1) & (len(q.buf) - 1)
+	q.count++
+}
+
+// Peek returns the element at the head of the queue. This call panics
+// if the queue is empty.
+func (q *Queue) Peek() interface{} {
+	if q.count <= 0 {
+		panic("queue: Peek() called on empty queue")
+	}
+	return q.buf[q.head]
+}
+
+// Get returns the element at index i in the queue. If the index is
+// invalid, the call will panic. This method accepts both positive and
+// negative index values. Index 0 refers to the first element, and
+// index -1 refers to the last.
+func (q *Queue) Get(i int) interface{} {
+	// If indexing backwards, convert to positive index.
+	if i < 0 {
+		i += q.count
+	}
+	if i < 0 || i >= q.count {
+		panic("queue: Get() called with index out of range")
+	}
+	// bitwise modulus
+	return q.buf[(q.head+i)&(len(q.buf)-1)]
+}
+
+// Remove removes and returns the element from the front of the queue. If the
+// queue is empty, the call will panic.
+func (q *Queue) Remove() interface{} {
+	if q.count <= 0 {
+		panic("queue: Remove() called on empty queue")
+	}
+	ret := q.buf[q.head]
+	q.buf[q.head] = nil
+	// bitwise modulus
+	q.head = (q.head + 1) & (len(q.buf) - 1)
+	q.count--
+	// Resize down if buffer 1/4 full.
+	if len(q.buf) > minQueueLen && (q.count<<2) == len(q.buf) {
+		q.resize()
+	}
+	return ret
+}
diff --git a/vendor/github.com/gogo/protobuf/AUTHORS b/vendor/github.com/gogo/protobuf/AUTHORS
new file mode 100644
index 0000000..3d97fc7
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/AUTHORS
@@ -0,0 +1,15 @@
+# This is the official list of GoGo authors for copyright purposes.
+# This file is distinct from the CONTRIBUTORS file, which
+# lists people.  For example, employees are listed in CONTRIBUTORS,
+# but not in AUTHORS, because the employer holds the copyright.
+
+# Names should be added to this file as one of
+#     Organization's name
+#     Individual's name <submission email address>
+#     Individual's name <submission email address> <email2> <emailN>
+
+# Please keep the list sorted.
+
+Sendgrid, Inc
+Vastech SA (PTY) LTD
+Walter Schulze <awalterschulze@gmail.com>
diff --git a/vendor/github.com/gogo/protobuf/CONTRIBUTORS b/vendor/github.com/gogo/protobuf/CONTRIBUTORS
new file mode 100644
index 0000000..1b4f6c2
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/CONTRIBUTORS
@@ -0,0 +1,23 @@
+Anton Povarov <anton.povarov@gmail.com>
+Brian Goff <cpuguy83@gmail.com>
+Clayton Coleman <ccoleman@redhat.com>
+Denis Smirnov <denis.smirnov.91@gmail.com>
+DongYun Kang <ceram1000@gmail.com>
+Dwayne Schultz <dschultz@pivotal.io>
+Georg Apitz <gapitz@pivotal.io>
+Gustav Paul <gustav.paul@gmail.com>
+Johan Brandhorst <johan.brandhorst@gmail.com>
+John Shahid <jvshahid@gmail.com>
+John Tuley <john@tuley.org>
+Laurent <laurent@adyoulike.com>
+Patrick Lee <patrick@dropbox.com>
+Peter Edge <peter.edge@gmail.com>
+Roger Johansson <rogeralsing@gmail.com>
+Sam Nguyen <sam.nguyen@sendgrid.com>
+Sergio Arbeo <serabe@gmail.com>
+Stephen J Day <stephen.day@docker.com>
+Tamir Duberstein <tamird@gmail.com>
+Todd Eisenberger <teisenberger@dropbox.com>
+Tormod Erevik Lea <tormodlea@gmail.com>
+Vyacheslav Kim <kane@sendgrid.com>
+Walter Schulze <awalterschulze@gmail.com>
diff --git a/vendor/github.com/gogo/protobuf/GOLANG_CONTRIBUTORS b/vendor/github.com/gogo/protobuf/GOLANG_CONTRIBUTORS
new file mode 100644
index 0000000..b368efb
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/GOLANG_CONTRIBUTORS
@@ -0,0 +1,5 @@
+The contributors to the Go protobuf repository:
+
+# This source code was written by the Go contributors.
+# The master list of contributors is in the main Go distribution,
+# visible at http://tip.golang.org/CONTRIBUTORS.
\ No newline at end of file
diff --git a/vendor/github.com/gogo/protobuf/LICENSE b/vendor/github.com/gogo/protobuf/LICENSE
new file mode 100644
index 0000000..f57de90
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/LICENSE
@@ -0,0 +1,35 @@
+Copyright (c) 2013, The GoGo Authors. All rights reserved.
+
+Protocol Buffers for Go with Gadgets
+
+Go support for Protocol Buffers - Google's data interchange format
+
+Copyright 2010 The Go Authors.  All rights reserved.
+https://github.com/golang/protobuf
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/Makefile b/vendor/github.com/gogo/protobuf/gogoproto/Makefile
new file mode 100644
index 0000000..0b4659b
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/Makefile
@@ -0,0 +1,37 @@
+# Protocol Buffers for Go with Gadgets
+#
+# Copyright (c) 2013, The GoGo Authors. All rights reserved.
+# http://github.com/gogo/protobuf
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+regenerate:
+	go install github.com/gogo/protobuf/protoc-gen-gogo
+	protoc --gogo_out=Mgoogle/protobuf/descriptor.proto=github.com/gogo/protobuf/protoc-gen-gogo/descriptor:../../../../ --proto_path=../../../../:../protobuf/:. *.proto
+
+restore:
+	cp gogo.pb.golden gogo.pb.go
+
+preserve:
+	cp gogo.pb.go gogo.pb.golden
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/doc.go b/vendor/github.com/gogo/protobuf/gogoproto/doc.go
new file mode 100644
index 0000000..081c86f
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/doc.go
@@ -0,0 +1,169 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+/*
+Package gogoproto provides extensions for protocol buffers to achieve:
+
+  - fast marshalling and unmarshalling.
+  - peace of mind by optionally generating test and benchmark code.
+  - more canonical Go structures.
+  - less typing by optionally generating extra helper code.
+  - goprotobuf compatibility
+
+More Canonical Go Structures
+
+A lot of time working with a goprotobuf struct will lead you to a place where you create another struct that is easier to work with and then have a function to copy the values between the two structs.
+You might also find that basic structs that started their life as part of an API need to be sent over the wire. With gob, you could just send it. With goprotobuf, you need to make a parallel struct.
+Gogoprotobuf tries to fix these problems with the nullable, embed, customtype and customname field extensions.
+
+  - nullable, if false, a field is generated without a pointer (see warning below).
+  - embed, if true, the field is generated as an embedded field.
+  - customtype, It works with the Marshal and Unmarshal methods, to allow you to have your own types in your struct, but marshal to bytes. For example, custom.Uuid or custom.Fixed128
+  - customname (beta), Changes the generated fieldname. This is especially useful when generated methods conflict with fieldnames.
+  - casttype (beta), Changes the generated fieldtype.  All generated code assumes that this type is castable to the protocol buffer field type.  It does not work for structs or enums.
+  - castkey (beta), Changes the generated fieldtype for a map key.  All generated code assumes that this type is castable to the protocol buffer field type.  Only supported on maps.
+  - castvalue (beta), Changes the generated fieldtype for a map value.  All generated code assumes that this type is castable to the protocol buffer field type.  Only supported on maps.
+
+Warning about nullable: According to the Protocol Buffer specification, you should be able to tell whether a field is set or unset. With the option nullable=false this feature is lost, since your non-nullable fields will always be set. It can be seen as a layer on top of Protocol Buffers, where before and after marshalling all non-nullable fields are set and they cannot be unset.
+
+Let us look at:
+
+	github.com/gogo/protobuf/test/example/example.proto
+
+for a quicker overview.
+
+The following message:
+
+  package test;
+
+  import "github.com/gogo/protobuf/gogoproto/gogo.proto";
+
+	message A {
+		optional string Description = 1 [(gogoproto.nullable) = false];
+		optional int64 Number = 2 [(gogoproto.nullable) = false];
+		optional bytes Id = 3 [(gogoproto.customtype) = "github.com/gogo/protobuf/test/custom.Uuid", (gogoproto.nullable) = false];
+	}
+
+Will generate a go struct which looks a lot like this:
+
+	type A struct {
+		Description string
+		Number      int64
+		Id          github_com_gogo_protobuf_test_custom.Uuid
+	}
+
+You will see there are no pointers, since all fields are non-nullable.
+You will also see a custom type which marshals to a string.
+Be warned it is your responsibility to test your custom types thoroughly.
+You should think of every possible empty and nil case for your marshaling, unmarshaling and size methods.
+
+Next we will embed the message A in message B.
+
+	message B {
+		optional A A = 1 [(gogoproto.nullable) = false, (gogoproto.embed) = true];
+		repeated bytes G = 2 [(gogoproto.customtype) = "github.com/gogo/protobuf/test/custom.Uint128", (gogoproto.nullable) = false];
+	}
+
+See below that A is embedded in B.
+
+	type B struct {
+		A
+		G []github_com_gogo_protobuf_test_custom.Uint128
+	}
+
+Also see the repeated custom type.
+
+	type Uint128 [2]uint64
+
+Next we will create a custom name for one of our fields.
+
+	message C {
+		optional int64 size = 1 [(gogoproto.customname) = "MySize"];
+	}
+
+See below that the field's name is MySize and not Size.
+
+	type C struct {
+		MySize		*int64
+	}
+
+The is useful when having a protocol buffer message with a field name which conflicts with a generated method.
+As an example, having a field name size and using the sizer plugin to generate a Size method will cause a go compiler error.
+Using customname you can fix this error without changing the field name.
+This is typically useful when working with a protocol buffer that was designed before these methods and/or the go language were avialable.
+
+Gogoprotobuf also has some more subtle changes, these could be changed back:
+
+  - the generated package name for imports do not have the extra /filename.pb,
+  but are actually the imports specified in the .proto file.
+
+Gogoprotobuf also has lost some features which should be brought back with time:
+
+  - Marshalling and unmarshalling with reflect and without the unsafe package,
+  this requires work in pointer_reflect.go
+
+Why does nullable break protocol buffer specifications:
+
+The protocol buffer specification states, somewhere, that you should be able to tell whether a
+field is set or unset.  With the option nullable=false this feature is lost,
+since your non-nullable fields will always be set.  It can be seen as a layer on top of
+protocol buffers, where before and after marshalling all non-nullable fields are set
+and they cannot be unset.
+
+Goprotobuf Compatibility:
+
+Gogoprotobuf is compatible with Goprotobuf, because it is compatible with protocol buffers.
+Gogoprotobuf generates the same code as goprotobuf if no extensions are used.
+The enumprefix, getters and stringer extensions can be used to remove some of the unnecessary code generated by goprotobuf:
+
+  - gogoproto_import, if false, the generated code imports github.com/golang/protobuf/proto instead of github.com/gogo/protobuf/proto.
+  - goproto_enum_prefix, if false, generates the enum constant names without the messagetype prefix
+  - goproto_enum_stringer (experimental), if false, the enum is generated without the default string method, this is useful for rather using enum_stringer, or allowing you to write your own string method.
+  - goproto_getters, if false, the message is generated without get methods, this is useful when you would rather want to use face
+  - goproto_stringer, if false, the message is generated without the default string method, this is useful for rather using stringer, or allowing you to write your own string method.
+  - goproto_extensions_map (beta), if false, the extensions field is generated as type []byte instead of type map[int32]proto.Extension
+  - goproto_unrecognized (beta), if false, XXX_unrecognized field is not generated. This is useful in conjunction with gogoproto.nullable=false, to generate structures completely devoid of pointers and reduce GC pressure at the cost of losing information about unrecognized fields.
+  - goproto_registration (beta), if true, the generated files will register all messages and types against both gogo/protobuf and golang/protobuf. This is necessary when using third-party packages which read registrations from golang/protobuf (such as the grpc-gateway).
+
+Less Typing and Peace of Mind is explained in their specific plugin folders godoc:
+
+	- github.com/gogo/protobuf/plugin/<extension_name>
+
+If you do not use any of these extension the code that is generated
+will be the same as if goprotobuf has generated it.
+
+The most complete way to see examples is to look at
+
+	github.com/gogo/protobuf/test/thetest.proto
+
+Gogoprototest is a seperate project,
+because we want to keep gogoprotobuf independent of goprotobuf,
+but we still want to test it thoroughly.
+
+*/
+package gogoproto
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.go b/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.go
new file mode 100644
index 0000000..e352808
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.go
@@ -0,0 +1,874 @@
+// Code generated by protoc-gen-gogo. DO NOT EDIT.
+// source: gogo.proto
+
+package gogoproto
+
+import (
+	fmt "fmt"
+	proto "github.com/gogo/protobuf/proto"
+	descriptor "github.com/gogo/protobuf/protoc-gen-gogo/descriptor"
+	math "math"
+)
+
+// Reference imports to suppress errors if they are not otherwise used.
+var _ = proto.Marshal
+var _ = fmt.Errorf
+var _ = math.Inf
+
+// This is a compile-time assertion to ensure that this generated file
+// is compatible with the proto package it is being compiled against.
+// A compilation error at this line likely means your copy of the
+// proto package needs to be updated.
+const _ = proto.GoGoProtoPackageIsVersion2 // please upgrade the proto package
+
+var E_GoprotoEnumPrefix = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         62001,
+	Name:          "gogoproto.goproto_enum_prefix",
+	Tag:           "varint,62001,opt,name=goproto_enum_prefix",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoEnumStringer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         62021,
+	Name:          "gogoproto.goproto_enum_stringer",
+	Tag:           "varint,62021,opt,name=goproto_enum_stringer",
+	Filename:      "gogo.proto",
+}
+
+var E_EnumStringer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         62022,
+	Name:          "gogoproto.enum_stringer",
+	Tag:           "varint,62022,opt,name=enum_stringer",
+	Filename:      "gogo.proto",
+}
+
+var E_EnumCustomname = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         62023,
+	Name:          "gogoproto.enum_customname",
+	Tag:           "bytes,62023,opt,name=enum_customname",
+	Filename:      "gogo.proto",
+}
+
+var E_Enumdecl = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         62024,
+	Name:          "gogoproto.enumdecl",
+	Tag:           "varint,62024,opt,name=enumdecl",
+	Filename:      "gogo.proto",
+}
+
+var E_EnumvalueCustomname = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.EnumValueOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         66001,
+	Name:          "gogoproto.enumvalue_customname",
+	Tag:           "bytes,66001,opt,name=enumvalue_customname",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoGettersAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63001,
+	Name:          "gogoproto.goproto_getters_all",
+	Tag:           "varint,63001,opt,name=goproto_getters_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoEnumPrefixAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63002,
+	Name:          "gogoproto.goproto_enum_prefix_all",
+	Tag:           "varint,63002,opt,name=goproto_enum_prefix_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoStringerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63003,
+	Name:          "gogoproto.goproto_stringer_all",
+	Tag:           "varint,63003,opt,name=goproto_stringer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_VerboseEqualAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63004,
+	Name:          "gogoproto.verbose_equal_all",
+	Tag:           "varint,63004,opt,name=verbose_equal_all",
+	Filename:      "gogo.proto",
+}
+
+var E_FaceAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63005,
+	Name:          "gogoproto.face_all",
+	Tag:           "varint,63005,opt,name=face_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GostringAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63006,
+	Name:          "gogoproto.gostring_all",
+	Tag:           "varint,63006,opt,name=gostring_all",
+	Filename:      "gogo.proto",
+}
+
+var E_PopulateAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63007,
+	Name:          "gogoproto.populate_all",
+	Tag:           "varint,63007,opt,name=populate_all",
+	Filename:      "gogo.proto",
+}
+
+var E_StringerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63008,
+	Name:          "gogoproto.stringer_all",
+	Tag:           "varint,63008,opt,name=stringer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_OnlyoneAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63009,
+	Name:          "gogoproto.onlyone_all",
+	Tag:           "varint,63009,opt,name=onlyone_all",
+	Filename:      "gogo.proto",
+}
+
+var E_EqualAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63013,
+	Name:          "gogoproto.equal_all",
+	Tag:           "varint,63013,opt,name=equal_all",
+	Filename:      "gogo.proto",
+}
+
+var E_DescriptionAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63014,
+	Name:          "gogoproto.description_all",
+	Tag:           "varint,63014,opt,name=description_all",
+	Filename:      "gogo.proto",
+}
+
+var E_TestgenAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63015,
+	Name:          "gogoproto.testgen_all",
+	Tag:           "varint,63015,opt,name=testgen_all",
+	Filename:      "gogo.proto",
+}
+
+var E_BenchgenAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63016,
+	Name:          "gogoproto.benchgen_all",
+	Tag:           "varint,63016,opt,name=benchgen_all",
+	Filename:      "gogo.proto",
+}
+
+var E_MarshalerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63017,
+	Name:          "gogoproto.marshaler_all",
+	Tag:           "varint,63017,opt,name=marshaler_all",
+	Filename:      "gogo.proto",
+}
+
+var E_UnmarshalerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63018,
+	Name:          "gogoproto.unmarshaler_all",
+	Tag:           "varint,63018,opt,name=unmarshaler_all",
+	Filename:      "gogo.proto",
+}
+
+var E_StableMarshalerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63019,
+	Name:          "gogoproto.stable_marshaler_all",
+	Tag:           "varint,63019,opt,name=stable_marshaler_all",
+	Filename:      "gogo.proto",
+}
+
+var E_SizerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63020,
+	Name:          "gogoproto.sizer_all",
+	Tag:           "varint,63020,opt,name=sizer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoEnumStringerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63021,
+	Name:          "gogoproto.goproto_enum_stringer_all",
+	Tag:           "varint,63021,opt,name=goproto_enum_stringer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_EnumStringerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63022,
+	Name:          "gogoproto.enum_stringer_all",
+	Tag:           "varint,63022,opt,name=enum_stringer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_UnsafeMarshalerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63023,
+	Name:          "gogoproto.unsafe_marshaler_all",
+	Tag:           "varint,63023,opt,name=unsafe_marshaler_all",
+	Filename:      "gogo.proto",
+}
+
+var E_UnsafeUnmarshalerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63024,
+	Name:          "gogoproto.unsafe_unmarshaler_all",
+	Tag:           "varint,63024,opt,name=unsafe_unmarshaler_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoExtensionsMapAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63025,
+	Name:          "gogoproto.goproto_extensions_map_all",
+	Tag:           "varint,63025,opt,name=goproto_extensions_map_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoUnrecognizedAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63026,
+	Name:          "gogoproto.goproto_unrecognized_all",
+	Tag:           "varint,63026,opt,name=goproto_unrecognized_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GogoprotoImport = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63027,
+	Name:          "gogoproto.gogoproto_import",
+	Tag:           "varint,63027,opt,name=gogoproto_import",
+	Filename:      "gogo.proto",
+}
+
+var E_ProtosizerAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63028,
+	Name:          "gogoproto.protosizer_all",
+	Tag:           "varint,63028,opt,name=protosizer_all",
+	Filename:      "gogo.proto",
+}
+
+var E_CompareAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63029,
+	Name:          "gogoproto.compare_all",
+	Tag:           "varint,63029,opt,name=compare_all",
+	Filename:      "gogo.proto",
+}
+
+var E_TypedeclAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63030,
+	Name:          "gogoproto.typedecl_all",
+	Tag:           "varint,63030,opt,name=typedecl_all",
+	Filename:      "gogo.proto",
+}
+
+var E_EnumdeclAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63031,
+	Name:          "gogoproto.enumdecl_all",
+	Tag:           "varint,63031,opt,name=enumdecl_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoRegistration = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63032,
+	Name:          "gogoproto.goproto_registration",
+	Tag:           "varint,63032,opt,name=goproto_registration",
+	Filename:      "gogo.proto",
+}
+
+var E_MessagenameAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63033,
+	Name:          "gogoproto.messagename_all",
+	Tag:           "varint,63033,opt,name=messagename_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoSizecacheAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63034,
+	Name:          "gogoproto.goproto_sizecache_all",
+	Tag:           "varint,63034,opt,name=goproto_sizecache_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoUnkeyedAll = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FileOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         63035,
+	Name:          "gogoproto.goproto_unkeyed_all",
+	Tag:           "varint,63035,opt,name=goproto_unkeyed_all",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoGetters = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64001,
+	Name:          "gogoproto.goproto_getters",
+	Tag:           "varint,64001,opt,name=goproto_getters",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoStringer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64003,
+	Name:          "gogoproto.goproto_stringer",
+	Tag:           "varint,64003,opt,name=goproto_stringer",
+	Filename:      "gogo.proto",
+}
+
+var E_VerboseEqual = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64004,
+	Name:          "gogoproto.verbose_equal",
+	Tag:           "varint,64004,opt,name=verbose_equal",
+	Filename:      "gogo.proto",
+}
+
+var E_Face = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64005,
+	Name:          "gogoproto.face",
+	Tag:           "varint,64005,opt,name=face",
+	Filename:      "gogo.proto",
+}
+
+var E_Gostring = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64006,
+	Name:          "gogoproto.gostring",
+	Tag:           "varint,64006,opt,name=gostring",
+	Filename:      "gogo.proto",
+}
+
+var E_Populate = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64007,
+	Name:          "gogoproto.populate",
+	Tag:           "varint,64007,opt,name=populate",
+	Filename:      "gogo.proto",
+}
+
+var E_Stringer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         67008,
+	Name:          "gogoproto.stringer",
+	Tag:           "varint,67008,opt,name=stringer",
+	Filename:      "gogo.proto",
+}
+
+var E_Onlyone = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64009,
+	Name:          "gogoproto.onlyone",
+	Tag:           "varint,64009,opt,name=onlyone",
+	Filename:      "gogo.proto",
+}
+
+var E_Equal = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64013,
+	Name:          "gogoproto.equal",
+	Tag:           "varint,64013,opt,name=equal",
+	Filename:      "gogo.proto",
+}
+
+var E_Description = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64014,
+	Name:          "gogoproto.description",
+	Tag:           "varint,64014,opt,name=description",
+	Filename:      "gogo.proto",
+}
+
+var E_Testgen = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64015,
+	Name:          "gogoproto.testgen",
+	Tag:           "varint,64015,opt,name=testgen",
+	Filename:      "gogo.proto",
+}
+
+var E_Benchgen = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64016,
+	Name:          "gogoproto.benchgen",
+	Tag:           "varint,64016,opt,name=benchgen",
+	Filename:      "gogo.proto",
+}
+
+var E_Marshaler = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64017,
+	Name:          "gogoproto.marshaler",
+	Tag:           "varint,64017,opt,name=marshaler",
+	Filename:      "gogo.proto",
+}
+
+var E_Unmarshaler = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64018,
+	Name:          "gogoproto.unmarshaler",
+	Tag:           "varint,64018,opt,name=unmarshaler",
+	Filename:      "gogo.proto",
+}
+
+var E_StableMarshaler = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64019,
+	Name:          "gogoproto.stable_marshaler",
+	Tag:           "varint,64019,opt,name=stable_marshaler",
+	Filename:      "gogo.proto",
+}
+
+var E_Sizer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64020,
+	Name:          "gogoproto.sizer",
+	Tag:           "varint,64020,opt,name=sizer",
+	Filename:      "gogo.proto",
+}
+
+var E_UnsafeMarshaler = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64023,
+	Name:          "gogoproto.unsafe_marshaler",
+	Tag:           "varint,64023,opt,name=unsafe_marshaler",
+	Filename:      "gogo.proto",
+}
+
+var E_UnsafeUnmarshaler = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64024,
+	Name:          "gogoproto.unsafe_unmarshaler",
+	Tag:           "varint,64024,opt,name=unsafe_unmarshaler",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoExtensionsMap = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64025,
+	Name:          "gogoproto.goproto_extensions_map",
+	Tag:           "varint,64025,opt,name=goproto_extensions_map",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoUnrecognized = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64026,
+	Name:          "gogoproto.goproto_unrecognized",
+	Tag:           "varint,64026,opt,name=goproto_unrecognized",
+	Filename:      "gogo.proto",
+}
+
+var E_Protosizer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64028,
+	Name:          "gogoproto.protosizer",
+	Tag:           "varint,64028,opt,name=protosizer",
+	Filename:      "gogo.proto",
+}
+
+var E_Compare = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64029,
+	Name:          "gogoproto.compare",
+	Tag:           "varint,64029,opt,name=compare",
+	Filename:      "gogo.proto",
+}
+
+var E_Typedecl = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64030,
+	Name:          "gogoproto.typedecl",
+	Tag:           "varint,64030,opt,name=typedecl",
+	Filename:      "gogo.proto",
+}
+
+var E_Messagename = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64033,
+	Name:          "gogoproto.messagename",
+	Tag:           "varint,64033,opt,name=messagename",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoSizecache = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64034,
+	Name:          "gogoproto.goproto_sizecache",
+	Tag:           "varint,64034,opt,name=goproto_sizecache",
+	Filename:      "gogo.proto",
+}
+
+var E_GoprotoUnkeyed = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.MessageOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         64035,
+	Name:          "gogoproto.goproto_unkeyed",
+	Tag:           "varint,64035,opt,name=goproto_unkeyed",
+	Filename:      "gogo.proto",
+}
+
+var E_Nullable = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         65001,
+	Name:          "gogoproto.nullable",
+	Tag:           "varint,65001,opt,name=nullable",
+	Filename:      "gogo.proto",
+}
+
+var E_Embed = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         65002,
+	Name:          "gogoproto.embed",
+	Tag:           "varint,65002,opt,name=embed",
+	Filename:      "gogo.proto",
+}
+
+var E_Customtype = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65003,
+	Name:          "gogoproto.customtype",
+	Tag:           "bytes,65003,opt,name=customtype",
+	Filename:      "gogo.proto",
+}
+
+var E_Customname = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65004,
+	Name:          "gogoproto.customname",
+	Tag:           "bytes,65004,opt,name=customname",
+	Filename:      "gogo.proto",
+}
+
+var E_Jsontag = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65005,
+	Name:          "gogoproto.jsontag",
+	Tag:           "bytes,65005,opt,name=jsontag",
+	Filename:      "gogo.proto",
+}
+
+var E_Moretags = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65006,
+	Name:          "gogoproto.moretags",
+	Tag:           "bytes,65006,opt,name=moretags",
+	Filename:      "gogo.proto",
+}
+
+var E_Casttype = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65007,
+	Name:          "gogoproto.casttype",
+	Tag:           "bytes,65007,opt,name=casttype",
+	Filename:      "gogo.proto",
+}
+
+var E_Castkey = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65008,
+	Name:          "gogoproto.castkey",
+	Tag:           "bytes,65008,opt,name=castkey",
+	Filename:      "gogo.proto",
+}
+
+var E_Castvalue = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         65009,
+	Name:          "gogoproto.castvalue",
+	Tag:           "bytes,65009,opt,name=castvalue",
+	Filename:      "gogo.proto",
+}
+
+var E_Stdtime = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         65010,
+	Name:          "gogoproto.stdtime",
+	Tag:           "varint,65010,opt,name=stdtime",
+	Filename:      "gogo.proto",
+}
+
+var E_Stdduration = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         65011,
+	Name:          "gogoproto.stdduration",
+	Tag:           "varint,65011,opt,name=stdduration",
+	Filename:      "gogo.proto",
+}
+
+var E_Wktpointer = &proto.ExtensionDesc{
+	ExtendedType:  (*descriptor.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         65012,
+	Name:          "gogoproto.wktpointer",
+	Tag:           "varint,65012,opt,name=wktpointer",
+	Filename:      "gogo.proto",
+}
+
+func init() {
+	proto.RegisterExtension(E_GoprotoEnumPrefix)
+	proto.RegisterExtension(E_GoprotoEnumStringer)
+	proto.RegisterExtension(E_EnumStringer)
+	proto.RegisterExtension(E_EnumCustomname)
+	proto.RegisterExtension(E_Enumdecl)
+	proto.RegisterExtension(E_EnumvalueCustomname)
+	proto.RegisterExtension(E_GoprotoGettersAll)
+	proto.RegisterExtension(E_GoprotoEnumPrefixAll)
+	proto.RegisterExtension(E_GoprotoStringerAll)
+	proto.RegisterExtension(E_VerboseEqualAll)
+	proto.RegisterExtension(E_FaceAll)
+	proto.RegisterExtension(E_GostringAll)
+	proto.RegisterExtension(E_PopulateAll)
+	proto.RegisterExtension(E_StringerAll)
+	proto.RegisterExtension(E_OnlyoneAll)
+	proto.RegisterExtension(E_EqualAll)
+	proto.RegisterExtension(E_DescriptionAll)
+	proto.RegisterExtension(E_TestgenAll)
+	proto.RegisterExtension(E_BenchgenAll)
+	proto.RegisterExtension(E_MarshalerAll)
+	proto.RegisterExtension(E_UnmarshalerAll)
+	proto.RegisterExtension(E_StableMarshalerAll)
+	proto.RegisterExtension(E_SizerAll)
+	proto.RegisterExtension(E_GoprotoEnumStringerAll)
+	proto.RegisterExtension(E_EnumStringerAll)
+	proto.RegisterExtension(E_UnsafeMarshalerAll)
+	proto.RegisterExtension(E_UnsafeUnmarshalerAll)
+	proto.RegisterExtension(E_GoprotoExtensionsMapAll)
+	proto.RegisterExtension(E_GoprotoUnrecognizedAll)
+	proto.RegisterExtension(E_GogoprotoImport)
+	proto.RegisterExtension(E_ProtosizerAll)
+	proto.RegisterExtension(E_CompareAll)
+	proto.RegisterExtension(E_TypedeclAll)
+	proto.RegisterExtension(E_EnumdeclAll)
+	proto.RegisterExtension(E_GoprotoRegistration)
+	proto.RegisterExtension(E_MessagenameAll)
+	proto.RegisterExtension(E_GoprotoSizecacheAll)
+	proto.RegisterExtension(E_GoprotoUnkeyedAll)
+	proto.RegisterExtension(E_GoprotoGetters)
+	proto.RegisterExtension(E_GoprotoStringer)
+	proto.RegisterExtension(E_VerboseEqual)
+	proto.RegisterExtension(E_Face)
+	proto.RegisterExtension(E_Gostring)
+	proto.RegisterExtension(E_Populate)
+	proto.RegisterExtension(E_Stringer)
+	proto.RegisterExtension(E_Onlyone)
+	proto.RegisterExtension(E_Equal)
+	proto.RegisterExtension(E_Description)
+	proto.RegisterExtension(E_Testgen)
+	proto.RegisterExtension(E_Benchgen)
+	proto.RegisterExtension(E_Marshaler)
+	proto.RegisterExtension(E_Unmarshaler)
+	proto.RegisterExtension(E_StableMarshaler)
+	proto.RegisterExtension(E_Sizer)
+	proto.RegisterExtension(E_UnsafeMarshaler)
+	proto.RegisterExtension(E_UnsafeUnmarshaler)
+	proto.RegisterExtension(E_GoprotoExtensionsMap)
+	proto.RegisterExtension(E_GoprotoUnrecognized)
+	proto.RegisterExtension(E_Protosizer)
+	proto.RegisterExtension(E_Compare)
+	proto.RegisterExtension(E_Typedecl)
+	proto.RegisterExtension(E_Messagename)
+	proto.RegisterExtension(E_GoprotoSizecache)
+	proto.RegisterExtension(E_GoprotoUnkeyed)
+	proto.RegisterExtension(E_Nullable)
+	proto.RegisterExtension(E_Embed)
+	proto.RegisterExtension(E_Customtype)
+	proto.RegisterExtension(E_Customname)
+	proto.RegisterExtension(E_Jsontag)
+	proto.RegisterExtension(E_Moretags)
+	proto.RegisterExtension(E_Casttype)
+	proto.RegisterExtension(E_Castkey)
+	proto.RegisterExtension(E_Castvalue)
+	proto.RegisterExtension(E_Stdtime)
+	proto.RegisterExtension(E_Stdduration)
+	proto.RegisterExtension(E_Wktpointer)
+}
+
+func init() { proto.RegisterFile("gogo.proto", fileDescriptor_592445b5231bc2b9) }
+
+var fileDescriptor_592445b5231bc2b9 = []byte{
+	// 1328 bytes of a gzipped FileDescriptorProto
+	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0x94, 0x98, 0x49, 0x6f, 0x1c, 0x45,
+	0x14, 0x80, 0x85, 0x48, 0x64, 0x4f, 0x79, 0x8b, 0xc7, 0xc6, 0x84, 0x08, 0x44, 0xe0, 0xc4, 0xc9,
+	0x3e, 0x45, 0x28, 0x65, 0x45, 0x96, 0x63, 0x39, 0x56, 0x10, 0x0e, 0xc6, 0x89, 0xc3, 0x76, 0x18,
+	0xf5, 0xf4, 0x94, 0xdb, 0x8d, 0xbb, 0xbb, 0x9a, 0xee, 0xea, 0x10, 0xe7, 0x86, 0xc2, 0x22, 0x84,
+	0xd8, 0x91, 0x20, 0x21, 0x09, 0x04, 0xc4, 0xbe, 0x86, 0x7d, 0xb9, 0x70, 0x61, 0xb9, 0xf2, 0x1f,
+	0xb8, 0x00, 0x66, 0xf7, 0xcd, 0x17, 0xf4, 0xba, 0xdf, 0xeb, 0xa9, 0x69, 0x8f, 0x54, 0x35, 0xb7,
+	0xf6, 0xb8, 0xbe, 0x6f, 0xaa, 0xdf, 0xeb, 0x7a, 0xef, 0x4d, 0x33, 0xe6, 0x49, 0x4f, 0x4e, 0xc6,
+	0x89, 0x54, 0xb2, 0x5e, 0x83, 0xeb, 0xfc, 0x72, 0xdf, 0x7e, 0x4f, 0x4a, 0x2f, 0x10, 0x53, 0xf9,
+	0x5f, 0xcd, 0x6c, 0x75, 0xaa, 0x25, 0x52, 0x37, 0xf1, 0x63, 0x25, 0x93, 0x62, 0x31, 0x3f, 0xc6,
+	0xc6, 0x70, 0x71, 0x43, 0x44, 0x59, 0xd8, 0x88, 0x13, 0xb1, 0xea, 0x9f, 0xae, 0x5f, 0x3f, 0x59,
+	0x90, 0x93, 0x44, 0x4e, 0xce, 0x47, 0x59, 0x78, 0x47, 0xac, 0x7c, 0x19, 0xa5, 0x7b, 0xaf, 0xfc,
+	0x72, 0xf5, 0xfe, 0xab, 0x6e, 0xe9, 0x5f, 0x1e, 0x45, 0x14, 0xfe, 0xb7, 0x94, 0x83, 0x7c, 0x99,
+	0x5d, 0xd3, 0xe1, 0x4b, 0x55, 0xe2, 0x47, 0x9e, 0x48, 0x0c, 0xc6, 0xef, 0xd1, 0x38, 0xa6, 0x19,
+	0x8f, 0x23, 0xca, 0xe7, 0xd8, 0x50, 0x2f, 0xae, 0x1f, 0xd0, 0x35, 0x28, 0x74, 0xc9, 0x02, 0x1b,
+	0xc9, 0x25, 0x6e, 0x96, 0x2a, 0x19, 0x46, 0x4e, 0x28, 0x0c, 0x9a, 0x1f, 0x73, 0x4d, 0x6d, 0x79,
+	0x18, 0xb0, 0xb9, 0x92, 0xe2, 0x9c, 0xf5, 0xc3, 0x27, 0x2d, 0xe1, 0x06, 0x06, 0xc3, 0x4f, 0xb8,
+	0x91, 0x72, 0x3d, 0x3f, 0xc9, 0xc6, 0xe1, 0xfa, 0x94, 0x13, 0x64, 0x42, 0xdf, 0xc9, 0x4d, 0x5d,
+	0x3d, 0x27, 0x61, 0x19, 0xc9, 0x7e, 0x3e, 0xbb, 0x2b, 0xdf, 0xce, 0x58, 0x29, 0xd0, 0xf6, 0xa4,
+	0x65, 0xd1, 0x13, 0x4a, 0x89, 0x24, 0x6d, 0x38, 0x41, 0xb7, 0xed, 0x1d, 0xf1, 0x83, 0xd2, 0x78,
+	0x6e, 0xb3, 0x33, 0x8b, 0x0b, 0x05, 0x39, 0x1b, 0x04, 0x7c, 0x85, 0x5d, 0xdb, 0xe5, 0xa9, 0xb0,
+	0x70, 0x9e, 0x47, 0xe7, 0xf8, 0x8e, 0x27, 0x03, 0xb4, 0x4b, 0x8c, 0x3e, 0x2f, 0x73, 0x69, 0xe1,
+	0x7c, 0x19, 0x9d, 0x75, 0x64, 0x29, 0xa5, 0x60, 0xbc, 0x8d, 0x8d, 0x9e, 0x12, 0x49, 0x53, 0xa6,
+	0xa2, 0x21, 0x1e, 0xc8, 0x9c, 0xc0, 0x42, 0x77, 0x01, 0x75, 0x23, 0x08, 0xce, 0x03, 0x07, 0xae,
+	0x83, 0xac, 0x7f, 0xd5, 0x71, 0x85, 0x85, 0xe2, 0x22, 0x2a, 0xfa, 0x60, 0x3d, 0xa0, 0xb3, 0x6c,
+	0xd0, 0x93, 0xc5, 0x2d, 0x59, 0xe0, 0x97, 0x10, 0x1f, 0x20, 0x06, 0x15, 0xb1, 0x8c, 0xb3, 0xc0,
+	0x51, 0x36, 0x3b, 0x78, 0x85, 0x14, 0xc4, 0xa0, 0xa2, 0x87, 0xb0, 0xbe, 0x4a, 0x8a, 0x54, 0x8b,
+	0xe7, 0x0c, 0x1b, 0x90, 0x51, 0xb0, 0x21, 0x23, 0x9b, 0x4d, 0x5c, 0x46, 0x03, 0x43, 0x04, 0x04,
+	0xd3, 0xac, 0x66, 0x9b, 0x88, 0x37, 0x36, 0xe9, 0x78, 0x50, 0x06, 0x16, 0xd8, 0x08, 0x15, 0x28,
+	0x5f, 0x46, 0x16, 0x8a, 0x37, 0x51, 0x31, 0xac, 0x61, 0x78, 0x1b, 0x4a, 0xa4, 0xca, 0x13, 0x36,
+	0x92, 0xb7, 0xe8, 0x36, 0x10, 0xc1, 0x50, 0x36, 0x45, 0xe4, 0xae, 0xd9, 0x19, 0xde, 0xa6, 0x50,
+	0x12, 0x03, 0x8a, 0x39, 0x36, 0x14, 0x3a, 0x49, 0xba, 0xe6, 0x04, 0x56, 0xe9, 0x78, 0x07, 0x1d,
+	0x83, 0x25, 0x84, 0x11, 0xc9, 0xa2, 0x5e, 0x34, 0xef, 0x52, 0x44, 0x34, 0x0c, 0x8f, 0x5e, 0xaa,
+	0x9c, 0x66, 0x20, 0x1a, 0xbd, 0xd8, 0xde, 0xa3, 0xa3, 0x57, 0xb0, 0x8b, 0xba, 0x71, 0x9a, 0xd5,
+	0x52, 0xff, 0x8c, 0x95, 0xe6, 0x7d, 0xca, 0x74, 0x0e, 0x00, 0x7c, 0x0f, 0xbb, 0xae, 0x6b, 0x9b,
+	0xb0, 0x90, 0x7d, 0x80, 0xb2, 0x89, 0x2e, 0xad, 0x02, 0x4b, 0x42, 0xaf, 0xca, 0x0f, 0xa9, 0x24,
+	0x88, 0x8a, 0x6b, 0x89, 0x8d, 0x67, 0x51, 0xea, 0xac, 0xf6, 0x16, 0xb5, 0x8f, 0x28, 0x6a, 0x05,
+	0xdb, 0x11, 0xb5, 0x13, 0x6c, 0x02, 0x8d, 0xbd, 0xe5, 0xf5, 0x63, 0x2a, 0xac, 0x05, 0xbd, 0xd2,
+	0x99, 0xdd, 0xfb, 0xd8, 0xbe, 0x32, 0x9c, 0xa7, 0x95, 0x88, 0x52, 0x60, 0x1a, 0xa1, 0x13, 0x5b,
+	0x98, 0xaf, 0xa0, 0x99, 0x2a, 0xfe, 0x7c, 0x29, 0x58, 0x74, 0x62, 0x90, 0xdf, 0xcd, 0xf6, 0x92,
+	0x3c, 0x8b, 0x12, 0xe1, 0x4a, 0x2f, 0xf2, 0xcf, 0x88, 0x96, 0x85, 0xfa, 0x93, 0x4a, 0xaa, 0x56,
+	0x34, 0x1c, 0xcc, 0x47, 0xd9, 0x9e, 0x72, 0x56, 0x69, 0xf8, 0x61, 0x2c, 0x13, 0x65, 0x30, 0x7e,
+	0x4a, 0x99, 0x2a, 0xb9, 0xa3, 0x39, 0xc6, 0xe7, 0xd9, 0x70, 0xfe, 0xa7, 0xed, 0x23, 0xf9, 0x19,
+	0x8a, 0x86, 0xda, 0x14, 0x16, 0x0e, 0x57, 0x86, 0xb1, 0x93, 0xd8, 0xd4, 0xbf, 0xcf, 0xa9, 0x70,
+	0x20, 0x82, 0x85, 0x43, 0x6d, 0xc4, 0x02, 0xba, 0xbd, 0x85, 0xe1, 0x0b, 0x2a, 0x1c, 0xc4, 0xa0,
+	0x82, 0x06, 0x06, 0x0b, 0xc5, 0x97, 0xa4, 0x20, 0x06, 0x14, 0x77, 0xb6, 0x1b, 0x6d, 0x22, 0x3c,
+	0x3f, 0x55, 0x89, 0x03, 0xab, 0x0d, 0xaa, 0xaf, 0x36, 0x3b, 0x87, 0xb0, 0x65, 0x0d, 0x85, 0x4a,
+	0x14, 0x8a, 0x34, 0x75, 0x3c, 0x01, 0x13, 0x87, 0xc5, 0xc6, 0xbe, 0xa6, 0x4a, 0xa4, 0x61, 0xb0,
+	0x37, 0x6d, 0x42, 0x84, 0xb0, 0xbb, 0x8e, 0xbb, 0x66, 0xa3, 0xfb, 0xa6, 0xb2, 0xb9, 0xe3, 0xc4,
+	0x82, 0x53, 0x9b, 0x7f, 0xb2, 0x68, 0x5d, 0x6c, 0x58, 0x3d, 0x9d, 0xdf, 0x56, 0xe6, 0x9f, 0x95,
+	0x82, 0x2c, 0x6a, 0xc8, 0x48, 0x65, 0x9e, 0xaa, 0xdf, 0xb8, 0xc3, 0xb5, 0x58, 0xdc, 0x17, 0xe9,
+	0x1e, 0xda, 0xc2, 0xfb, 0xed, 0x1c, 0xa7, 0xf8, 0xed, 0xf0, 0x90, 0x77, 0x0e, 0x3d, 0x66, 0xd9,
+	0xd9, 0xad, 0xf2, 0x39, 0xef, 0x98, 0x79, 0xf8, 0x11, 0x36, 0xd4, 0x31, 0xf0, 0x98, 0x55, 0x0f,
+	0xa3, 0x6a, 0x50, 0x9f, 0x77, 0xf8, 0x01, 0xb6, 0x0b, 0x86, 0x17, 0x33, 0xfe, 0x08, 0xe2, 0xf9,
+	0x72, 0x7e, 0x88, 0xf5, 0xd3, 0xd0, 0x62, 0x46, 0x1f, 0x45, 0xb4, 0x44, 0x00, 0xa7, 0x81, 0xc5,
+	0x8c, 0x3f, 0x46, 0x38, 0x21, 0x80, 0xdb, 0x87, 0xf0, 0xbb, 0x27, 0x76, 0x61, 0xd3, 0xa1, 0xd8,
+	0x4d, 0xb3, 0x3e, 0x9c, 0x54, 0xcc, 0xf4, 0xe3, 0xf8, 0xe5, 0x44, 0xf0, 0x5b, 0xd9, 0x6e, 0xcb,
+	0x80, 0x3f, 0x89, 0x68, 0xb1, 0x9e, 0xcf, 0xb1, 0x01, 0x6d, 0x3a, 0x31, 0xe3, 0x4f, 0x21, 0xae,
+	0x53, 0xb0, 0x75, 0x9c, 0x4e, 0xcc, 0x82, 0xa7, 0x69, 0xeb, 0x48, 0x40, 0xd8, 0x68, 0x30, 0x31,
+	0xd3, 0xcf, 0x50, 0xd4, 0x09, 0xe1, 0x33, 0xac, 0x56, 0x36, 0x1b, 0x33, 0xff, 0x2c, 0xf2, 0x6d,
+	0x06, 0x22, 0xa0, 0x35, 0x3b, 0xb3, 0xe2, 0x39, 0x8a, 0x80, 0x46, 0xc1, 0x31, 0xaa, 0x0e, 0x30,
+	0x66, 0xd3, 0xf3, 0x74, 0x8c, 0x2a, 0xf3, 0x0b, 0x64, 0x33, 0xaf, 0xf9, 0x66, 0xc5, 0x0b, 0x94,
+	0xcd, 0x7c, 0x3d, 0x6c, 0xa3, 0x3a, 0x11, 0x98, 0x1d, 0x2f, 0xd2, 0x36, 0x2a, 0x03, 0x01, 0x5f,
+	0x62, 0xf5, 0x9d, 0xd3, 0x80, 0xd9, 0xf7, 0x12, 0xfa, 0x46, 0x77, 0x0c, 0x03, 0xfc, 0x2e, 0x36,
+	0xd1, 0x7d, 0x12, 0x30, 0x5b, 0xcf, 0x6d, 0x55, 0x7e, 0xbb, 0xe9, 0x83, 0x00, 0x3f, 0xd1, 0x6e,
+	0x29, 0xfa, 0x14, 0x60, 0xd6, 0x9e, 0xdf, 0xea, 0x2c, 0xdc, 0xfa, 0x10, 0xc0, 0x67, 0x19, 0x6b,
+	0x37, 0x60, 0xb3, 0xeb, 0x02, 0xba, 0x34, 0x08, 0x8e, 0x06, 0xf6, 0x5f, 0x33, 0x7f, 0x91, 0x8e,
+	0x06, 0x12, 0x70, 0x34, 0xa8, 0xf5, 0x9a, 0xe9, 0x4b, 0x74, 0x34, 0x08, 0x81, 0x27, 0x5b, 0xeb,
+	0x6e, 0x66, 0xc3, 0x65, 0x7a, 0xb2, 0x35, 0x8a, 0x1f, 0x63, 0xa3, 0x3b, 0x1a, 0xa2, 0x59, 0xf5,
+	0x1a, 0xaa, 0xf6, 0x54, 0xfb, 0xa1, 0xde, 0xbc, 0xb0, 0x19, 0x9a, 0x6d, 0xaf, 0x57, 0x9a, 0x17,
+	0xf6, 0x42, 0x3e, 0xcd, 0xfa, 0xa3, 0x2c, 0x08, 0xe0, 0xf0, 0xd4, 0x6f, 0xe8, 0xd2, 0x4d, 0x45,
+	0xd0, 0x22, 0xc5, 0xaf, 0xdb, 0x18, 0x1d, 0x02, 0xf8, 0x01, 0xb6, 0x5b, 0x84, 0x4d, 0xd1, 0x32,
+	0x91, 0xbf, 0x6d, 0x53, 0xc1, 0x84, 0xd5, 0x7c, 0x86, 0xb1, 0xe2, 0xd5, 0x08, 0x84, 0xd9, 0xc4,
+	0xfe, 0xbe, 0x5d, 0xbc, 0xa5, 0xd1, 0x90, 0xb6, 0x20, 0x4f, 0x8a, 0x41, 0xb0, 0xd9, 0x29, 0xc8,
+	0x33, 0x72, 0x90, 0xf5, 0xdd, 0x9f, 0xca, 0x48, 0x39, 0x9e, 0x89, 0xfe, 0x03, 0x69, 0x5a, 0x0f,
+	0x01, 0x0b, 0x65, 0x22, 0x94, 0xe3, 0xa5, 0x26, 0xf6, 0x4f, 0x64, 0x4b, 0x00, 0x60, 0xd7, 0x49,
+	0x95, 0xcd, 0x7d, 0xff, 0x45, 0x30, 0x01, 0xb0, 0x69, 0xb8, 0x5e, 0x17, 0x1b, 0x26, 0xf6, 0x6f,
+	0xda, 0x34, 0xae, 0xe7, 0x87, 0x58, 0x0d, 0x2e, 0xf3, 0xb7, 0x4a, 0x26, 0xf8, 0x1f, 0x84, 0xdb,
+	0x04, 0x7c, 0x73, 0xaa, 0x5a, 0xca, 0x37, 0x07, 0xfb, 0x5f, 0xcc, 0x34, 0xad, 0xe7, 0xb3, 0x6c,
+	0x20, 0x55, 0xad, 0x56, 0x86, 0xf3, 0xa9, 0x01, 0xff, 0x6f, 0xbb, 0x7c, 0x65, 0x51, 0x32, 0x90,
+	0xed, 0x07, 0xd7, 0x55, 0x2c, 0xfd, 0x48, 0x89, 0xc4, 0x64, 0xd8, 0x42, 0x83, 0x86, 0x1c, 0x9e,
+	0x67, 0x63, 0xae, 0x0c, 0xab, 0xdc, 0x61, 0xb6, 0x20, 0x17, 0xe4, 0x52, 0x5e, 0x67, 0xee, 0xbd,
+	0xd9, 0xf3, 0xd5, 0x5a, 0xd6, 0x9c, 0x74, 0x65, 0x38, 0x05, 0xbf, 0x3c, 0xda, 0x2f, 0x54, 0xcb,
+	0xdf, 0x21, 0xff, 0x07, 0x00, 0x00, 0xff, 0xff, 0x9c, 0xaf, 0x70, 0x4e, 0x83, 0x15, 0x00, 0x00,
+}
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.golden b/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.golden
new file mode 100644
index 0000000..f6502e4
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/gogo.pb.golden
@@ -0,0 +1,45 @@
+// Code generated by protoc-gen-go.
+// source: gogo.proto
+// DO NOT EDIT!
+
+package gogoproto
+
+import proto "github.com/gogo/protobuf/proto"
+import json "encoding/json"
+import math "math"
+import google_protobuf "github.com/gogo/protobuf/protoc-gen-gogo/descriptor"
+
+// Reference proto, json, and math imports to suppress error if they are not otherwise used.
+var _ = proto.Marshal
+var _ = &json.SyntaxError{}
+var _ = math.Inf
+
+var E_Nullable = &proto.ExtensionDesc{
+	ExtendedType:  (*google_protobuf.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         51235,
+	Name:          "gogoproto.nullable",
+	Tag:           "varint,51235,opt,name=nullable",
+}
+
+var E_Embed = &proto.ExtensionDesc{
+	ExtendedType:  (*google_protobuf.FieldOptions)(nil),
+	ExtensionType: (*bool)(nil),
+	Field:         51236,
+	Name:          "gogoproto.embed",
+	Tag:           "varint,51236,opt,name=embed",
+}
+
+var E_Customtype = &proto.ExtensionDesc{
+	ExtendedType:  (*google_protobuf.FieldOptions)(nil),
+	ExtensionType: (*string)(nil),
+	Field:         51237,
+	Name:          "gogoproto.customtype",
+	Tag:           "bytes,51237,opt,name=customtype",
+}
+
+func init() {
+	proto.RegisterExtension(E_Nullable)
+	proto.RegisterExtension(E_Embed)
+	proto.RegisterExtension(E_Customtype)
+}
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/gogo.proto b/vendor/github.com/gogo/protobuf/gogoproto/gogo.proto
new file mode 100644
index 0000000..b80c856
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/gogo.proto
@@ -0,0 +1,144 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+package gogoproto;
+
+import "google/protobuf/descriptor.proto";
+
+option java_package = "com.google.protobuf";
+option java_outer_classname = "GoGoProtos";
+option go_package = "github.com/gogo/protobuf/gogoproto";
+
+extend google.protobuf.EnumOptions {
+	optional bool goproto_enum_prefix = 62001;
+	optional bool goproto_enum_stringer = 62021;
+	optional bool enum_stringer = 62022;
+	optional string enum_customname = 62023;
+	optional bool enumdecl = 62024;
+}
+
+extend google.protobuf.EnumValueOptions {
+	optional string enumvalue_customname = 66001;
+}
+
+extend google.protobuf.FileOptions {
+	optional bool goproto_getters_all = 63001;
+	optional bool goproto_enum_prefix_all = 63002;
+	optional bool goproto_stringer_all = 63003;
+	optional bool verbose_equal_all = 63004;
+	optional bool face_all = 63005;
+	optional bool gostring_all = 63006;
+	optional bool populate_all = 63007;
+	optional bool stringer_all = 63008;
+	optional bool onlyone_all = 63009;
+
+	optional bool equal_all = 63013;
+	optional bool description_all = 63014;
+	optional bool testgen_all = 63015;
+	optional bool benchgen_all = 63016;
+	optional bool marshaler_all = 63017;
+	optional bool unmarshaler_all = 63018;
+	optional bool stable_marshaler_all = 63019;
+
+	optional bool sizer_all = 63020;
+
+	optional bool goproto_enum_stringer_all = 63021;
+	optional bool enum_stringer_all = 63022;
+
+	optional bool unsafe_marshaler_all = 63023;
+	optional bool unsafe_unmarshaler_all = 63024;
+
+	optional bool goproto_extensions_map_all = 63025;
+	optional bool goproto_unrecognized_all = 63026;
+	optional bool gogoproto_import = 63027;
+	optional bool protosizer_all = 63028;
+	optional bool compare_all = 63029;
+    optional bool typedecl_all = 63030;
+    optional bool enumdecl_all = 63031;
+
+	optional bool goproto_registration = 63032;
+	optional bool messagename_all = 63033;
+
+	optional bool goproto_sizecache_all = 63034;
+	optional bool goproto_unkeyed_all = 63035;
+}
+
+extend google.protobuf.MessageOptions {
+	optional bool goproto_getters = 64001;
+	optional bool goproto_stringer = 64003;
+	optional bool verbose_equal = 64004;
+	optional bool face = 64005;
+	optional bool gostring = 64006;
+	optional bool populate = 64007;
+	optional bool stringer = 67008;
+	optional bool onlyone = 64009;
+
+	optional bool equal = 64013;
+	optional bool description = 64014;
+	optional bool testgen = 64015;
+	optional bool benchgen = 64016;
+	optional bool marshaler = 64017;
+	optional bool unmarshaler = 64018;
+	optional bool stable_marshaler = 64019;
+
+	optional bool sizer = 64020;
+
+	optional bool unsafe_marshaler = 64023;
+	optional bool unsafe_unmarshaler = 64024;
+
+	optional bool goproto_extensions_map = 64025;
+	optional bool goproto_unrecognized = 64026;
+
+	optional bool protosizer = 64028;
+	optional bool compare = 64029;
+
+	optional bool typedecl = 64030;
+
+	optional bool messagename = 64033;
+
+	optional bool goproto_sizecache = 64034;
+	optional bool goproto_unkeyed = 64035;
+}
+
+extend google.protobuf.FieldOptions {
+	optional bool nullable = 65001;
+	optional bool embed = 65002;
+	optional string customtype = 65003;
+	optional string customname = 65004;
+	optional string jsontag = 65005;
+	optional string moretags = 65006;
+	optional string casttype = 65007;
+	optional string castkey = 65008;
+	optional string castvalue = 65009;
+
+	optional bool stdtime = 65010;
+	optional bool stdduration = 65011;
+	optional bool wktpointer = 65012;
+
+}
diff --git a/vendor/github.com/gogo/protobuf/gogoproto/helper.go b/vendor/github.com/gogo/protobuf/gogoproto/helper.go
new file mode 100644
index 0000000..390d4e4
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/gogoproto/helper.go
@@ -0,0 +1,415 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package gogoproto
+
+import google_protobuf "github.com/gogo/protobuf/protoc-gen-gogo/descriptor"
+import proto "github.com/gogo/protobuf/proto"
+
+func IsEmbed(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Embed, false)
+}
+
+func IsNullable(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Nullable, true)
+}
+
+func IsStdTime(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Stdtime, false)
+}
+
+func IsStdDuration(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Stdduration, false)
+}
+
+func IsStdDouble(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.DoubleValue"
+}
+
+func IsStdFloat(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.FloatValue"
+}
+
+func IsStdInt64(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.Int64Value"
+}
+
+func IsStdUInt64(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.UInt64Value"
+}
+
+func IsStdInt32(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.Int32Value"
+}
+
+func IsStdUInt32(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.UInt32Value"
+}
+
+func IsStdBool(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.BoolValue"
+}
+
+func IsStdString(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.StringValue"
+}
+
+func IsStdBytes(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false) && *field.TypeName == ".google.protobuf.BytesValue"
+}
+
+func IsStdType(field *google_protobuf.FieldDescriptorProto) bool {
+	return (IsStdTime(field) || IsStdDuration(field) ||
+		IsStdDouble(field) || IsStdFloat(field) ||
+		IsStdInt64(field) || IsStdUInt64(field) ||
+		IsStdInt32(field) || IsStdUInt32(field) ||
+		IsStdBool(field) ||
+		IsStdString(field) || IsStdBytes(field))
+}
+
+func IsWktPtr(field *google_protobuf.FieldDescriptorProto) bool {
+	return proto.GetBoolExtension(field.Options, E_Wktpointer, false)
+}
+
+func NeedsNilCheck(proto3 bool, field *google_protobuf.FieldDescriptorProto) bool {
+	nullable := IsNullable(field)
+	if field.IsMessage() || IsCustomType(field) {
+		return nullable
+	}
+	if proto3 {
+		return false
+	}
+	return nullable || *field.Type == google_protobuf.FieldDescriptorProto_TYPE_BYTES
+}
+
+func IsCustomType(field *google_protobuf.FieldDescriptorProto) bool {
+	typ := GetCustomType(field)
+	if len(typ) > 0 {
+		return true
+	}
+	return false
+}
+
+func IsCastType(field *google_protobuf.FieldDescriptorProto) bool {
+	typ := GetCastType(field)
+	if len(typ) > 0 {
+		return true
+	}
+	return false
+}
+
+func IsCastKey(field *google_protobuf.FieldDescriptorProto) bool {
+	typ := GetCastKey(field)
+	if len(typ) > 0 {
+		return true
+	}
+	return false
+}
+
+func IsCastValue(field *google_protobuf.FieldDescriptorProto) bool {
+	typ := GetCastValue(field)
+	if len(typ) > 0 {
+		return true
+	}
+	return false
+}
+
+func HasEnumDecl(file *google_protobuf.FileDescriptorProto, enum *google_protobuf.EnumDescriptorProto) bool {
+	return proto.GetBoolExtension(enum.Options, E_Enumdecl, proto.GetBoolExtension(file.Options, E_EnumdeclAll, true))
+}
+
+func HasTypeDecl(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Typedecl, proto.GetBoolExtension(file.Options, E_TypedeclAll, true))
+}
+
+func GetCustomType(field *google_protobuf.FieldDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Customtype)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetCastType(field *google_protobuf.FieldDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Casttype)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetCastKey(field *google_protobuf.FieldDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Castkey)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetCastValue(field *google_protobuf.FieldDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Castvalue)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func IsCustomName(field *google_protobuf.FieldDescriptorProto) bool {
+	name := GetCustomName(field)
+	if len(name) > 0 {
+		return true
+	}
+	return false
+}
+
+func IsEnumCustomName(field *google_protobuf.EnumDescriptorProto) bool {
+	name := GetEnumCustomName(field)
+	if len(name) > 0 {
+		return true
+	}
+	return false
+}
+
+func IsEnumValueCustomName(field *google_protobuf.EnumValueDescriptorProto) bool {
+	name := GetEnumValueCustomName(field)
+	if len(name) > 0 {
+		return true
+	}
+	return false
+}
+
+func GetCustomName(field *google_protobuf.FieldDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Customname)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetEnumCustomName(field *google_protobuf.EnumDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_EnumCustomname)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetEnumValueCustomName(field *google_protobuf.EnumValueDescriptorProto) string {
+	if field == nil {
+		return ""
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_EnumvalueCustomname)
+		if err == nil && v.(*string) != nil {
+			return *(v.(*string))
+		}
+	}
+	return ""
+}
+
+func GetJsonTag(field *google_protobuf.FieldDescriptorProto) *string {
+	if field == nil {
+		return nil
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Jsontag)
+		if err == nil && v.(*string) != nil {
+			return (v.(*string))
+		}
+	}
+	return nil
+}
+
+func GetMoreTags(field *google_protobuf.FieldDescriptorProto) *string {
+	if field == nil {
+		return nil
+	}
+	if field.Options != nil {
+		v, err := proto.GetExtension(field.Options, E_Moretags)
+		if err == nil && v.(*string) != nil {
+			return (v.(*string))
+		}
+	}
+	return nil
+}
+
+type EnableFunc func(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool
+
+func EnabledGoEnumPrefix(file *google_protobuf.FileDescriptorProto, enum *google_protobuf.EnumDescriptorProto) bool {
+	return proto.GetBoolExtension(enum.Options, E_GoprotoEnumPrefix, proto.GetBoolExtension(file.Options, E_GoprotoEnumPrefixAll, true))
+}
+
+func EnabledGoStringer(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoStringer, proto.GetBoolExtension(file.Options, E_GoprotoStringerAll, true))
+}
+
+func HasGoGetters(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoGetters, proto.GetBoolExtension(file.Options, E_GoprotoGettersAll, true))
+}
+
+func IsUnion(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Onlyone, proto.GetBoolExtension(file.Options, E_OnlyoneAll, false))
+}
+
+func HasGoString(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Gostring, proto.GetBoolExtension(file.Options, E_GostringAll, false))
+}
+
+func HasEqual(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Equal, proto.GetBoolExtension(file.Options, E_EqualAll, false))
+}
+
+func HasVerboseEqual(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_VerboseEqual, proto.GetBoolExtension(file.Options, E_VerboseEqualAll, false))
+}
+
+func IsStringer(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Stringer, proto.GetBoolExtension(file.Options, E_StringerAll, false))
+}
+
+func IsFace(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Face, proto.GetBoolExtension(file.Options, E_FaceAll, false))
+}
+
+func HasDescription(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Description, proto.GetBoolExtension(file.Options, E_DescriptionAll, false))
+}
+
+func HasPopulate(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Populate, proto.GetBoolExtension(file.Options, E_PopulateAll, false))
+}
+
+func HasTestGen(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Testgen, proto.GetBoolExtension(file.Options, E_TestgenAll, false))
+}
+
+func HasBenchGen(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Benchgen, proto.GetBoolExtension(file.Options, E_BenchgenAll, false))
+}
+
+func IsMarshaler(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Marshaler, proto.GetBoolExtension(file.Options, E_MarshalerAll, false))
+}
+
+func IsUnmarshaler(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Unmarshaler, proto.GetBoolExtension(file.Options, E_UnmarshalerAll, false))
+}
+
+func IsStableMarshaler(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_StableMarshaler, proto.GetBoolExtension(file.Options, E_StableMarshalerAll, false))
+}
+
+func IsSizer(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Sizer, proto.GetBoolExtension(file.Options, E_SizerAll, false))
+}
+
+func IsProtoSizer(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Protosizer, proto.GetBoolExtension(file.Options, E_ProtosizerAll, false))
+}
+
+func IsGoEnumStringer(file *google_protobuf.FileDescriptorProto, enum *google_protobuf.EnumDescriptorProto) bool {
+	return proto.GetBoolExtension(enum.Options, E_GoprotoEnumStringer, proto.GetBoolExtension(file.Options, E_GoprotoEnumStringerAll, true))
+}
+
+func IsEnumStringer(file *google_protobuf.FileDescriptorProto, enum *google_protobuf.EnumDescriptorProto) bool {
+	return proto.GetBoolExtension(enum.Options, E_EnumStringer, proto.GetBoolExtension(file.Options, E_EnumStringerAll, false))
+}
+
+func IsUnsafeMarshaler(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_UnsafeMarshaler, proto.GetBoolExtension(file.Options, E_UnsafeMarshalerAll, false))
+}
+
+func IsUnsafeUnmarshaler(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_UnsafeUnmarshaler, proto.GetBoolExtension(file.Options, E_UnsafeUnmarshalerAll, false))
+}
+
+func HasExtensionsMap(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoExtensionsMap, proto.GetBoolExtension(file.Options, E_GoprotoExtensionsMapAll, true))
+}
+
+func HasUnrecognized(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoUnrecognized, proto.GetBoolExtension(file.Options, E_GoprotoUnrecognizedAll, true))
+}
+
+func IsProto3(file *google_protobuf.FileDescriptorProto) bool {
+	return file.GetSyntax() == "proto3"
+}
+
+func ImportsGoGoProto(file *google_protobuf.FileDescriptorProto) bool {
+	return proto.GetBoolExtension(file.Options, E_GogoprotoImport, true)
+}
+
+func HasCompare(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Compare, proto.GetBoolExtension(file.Options, E_CompareAll, false))
+}
+
+func RegistersGolangProto(file *google_protobuf.FileDescriptorProto) bool {
+	return proto.GetBoolExtension(file.Options, E_GoprotoRegistration, false)
+}
+
+func HasMessageName(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_Messagename, proto.GetBoolExtension(file.Options, E_MessagenameAll, false))
+}
+
+func HasSizecache(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoSizecache, proto.GetBoolExtension(file.Options, E_GoprotoSizecacheAll, true))
+}
+
+func HasUnkeyed(file *google_protobuf.FileDescriptorProto, message *google_protobuf.DescriptorProto) bool {
+	return proto.GetBoolExtension(message.Options, E_GoprotoUnkeyed, proto.GetBoolExtension(file.Options, E_GoprotoUnkeyedAll, true))
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/Makefile b/vendor/github.com/gogo/protobuf/proto/Makefile
new file mode 100644
index 0000000..00d65f3
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/Makefile
@@ -0,0 +1,43 @@
+# Go support for Protocol Buffers - Google's data interchange format
+#
+# Copyright 2010 The Go Authors.  All rights reserved.
+# https://github.com/golang/protobuf
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+install:
+	go install
+
+test: install generate-test-pbs
+	go test
+
+
+generate-test-pbs:
+	make install
+	make -C test_proto
+	make -C proto3_proto
+	make
diff --git a/vendor/github.com/gogo/protobuf/proto/clone.go b/vendor/github.com/gogo/protobuf/proto/clone.go
new file mode 100644
index 0000000..a26b046
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/clone.go
@@ -0,0 +1,258 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2011 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Protocol buffer deep copy and merge.
+// TODO: RawMessage.
+
+package proto
+
+import (
+	"fmt"
+	"log"
+	"reflect"
+	"strings"
+)
+
+// Clone returns a deep copy of a protocol buffer.
+func Clone(src Message) Message {
+	in := reflect.ValueOf(src)
+	if in.IsNil() {
+		return src
+	}
+	out := reflect.New(in.Type().Elem())
+	dst := out.Interface().(Message)
+	Merge(dst, src)
+	return dst
+}
+
+// Merger is the interface representing objects that can merge messages of the same type.
+type Merger interface {
+	// Merge merges src into this message.
+	// Required and optional fields that are set in src will be set to that value in dst.
+	// Elements of repeated fields will be appended.
+	//
+	// Merge may panic if called with a different argument type than the receiver.
+	Merge(src Message)
+}
+
+// generatedMerger is the custom merge method that generated protos will have.
+// We must add this method since a generate Merge method will conflict with
+// many existing protos that have a Merge data field already defined.
+type generatedMerger interface {
+	XXX_Merge(src Message)
+}
+
+// Merge merges src into dst.
+// Required and optional fields that are set in src will be set to that value in dst.
+// Elements of repeated fields will be appended.
+// Merge panics if src and dst are not the same type, or if dst is nil.
+func Merge(dst, src Message) {
+	if m, ok := dst.(Merger); ok {
+		m.Merge(src)
+		return
+	}
+
+	in := reflect.ValueOf(src)
+	out := reflect.ValueOf(dst)
+	if out.IsNil() {
+		panic("proto: nil destination")
+	}
+	if in.Type() != out.Type() {
+		panic(fmt.Sprintf("proto.Merge(%T, %T) type mismatch", dst, src))
+	}
+	if in.IsNil() {
+		return // Merge from nil src is a noop
+	}
+	if m, ok := dst.(generatedMerger); ok {
+		m.XXX_Merge(src)
+		return
+	}
+	mergeStruct(out.Elem(), in.Elem())
+}
+
+func mergeStruct(out, in reflect.Value) {
+	sprop := GetProperties(in.Type())
+	for i := 0; i < in.NumField(); i++ {
+		f := in.Type().Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		mergeAny(out.Field(i), in.Field(i), false, sprop.Prop[i])
+	}
+
+	if emIn, ok := in.Addr().Interface().(extensionsBytes); ok {
+		emOut := out.Addr().Interface().(extensionsBytes)
+		bIn := emIn.GetExtensions()
+		bOut := emOut.GetExtensions()
+		*bOut = append(*bOut, *bIn...)
+	} else if emIn, err := extendable(in.Addr().Interface()); err == nil {
+		emOut, _ := extendable(out.Addr().Interface())
+		mIn, muIn := emIn.extensionsRead()
+		if mIn != nil {
+			mOut := emOut.extensionsWrite()
+			muIn.Lock()
+			mergeExtension(mOut, mIn)
+			muIn.Unlock()
+		}
+	}
+
+	uf := in.FieldByName("XXX_unrecognized")
+	if !uf.IsValid() {
+		return
+	}
+	uin := uf.Bytes()
+	if len(uin) > 0 {
+		out.FieldByName("XXX_unrecognized").SetBytes(append([]byte(nil), uin...))
+	}
+}
+
+// mergeAny performs a merge between two values of the same type.
+// viaPtr indicates whether the values were indirected through a pointer (implying proto2).
+// prop is set if this is a struct field (it may be nil).
+func mergeAny(out, in reflect.Value, viaPtr bool, prop *Properties) {
+	if in.Type() == protoMessageType {
+		if !in.IsNil() {
+			if out.IsNil() {
+				out.Set(reflect.ValueOf(Clone(in.Interface().(Message))))
+			} else {
+				Merge(out.Interface().(Message), in.Interface().(Message))
+			}
+		}
+		return
+	}
+	switch in.Kind() {
+	case reflect.Bool, reflect.Float32, reflect.Float64, reflect.Int32, reflect.Int64,
+		reflect.String, reflect.Uint32, reflect.Uint64:
+		if !viaPtr && isProto3Zero(in) {
+			return
+		}
+		out.Set(in)
+	case reflect.Interface:
+		// Probably a oneof field; copy non-nil values.
+		if in.IsNil() {
+			return
+		}
+		// Allocate destination if it is not set, or set to a different type.
+		// Otherwise we will merge as normal.
+		if out.IsNil() || out.Elem().Type() != in.Elem().Type() {
+			out.Set(reflect.New(in.Elem().Elem().Type())) // interface -> *T -> T -> new(T)
+		}
+		mergeAny(out.Elem(), in.Elem(), false, nil)
+	case reflect.Map:
+		if in.Len() == 0 {
+			return
+		}
+		if out.IsNil() {
+			out.Set(reflect.MakeMap(in.Type()))
+		}
+		// For maps with value types of *T or []byte we need to deep copy each value.
+		elemKind := in.Type().Elem().Kind()
+		for _, key := range in.MapKeys() {
+			var val reflect.Value
+			switch elemKind {
+			case reflect.Ptr:
+				val = reflect.New(in.Type().Elem().Elem())
+				mergeAny(val, in.MapIndex(key), false, nil)
+			case reflect.Slice:
+				val = in.MapIndex(key)
+				val = reflect.ValueOf(append([]byte{}, val.Bytes()...))
+			default:
+				val = in.MapIndex(key)
+			}
+			out.SetMapIndex(key, val)
+		}
+	case reflect.Ptr:
+		if in.IsNil() {
+			return
+		}
+		if out.IsNil() {
+			out.Set(reflect.New(in.Elem().Type()))
+		}
+		mergeAny(out.Elem(), in.Elem(), true, nil)
+	case reflect.Slice:
+		if in.IsNil() {
+			return
+		}
+		if in.Type().Elem().Kind() == reflect.Uint8 {
+			// []byte is a scalar bytes field, not a repeated field.
+
+			// Edge case: if this is in a proto3 message, a zero length
+			// bytes field is considered the zero value, and should not
+			// be merged.
+			if prop != nil && prop.proto3 && in.Len() == 0 {
+				return
+			}
+
+			// Make a deep copy.
+			// Append to []byte{} instead of []byte(nil) so that we never end up
+			// with a nil result.
+			out.SetBytes(append([]byte{}, in.Bytes()...))
+			return
+		}
+		n := in.Len()
+		if out.IsNil() {
+			out.Set(reflect.MakeSlice(in.Type(), 0, n))
+		}
+		switch in.Type().Elem().Kind() {
+		case reflect.Bool, reflect.Float32, reflect.Float64, reflect.Int32, reflect.Int64,
+			reflect.String, reflect.Uint32, reflect.Uint64:
+			out.Set(reflect.AppendSlice(out, in))
+		default:
+			for i := 0; i < n; i++ {
+				x := reflect.Indirect(reflect.New(in.Type().Elem()))
+				mergeAny(x, in.Index(i), false, nil)
+				out.Set(reflect.Append(out, x))
+			}
+		}
+	case reflect.Struct:
+		mergeStruct(out, in)
+	default:
+		// unknown type, so not a protocol buffer
+		log.Printf("proto: don't know how to copy %v", in)
+	}
+}
+
+func mergeExtension(out, in map[int32]Extension) {
+	for extNum, eIn := range in {
+		eOut := Extension{desc: eIn.desc}
+		if eIn.value != nil {
+			v := reflect.New(reflect.TypeOf(eIn.value)).Elem()
+			mergeAny(v, reflect.ValueOf(eIn.value), false, nil)
+			eOut.value = v.Interface()
+		}
+		if eIn.enc != nil {
+			eOut.enc = make([]byte, len(eIn.enc))
+			copy(eOut.enc, eIn.enc)
+		}
+
+		out[extNum] = eOut
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/custom_gogo.go b/vendor/github.com/gogo/protobuf/proto/custom_gogo.go
new file mode 100644
index 0000000..2455248
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/custom_gogo.go
@@ -0,0 +1,39 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import "reflect"
+
+type custom interface {
+	Marshal() ([]byte, error)
+	Unmarshal(data []byte) error
+	Size() int
+}
+
+var customType = reflect.TypeOf((*custom)(nil)).Elem()
diff --git a/vendor/github.com/gogo/protobuf/proto/decode.go b/vendor/github.com/gogo/protobuf/proto/decode.go
new file mode 100644
index 0000000..63b0f08
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/decode.go
@@ -0,0 +1,427 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for decoding protocol buffer data to construct in-memory representations.
+ */
+
+import (
+	"errors"
+	"fmt"
+	"io"
+)
+
+// errOverflow is returned when an integer is too large to be represented.
+var errOverflow = errors.New("proto: integer overflow")
+
+// ErrInternalBadWireType is returned by generated code when an incorrect
+// wire type is encountered. It does not get returned to user code.
+var ErrInternalBadWireType = errors.New("proto: internal error: bad wiretype for oneof")
+
+// DecodeVarint reads a varint-encoded integer from the slice.
+// It returns the integer and the number of bytes consumed, or
+// zero if there is not enough.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func DecodeVarint(buf []byte) (x uint64, n int) {
+	for shift := uint(0); shift < 64; shift += 7 {
+		if n >= len(buf) {
+			return 0, 0
+		}
+		b := uint64(buf[n])
+		n++
+		x |= (b & 0x7F) << shift
+		if (b & 0x80) == 0 {
+			return x, n
+		}
+	}
+
+	// The number is too large to represent in a 64-bit value.
+	return 0, 0
+}
+
+func (p *Buffer) decodeVarintSlow() (x uint64, err error) {
+	i := p.index
+	l := len(p.buf)
+
+	for shift := uint(0); shift < 64; shift += 7 {
+		if i >= l {
+			err = io.ErrUnexpectedEOF
+			return
+		}
+		b := p.buf[i]
+		i++
+		x |= (uint64(b) & 0x7F) << shift
+		if b < 0x80 {
+			p.index = i
+			return
+		}
+	}
+
+	// The number is too large to represent in a 64-bit value.
+	err = errOverflow
+	return
+}
+
+// DecodeVarint reads a varint-encoded integer from the Buffer.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func (p *Buffer) DecodeVarint() (x uint64, err error) {
+	i := p.index
+	buf := p.buf
+
+	if i >= len(buf) {
+		return 0, io.ErrUnexpectedEOF
+	} else if buf[i] < 0x80 {
+		p.index++
+		return uint64(buf[i]), nil
+	} else if len(buf)-i < 10 {
+		return p.decodeVarintSlow()
+	}
+
+	var b uint64
+	// we already checked the first byte
+	x = uint64(buf[i]) - 0x80
+	i++
+
+	b = uint64(buf[i])
+	i++
+	x += b << 7
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 7
+
+	b = uint64(buf[i])
+	i++
+	x += b << 14
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 14
+
+	b = uint64(buf[i])
+	i++
+	x += b << 21
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 21
+
+	b = uint64(buf[i])
+	i++
+	x += b << 28
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 28
+
+	b = uint64(buf[i])
+	i++
+	x += b << 35
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 35
+
+	b = uint64(buf[i])
+	i++
+	x += b << 42
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 42
+
+	b = uint64(buf[i])
+	i++
+	x += b << 49
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 49
+
+	b = uint64(buf[i])
+	i++
+	x += b << 56
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 56
+
+	b = uint64(buf[i])
+	i++
+	x += b << 63
+	if b&0x80 == 0 {
+		goto done
+	}
+
+	return 0, errOverflow
+
+done:
+	p.index = i
+	return x, nil
+}
+
+// DecodeFixed64 reads a 64-bit integer from the Buffer.
+// This is the format for the
+// fixed64, sfixed64, and double protocol buffer types.
+func (p *Buffer) DecodeFixed64() (x uint64, err error) {
+	// x, err already 0
+	i := p.index + 8
+	if i < 0 || i > len(p.buf) {
+		err = io.ErrUnexpectedEOF
+		return
+	}
+	p.index = i
+
+	x = uint64(p.buf[i-8])
+	x |= uint64(p.buf[i-7]) << 8
+	x |= uint64(p.buf[i-6]) << 16
+	x |= uint64(p.buf[i-5]) << 24
+	x |= uint64(p.buf[i-4]) << 32
+	x |= uint64(p.buf[i-3]) << 40
+	x |= uint64(p.buf[i-2]) << 48
+	x |= uint64(p.buf[i-1]) << 56
+	return
+}
+
+// DecodeFixed32 reads a 32-bit integer from the Buffer.
+// This is the format for the
+// fixed32, sfixed32, and float protocol buffer types.
+func (p *Buffer) DecodeFixed32() (x uint64, err error) {
+	// x, err already 0
+	i := p.index + 4
+	if i < 0 || i > len(p.buf) {
+		err = io.ErrUnexpectedEOF
+		return
+	}
+	p.index = i
+
+	x = uint64(p.buf[i-4])
+	x |= uint64(p.buf[i-3]) << 8
+	x |= uint64(p.buf[i-2]) << 16
+	x |= uint64(p.buf[i-1]) << 24
+	return
+}
+
+// DecodeZigzag64 reads a zigzag-encoded 64-bit integer
+// from the Buffer.
+// This is the format used for the sint64 protocol buffer type.
+func (p *Buffer) DecodeZigzag64() (x uint64, err error) {
+	x, err = p.DecodeVarint()
+	if err != nil {
+		return
+	}
+	x = (x >> 1) ^ uint64((int64(x&1)<<63)>>63)
+	return
+}
+
+// DecodeZigzag32 reads a zigzag-encoded 32-bit integer
+// from  the Buffer.
+// This is the format used for the sint32 protocol buffer type.
+func (p *Buffer) DecodeZigzag32() (x uint64, err error) {
+	x, err = p.DecodeVarint()
+	if err != nil {
+		return
+	}
+	x = uint64((uint32(x) >> 1) ^ uint32((int32(x&1)<<31)>>31))
+	return
+}
+
+// DecodeRawBytes reads a count-delimited byte buffer from the Buffer.
+// This is the format used for the bytes protocol buffer
+// type and for embedded messages.
+func (p *Buffer) DecodeRawBytes(alloc bool) (buf []byte, err error) {
+	n, err := p.DecodeVarint()
+	if err != nil {
+		return nil, err
+	}
+
+	nb := int(n)
+	if nb < 0 {
+		return nil, fmt.Errorf("proto: bad byte length %d", nb)
+	}
+	end := p.index + nb
+	if end < p.index || end > len(p.buf) {
+		return nil, io.ErrUnexpectedEOF
+	}
+
+	if !alloc {
+		// todo: check if can get more uses of alloc=false
+		buf = p.buf[p.index:end]
+		p.index += nb
+		return
+	}
+
+	buf = make([]byte, nb)
+	copy(buf, p.buf[p.index:])
+	p.index += nb
+	return
+}
+
+// DecodeStringBytes reads an encoded string from the Buffer.
+// This is the format used for the proto2 string type.
+func (p *Buffer) DecodeStringBytes() (s string, err error) {
+	buf, err := p.DecodeRawBytes(false)
+	if err != nil {
+		return
+	}
+	return string(buf), nil
+}
+
+// Unmarshaler is the interface representing objects that can
+// unmarshal themselves.  The argument points to data that may be
+// overwritten, so implementations should not keep references to the
+// buffer.
+// Unmarshal implementations should not clear the receiver.
+// Any unmarshaled data should be merged into the receiver.
+// Callers of Unmarshal that do not want to retain existing data
+// should Reset the receiver before calling Unmarshal.
+type Unmarshaler interface {
+	Unmarshal([]byte) error
+}
+
+// newUnmarshaler is the interface representing objects that can
+// unmarshal themselves. The semantics are identical to Unmarshaler.
+//
+// This exists to support protoc-gen-go generated messages.
+// The proto package will stop type-asserting to this interface in the future.
+//
+// DO NOT DEPEND ON THIS.
+type newUnmarshaler interface {
+	XXX_Unmarshal([]byte) error
+}
+
+// Unmarshal parses the protocol buffer representation in buf and places the
+// decoded result in pb.  If the struct underlying pb does not match
+// the data in buf, the results can be unpredictable.
+//
+// Unmarshal resets pb before starting to unmarshal, so any
+// existing data in pb is always removed. Use UnmarshalMerge
+// to preserve and append to existing data.
+func Unmarshal(buf []byte, pb Message) error {
+	pb.Reset()
+	if u, ok := pb.(newUnmarshaler); ok {
+		return u.XXX_Unmarshal(buf)
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		return u.Unmarshal(buf)
+	}
+	return NewBuffer(buf).Unmarshal(pb)
+}
+
+// UnmarshalMerge parses the protocol buffer representation in buf and
+// writes the decoded result to pb.  If the struct underlying pb does not match
+// the data in buf, the results can be unpredictable.
+//
+// UnmarshalMerge merges into existing data in pb.
+// Most code should use Unmarshal instead.
+func UnmarshalMerge(buf []byte, pb Message) error {
+	if u, ok := pb.(newUnmarshaler); ok {
+		return u.XXX_Unmarshal(buf)
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		// NOTE: The history of proto have unfortunately been inconsistent
+		// whether Unmarshaler should or should not implicitly clear itself.
+		// Some implementations do, most do not.
+		// Thus, calling this here may or may not do what people want.
+		//
+		// See https://github.com/golang/protobuf/issues/424
+		return u.Unmarshal(buf)
+	}
+	return NewBuffer(buf).Unmarshal(pb)
+}
+
+// DecodeMessage reads a count-delimited message from the Buffer.
+func (p *Buffer) DecodeMessage(pb Message) error {
+	enc, err := p.DecodeRawBytes(false)
+	if err != nil {
+		return err
+	}
+	return NewBuffer(enc).Unmarshal(pb)
+}
+
+// DecodeGroup reads a tag-delimited group from the Buffer.
+// StartGroup tag is already consumed. This function consumes
+// EndGroup tag.
+func (p *Buffer) DecodeGroup(pb Message) error {
+	b := p.buf[p.index:]
+	x, y := findEndGroup(b)
+	if x < 0 {
+		return io.ErrUnexpectedEOF
+	}
+	err := Unmarshal(b[:x], pb)
+	p.index += y
+	return err
+}
+
+// Unmarshal parses the protocol buffer representation in the
+// Buffer and places the decoded result in pb.  If the struct
+// underlying pb does not match the data in the buffer, the results can be
+// unpredictable.
+//
+// Unlike proto.Unmarshal, this does not reset pb before starting to unmarshal.
+func (p *Buffer) Unmarshal(pb Message) error {
+	// If the object can unmarshal itself, let it.
+	if u, ok := pb.(newUnmarshaler); ok {
+		err := u.XXX_Unmarshal(p.buf[p.index:])
+		p.index = len(p.buf)
+		return err
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		// NOTE: The history of proto have unfortunately been inconsistent
+		// whether Unmarshaler should or should not implicitly clear itself.
+		// Some implementations do, most do not.
+		// Thus, calling this here may or may not do what people want.
+		//
+		// See https://github.com/golang/protobuf/issues/424
+		err := u.Unmarshal(p.buf[p.index:])
+		p.index = len(p.buf)
+		return err
+	}
+
+	// Slow workaround for messages that aren't Unmarshalers.
+	// This includes some hand-coded .pb.go files and
+	// bootstrap protos.
+	// TODO: fix all of those and then add Unmarshal to
+	// the Message interface. Then:
+	// The cast above and code below can be deleted.
+	// The old unmarshaler can be deleted.
+	// Clients can call Unmarshal directly (can already do that, actually).
+	var info InternalMessageInfo
+	err := info.Unmarshal(pb, p.buf[p.index:])
+	p.index = len(p.buf)
+	return err
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/deprecated.go b/vendor/github.com/gogo/protobuf/proto/deprecated.go
new file mode 100644
index 0000000..35b882c
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/deprecated.go
@@ -0,0 +1,63 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2018 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import "errors"
+
+// Deprecated: do not use.
+type Stats struct{ Emalloc, Dmalloc, Encode, Decode, Chit, Cmiss, Size uint64 }
+
+// Deprecated: do not use.
+func GetStats() Stats { return Stats{} }
+
+// Deprecated: do not use.
+func MarshalMessageSet(interface{}) ([]byte, error) {
+	return nil, errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func UnmarshalMessageSet([]byte, interface{}) error {
+	return errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func MarshalMessageSetJSON(interface{}) ([]byte, error) {
+	return nil, errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func UnmarshalMessageSetJSON([]byte, interface{}) error {
+	return errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func RegisterMessageSetType(Message, int32, string) {}
diff --git a/vendor/github.com/gogo/protobuf/proto/discard.go b/vendor/github.com/gogo/protobuf/proto/discard.go
new file mode 100644
index 0000000..fe1bd7d
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/discard.go
@@ -0,0 +1,350 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2017 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+	"sync"
+	"sync/atomic"
+)
+
+type generatedDiscarder interface {
+	XXX_DiscardUnknown()
+}
+
+// DiscardUnknown recursively discards all unknown fields from this message
+// and all embedded messages.
+//
+// When unmarshaling a message with unrecognized fields, the tags and values
+// of such fields are preserved in the Message. This allows a later call to
+// marshal to be able to produce a message that continues to have those
+// unrecognized fields. To avoid this, DiscardUnknown is used to
+// explicitly clear the unknown fields after unmarshaling.
+//
+// For proto2 messages, the unknown fields of message extensions are only
+// discarded from messages that have been accessed via GetExtension.
+func DiscardUnknown(m Message) {
+	if m, ok := m.(generatedDiscarder); ok {
+		m.XXX_DiscardUnknown()
+		return
+	}
+	// TODO: Dynamically populate a InternalMessageInfo for legacy messages,
+	// but the master branch has no implementation for InternalMessageInfo,
+	// so it would be more work to replicate that approach.
+	discardLegacy(m)
+}
+
+// DiscardUnknown recursively discards all unknown fields.
+func (a *InternalMessageInfo) DiscardUnknown(m Message) {
+	di := atomicLoadDiscardInfo(&a.discard)
+	if di == nil {
+		di = getDiscardInfo(reflect.TypeOf(m).Elem())
+		atomicStoreDiscardInfo(&a.discard, di)
+	}
+	di.discard(toPointer(&m))
+}
+
+type discardInfo struct {
+	typ reflect.Type
+
+	initialized int32 // 0: only typ is valid, 1: everything is valid
+	lock        sync.Mutex
+
+	fields       []discardFieldInfo
+	unrecognized field
+}
+
+type discardFieldInfo struct {
+	field   field // Offset of field, guaranteed to be valid
+	discard func(src pointer)
+}
+
+var (
+	discardInfoMap  = map[reflect.Type]*discardInfo{}
+	discardInfoLock sync.Mutex
+)
+
+func getDiscardInfo(t reflect.Type) *discardInfo {
+	discardInfoLock.Lock()
+	defer discardInfoLock.Unlock()
+	di := discardInfoMap[t]
+	if di == nil {
+		di = &discardInfo{typ: t}
+		discardInfoMap[t] = di
+	}
+	return di
+}
+
+func (di *discardInfo) discard(src pointer) {
+	if src.isNil() {
+		return // Nothing to do.
+	}
+
+	if atomic.LoadInt32(&di.initialized) == 0 {
+		di.computeDiscardInfo()
+	}
+
+	for _, fi := range di.fields {
+		sfp := src.offset(fi.field)
+		fi.discard(sfp)
+	}
+
+	// For proto2 messages, only discard unknown fields in message extensions
+	// that have been accessed via GetExtension.
+	if em, err := extendable(src.asPointerTo(di.typ).Interface()); err == nil {
+		// Ignore lock since DiscardUnknown is not concurrency safe.
+		emm, _ := em.extensionsRead()
+		for _, mx := range emm {
+			if m, ok := mx.value.(Message); ok {
+				DiscardUnknown(m)
+			}
+		}
+	}
+
+	if di.unrecognized.IsValid() {
+		*src.offset(di.unrecognized).toBytes() = nil
+	}
+}
+
+func (di *discardInfo) computeDiscardInfo() {
+	di.lock.Lock()
+	defer di.lock.Unlock()
+	if di.initialized != 0 {
+		return
+	}
+	t := di.typ
+	n := t.NumField()
+
+	for i := 0; i < n; i++ {
+		f := t.Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+
+		dfi := discardFieldInfo{field: toField(&f)}
+		tf := f.Type
+
+		// Unwrap tf to get its most basic type.
+		var isPointer, isSlice bool
+		if tf.Kind() == reflect.Slice && tf.Elem().Kind() != reflect.Uint8 {
+			isSlice = true
+			tf = tf.Elem()
+		}
+		if tf.Kind() == reflect.Ptr {
+			isPointer = true
+			tf = tf.Elem()
+		}
+		if isPointer && isSlice && tf.Kind() != reflect.Struct {
+			panic(fmt.Sprintf("%v.%s cannot be a slice of pointers to primitive types", t, f.Name))
+		}
+
+		switch tf.Kind() {
+		case reflect.Struct:
+			switch {
+			case !isPointer:
+				panic(fmt.Sprintf("%v.%s cannot be a direct struct value", t, f.Name))
+			case isSlice: // E.g., []*pb.T
+				discardInfo := getDiscardInfo(tf)
+				dfi.discard = func(src pointer) {
+					sps := src.getPointerSlice()
+					for _, sp := range sps {
+						if !sp.isNil() {
+							discardInfo.discard(sp)
+						}
+					}
+				}
+			default: // E.g., *pb.T
+				discardInfo := getDiscardInfo(tf)
+				dfi.discard = func(src pointer) {
+					sp := src.getPointer()
+					if !sp.isNil() {
+						discardInfo.discard(sp)
+					}
+				}
+			}
+		case reflect.Map:
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%v.%s cannot be a pointer to a map or a slice of map values", t, f.Name))
+			default: // E.g., map[K]V
+				if tf.Elem().Kind() == reflect.Ptr { // Proto struct (e.g., *T)
+					dfi.discard = func(src pointer) {
+						sm := src.asPointerTo(tf).Elem()
+						if sm.Len() == 0 {
+							return
+						}
+						for _, key := range sm.MapKeys() {
+							val := sm.MapIndex(key)
+							DiscardUnknown(val.Interface().(Message))
+						}
+					}
+				} else {
+					dfi.discard = func(pointer) {} // Noop
+				}
+			}
+		case reflect.Interface:
+			// Must be oneof field.
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%v.%s cannot be a pointer to a interface or a slice of interface values", t, f.Name))
+			default: // E.g., interface{}
+				// TODO: Make this faster?
+				dfi.discard = func(src pointer) {
+					su := src.asPointerTo(tf).Elem()
+					if !su.IsNil() {
+						sv := su.Elem().Elem().Field(0)
+						if sv.Kind() == reflect.Ptr && sv.IsNil() {
+							return
+						}
+						switch sv.Type().Kind() {
+						case reflect.Ptr: // Proto struct (e.g., *T)
+							DiscardUnknown(sv.Interface().(Message))
+						}
+					}
+				}
+			}
+		default:
+			continue
+		}
+		di.fields = append(di.fields, dfi)
+	}
+
+	di.unrecognized = invalidField
+	if f, ok := t.FieldByName("XXX_unrecognized"); ok {
+		if f.Type != reflect.TypeOf([]byte{}) {
+			panic("expected XXX_unrecognized to be of type []byte")
+		}
+		di.unrecognized = toField(&f)
+	}
+
+	atomic.StoreInt32(&di.initialized, 1)
+}
+
+func discardLegacy(m Message) {
+	v := reflect.ValueOf(m)
+	if v.Kind() != reflect.Ptr || v.IsNil() {
+		return
+	}
+	v = v.Elem()
+	if v.Kind() != reflect.Struct {
+		return
+	}
+	t := v.Type()
+
+	for i := 0; i < v.NumField(); i++ {
+		f := t.Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		vf := v.Field(i)
+		tf := f.Type
+
+		// Unwrap tf to get its most basic type.
+		var isPointer, isSlice bool
+		if tf.Kind() == reflect.Slice && tf.Elem().Kind() != reflect.Uint8 {
+			isSlice = true
+			tf = tf.Elem()
+		}
+		if tf.Kind() == reflect.Ptr {
+			isPointer = true
+			tf = tf.Elem()
+		}
+		if isPointer && isSlice && tf.Kind() != reflect.Struct {
+			panic(fmt.Sprintf("%T.%s cannot be a slice of pointers to primitive types", m, f.Name))
+		}
+
+		switch tf.Kind() {
+		case reflect.Struct:
+			switch {
+			case !isPointer:
+				panic(fmt.Sprintf("%T.%s cannot be a direct struct value", m, f.Name))
+			case isSlice: // E.g., []*pb.T
+				for j := 0; j < vf.Len(); j++ {
+					discardLegacy(vf.Index(j).Interface().(Message))
+				}
+			default: // E.g., *pb.T
+				discardLegacy(vf.Interface().(Message))
+			}
+		case reflect.Map:
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%T.%s cannot be a pointer to a map or a slice of map values", m, f.Name))
+			default: // E.g., map[K]V
+				tv := vf.Type().Elem()
+				if tv.Kind() == reflect.Ptr && tv.Implements(protoMessageType) { // Proto struct (e.g., *T)
+					for _, key := range vf.MapKeys() {
+						val := vf.MapIndex(key)
+						discardLegacy(val.Interface().(Message))
+					}
+				}
+			}
+		case reflect.Interface:
+			// Must be oneof field.
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%T.%s cannot be a pointer to a interface or a slice of interface values", m, f.Name))
+			default: // E.g., test_proto.isCommunique_Union interface
+				if !vf.IsNil() && f.Tag.Get("protobuf_oneof") != "" {
+					vf = vf.Elem() // E.g., *test_proto.Communique_Msg
+					if !vf.IsNil() {
+						vf = vf.Elem()   // E.g., test_proto.Communique_Msg
+						vf = vf.Field(0) // E.g., Proto struct (e.g., *T) or primitive value
+						if vf.Kind() == reflect.Ptr {
+							discardLegacy(vf.Interface().(Message))
+						}
+					}
+				}
+			}
+		}
+	}
+
+	if vf := v.FieldByName("XXX_unrecognized"); vf.IsValid() {
+		if vf.Type() != reflect.TypeOf([]byte{}) {
+			panic("expected XXX_unrecognized to be of type []byte")
+		}
+		vf.Set(reflect.ValueOf([]byte(nil)))
+	}
+
+	// For proto2 messages, only discard unknown fields in message extensions
+	// that have been accessed via GetExtension.
+	if em, err := extendable(m); err == nil {
+		// Ignore lock since discardLegacy is not concurrency safe.
+		emm, _ := em.extensionsRead()
+		for _, mx := range emm {
+			if m, ok := mx.value.(Message); ok {
+				discardLegacy(m)
+			}
+		}
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/duration.go b/vendor/github.com/gogo/protobuf/proto/duration.go
new file mode 100644
index 0000000..93464c9
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/duration.go
@@ -0,0 +1,100 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+// This file implements conversions between google.protobuf.Duration
+// and time.Duration.
+
+import (
+	"errors"
+	"fmt"
+	"time"
+)
+
+const (
+	// Range of a Duration in seconds, as specified in
+	// google/protobuf/duration.proto. This is about 10,000 years in seconds.
+	maxSeconds = int64(10000 * 365.25 * 24 * 60 * 60)
+	minSeconds = -maxSeconds
+)
+
+// validateDuration determines whether the Duration is valid according to the
+// definition in google/protobuf/duration.proto. A valid Duration
+// may still be too large to fit into a time.Duration (the range of Duration
+// is about 10,000 years, and the range of time.Duration is about 290).
+func validateDuration(d *duration) error {
+	if d == nil {
+		return errors.New("duration: nil Duration")
+	}
+	if d.Seconds < minSeconds || d.Seconds > maxSeconds {
+		return fmt.Errorf("duration: %#v: seconds out of range", d)
+	}
+	if d.Nanos <= -1e9 || d.Nanos >= 1e9 {
+		return fmt.Errorf("duration: %#v: nanos out of range", d)
+	}
+	// Seconds and Nanos must have the same sign, unless d.Nanos is zero.
+	if (d.Seconds < 0 && d.Nanos > 0) || (d.Seconds > 0 && d.Nanos < 0) {
+		return fmt.Errorf("duration: %#v: seconds and nanos have different signs", d)
+	}
+	return nil
+}
+
+// DurationFromProto converts a Duration to a time.Duration. DurationFromProto
+// returns an error if the Duration is invalid or is too large to be
+// represented in a time.Duration.
+func durationFromProto(p *duration) (time.Duration, error) {
+	if err := validateDuration(p); err != nil {
+		return 0, err
+	}
+	d := time.Duration(p.Seconds) * time.Second
+	if int64(d/time.Second) != p.Seconds {
+		return 0, fmt.Errorf("duration: %#v is out of range for time.Duration", p)
+	}
+	if p.Nanos != 0 {
+		d += time.Duration(p.Nanos)
+		if (d < 0) != (p.Nanos < 0) {
+			return 0, fmt.Errorf("duration: %#v is out of range for time.Duration", p)
+		}
+	}
+	return d, nil
+}
+
+// DurationProto converts a time.Duration to a Duration.
+func durationProto(d time.Duration) *duration {
+	nanos := d.Nanoseconds()
+	secs := nanos / 1e9
+	nanos -= secs * 1e9
+	return &duration{
+		Seconds: secs,
+		Nanos:   int32(nanos),
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/duration_gogo.go b/vendor/github.com/gogo/protobuf/proto/duration_gogo.go
new file mode 100644
index 0000000..e748e17
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/duration_gogo.go
@@ -0,0 +1,49 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2016, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"reflect"
+	"time"
+)
+
+var durationType = reflect.TypeOf((*time.Duration)(nil)).Elem()
+
+type duration struct {
+	Seconds int64 `protobuf:"varint,1,opt,name=seconds,proto3" json:"seconds,omitempty"`
+	Nanos   int32 `protobuf:"varint,2,opt,name=nanos,proto3" json:"nanos,omitempty"`
+}
+
+func (m *duration) Reset()       { *m = duration{} }
+func (*duration) ProtoMessage()  {}
+func (*duration) String() string { return "duration<string>" }
+
+func init() {
+	RegisterType((*duration)(nil), "gogo.protobuf.proto.duration")
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/encode.go b/vendor/github.com/gogo/protobuf/proto/encode.go
new file mode 100644
index 0000000..3abfed2
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/encode.go
@@ -0,0 +1,203 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for encoding data into the wire format for protocol buffers.
+ */
+
+import (
+	"errors"
+	"reflect"
+)
+
+var (
+	// errRepeatedHasNil is the error returned if Marshal is called with
+	// a struct with a repeated field containing a nil element.
+	errRepeatedHasNil = errors.New("proto: repeated field has nil element")
+
+	// errOneofHasNil is the error returned if Marshal is called with
+	// a struct with a oneof field containing a nil element.
+	errOneofHasNil = errors.New("proto: oneof field has nil value")
+
+	// ErrNil is the error returned if Marshal is called with nil.
+	ErrNil = errors.New("proto: Marshal called with nil")
+
+	// ErrTooLarge is the error returned if Marshal is called with a
+	// message that encodes to >2GB.
+	ErrTooLarge = errors.New("proto: message encodes to over 2 GB")
+)
+
+// The fundamental encoders that put bytes on the wire.
+// Those that take integer types all accept uint64 and are
+// therefore of type valueEncoder.
+
+const maxVarintBytes = 10 // maximum length of a varint
+
+// EncodeVarint returns the varint encoding of x.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+// Not used by the package itself, but helpful to clients
+// wishing to use the same encoding.
+func EncodeVarint(x uint64) []byte {
+	var buf [maxVarintBytes]byte
+	var n int
+	for n = 0; x > 127; n++ {
+		buf[n] = 0x80 | uint8(x&0x7F)
+		x >>= 7
+	}
+	buf[n] = uint8(x)
+	n++
+	return buf[0:n]
+}
+
+// EncodeVarint writes a varint-encoded integer to the Buffer.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func (p *Buffer) EncodeVarint(x uint64) error {
+	for x >= 1<<7 {
+		p.buf = append(p.buf, uint8(x&0x7f|0x80))
+		x >>= 7
+	}
+	p.buf = append(p.buf, uint8(x))
+	return nil
+}
+
+// SizeVarint returns the varint encoding size of an integer.
+func SizeVarint(x uint64) int {
+	switch {
+	case x < 1<<7:
+		return 1
+	case x < 1<<14:
+		return 2
+	case x < 1<<21:
+		return 3
+	case x < 1<<28:
+		return 4
+	case x < 1<<35:
+		return 5
+	case x < 1<<42:
+		return 6
+	case x < 1<<49:
+		return 7
+	case x < 1<<56:
+		return 8
+	case x < 1<<63:
+		return 9
+	}
+	return 10
+}
+
+// EncodeFixed64 writes a 64-bit integer to the Buffer.
+// This is the format for the
+// fixed64, sfixed64, and double protocol buffer types.
+func (p *Buffer) EncodeFixed64(x uint64) error {
+	p.buf = append(p.buf,
+		uint8(x),
+		uint8(x>>8),
+		uint8(x>>16),
+		uint8(x>>24),
+		uint8(x>>32),
+		uint8(x>>40),
+		uint8(x>>48),
+		uint8(x>>56))
+	return nil
+}
+
+// EncodeFixed32 writes a 32-bit integer to the Buffer.
+// This is the format for the
+// fixed32, sfixed32, and float protocol buffer types.
+func (p *Buffer) EncodeFixed32(x uint64) error {
+	p.buf = append(p.buf,
+		uint8(x),
+		uint8(x>>8),
+		uint8(x>>16),
+		uint8(x>>24))
+	return nil
+}
+
+// EncodeZigzag64 writes a zigzag-encoded 64-bit integer
+// to the Buffer.
+// This is the format used for the sint64 protocol buffer type.
+func (p *Buffer) EncodeZigzag64(x uint64) error {
+	// use signed number to get arithmetic right shift.
+	return p.EncodeVarint(uint64((x << 1) ^ uint64((int64(x) >> 63))))
+}
+
+// EncodeZigzag32 writes a zigzag-encoded 32-bit integer
+// to the Buffer.
+// This is the format used for the sint32 protocol buffer type.
+func (p *Buffer) EncodeZigzag32(x uint64) error {
+	// use signed number to get arithmetic right shift.
+	return p.EncodeVarint(uint64((uint32(x) << 1) ^ uint32((int32(x) >> 31))))
+}
+
+// EncodeRawBytes writes a count-delimited byte buffer to the Buffer.
+// This is the format used for the bytes protocol buffer
+// type and for embedded messages.
+func (p *Buffer) EncodeRawBytes(b []byte) error {
+	p.EncodeVarint(uint64(len(b)))
+	p.buf = append(p.buf, b...)
+	return nil
+}
+
+// EncodeStringBytes writes an encoded string to the Buffer.
+// This is the format used for the proto2 string type.
+func (p *Buffer) EncodeStringBytes(s string) error {
+	p.EncodeVarint(uint64(len(s)))
+	p.buf = append(p.buf, s...)
+	return nil
+}
+
+// Marshaler is the interface representing objects that can marshal themselves.
+type Marshaler interface {
+	Marshal() ([]byte, error)
+}
+
+// EncodeMessage writes the protocol buffer to the Buffer,
+// prefixed by a varint-encoded length.
+func (p *Buffer) EncodeMessage(pb Message) error {
+	siz := Size(pb)
+	p.EncodeVarint(uint64(siz))
+	return p.Marshal(pb)
+}
+
+// All protocol buffer fields are nillable, but be careful.
+func isNil(v reflect.Value) bool {
+	switch v.Kind() {
+	case reflect.Interface, reflect.Map, reflect.Ptr, reflect.Slice:
+		return v.IsNil()
+	}
+	return false
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/encode_gogo.go b/vendor/github.com/gogo/protobuf/proto/encode_gogo.go
new file mode 100644
index 0000000..0f5fb17
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/encode_gogo.go
@@ -0,0 +1,33 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+func NewRequiredNotSetError(field string) *RequiredNotSetError {
+	return &RequiredNotSetError{field}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/equal.go b/vendor/github.com/gogo/protobuf/proto/equal.go
new file mode 100644
index 0000000..d4db5a1
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/equal.go
@@ -0,0 +1,300 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2011 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Protocol buffer comparison.
+
+package proto
+
+import (
+	"bytes"
+	"log"
+	"reflect"
+	"strings"
+)
+
+/*
+Equal returns true iff protocol buffers a and b are equal.
+The arguments must both be pointers to protocol buffer structs.
+
+Equality is defined in this way:
+  - Two messages are equal iff they are the same type,
+    corresponding fields are equal, unknown field sets
+    are equal, and extensions sets are equal.
+  - Two set scalar fields are equal iff their values are equal.
+    If the fields are of a floating-point type, remember that
+    NaN != x for all x, including NaN. If the message is defined
+    in a proto3 .proto file, fields are not "set"; specifically,
+    zero length proto3 "bytes" fields are equal (nil == {}).
+  - Two repeated fields are equal iff their lengths are the same,
+    and their corresponding elements are equal. Note a "bytes" field,
+    although represented by []byte, is not a repeated field and the
+    rule for the scalar fields described above applies.
+  - Two unset fields are equal.
+  - Two unknown field sets are equal if their current
+    encoded state is equal.
+  - Two extension sets are equal iff they have corresponding
+    elements that are pairwise equal.
+  - Two map fields are equal iff their lengths are the same,
+    and they contain the same set of elements. Zero-length map
+    fields are equal.
+  - Every other combination of things are not equal.
+
+The return value is undefined if a and b are not protocol buffers.
+*/
+func Equal(a, b Message) bool {
+	if a == nil || b == nil {
+		return a == b
+	}
+	v1, v2 := reflect.ValueOf(a), reflect.ValueOf(b)
+	if v1.Type() != v2.Type() {
+		return false
+	}
+	if v1.Kind() == reflect.Ptr {
+		if v1.IsNil() {
+			return v2.IsNil()
+		}
+		if v2.IsNil() {
+			return false
+		}
+		v1, v2 = v1.Elem(), v2.Elem()
+	}
+	if v1.Kind() != reflect.Struct {
+		return false
+	}
+	return equalStruct(v1, v2)
+}
+
+// v1 and v2 are known to have the same type.
+func equalStruct(v1, v2 reflect.Value) bool {
+	sprop := GetProperties(v1.Type())
+	for i := 0; i < v1.NumField(); i++ {
+		f := v1.Type().Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		f1, f2 := v1.Field(i), v2.Field(i)
+		if f.Type.Kind() == reflect.Ptr {
+			if n1, n2 := f1.IsNil(), f2.IsNil(); n1 && n2 {
+				// both unset
+				continue
+			} else if n1 != n2 {
+				// set/unset mismatch
+				return false
+			}
+			f1, f2 = f1.Elem(), f2.Elem()
+		}
+		if !equalAny(f1, f2, sprop.Prop[i]) {
+			return false
+		}
+	}
+
+	if em1 := v1.FieldByName("XXX_InternalExtensions"); em1.IsValid() {
+		em2 := v2.FieldByName("XXX_InternalExtensions")
+		if !equalExtensions(v1.Type(), em1.Interface().(XXX_InternalExtensions), em2.Interface().(XXX_InternalExtensions)) {
+			return false
+		}
+	}
+
+	if em1 := v1.FieldByName("XXX_extensions"); em1.IsValid() {
+		em2 := v2.FieldByName("XXX_extensions")
+		if !equalExtMap(v1.Type(), em1.Interface().(map[int32]Extension), em2.Interface().(map[int32]Extension)) {
+			return false
+		}
+	}
+
+	uf := v1.FieldByName("XXX_unrecognized")
+	if !uf.IsValid() {
+		return true
+	}
+
+	u1 := uf.Bytes()
+	u2 := v2.FieldByName("XXX_unrecognized").Bytes()
+	return bytes.Equal(u1, u2)
+}
+
+// v1 and v2 are known to have the same type.
+// prop may be nil.
+func equalAny(v1, v2 reflect.Value, prop *Properties) bool {
+	if v1.Type() == protoMessageType {
+		m1, _ := v1.Interface().(Message)
+		m2, _ := v2.Interface().(Message)
+		return Equal(m1, m2)
+	}
+	switch v1.Kind() {
+	case reflect.Bool:
+		return v1.Bool() == v2.Bool()
+	case reflect.Float32, reflect.Float64:
+		return v1.Float() == v2.Float()
+	case reflect.Int32, reflect.Int64:
+		return v1.Int() == v2.Int()
+	case reflect.Interface:
+		// Probably a oneof field; compare the inner values.
+		n1, n2 := v1.IsNil(), v2.IsNil()
+		if n1 || n2 {
+			return n1 == n2
+		}
+		e1, e2 := v1.Elem(), v2.Elem()
+		if e1.Type() != e2.Type() {
+			return false
+		}
+		return equalAny(e1, e2, nil)
+	case reflect.Map:
+		if v1.Len() != v2.Len() {
+			return false
+		}
+		for _, key := range v1.MapKeys() {
+			val2 := v2.MapIndex(key)
+			if !val2.IsValid() {
+				// This key was not found in the second map.
+				return false
+			}
+			if !equalAny(v1.MapIndex(key), val2, nil) {
+				return false
+			}
+		}
+		return true
+	case reflect.Ptr:
+		// Maps may have nil values in them, so check for nil.
+		if v1.IsNil() && v2.IsNil() {
+			return true
+		}
+		if v1.IsNil() != v2.IsNil() {
+			return false
+		}
+		return equalAny(v1.Elem(), v2.Elem(), prop)
+	case reflect.Slice:
+		if v1.Type().Elem().Kind() == reflect.Uint8 {
+			// short circuit: []byte
+
+			// Edge case: if this is in a proto3 message, a zero length
+			// bytes field is considered the zero value.
+			if prop != nil && prop.proto3 && v1.Len() == 0 && v2.Len() == 0 {
+				return true
+			}
+			if v1.IsNil() != v2.IsNil() {
+				return false
+			}
+			return bytes.Equal(v1.Interface().([]byte), v2.Interface().([]byte))
+		}
+
+		if v1.Len() != v2.Len() {
+			return false
+		}
+		for i := 0; i < v1.Len(); i++ {
+			if !equalAny(v1.Index(i), v2.Index(i), prop) {
+				return false
+			}
+		}
+		return true
+	case reflect.String:
+		return v1.Interface().(string) == v2.Interface().(string)
+	case reflect.Struct:
+		return equalStruct(v1, v2)
+	case reflect.Uint32, reflect.Uint64:
+		return v1.Uint() == v2.Uint()
+	}
+
+	// unknown type, so not a protocol buffer
+	log.Printf("proto: don't know how to compare %v", v1)
+	return false
+}
+
+// base is the struct type that the extensions are based on.
+// x1 and x2 are InternalExtensions.
+func equalExtensions(base reflect.Type, x1, x2 XXX_InternalExtensions) bool {
+	em1, _ := x1.extensionsRead()
+	em2, _ := x2.extensionsRead()
+	return equalExtMap(base, em1, em2)
+}
+
+func equalExtMap(base reflect.Type, em1, em2 map[int32]Extension) bool {
+	if len(em1) != len(em2) {
+		return false
+	}
+
+	for extNum, e1 := range em1 {
+		e2, ok := em2[extNum]
+		if !ok {
+			return false
+		}
+
+		m1, m2 := e1.value, e2.value
+
+		if m1 == nil && m2 == nil {
+			// Both have only encoded form.
+			if bytes.Equal(e1.enc, e2.enc) {
+				continue
+			}
+			// The bytes are different, but the extensions might still be
+			// equal. We need to decode them to compare.
+		}
+
+		if m1 != nil && m2 != nil {
+			// Both are unencoded.
+			if !equalAny(reflect.ValueOf(m1), reflect.ValueOf(m2), nil) {
+				return false
+			}
+			continue
+		}
+
+		// At least one is encoded. To do a semantically correct comparison
+		// we need to unmarshal them first.
+		var desc *ExtensionDesc
+		if m := extensionMaps[base]; m != nil {
+			desc = m[extNum]
+		}
+		if desc == nil {
+			// If both have only encoded form and the bytes are the same,
+			// it is handled above. We get here when the bytes are different.
+			// We don't know how to decode it, so just compare them as byte
+			// slices.
+			log.Printf("proto: don't know how to compare extension %d of %v", extNum, base)
+			return false
+		}
+		var err error
+		if m1 == nil {
+			m1, err = decodeExtension(e1.enc, desc)
+		}
+		if m2 == nil && err == nil {
+			m2, err = decodeExtension(e2.enc, desc)
+		}
+		if err != nil {
+			// The encoded form is invalid.
+			log.Printf("proto: badly encoded extension %d of %v: %v", extNum, base, err)
+			return false
+		}
+		if !equalAny(reflect.ValueOf(m1), reflect.ValueOf(m2), nil) {
+			return false
+		}
+	}
+
+	return true
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/extensions.go b/vendor/github.com/gogo/protobuf/proto/extensions.go
new file mode 100644
index 0000000..686bd2a
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/extensions.go
@@ -0,0 +1,604 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Types and routines for supporting protocol buffer extensions.
+ */
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"reflect"
+	"strconv"
+	"sync"
+)
+
+// ErrMissingExtension is the error returned by GetExtension if the named extension is not in the message.
+var ErrMissingExtension = errors.New("proto: missing extension")
+
+// ExtensionRange represents a range of message extensions for a protocol buffer.
+// Used in code generated by the protocol compiler.
+type ExtensionRange struct {
+	Start, End int32 // both inclusive
+}
+
+// extendableProto is an interface implemented by any protocol buffer generated by the current
+// proto compiler that may be extended.
+type extendableProto interface {
+	Message
+	ExtensionRangeArray() []ExtensionRange
+	extensionsWrite() map[int32]Extension
+	extensionsRead() (map[int32]Extension, sync.Locker)
+}
+
+// extendableProtoV1 is an interface implemented by a protocol buffer generated by the previous
+// version of the proto compiler that may be extended.
+type extendableProtoV1 interface {
+	Message
+	ExtensionRangeArray() []ExtensionRange
+	ExtensionMap() map[int32]Extension
+}
+
+// extensionAdapter is a wrapper around extendableProtoV1 that implements extendableProto.
+type extensionAdapter struct {
+	extendableProtoV1
+}
+
+func (e extensionAdapter) extensionsWrite() map[int32]Extension {
+	return e.ExtensionMap()
+}
+
+func (e extensionAdapter) extensionsRead() (map[int32]Extension, sync.Locker) {
+	return e.ExtensionMap(), notLocker{}
+}
+
+// notLocker is a sync.Locker whose Lock and Unlock methods are nops.
+type notLocker struct{}
+
+func (n notLocker) Lock()   {}
+func (n notLocker) Unlock() {}
+
+// extendable returns the extendableProto interface for the given generated proto message.
+// If the proto message has the old extension format, it returns a wrapper that implements
+// the extendableProto interface.
+func extendable(p interface{}) (extendableProto, error) {
+	switch p := p.(type) {
+	case extendableProto:
+		if isNilPtr(p) {
+			return nil, fmt.Errorf("proto: nil %T is not extendable", p)
+		}
+		return p, nil
+	case extendableProtoV1:
+		if isNilPtr(p) {
+			return nil, fmt.Errorf("proto: nil %T is not extendable", p)
+		}
+		return extensionAdapter{p}, nil
+	case extensionsBytes:
+		return slowExtensionAdapter{p}, nil
+	}
+	// Don't allocate a specific error containing %T:
+	// this is the hot path for Clone and MarshalText.
+	return nil, errNotExtendable
+}
+
+var errNotExtendable = errors.New("proto: not an extendable proto.Message")
+
+func isNilPtr(x interface{}) bool {
+	v := reflect.ValueOf(x)
+	return v.Kind() == reflect.Ptr && v.IsNil()
+}
+
+// XXX_InternalExtensions is an internal representation of proto extensions.
+//
+// Each generated message struct type embeds an anonymous XXX_InternalExtensions field,
+// thus gaining the unexported 'extensions' method, which can be called only from the proto package.
+//
+// The methods of XXX_InternalExtensions are not concurrency safe in general,
+// but calls to logically read-only methods such as has and get may be executed concurrently.
+type XXX_InternalExtensions struct {
+	// The struct must be indirect so that if a user inadvertently copies a
+	// generated message and its embedded XXX_InternalExtensions, they
+	// avoid the mayhem of a copied mutex.
+	//
+	// The mutex serializes all logically read-only operations to p.extensionMap.
+	// It is up to the client to ensure that write operations to p.extensionMap are
+	// mutually exclusive with other accesses.
+	p *struct {
+		mu           sync.Mutex
+		extensionMap map[int32]Extension
+	}
+}
+
+// extensionsWrite returns the extension map, creating it on first use.
+func (e *XXX_InternalExtensions) extensionsWrite() map[int32]Extension {
+	if e.p == nil {
+		e.p = new(struct {
+			mu           sync.Mutex
+			extensionMap map[int32]Extension
+		})
+		e.p.extensionMap = make(map[int32]Extension)
+	}
+	return e.p.extensionMap
+}
+
+// extensionsRead returns the extensions map for read-only use.  It may be nil.
+// The caller must hold the returned mutex's lock when accessing Elements within the map.
+func (e *XXX_InternalExtensions) extensionsRead() (map[int32]Extension, sync.Locker) {
+	if e.p == nil {
+		return nil, nil
+	}
+	return e.p.extensionMap, &e.p.mu
+}
+
+// ExtensionDesc represents an extension specification.
+// Used in generated code from the protocol compiler.
+type ExtensionDesc struct {
+	ExtendedType  Message     // nil pointer to the type that is being extended
+	ExtensionType interface{} // nil pointer to the extension type
+	Field         int32       // field number
+	Name          string      // fully-qualified name of extension, for text formatting
+	Tag           string      // protobuf tag style
+	Filename      string      // name of the file in which the extension is defined
+}
+
+func (ed *ExtensionDesc) repeated() bool {
+	t := reflect.TypeOf(ed.ExtensionType)
+	return t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8
+}
+
+// Extension represents an extension in a message.
+type Extension struct {
+	// When an extension is stored in a message using SetExtension
+	// only desc and value are set. When the message is marshaled
+	// enc will be set to the encoded form of the message.
+	//
+	// When a message is unmarshaled and contains extensions, each
+	// extension will have only enc set. When such an extension is
+	// accessed using GetExtension (or GetExtensions) desc and value
+	// will be set.
+	desc  *ExtensionDesc
+	value interface{}
+	enc   []byte
+}
+
+// SetRawExtension is for testing only.
+func SetRawExtension(base Message, id int32, b []byte) {
+	if ebase, ok := base.(extensionsBytes); ok {
+		clearExtension(base, id)
+		ext := ebase.GetExtensions()
+		*ext = append(*ext, b...)
+		return
+	}
+	epb, err := extendable(base)
+	if err != nil {
+		return
+	}
+	extmap := epb.extensionsWrite()
+	extmap[id] = Extension{enc: b}
+}
+
+// isExtensionField returns true iff the given field number is in an extension range.
+func isExtensionField(pb extendableProto, field int32) bool {
+	for _, er := range pb.ExtensionRangeArray() {
+		if er.Start <= field && field <= er.End {
+			return true
+		}
+	}
+	return false
+}
+
+// checkExtensionTypes checks that the given extension is valid for pb.
+func checkExtensionTypes(pb extendableProto, extension *ExtensionDesc) error {
+	var pbi interface{} = pb
+	// Check the extended type.
+	if ea, ok := pbi.(extensionAdapter); ok {
+		pbi = ea.extendableProtoV1
+	}
+	if ea, ok := pbi.(slowExtensionAdapter); ok {
+		pbi = ea.extensionsBytes
+	}
+	if a, b := reflect.TypeOf(pbi), reflect.TypeOf(extension.ExtendedType); a != b {
+		return fmt.Errorf("proto: bad extended type; %v does not extend %v", b, a)
+	}
+	// Check the range.
+	if !isExtensionField(pb, extension.Field) {
+		return errors.New("proto: bad extension number; not in declared ranges")
+	}
+	return nil
+}
+
+// extPropKey is sufficient to uniquely identify an extension.
+type extPropKey struct {
+	base  reflect.Type
+	field int32
+}
+
+var extProp = struct {
+	sync.RWMutex
+	m map[extPropKey]*Properties
+}{
+	m: make(map[extPropKey]*Properties),
+}
+
+func extensionProperties(ed *ExtensionDesc) *Properties {
+	key := extPropKey{base: reflect.TypeOf(ed.ExtendedType), field: ed.Field}
+
+	extProp.RLock()
+	if prop, ok := extProp.m[key]; ok {
+		extProp.RUnlock()
+		return prop
+	}
+	extProp.RUnlock()
+
+	extProp.Lock()
+	defer extProp.Unlock()
+	// Check again.
+	if prop, ok := extProp.m[key]; ok {
+		return prop
+	}
+
+	prop := new(Properties)
+	prop.Init(reflect.TypeOf(ed.ExtensionType), "unknown_name", ed.Tag, nil)
+	extProp.m[key] = prop
+	return prop
+}
+
+// HasExtension returns whether the given extension is present in pb.
+func HasExtension(pb Message, extension *ExtensionDesc) bool {
+	if epb, doki := pb.(extensionsBytes); doki {
+		ext := epb.GetExtensions()
+		buf := *ext
+		o := 0
+		for o < len(buf) {
+			tag, n := DecodeVarint(buf[o:])
+			fieldNum := int32(tag >> 3)
+			if int32(fieldNum) == extension.Field {
+				return true
+			}
+			wireType := int(tag & 0x7)
+			o += n
+			l, err := size(buf[o:], wireType)
+			if err != nil {
+				return false
+			}
+			o += l
+		}
+		return false
+	}
+	// TODO: Check types, field numbers, etc.?
+	epb, err := extendable(pb)
+	if err != nil {
+		return false
+	}
+	extmap, mu := epb.extensionsRead()
+	if extmap == nil {
+		return false
+	}
+	mu.Lock()
+	_, ok := extmap[extension.Field]
+	mu.Unlock()
+	return ok
+}
+
+// ClearExtension removes the given extension from pb.
+func ClearExtension(pb Message, extension *ExtensionDesc) {
+	clearExtension(pb, extension.Field)
+}
+
+func clearExtension(pb Message, fieldNum int32) {
+	if epb, ok := pb.(extensionsBytes); ok {
+		offset := 0
+		for offset != -1 {
+			offset = deleteExtension(epb, fieldNum, offset)
+		}
+		return
+	}
+	epb, err := extendable(pb)
+	if err != nil {
+		return
+	}
+	// TODO: Check types, field numbers, etc.?
+	extmap := epb.extensionsWrite()
+	delete(extmap, fieldNum)
+}
+
+// GetExtension retrieves a proto2 extended field from pb.
+//
+// If the descriptor is type complete (i.e., ExtensionDesc.ExtensionType is non-nil),
+// then GetExtension parses the encoded field and returns a Go value of the specified type.
+// If the field is not present, then the default value is returned (if one is specified),
+// otherwise ErrMissingExtension is reported.
+//
+// If the descriptor is not type complete (i.e., ExtensionDesc.ExtensionType is nil),
+// then GetExtension returns the raw encoded bytes of the field extension.
+func GetExtension(pb Message, extension *ExtensionDesc) (interface{}, error) {
+	if epb, doki := pb.(extensionsBytes); doki {
+		ext := epb.GetExtensions()
+		return decodeExtensionFromBytes(extension, *ext)
+	}
+
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+
+	if extension.ExtendedType != nil {
+		// can only check type if this is a complete descriptor
+		if cerr := checkExtensionTypes(epb, extension); cerr != nil {
+			return nil, cerr
+		}
+	}
+
+	emap, mu := epb.extensionsRead()
+	if emap == nil {
+		return defaultExtensionValue(extension)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	e, ok := emap[extension.Field]
+	if !ok {
+		// defaultExtensionValue returns the default value or
+		// ErrMissingExtension if there is no default.
+		return defaultExtensionValue(extension)
+	}
+
+	if e.value != nil {
+		// Already decoded. Check the descriptor, though.
+		if e.desc != extension {
+			// This shouldn't happen. If it does, it means that
+			// GetExtension was called twice with two different
+			// descriptors with the same field number.
+			return nil, errors.New("proto: descriptor conflict")
+		}
+		return e.value, nil
+	}
+
+	if extension.ExtensionType == nil {
+		// incomplete descriptor
+		return e.enc, nil
+	}
+
+	v, err := decodeExtension(e.enc, extension)
+	if err != nil {
+		return nil, err
+	}
+
+	// Remember the decoded version and drop the encoded version.
+	// That way it is safe to mutate what we return.
+	e.value = v
+	e.desc = extension
+	e.enc = nil
+	emap[extension.Field] = e
+	return e.value, nil
+}
+
+// defaultExtensionValue returns the default value for extension.
+// If no default for an extension is defined ErrMissingExtension is returned.
+func defaultExtensionValue(extension *ExtensionDesc) (interface{}, error) {
+	if extension.ExtensionType == nil {
+		// incomplete descriptor, so no default
+		return nil, ErrMissingExtension
+	}
+
+	t := reflect.TypeOf(extension.ExtensionType)
+	props := extensionProperties(extension)
+
+	sf, _, err := fieldDefault(t, props)
+	if err != nil {
+		return nil, err
+	}
+
+	if sf == nil || sf.value == nil {
+		// There is no default value.
+		return nil, ErrMissingExtension
+	}
+
+	if t.Kind() != reflect.Ptr {
+		// We do not need to return a Ptr, we can directly return sf.value.
+		return sf.value, nil
+	}
+
+	// We need to return an interface{} that is a pointer to sf.value.
+	value := reflect.New(t).Elem()
+	value.Set(reflect.New(value.Type().Elem()))
+	if sf.kind == reflect.Int32 {
+		// We may have an int32 or an enum, but the underlying data is int32.
+		// Since we can't set an int32 into a non int32 reflect.value directly
+		// set it as a int32.
+		value.Elem().SetInt(int64(sf.value.(int32)))
+	} else {
+		value.Elem().Set(reflect.ValueOf(sf.value))
+	}
+	return value.Interface(), nil
+}
+
+// decodeExtension decodes an extension encoded in b.
+func decodeExtension(b []byte, extension *ExtensionDesc) (interface{}, error) {
+	t := reflect.TypeOf(extension.ExtensionType)
+	unmarshal := typeUnmarshaler(t, extension.Tag)
+
+	// t is a pointer to a struct, pointer to basic type or a slice.
+	// Allocate space to store the pointer/slice.
+	value := reflect.New(t).Elem()
+
+	var err error
+	for {
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		wire := int(x) & 7
+
+		b, err = unmarshal(b, valToPointer(value.Addr()), wire)
+		if err != nil {
+			return nil, err
+		}
+
+		if len(b) == 0 {
+			break
+		}
+	}
+	return value.Interface(), nil
+}
+
+// GetExtensions returns a slice of the extensions present in pb that are also listed in es.
+// The returned slice has the same length as es; missing extensions will appear as nil elements.
+func GetExtensions(pb Message, es []*ExtensionDesc) (extensions []interface{}, err error) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+	extensions = make([]interface{}, len(es))
+	for i, e := range es {
+		extensions[i], err = GetExtension(epb, e)
+		if err == ErrMissingExtension {
+			err = nil
+		}
+		if err != nil {
+			return
+		}
+	}
+	return
+}
+
+// ExtensionDescs returns a new slice containing pb's extension descriptors, in undefined order.
+// For non-registered extensions, ExtensionDescs returns an incomplete descriptor containing
+// just the Field field, which defines the extension's field number.
+func ExtensionDescs(pb Message) ([]*ExtensionDesc, error) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+	registeredExtensions := RegisteredExtensions(pb)
+
+	emap, mu := epb.extensionsRead()
+	if emap == nil {
+		return nil, nil
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	extensions := make([]*ExtensionDesc, 0, len(emap))
+	for extid, e := range emap {
+		desc := e.desc
+		if desc == nil {
+			desc = registeredExtensions[extid]
+			if desc == nil {
+				desc = &ExtensionDesc{Field: extid}
+			}
+		}
+
+		extensions = append(extensions, desc)
+	}
+	return extensions, nil
+}
+
+// SetExtension sets the specified extension of pb to the specified value.
+func SetExtension(pb Message, extension *ExtensionDesc, value interface{}) error {
+	if epb, ok := pb.(extensionsBytes); ok {
+		newb, err := encodeExtension(extension, value)
+		if err != nil {
+			return err
+		}
+		bb := epb.GetExtensions()
+		*bb = append(*bb, newb...)
+		return nil
+	}
+	epb, err := extendable(pb)
+	if err != nil {
+		return err
+	}
+	if err := checkExtensionTypes(epb, extension); err != nil {
+		return err
+	}
+	typ := reflect.TypeOf(extension.ExtensionType)
+	if typ != reflect.TypeOf(value) {
+		return fmt.Errorf("proto: bad extension value type. got: %T, want: %T", value, extension.ExtensionType)
+	}
+	// nil extension values need to be caught early, because the
+	// encoder can't distinguish an ErrNil due to a nil extension
+	// from an ErrNil due to a missing field. Extensions are
+	// always optional, so the encoder would just swallow the error
+	// and drop all the extensions from the encoded message.
+	if reflect.ValueOf(value).IsNil() {
+		return fmt.Errorf("proto: SetExtension called with nil value of type %T", value)
+	}
+
+	extmap := epb.extensionsWrite()
+	extmap[extension.Field] = Extension{desc: extension, value: value}
+	return nil
+}
+
+// ClearAllExtensions clears all extensions from pb.
+func ClearAllExtensions(pb Message) {
+	if epb, doki := pb.(extensionsBytes); doki {
+		ext := epb.GetExtensions()
+		*ext = []byte{}
+		return
+	}
+	epb, err := extendable(pb)
+	if err != nil {
+		return
+	}
+	m := epb.extensionsWrite()
+	for k := range m {
+		delete(m, k)
+	}
+}
+
+// A global registry of extensions.
+// The generated code will register the generated descriptors by calling RegisterExtension.
+
+var extensionMaps = make(map[reflect.Type]map[int32]*ExtensionDesc)
+
+// RegisterExtension is called from the generated code.
+func RegisterExtension(desc *ExtensionDesc) {
+	st := reflect.TypeOf(desc.ExtendedType).Elem()
+	m := extensionMaps[st]
+	if m == nil {
+		m = make(map[int32]*ExtensionDesc)
+		extensionMaps[st] = m
+	}
+	if _, ok := m[desc.Field]; ok {
+		panic("proto: duplicate extension registered: " + st.String() + " " + strconv.Itoa(int(desc.Field)))
+	}
+	m[desc.Field] = desc
+}
+
+// RegisteredExtensions returns a map of the registered extensions of a
+// protocol buffer struct, indexed by the extension number.
+// The argument pb should be a nil pointer to the struct type.
+func RegisteredExtensions(pb Message) map[int32]*ExtensionDesc {
+	return extensionMaps[reflect.TypeOf(pb).Elem()]
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/extensions_gogo.go b/vendor/github.com/gogo/protobuf/proto/extensions_gogo.go
new file mode 100644
index 0000000..53ebd8c
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/extensions_gogo.go
@@ -0,0 +1,368 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"io"
+	"reflect"
+	"sort"
+	"strings"
+	"sync"
+)
+
+type extensionsBytes interface {
+	Message
+	ExtensionRangeArray() []ExtensionRange
+	GetExtensions() *[]byte
+}
+
+type slowExtensionAdapter struct {
+	extensionsBytes
+}
+
+func (s slowExtensionAdapter) extensionsWrite() map[int32]Extension {
+	panic("Please report a bug to github.com/gogo/protobuf if you see this message: Writing extensions is not supported for extensions stored in a byte slice field.")
+}
+
+func (s slowExtensionAdapter) extensionsRead() (map[int32]Extension, sync.Locker) {
+	b := s.GetExtensions()
+	m, err := BytesToExtensionsMap(*b)
+	if err != nil {
+		panic(err)
+	}
+	return m, notLocker{}
+}
+
+func GetBoolExtension(pb Message, extension *ExtensionDesc, ifnotset bool) bool {
+	if reflect.ValueOf(pb).IsNil() {
+		return ifnotset
+	}
+	value, err := GetExtension(pb, extension)
+	if err != nil {
+		return ifnotset
+	}
+	if value == nil {
+		return ifnotset
+	}
+	if value.(*bool) == nil {
+		return ifnotset
+	}
+	return *(value.(*bool))
+}
+
+func (this *Extension) Equal(that *Extension) bool {
+	if err := this.Encode(); err != nil {
+		return false
+	}
+	if err := that.Encode(); err != nil {
+		return false
+	}
+	return bytes.Equal(this.enc, that.enc)
+}
+
+func (this *Extension) Compare(that *Extension) int {
+	if err := this.Encode(); err != nil {
+		return 1
+	}
+	if err := that.Encode(); err != nil {
+		return -1
+	}
+	return bytes.Compare(this.enc, that.enc)
+}
+
+func SizeOfInternalExtension(m extendableProto) (n int) {
+	info := getMarshalInfo(reflect.TypeOf(m))
+	return info.sizeV1Extensions(m.extensionsWrite())
+}
+
+type sortableMapElem struct {
+	field int32
+	ext   Extension
+}
+
+func newSortableExtensionsFromMap(m map[int32]Extension) sortableExtensions {
+	s := make(sortableExtensions, 0, len(m))
+	for k, v := range m {
+		s = append(s, &sortableMapElem{field: k, ext: v})
+	}
+	return s
+}
+
+type sortableExtensions []*sortableMapElem
+
+func (this sortableExtensions) Len() int { return len(this) }
+
+func (this sortableExtensions) Swap(i, j int) { this[i], this[j] = this[j], this[i] }
+
+func (this sortableExtensions) Less(i, j int) bool { return this[i].field < this[j].field }
+
+func (this sortableExtensions) String() string {
+	sort.Sort(this)
+	ss := make([]string, len(this))
+	for i := range this {
+		ss[i] = fmt.Sprintf("%d: %v", this[i].field, this[i].ext)
+	}
+	return "map[" + strings.Join(ss, ",") + "]"
+}
+
+func StringFromInternalExtension(m extendableProto) string {
+	return StringFromExtensionsMap(m.extensionsWrite())
+}
+
+func StringFromExtensionsMap(m map[int32]Extension) string {
+	return newSortableExtensionsFromMap(m).String()
+}
+
+func StringFromExtensionsBytes(ext []byte) string {
+	m, err := BytesToExtensionsMap(ext)
+	if err != nil {
+		panic(err)
+	}
+	return StringFromExtensionsMap(m)
+}
+
+func EncodeInternalExtension(m extendableProto, data []byte) (n int, err error) {
+	return EncodeExtensionMap(m.extensionsWrite(), data)
+}
+
+func EncodeExtensionMap(m map[int32]Extension, data []byte) (n int, err error) {
+	o := 0
+	for _, e := range m {
+		if err := e.Encode(); err != nil {
+			return 0, err
+		}
+		n := copy(data[o:], e.enc)
+		if n != len(e.enc) {
+			return 0, io.ErrShortBuffer
+		}
+		o += n
+	}
+	return o, nil
+}
+
+func GetRawExtension(m map[int32]Extension, id int32) ([]byte, error) {
+	e := m[id]
+	if err := e.Encode(); err != nil {
+		return nil, err
+	}
+	return e.enc, nil
+}
+
+func size(buf []byte, wire int) (int, error) {
+	switch wire {
+	case WireVarint:
+		_, n := DecodeVarint(buf)
+		return n, nil
+	case WireFixed64:
+		return 8, nil
+	case WireBytes:
+		v, n := DecodeVarint(buf)
+		return int(v) + n, nil
+	case WireFixed32:
+		return 4, nil
+	case WireStartGroup:
+		offset := 0
+		for {
+			u, n := DecodeVarint(buf[offset:])
+			fwire := int(u & 0x7)
+			offset += n
+			if fwire == WireEndGroup {
+				return offset, nil
+			}
+			s, err := size(buf[offset:], wire)
+			if err != nil {
+				return 0, err
+			}
+			offset += s
+		}
+	}
+	return 0, fmt.Errorf("proto: can't get size for unknown wire type %d", wire)
+}
+
+func BytesToExtensionsMap(buf []byte) (map[int32]Extension, error) {
+	m := make(map[int32]Extension)
+	i := 0
+	for i < len(buf) {
+		tag, n := DecodeVarint(buf[i:])
+		if n <= 0 {
+			return nil, fmt.Errorf("unable to decode varint")
+		}
+		fieldNum := int32(tag >> 3)
+		wireType := int(tag & 0x7)
+		l, err := size(buf[i+n:], wireType)
+		if err != nil {
+			return nil, err
+		}
+		end := i + int(l) + n
+		m[int32(fieldNum)] = Extension{enc: buf[i:end]}
+		i = end
+	}
+	return m, nil
+}
+
+func NewExtension(e []byte) Extension {
+	ee := Extension{enc: make([]byte, len(e))}
+	copy(ee.enc, e)
+	return ee
+}
+
+func AppendExtension(e Message, tag int32, buf []byte) {
+	if ee, eok := e.(extensionsBytes); eok {
+		ext := ee.GetExtensions()
+		*ext = append(*ext, buf...)
+		return
+	}
+	if ee, eok := e.(extendableProto); eok {
+		m := ee.extensionsWrite()
+		ext := m[int32(tag)] // may be missing
+		ext.enc = append(ext.enc, buf...)
+		m[int32(tag)] = ext
+	}
+}
+
+func encodeExtension(extension *ExtensionDesc, value interface{}) ([]byte, error) {
+	u := getMarshalInfo(reflect.TypeOf(extension.ExtendedType))
+	ei := u.getExtElemInfo(extension)
+	v := value
+	p := toAddrPointer(&v, ei.isptr)
+	siz := ei.sizer(p, SizeVarint(ei.wiretag))
+	buf := make([]byte, 0, siz)
+	return ei.marshaler(buf, p, ei.wiretag, false)
+}
+
+func decodeExtensionFromBytes(extension *ExtensionDesc, buf []byte) (interface{}, error) {
+	o := 0
+	for o < len(buf) {
+		tag, n := DecodeVarint((buf)[o:])
+		fieldNum := int32(tag >> 3)
+		wireType := int(tag & 0x7)
+		if o+n > len(buf) {
+			return nil, fmt.Errorf("unable to decode extension")
+		}
+		l, err := size((buf)[o+n:], wireType)
+		if err != nil {
+			return nil, err
+		}
+		if int32(fieldNum) == extension.Field {
+			if o+n+l > len(buf) {
+				return nil, fmt.Errorf("unable to decode extension")
+			}
+			v, err := decodeExtension((buf)[o:o+n+l], extension)
+			if err != nil {
+				return nil, err
+			}
+			return v, nil
+		}
+		o += n + l
+	}
+	return defaultExtensionValue(extension)
+}
+
+func (this *Extension) Encode() error {
+	if this.enc == nil {
+		var err error
+		this.enc, err = encodeExtension(this.desc, this.value)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (this Extension) GoString() string {
+	if err := this.Encode(); err != nil {
+		return fmt.Sprintf("error encoding extension: %v", err)
+	}
+	return fmt.Sprintf("proto.NewExtension(%#v)", this.enc)
+}
+
+func SetUnsafeExtension(pb Message, fieldNum int32, value interface{}) error {
+	typ := reflect.TypeOf(pb).Elem()
+	ext, ok := extensionMaps[typ]
+	if !ok {
+		return fmt.Errorf("proto: bad extended type; %s is not extendable", typ.String())
+	}
+	desc, ok := ext[fieldNum]
+	if !ok {
+		return errors.New("proto: bad extension number; not in declared ranges")
+	}
+	return SetExtension(pb, desc, value)
+}
+
+func GetUnsafeExtension(pb Message, fieldNum int32) (interface{}, error) {
+	typ := reflect.TypeOf(pb).Elem()
+	ext, ok := extensionMaps[typ]
+	if !ok {
+		return nil, fmt.Errorf("proto: bad extended type; %s is not extendable", typ.String())
+	}
+	desc, ok := ext[fieldNum]
+	if !ok {
+		return nil, fmt.Errorf("unregistered field number %d", fieldNum)
+	}
+	return GetExtension(pb, desc)
+}
+
+func NewUnsafeXXX_InternalExtensions(m map[int32]Extension) XXX_InternalExtensions {
+	x := &XXX_InternalExtensions{
+		p: new(struct {
+			mu           sync.Mutex
+			extensionMap map[int32]Extension
+		}),
+	}
+	x.p.extensionMap = m
+	return *x
+}
+
+func GetUnsafeExtensionsMap(extendable Message) map[int32]Extension {
+	pb := extendable.(extendableProto)
+	return pb.extensionsWrite()
+}
+
+func deleteExtension(pb extensionsBytes, theFieldNum int32, offset int) int {
+	ext := pb.GetExtensions()
+	for offset < len(*ext) {
+		tag, n1 := DecodeVarint((*ext)[offset:])
+		fieldNum := int32(tag >> 3)
+		wireType := int(tag & 0x7)
+		n2, err := size((*ext)[offset+n1:], wireType)
+		if err != nil {
+			panic(err)
+		}
+		newOffset := offset + n1 + n2
+		if fieldNum == theFieldNum {
+			*ext = append((*ext)[:offset], (*ext)[newOffset:]...)
+			return offset
+		}
+		offset = newOffset
+	}
+	return -1
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/lib.go b/vendor/github.com/gogo/protobuf/proto/lib.go
new file mode 100644
index 0000000..d17f802
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/lib.go
@@ -0,0 +1,967 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+/*
+Package proto converts data structures to and from the wire format of
+protocol buffers.  It works in concert with the Go source code generated
+for .proto files by the protocol compiler.
+
+A summary of the properties of the protocol buffer interface
+for a protocol buffer variable v:
+
+  - Names are turned from camel_case to CamelCase for export.
+  - There are no methods on v to set fields; just treat
+	them as structure fields.
+  - There are getters that return a field's value if set,
+	and return the field's default value if unset.
+	The getters work even if the receiver is a nil message.
+  - The zero value for a struct is its correct initialization state.
+	All desired fields must be set before marshaling.
+  - A Reset() method will restore a protobuf struct to its zero state.
+  - Non-repeated fields are pointers to the values; nil means unset.
+	That is, optional or required field int32 f becomes F *int32.
+  - Repeated fields are slices.
+  - Helper functions are available to aid the setting of fields.
+	msg.Foo = proto.String("hello") // set field
+  - Constants are defined to hold the default values of all fields that
+	have them.  They have the form Default_StructName_FieldName.
+	Because the getter methods handle defaulted values,
+	direct use of these constants should be rare.
+  - Enums are given type names and maps from names to values.
+	Enum values are prefixed by the enclosing message's name, or by the
+	enum's type name if it is a top-level enum. Enum types have a String
+	method, and a Enum method to assist in message construction.
+  - Nested messages, groups and enums have type names prefixed with the name of
+	the surrounding message type.
+  - Extensions are given descriptor names that start with E_,
+	followed by an underscore-delimited list of the nested messages
+	that contain it (if any) followed by the CamelCased name of the
+	extension field itself.  HasExtension, ClearExtension, GetExtension
+	and SetExtension are functions for manipulating extensions.
+  - Oneof field sets are given a single field in their message,
+	with distinguished wrapper types for each possible field value.
+  - Marshal and Unmarshal are functions to encode and decode the wire format.
+
+When the .proto file specifies `syntax="proto3"`, there are some differences:
+
+  - Non-repeated fields of non-message type are values instead of pointers.
+  - Enum types do not get an Enum method.
+
+The simplest way to describe this is to see an example.
+Given file test.proto, containing
+
+	package example;
+
+	enum FOO { X = 17; }
+
+	message Test {
+	  required string label = 1;
+	  optional int32 type = 2 [default=77];
+	  repeated int64 reps = 3;
+	  optional group OptionalGroup = 4 {
+	    required string RequiredField = 5;
+	  }
+	  oneof union {
+	    int32 number = 6;
+	    string name = 7;
+	  }
+	}
+
+The resulting file, test.pb.go, is:
+
+	package example
+
+	import proto "github.com/gogo/protobuf/proto"
+	import math "math"
+
+	type FOO int32
+	const (
+		FOO_X FOO = 17
+	)
+	var FOO_name = map[int32]string{
+		17: "X",
+	}
+	var FOO_value = map[string]int32{
+		"X": 17,
+	}
+
+	func (x FOO) Enum() *FOO {
+		p := new(FOO)
+		*p = x
+		return p
+	}
+	func (x FOO) String() string {
+		return proto.EnumName(FOO_name, int32(x))
+	}
+	func (x *FOO) UnmarshalJSON(data []byte) error {
+		value, err := proto.UnmarshalJSONEnum(FOO_value, data)
+		if err != nil {
+			return err
+		}
+		*x = FOO(value)
+		return nil
+	}
+
+	type Test struct {
+		Label         *string             `protobuf:"bytes,1,req,name=label" json:"label,omitempty"`
+		Type          *int32              `protobuf:"varint,2,opt,name=type,def=77" json:"type,omitempty"`
+		Reps          []int64             `protobuf:"varint,3,rep,name=reps" json:"reps,omitempty"`
+		Optionalgroup *Test_OptionalGroup `protobuf:"group,4,opt,name=OptionalGroup" json:"optionalgroup,omitempty"`
+		// Types that are valid to be assigned to Union:
+		//	*Test_Number
+		//	*Test_Name
+		Union            isTest_Union `protobuf_oneof:"union"`
+		XXX_unrecognized []byte       `json:"-"`
+	}
+	func (m *Test) Reset()         { *m = Test{} }
+	func (m *Test) String() string { return proto.CompactTextString(m) }
+	func (*Test) ProtoMessage() {}
+
+	type isTest_Union interface {
+		isTest_Union()
+	}
+
+	type Test_Number struct {
+		Number int32 `protobuf:"varint,6,opt,name=number"`
+	}
+	type Test_Name struct {
+		Name string `protobuf:"bytes,7,opt,name=name"`
+	}
+
+	func (*Test_Number) isTest_Union() {}
+	func (*Test_Name) isTest_Union()   {}
+
+	func (m *Test) GetUnion() isTest_Union {
+		if m != nil {
+			return m.Union
+		}
+		return nil
+	}
+	const Default_Test_Type int32 = 77
+
+	func (m *Test) GetLabel() string {
+		if m != nil && m.Label != nil {
+			return *m.Label
+		}
+		return ""
+	}
+
+	func (m *Test) GetType() int32 {
+		if m != nil && m.Type != nil {
+			return *m.Type
+		}
+		return Default_Test_Type
+	}
+
+	func (m *Test) GetOptionalgroup() *Test_OptionalGroup {
+		if m != nil {
+			return m.Optionalgroup
+		}
+		return nil
+	}
+
+	type Test_OptionalGroup struct {
+		RequiredField *string `protobuf:"bytes,5,req" json:"RequiredField,omitempty"`
+	}
+	func (m *Test_OptionalGroup) Reset()         { *m = Test_OptionalGroup{} }
+	func (m *Test_OptionalGroup) String() string { return proto.CompactTextString(m) }
+
+	func (m *Test_OptionalGroup) GetRequiredField() string {
+		if m != nil && m.RequiredField != nil {
+			return *m.RequiredField
+		}
+		return ""
+	}
+
+	func (m *Test) GetNumber() int32 {
+		if x, ok := m.GetUnion().(*Test_Number); ok {
+			return x.Number
+		}
+		return 0
+	}
+
+	func (m *Test) GetName() string {
+		if x, ok := m.GetUnion().(*Test_Name); ok {
+			return x.Name
+		}
+		return ""
+	}
+
+	func init() {
+		proto.RegisterEnum("example.FOO", FOO_name, FOO_value)
+	}
+
+To create and play with a Test object:
+
+	package main
+
+	import (
+		"log"
+
+		"github.com/gogo/protobuf/proto"
+		pb "./example.pb"
+	)
+
+	func main() {
+		test := &pb.Test{
+			Label: proto.String("hello"),
+			Type:  proto.Int32(17),
+			Reps:  []int64{1, 2, 3},
+			Optionalgroup: &pb.Test_OptionalGroup{
+				RequiredField: proto.String("good bye"),
+			},
+			Union: &pb.Test_Name{"fred"},
+		}
+		data, err := proto.Marshal(test)
+		if err != nil {
+			log.Fatal("marshaling error: ", err)
+		}
+		newTest := &pb.Test{}
+		err = proto.Unmarshal(data, newTest)
+		if err != nil {
+			log.Fatal("unmarshaling error: ", err)
+		}
+		// Now test and newTest contain the same data.
+		if test.GetLabel() != newTest.GetLabel() {
+			log.Fatalf("data mismatch %q != %q", test.GetLabel(), newTest.GetLabel())
+		}
+		// Use a type switch to determine which oneof was set.
+		switch u := test.Union.(type) {
+		case *pb.Test_Number: // u.Number contains the number.
+		case *pb.Test_Name: // u.Name contains the string.
+		}
+		// etc.
+	}
+*/
+package proto
+
+import (
+	"encoding/json"
+	"fmt"
+	"log"
+	"reflect"
+	"sort"
+	"strconv"
+	"sync"
+)
+
+// RequiredNotSetError is an error type returned by either Marshal or Unmarshal.
+// Marshal reports this when a required field is not initialized.
+// Unmarshal reports this when a required field is missing from the wire data.
+type RequiredNotSetError struct{ field string }
+
+func (e *RequiredNotSetError) Error() string {
+	if e.field == "" {
+		return fmt.Sprintf("proto: required field not set")
+	}
+	return fmt.Sprintf("proto: required field %q not set", e.field)
+}
+func (e *RequiredNotSetError) RequiredNotSet() bool {
+	return true
+}
+
+type invalidUTF8Error struct{ field string }
+
+func (e *invalidUTF8Error) Error() string {
+	if e.field == "" {
+		return "proto: invalid UTF-8 detected"
+	}
+	return fmt.Sprintf("proto: field %q contains invalid UTF-8", e.field)
+}
+func (e *invalidUTF8Error) InvalidUTF8() bool {
+	return true
+}
+
+// errInvalidUTF8 is a sentinel error to identify fields with invalid UTF-8.
+// This error should not be exposed to the external API as such errors should
+// be recreated with the field information.
+var errInvalidUTF8 = &invalidUTF8Error{}
+
+// isNonFatal reports whether the error is either a RequiredNotSet error
+// or a InvalidUTF8 error.
+func isNonFatal(err error) bool {
+	if re, ok := err.(interface{ RequiredNotSet() bool }); ok && re.RequiredNotSet() {
+		return true
+	}
+	if re, ok := err.(interface{ InvalidUTF8() bool }); ok && re.InvalidUTF8() {
+		return true
+	}
+	return false
+}
+
+type nonFatal struct{ E error }
+
+// Merge merges err into nf and reports whether it was successful.
+// Otherwise it returns false for any fatal non-nil errors.
+func (nf *nonFatal) Merge(err error) (ok bool) {
+	if err == nil {
+		return true // not an error
+	}
+	if !isNonFatal(err) {
+		return false // fatal error
+	}
+	if nf.E == nil {
+		nf.E = err // store first instance of non-fatal error
+	}
+	return true
+}
+
+// Message is implemented by generated protocol buffer messages.
+type Message interface {
+	Reset()
+	String() string
+	ProtoMessage()
+}
+
+// A Buffer is a buffer manager for marshaling and unmarshaling
+// protocol buffers.  It may be reused between invocations to
+// reduce memory usage.  It is not necessary to use a Buffer;
+// the global functions Marshal and Unmarshal create a
+// temporary Buffer and are fine for most applications.
+type Buffer struct {
+	buf   []byte // encode/decode byte stream
+	index int    // read point
+
+	deterministic bool
+}
+
+// NewBuffer allocates a new Buffer and initializes its internal data to
+// the contents of the argument slice.
+func NewBuffer(e []byte) *Buffer {
+	return &Buffer{buf: e}
+}
+
+// Reset resets the Buffer, ready for marshaling a new protocol buffer.
+func (p *Buffer) Reset() {
+	p.buf = p.buf[0:0] // for reading/writing
+	p.index = 0        // for reading
+}
+
+// SetBuf replaces the internal buffer with the slice,
+// ready for unmarshaling the contents of the slice.
+func (p *Buffer) SetBuf(s []byte) {
+	p.buf = s
+	p.index = 0
+}
+
+// Bytes returns the contents of the Buffer.
+func (p *Buffer) Bytes() []byte { return p.buf }
+
+// SetDeterministic sets whether to use deterministic serialization.
+//
+// Deterministic serialization guarantees that for a given binary, equal
+// messages will always be serialized to the same bytes. This implies:
+//
+//   - Repeated serialization of a message will return the same bytes.
+//   - Different processes of the same binary (which may be executing on
+//     different machines) will serialize equal messages to the same bytes.
+//
+// Note that the deterministic serialization is NOT canonical across
+// languages. It is not guaranteed to remain stable over time. It is unstable
+// across different builds with schema changes due to unknown fields.
+// Users who need canonical serialization (e.g., persistent storage in a
+// canonical form, fingerprinting, etc.) should define their own
+// canonicalization specification and implement their own serializer rather
+// than relying on this API.
+//
+// If deterministic serialization is requested, map entries will be sorted
+// by keys in lexographical order. This is an implementation detail and
+// subject to change.
+func (p *Buffer) SetDeterministic(deterministic bool) {
+	p.deterministic = deterministic
+}
+
+/*
+ * Helper routines for simplifying the creation of optional fields of basic type.
+ */
+
+// Bool is a helper routine that allocates a new bool value
+// to store v and returns a pointer to it.
+func Bool(v bool) *bool {
+	return &v
+}
+
+// Int32 is a helper routine that allocates a new int32 value
+// to store v and returns a pointer to it.
+func Int32(v int32) *int32 {
+	return &v
+}
+
+// Int is a helper routine that allocates a new int32 value
+// to store v and returns a pointer to it, but unlike Int32
+// its argument value is an int.
+func Int(v int) *int32 {
+	p := new(int32)
+	*p = int32(v)
+	return p
+}
+
+// Int64 is a helper routine that allocates a new int64 value
+// to store v and returns a pointer to it.
+func Int64(v int64) *int64 {
+	return &v
+}
+
+// Float32 is a helper routine that allocates a new float32 value
+// to store v and returns a pointer to it.
+func Float32(v float32) *float32 {
+	return &v
+}
+
+// Float64 is a helper routine that allocates a new float64 value
+// to store v and returns a pointer to it.
+func Float64(v float64) *float64 {
+	return &v
+}
+
+// Uint32 is a helper routine that allocates a new uint32 value
+// to store v and returns a pointer to it.
+func Uint32(v uint32) *uint32 {
+	return &v
+}
+
+// Uint64 is a helper routine that allocates a new uint64 value
+// to store v and returns a pointer to it.
+func Uint64(v uint64) *uint64 {
+	return &v
+}
+
+// String is a helper routine that allocates a new string value
+// to store v and returns a pointer to it.
+func String(v string) *string {
+	return &v
+}
+
+// EnumName is a helper function to simplify printing protocol buffer enums
+// by name.  Given an enum map and a value, it returns a useful string.
+func EnumName(m map[int32]string, v int32) string {
+	s, ok := m[v]
+	if ok {
+		return s
+	}
+	return strconv.Itoa(int(v))
+}
+
+// UnmarshalJSONEnum is a helper function to simplify recovering enum int values
+// from their JSON-encoded representation. Given a map from the enum's symbolic
+// names to its int values, and a byte buffer containing the JSON-encoded
+// value, it returns an int32 that can be cast to the enum type by the caller.
+//
+// The function can deal with both JSON representations, numeric and symbolic.
+func UnmarshalJSONEnum(m map[string]int32, data []byte, enumName string) (int32, error) {
+	if data[0] == '"' {
+		// New style: enums are strings.
+		var repr string
+		if err := json.Unmarshal(data, &repr); err != nil {
+			return -1, err
+		}
+		val, ok := m[repr]
+		if !ok {
+			return 0, fmt.Errorf("unrecognized enum %s value %q", enumName, repr)
+		}
+		return val, nil
+	}
+	// Old style: enums are ints.
+	var val int32
+	if err := json.Unmarshal(data, &val); err != nil {
+		return 0, fmt.Errorf("cannot unmarshal %#q into enum %s", data, enumName)
+	}
+	return val, nil
+}
+
+// DebugPrint dumps the encoded data in b in a debugging format with a header
+// including the string s. Used in testing but made available for general debugging.
+func (p *Buffer) DebugPrint(s string, b []byte) {
+	var u uint64
+
+	obuf := p.buf
+	sindex := p.index
+	p.buf = b
+	p.index = 0
+	depth := 0
+
+	fmt.Printf("\n--- %s ---\n", s)
+
+out:
+	for {
+		for i := 0; i < depth; i++ {
+			fmt.Print("  ")
+		}
+
+		index := p.index
+		if index == len(p.buf) {
+			break
+		}
+
+		op, err := p.DecodeVarint()
+		if err != nil {
+			fmt.Printf("%3d: fetching op err %v\n", index, err)
+			break out
+		}
+		tag := op >> 3
+		wire := op & 7
+
+		switch wire {
+		default:
+			fmt.Printf("%3d: t=%3d unknown wire=%d\n",
+				index, tag, wire)
+			break out
+
+		case WireBytes:
+			var r []byte
+
+			r, err = p.DecodeRawBytes(false)
+			if err != nil {
+				break out
+			}
+			fmt.Printf("%3d: t=%3d bytes [%d]", index, tag, len(r))
+			if len(r) <= 6 {
+				for i := 0; i < len(r); i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+			} else {
+				for i := 0; i < 3; i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+				fmt.Printf(" ..")
+				for i := len(r) - 3; i < len(r); i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+			}
+			fmt.Printf("\n")
+
+		case WireFixed32:
+			u, err = p.DecodeFixed32()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d fix32 err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d fix32 %d\n", index, tag, u)
+
+		case WireFixed64:
+			u, err = p.DecodeFixed64()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d fix64 err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d fix64 %d\n", index, tag, u)
+
+		case WireVarint:
+			u, err = p.DecodeVarint()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d varint err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d varint %d\n", index, tag, u)
+
+		case WireStartGroup:
+			fmt.Printf("%3d: t=%3d start\n", index, tag)
+			depth++
+
+		case WireEndGroup:
+			depth--
+			fmt.Printf("%3d: t=%3d end\n", index, tag)
+		}
+	}
+
+	if depth != 0 {
+		fmt.Printf("%3d: start-end not balanced %d\n", p.index, depth)
+	}
+	fmt.Printf("\n")
+
+	p.buf = obuf
+	p.index = sindex
+}
+
+// SetDefaults sets unset protocol buffer fields to their default values.
+// It only modifies fields that are both unset and have defined defaults.
+// It recursively sets default values in any non-nil sub-messages.
+func SetDefaults(pb Message) {
+	setDefaults(reflect.ValueOf(pb), true, false)
+}
+
+// v is a struct.
+func setDefaults(v reflect.Value, recur, zeros bool) {
+	if v.Kind() == reflect.Ptr {
+		v = v.Elem()
+	}
+
+	defaultMu.RLock()
+	dm, ok := defaults[v.Type()]
+	defaultMu.RUnlock()
+	if !ok {
+		dm = buildDefaultMessage(v.Type())
+		defaultMu.Lock()
+		defaults[v.Type()] = dm
+		defaultMu.Unlock()
+	}
+
+	for _, sf := range dm.scalars {
+		f := v.Field(sf.index)
+		if !f.IsNil() {
+			// field already set
+			continue
+		}
+		dv := sf.value
+		if dv == nil && !zeros {
+			// no explicit default, and don't want to set zeros
+			continue
+		}
+		fptr := f.Addr().Interface() // **T
+		// TODO: Consider batching the allocations we do here.
+		switch sf.kind {
+		case reflect.Bool:
+			b := new(bool)
+			if dv != nil {
+				*b = dv.(bool)
+			}
+			*(fptr.(**bool)) = b
+		case reflect.Float32:
+			f := new(float32)
+			if dv != nil {
+				*f = dv.(float32)
+			}
+			*(fptr.(**float32)) = f
+		case reflect.Float64:
+			f := new(float64)
+			if dv != nil {
+				*f = dv.(float64)
+			}
+			*(fptr.(**float64)) = f
+		case reflect.Int32:
+			// might be an enum
+			if ft := f.Type(); ft != int32PtrType {
+				// enum
+				f.Set(reflect.New(ft.Elem()))
+				if dv != nil {
+					f.Elem().SetInt(int64(dv.(int32)))
+				}
+			} else {
+				// int32 field
+				i := new(int32)
+				if dv != nil {
+					*i = dv.(int32)
+				}
+				*(fptr.(**int32)) = i
+			}
+		case reflect.Int64:
+			i := new(int64)
+			if dv != nil {
+				*i = dv.(int64)
+			}
+			*(fptr.(**int64)) = i
+		case reflect.String:
+			s := new(string)
+			if dv != nil {
+				*s = dv.(string)
+			}
+			*(fptr.(**string)) = s
+		case reflect.Uint8:
+			// exceptional case: []byte
+			var b []byte
+			if dv != nil {
+				db := dv.([]byte)
+				b = make([]byte, len(db))
+				copy(b, db)
+			} else {
+				b = []byte{}
+			}
+			*(fptr.(*[]byte)) = b
+		case reflect.Uint32:
+			u := new(uint32)
+			if dv != nil {
+				*u = dv.(uint32)
+			}
+			*(fptr.(**uint32)) = u
+		case reflect.Uint64:
+			u := new(uint64)
+			if dv != nil {
+				*u = dv.(uint64)
+			}
+			*(fptr.(**uint64)) = u
+		default:
+			log.Printf("proto: can't set default for field %v (sf.kind=%v)", f, sf.kind)
+		}
+	}
+
+	for _, ni := range dm.nested {
+		f := v.Field(ni)
+		// f is *T or T or []*T or []T
+		switch f.Kind() {
+		case reflect.Struct:
+			setDefaults(f, recur, zeros)
+
+		case reflect.Ptr:
+			if f.IsNil() {
+				continue
+			}
+			setDefaults(f, recur, zeros)
+
+		case reflect.Slice:
+			for i := 0; i < f.Len(); i++ {
+				e := f.Index(i)
+				if e.Kind() == reflect.Ptr && e.IsNil() {
+					continue
+				}
+				setDefaults(e, recur, zeros)
+			}
+
+		case reflect.Map:
+			for _, k := range f.MapKeys() {
+				e := f.MapIndex(k)
+				if e.IsNil() {
+					continue
+				}
+				setDefaults(e, recur, zeros)
+			}
+		}
+	}
+}
+
+var (
+	// defaults maps a protocol buffer struct type to a slice of the fields,
+	// with its scalar fields set to their proto-declared non-zero default values.
+	defaultMu sync.RWMutex
+	defaults  = make(map[reflect.Type]defaultMessage)
+
+	int32PtrType = reflect.TypeOf((*int32)(nil))
+)
+
+// defaultMessage represents information about the default values of a message.
+type defaultMessage struct {
+	scalars []scalarField
+	nested  []int // struct field index of nested messages
+}
+
+type scalarField struct {
+	index int          // struct field index
+	kind  reflect.Kind // element type (the T in *T or []T)
+	value interface{}  // the proto-declared default value, or nil
+}
+
+// t is a struct type.
+func buildDefaultMessage(t reflect.Type) (dm defaultMessage) {
+	sprop := GetProperties(t)
+	for _, prop := range sprop.Prop {
+		fi, ok := sprop.decoderTags.get(prop.Tag)
+		if !ok {
+			// XXX_unrecognized
+			continue
+		}
+		ft := t.Field(fi).Type
+
+		sf, nested, err := fieldDefault(ft, prop)
+		switch {
+		case err != nil:
+			log.Print(err)
+		case nested:
+			dm.nested = append(dm.nested, fi)
+		case sf != nil:
+			sf.index = fi
+			dm.scalars = append(dm.scalars, *sf)
+		}
+	}
+
+	return dm
+}
+
+// fieldDefault returns the scalarField for field type ft.
+// sf will be nil if the field can not have a default.
+// nestedMessage will be true if this is a nested message.
+// Note that sf.index is not set on return.
+func fieldDefault(ft reflect.Type, prop *Properties) (sf *scalarField, nestedMessage bool, err error) {
+	var canHaveDefault bool
+	switch ft.Kind() {
+	case reflect.Struct:
+		nestedMessage = true // non-nullable
+
+	case reflect.Ptr:
+		if ft.Elem().Kind() == reflect.Struct {
+			nestedMessage = true
+		} else {
+			canHaveDefault = true // proto2 scalar field
+		}
+
+	case reflect.Slice:
+		switch ft.Elem().Kind() {
+		case reflect.Ptr, reflect.Struct:
+			nestedMessage = true // repeated message
+		case reflect.Uint8:
+			canHaveDefault = true // bytes field
+		}
+
+	case reflect.Map:
+		if ft.Elem().Kind() == reflect.Ptr {
+			nestedMessage = true // map with message values
+		}
+	}
+
+	if !canHaveDefault {
+		if nestedMessage {
+			return nil, true, nil
+		}
+		return nil, false, nil
+	}
+
+	// We now know that ft is a pointer or slice.
+	sf = &scalarField{kind: ft.Elem().Kind()}
+
+	// scalar fields without defaults
+	if !prop.HasDefault {
+		return sf, false, nil
+	}
+
+	// a scalar field: either *T or []byte
+	switch ft.Elem().Kind() {
+	case reflect.Bool:
+		x, err := strconv.ParseBool(prop.Default)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default bool %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.Float32:
+		x, err := strconv.ParseFloat(prop.Default, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default float32 %q: %v", prop.Default, err)
+		}
+		sf.value = float32(x)
+	case reflect.Float64:
+		x, err := strconv.ParseFloat(prop.Default, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default float64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.Int32:
+		x, err := strconv.ParseInt(prop.Default, 10, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default int32 %q: %v", prop.Default, err)
+		}
+		sf.value = int32(x)
+	case reflect.Int64:
+		x, err := strconv.ParseInt(prop.Default, 10, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default int64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.String:
+		sf.value = prop.Default
+	case reflect.Uint8:
+		// []byte (not *uint8)
+		sf.value = []byte(prop.Default)
+	case reflect.Uint32:
+		x, err := strconv.ParseUint(prop.Default, 10, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default uint32 %q: %v", prop.Default, err)
+		}
+		sf.value = uint32(x)
+	case reflect.Uint64:
+		x, err := strconv.ParseUint(prop.Default, 10, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default uint64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	default:
+		return nil, false, fmt.Errorf("proto: unhandled def kind %v", ft.Elem().Kind())
+	}
+
+	return sf, false, nil
+}
+
+// mapKeys returns a sort.Interface to be used for sorting the map keys.
+// Map fields may have key types of non-float scalars, strings and enums.
+func mapKeys(vs []reflect.Value) sort.Interface {
+	s := mapKeySorter{vs: vs}
+
+	// Type specialization per https://developers.google.com/protocol-buffers/docs/proto#maps.
+	if len(vs) == 0 {
+		return s
+	}
+	switch vs[0].Kind() {
+	case reflect.Int32, reflect.Int64:
+		s.less = func(a, b reflect.Value) bool { return a.Int() < b.Int() }
+	case reflect.Uint32, reflect.Uint64:
+		s.less = func(a, b reflect.Value) bool { return a.Uint() < b.Uint() }
+	case reflect.Bool:
+		s.less = func(a, b reflect.Value) bool { return !a.Bool() && b.Bool() } // false < true
+	case reflect.String:
+		s.less = func(a, b reflect.Value) bool { return a.String() < b.String() }
+	default:
+		panic(fmt.Sprintf("unsupported map key type: %v", vs[0].Kind()))
+	}
+
+	return s
+}
+
+type mapKeySorter struct {
+	vs   []reflect.Value
+	less func(a, b reflect.Value) bool
+}
+
+func (s mapKeySorter) Len() int      { return len(s.vs) }
+func (s mapKeySorter) Swap(i, j int) { s.vs[i], s.vs[j] = s.vs[j], s.vs[i] }
+func (s mapKeySorter) Less(i, j int) bool {
+	return s.less(s.vs[i], s.vs[j])
+}
+
+// isProto3Zero reports whether v is a zero proto3 value.
+func isProto3Zero(v reflect.Value) bool {
+	switch v.Kind() {
+	case reflect.Bool:
+		return !v.Bool()
+	case reflect.Int32, reflect.Int64:
+		return v.Int() == 0
+	case reflect.Uint32, reflect.Uint64:
+		return v.Uint() == 0
+	case reflect.Float32, reflect.Float64:
+		return v.Float() == 0
+	case reflect.String:
+		return v.String() == ""
+	}
+	return false
+}
+
+// ProtoPackageIsVersion2 is referenced from generated protocol buffer files
+// to assert that that code is compatible with this version of the proto package.
+const GoGoProtoPackageIsVersion2 = true
+
+// ProtoPackageIsVersion1 is referenced from generated protocol buffer files
+// to assert that that code is compatible with this version of the proto package.
+const GoGoProtoPackageIsVersion1 = true
+
+// InternalMessageInfo is a type used internally by generated .pb.go files.
+// This type is not intended to be used by non-generated code.
+// This type is not subject to any compatibility guarantee.
+type InternalMessageInfo struct {
+	marshal   *marshalInfo
+	unmarshal *unmarshalInfo
+	merge     *mergeInfo
+	discard   *discardInfo
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/lib_gogo.go b/vendor/github.com/gogo/protobuf/proto/lib_gogo.go
new file mode 100644
index 0000000..b3aa391
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/lib_gogo.go
@@ -0,0 +1,50 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"encoding/json"
+	"strconv"
+)
+
+type Sizer interface {
+	Size() int
+}
+
+type ProtoSizer interface {
+	ProtoSize() int
+}
+
+func MarshalJSONEnum(m map[int32]string, value int32) ([]byte, error) {
+	s, ok := m[value]
+	if !ok {
+		s = strconv.Itoa(int(value))
+	}
+	return json.Marshal(s)
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/message_set.go b/vendor/github.com/gogo/protobuf/proto/message_set.go
new file mode 100644
index 0000000..f48a756
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/message_set.go
@@ -0,0 +1,181 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Support for message sets.
+ */
+
+import (
+	"errors"
+)
+
+// errNoMessageTypeID occurs when a protocol buffer does not have a message type ID.
+// A message type ID is required for storing a protocol buffer in a message set.
+var errNoMessageTypeID = errors.New("proto does not have a message type ID")
+
+// The first two types (_MessageSet_Item and messageSet)
+// model what the protocol compiler produces for the following protocol message:
+//   message MessageSet {
+//     repeated group Item = 1 {
+//       required int32 type_id = 2;
+//       required string message = 3;
+//     };
+//   }
+// That is the MessageSet wire format. We can't use a proto to generate these
+// because that would introduce a circular dependency between it and this package.
+
+type _MessageSet_Item struct {
+	TypeId  *int32 `protobuf:"varint,2,req,name=type_id"`
+	Message []byte `protobuf:"bytes,3,req,name=message"`
+}
+
+type messageSet struct {
+	Item             []*_MessageSet_Item `protobuf:"group,1,rep"`
+	XXX_unrecognized []byte
+	// TODO: caching?
+}
+
+// Make sure messageSet is a Message.
+var _ Message = (*messageSet)(nil)
+
+// messageTypeIder is an interface satisfied by a protocol buffer type
+// that may be stored in a MessageSet.
+type messageTypeIder interface {
+	MessageTypeId() int32
+}
+
+func (ms *messageSet) find(pb Message) *_MessageSet_Item {
+	mti, ok := pb.(messageTypeIder)
+	if !ok {
+		return nil
+	}
+	id := mti.MessageTypeId()
+	for _, item := range ms.Item {
+		if *item.TypeId == id {
+			return item
+		}
+	}
+	return nil
+}
+
+func (ms *messageSet) Has(pb Message) bool {
+	return ms.find(pb) != nil
+}
+
+func (ms *messageSet) Unmarshal(pb Message) error {
+	if item := ms.find(pb); item != nil {
+		return Unmarshal(item.Message, pb)
+	}
+	if _, ok := pb.(messageTypeIder); !ok {
+		return errNoMessageTypeID
+	}
+	return nil // TODO: return error instead?
+}
+
+func (ms *messageSet) Marshal(pb Message) error {
+	msg, err := Marshal(pb)
+	if err != nil {
+		return err
+	}
+	if item := ms.find(pb); item != nil {
+		// reuse existing item
+		item.Message = msg
+		return nil
+	}
+
+	mti, ok := pb.(messageTypeIder)
+	if !ok {
+		return errNoMessageTypeID
+	}
+
+	mtid := mti.MessageTypeId()
+	ms.Item = append(ms.Item, &_MessageSet_Item{
+		TypeId:  &mtid,
+		Message: msg,
+	})
+	return nil
+}
+
+func (ms *messageSet) Reset()         { *ms = messageSet{} }
+func (ms *messageSet) String() string { return CompactTextString(ms) }
+func (*messageSet) ProtoMessage()     {}
+
+// Support for the message_set_wire_format message option.
+
+func skipVarint(buf []byte) []byte {
+	i := 0
+	for ; buf[i]&0x80 != 0; i++ {
+	}
+	return buf[i+1:]
+}
+
+// unmarshalMessageSet decodes the extension map encoded in buf in the message set wire format.
+// It is called by Unmarshal methods on protocol buffer messages with the message_set_wire_format option.
+func unmarshalMessageSet(buf []byte, exts interface{}) error {
+	var m map[int32]Extension
+	switch exts := exts.(type) {
+	case *XXX_InternalExtensions:
+		m = exts.extensionsWrite()
+	case map[int32]Extension:
+		m = exts
+	default:
+		return errors.New("proto: not an extension map")
+	}
+
+	ms := new(messageSet)
+	if err := Unmarshal(buf, ms); err != nil {
+		return err
+	}
+	for _, item := range ms.Item {
+		id := *item.TypeId
+		msg := item.Message
+
+		// Restore wire type and field number varint, plus length varint.
+		// Be careful to preserve duplicate items.
+		b := EncodeVarint(uint64(id)<<3 | WireBytes)
+		if ext, ok := m[id]; ok {
+			// Existing data; rip off the tag and length varint
+			// so we join the new data correctly.
+			// We can assume that ext.enc is set because we are unmarshaling.
+			o := ext.enc[len(b):]   // skip wire type and field number
+			_, n := DecodeVarint(o) // calculate length of length varint
+			o = o[n:]               // skip length varint
+			msg = append(o, msg...) // join old data and new data
+		}
+		b = append(b, EncodeVarint(uint64(len(msg)))...)
+		b = append(b, msg...)
+
+		m[id] = Extension{enc: b}
+	}
+	return nil
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/pointer_reflect.go b/vendor/github.com/gogo/protobuf/proto/pointer_reflect.go
new file mode 100644
index 0000000..b6cad90
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/pointer_reflect.go
@@ -0,0 +1,357 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2012 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build purego appengine js
+
+// This file contains an implementation of proto field accesses using package reflect.
+// It is slower than the code in pointer_unsafe.go but it avoids package unsafe and can
+// be used on App Engine.
+
+package proto
+
+import (
+	"reflect"
+	"sync"
+)
+
+const unsafeAllowed = false
+
+// A field identifies a field in a struct, accessible from a pointer.
+// In this implementation, a field is identified by the sequence of field indices
+// passed to reflect's FieldByIndex.
+type field []int
+
+// toField returns a field equivalent to the given reflect field.
+func toField(f *reflect.StructField) field {
+	return f.Index
+}
+
+// invalidField is an invalid field identifier.
+var invalidField = field(nil)
+
+// zeroField is a noop when calling pointer.offset.
+var zeroField = field([]int{})
+
+// IsValid reports whether the field identifier is valid.
+func (f field) IsValid() bool { return f != nil }
+
+// The pointer type is for the table-driven decoder.
+// The implementation here uses a reflect.Value of pointer type to
+// create a generic pointer. In pointer_unsafe.go we use unsafe
+// instead of reflect to implement the same (but faster) interface.
+type pointer struct {
+	v reflect.Value
+}
+
+// toPointer converts an interface of pointer type to a pointer
+// that points to the same target.
+func toPointer(i *Message) pointer {
+	return pointer{v: reflect.ValueOf(*i)}
+}
+
+// toAddrPointer converts an interface to a pointer that points to
+// the interface data.
+func toAddrPointer(i *interface{}, isptr bool) pointer {
+	v := reflect.ValueOf(*i)
+	u := reflect.New(v.Type())
+	u.Elem().Set(v)
+	return pointer{v: u}
+}
+
+// valToPointer converts v to a pointer.  v must be of pointer type.
+func valToPointer(v reflect.Value) pointer {
+	return pointer{v: v}
+}
+
+// offset converts from a pointer to a structure to a pointer to
+// one of its fields.
+func (p pointer) offset(f field) pointer {
+	return pointer{v: p.v.Elem().FieldByIndex(f).Addr()}
+}
+
+func (p pointer) isNil() bool {
+	return p.v.IsNil()
+}
+
+// grow updates the slice s in place to make it one element longer.
+// s must be addressable.
+// Returns the (addressable) new element.
+func grow(s reflect.Value) reflect.Value {
+	n, m := s.Len(), s.Cap()
+	if n < m {
+		s.SetLen(n + 1)
+	} else {
+		s.Set(reflect.Append(s, reflect.Zero(s.Type().Elem())))
+	}
+	return s.Index(n)
+}
+
+func (p pointer) toInt64() *int64 {
+	return p.v.Interface().(*int64)
+}
+func (p pointer) toInt64Ptr() **int64 {
+	return p.v.Interface().(**int64)
+}
+func (p pointer) toInt64Slice() *[]int64 {
+	return p.v.Interface().(*[]int64)
+}
+
+var int32ptr = reflect.TypeOf((*int32)(nil))
+
+func (p pointer) toInt32() *int32 {
+	return p.v.Convert(int32ptr).Interface().(*int32)
+}
+
+// The toInt32Ptr/Slice methods don't work because of enums.
+// Instead, we must use set/get methods for the int32ptr/slice case.
+/*
+	func (p pointer) toInt32Ptr() **int32 {
+		return p.v.Interface().(**int32)
+}
+	func (p pointer) toInt32Slice() *[]int32 {
+		return p.v.Interface().(*[]int32)
+}
+*/
+func (p pointer) getInt32Ptr() *int32 {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		return p.v.Elem().Interface().(*int32)
+	}
+	// an enum
+	return p.v.Elem().Convert(int32PtrType).Interface().(*int32)
+}
+func (p pointer) setInt32Ptr(v int32) {
+	// Allocate value in a *int32. Possibly convert that to a *enum.
+	// Then assign it to a **int32 or **enum.
+	// Note: we can convert *int32 to *enum, but we can't convert
+	// **int32 to **enum!
+	p.v.Elem().Set(reflect.ValueOf(&v).Convert(p.v.Type().Elem()))
+}
+
+// getInt32Slice copies []int32 from p as a new slice.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) getInt32Slice() []int32 {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		return p.v.Elem().Interface().([]int32)
+	}
+	// an enum
+	// Allocate a []int32, then assign []enum's values into it.
+	// Note: we can't convert []enum to []int32.
+	slice := p.v.Elem()
+	s := make([]int32, slice.Len())
+	for i := 0; i < slice.Len(); i++ {
+		s[i] = int32(slice.Index(i).Int())
+	}
+	return s
+}
+
+// setInt32Slice copies []int32 into p as a new slice.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) setInt32Slice(v []int32) {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		p.v.Elem().Set(reflect.ValueOf(v))
+		return
+	}
+	// an enum
+	// Allocate a []enum, then assign []int32's values into it.
+	// Note: we can't convert []enum to []int32.
+	slice := reflect.MakeSlice(p.v.Type().Elem(), len(v), cap(v))
+	for i, x := range v {
+		slice.Index(i).SetInt(int64(x))
+	}
+	p.v.Elem().Set(slice)
+}
+func (p pointer) appendInt32Slice(v int32) {
+	grow(p.v.Elem()).SetInt(int64(v))
+}
+
+func (p pointer) toUint64() *uint64 {
+	return p.v.Interface().(*uint64)
+}
+func (p pointer) toUint64Ptr() **uint64 {
+	return p.v.Interface().(**uint64)
+}
+func (p pointer) toUint64Slice() *[]uint64 {
+	return p.v.Interface().(*[]uint64)
+}
+func (p pointer) toUint32() *uint32 {
+	return p.v.Interface().(*uint32)
+}
+func (p pointer) toUint32Ptr() **uint32 {
+	return p.v.Interface().(**uint32)
+}
+func (p pointer) toUint32Slice() *[]uint32 {
+	return p.v.Interface().(*[]uint32)
+}
+func (p pointer) toBool() *bool {
+	return p.v.Interface().(*bool)
+}
+func (p pointer) toBoolPtr() **bool {
+	return p.v.Interface().(**bool)
+}
+func (p pointer) toBoolSlice() *[]bool {
+	return p.v.Interface().(*[]bool)
+}
+func (p pointer) toFloat64() *float64 {
+	return p.v.Interface().(*float64)
+}
+func (p pointer) toFloat64Ptr() **float64 {
+	return p.v.Interface().(**float64)
+}
+func (p pointer) toFloat64Slice() *[]float64 {
+	return p.v.Interface().(*[]float64)
+}
+func (p pointer) toFloat32() *float32 {
+	return p.v.Interface().(*float32)
+}
+func (p pointer) toFloat32Ptr() **float32 {
+	return p.v.Interface().(**float32)
+}
+func (p pointer) toFloat32Slice() *[]float32 {
+	return p.v.Interface().(*[]float32)
+}
+func (p pointer) toString() *string {
+	return p.v.Interface().(*string)
+}
+func (p pointer) toStringPtr() **string {
+	return p.v.Interface().(**string)
+}
+func (p pointer) toStringSlice() *[]string {
+	return p.v.Interface().(*[]string)
+}
+func (p pointer) toBytes() *[]byte {
+	return p.v.Interface().(*[]byte)
+}
+func (p pointer) toBytesSlice() *[][]byte {
+	return p.v.Interface().(*[][]byte)
+}
+func (p pointer) toExtensions() *XXX_InternalExtensions {
+	return p.v.Interface().(*XXX_InternalExtensions)
+}
+func (p pointer) toOldExtensions() *map[int32]Extension {
+	return p.v.Interface().(*map[int32]Extension)
+}
+func (p pointer) getPointer() pointer {
+	return pointer{v: p.v.Elem()}
+}
+func (p pointer) setPointer(q pointer) {
+	p.v.Elem().Set(q.v)
+}
+func (p pointer) appendPointer(q pointer) {
+	grow(p.v.Elem()).Set(q.v)
+}
+
+// getPointerSlice copies []*T from p as a new []pointer.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) getPointerSlice() []pointer {
+	if p.v.IsNil() {
+		return nil
+	}
+	n := p.v.Elem().Len()
+	s := make([]pointer, n)
+	for i := 0; i < n; i++ {
+		s[i] = pointer{v: p.v.Elem().Index(i)}
+	}
+	return s
+}
+
+// setPointerSlice copies []pointer into p as a new []*T.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) setPointerSlice(v []pointer) {
+	if v == nil {
+		p.v.Elem().Set(reflect.New(p.v.Elem().Type()).Elem())
+		return
+	}
+	s := reflect.MakeSlice(p.v.Elem().Type(), 0, len(v))
+	for _, p := range v {
+		s = reflect.Append(s, p.v)
+	}
+	p.v.Elem().Set(s)
+}
+
+// getInterfacePointer returns a pointer that points to the
+// interface data of the interface pointed by p.
+func (p pointer) getInterfacePointer() pointer {
+	if p.v.Elem().IsNil() {
+		return pointer{v: p.v.Elem()}
+	}
+	return pointer{v: p.v.Elem().Elem().Elem().Field(0).Addr()} // *interface -> interface -> *struct -> struct
+}
+
+func (p pointer) asPointerTo(t reflect.Type) reflect.Value {
+	// TODO: check that p.v.Type().Elem() == t?
+	return p.v
+}
+
+func atomicLoadUnmarshalInfo(p **unmarshalInfo) *unmarshalInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreUnmarshalInfo(p **unmarshalInfo, v *unmarshalInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadMarshalInfo(p **marshalInfo) *marshalInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreMarshalInfo(p **marshalInfo, v *marshalInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadMergeInfo(p **mergeInfo) *mergeInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreMergeInfo(p **mergeInfo, v *mergeInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadDiscardInfo(p **discardInfo) *discardInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreDiscardInfo(p **discardInfo, v *discardInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+
+var atomicLock sync.Mutex
diff --git a/vendor/github.com/gogo/protobuf/proto/pointer_reflect_gogo.go b/vendor/github.com/gogo/protobuf/proto/pointer_reflect_gogo.go
new file mode 100644
index 0000000..7ffd3c2
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/pointer_reflect_gogo.go
@@ -0,0 +1,59 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build purego appengine js
+
+// This file contains an implementation of proto field accesses using package reflect.
+// It is slower than the code in pointer_unsafe.go but it avoids package unsafe and can
+// be used on App Engine.
+
+package proto
+
+import (
+	"reflect"
+)
+
+// TODO: untested, so probably incorrect.
+
+func (p pointer) getRef() pointer {
+	return pointer{v: p.v.Addr()}
+}
+
+func (p pointer) appendRef(v pointer, typ reflect.Type) {
+	slice := p.getSlice(typ)
+	elem := v.asPointerTo(typ).Elem()
+	newSlice := reflect.Append(slice, elem)
+	slice.Set(newSlice)
+}
+
+func (p pointer) getSlice(typ reflect.Type) reflect.Value {
+	sliceTyp := reflect.SliceOf(typ)
+	slice := p.asPointerTo(sliceTyp)
+	slice = slice.Elem()
+	return slice
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/pointer_unsafe.go b/vendor/github.com/gogo/protobuf/proto/pointer_unsafe.go
new file mode 100644
index 0000000..d55a335
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/pointer_unsafe.go
@@ -0,0 +1,308 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2012 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build !purego,!appengine,!js
+
+// This file contains the implementation of the proto field accesses using package unsafe.
+
+package proto
+
+import (
+	"reflect"
+	"sync/atomic"
+	"unsafe"
+)
+
+const unsafeAllowed = true
+
+// A field identifies a field in a struct, accessible from a pointer.
+// In this implementation, a field is identified by its byte offset from the start of the struct.
+type field uintptr
+
+// toField returns a field equivalent to the given reflect field.
+func toField(f *reflect.StructField) field {
+	return field(f.Offset)
+}
+
+// invalidField is an invalid field identifier.
+const invalidField = ^field(0)
+
+// zeroField is a noop when calling pointer.offset.
+const zeroField = field(0)
+
+// IsValid reports whether the field identifier is valid.
+func (f field) IsValid() bool {
+	return f != invalidField
+}
+
+// The pointer type below is for the new table-driven encoder/decoder.
+// The implementation here uses unsafe.Pointer to create a generic pointer.
+// In pointer_reflect.go we use reflect instead of unsafe to implement
+// the same (but slower) interface.
+type pointer struct {
+	p unsafe.Pointer
+}
+
+// size of pointer
+var ptrSize = unsafe.Sizeof(uintptr(0))
+
+// toPointer converts an interface of pointer type to a pointer
+// that points to the same target.
+func toPointer(i *Message) pointer {
+	// Super-tricky - read pointer out of data word of interface value.
+	// Saves ~25ns over the equivalent:
+	// return valToPointer(reflect.ValueOf(*i))
+	return pointer{p: (*[2]unsafe.Pointer)(unsafe.Pointer(i))[1]}
+}
+
+// toAddrPointer converts an interface to a pointer that points to
+// the interface data.
+func toAddrPointer(i *interface{}, isptr bool) pointer {
+	// Super-tricky - read or get the address of data word of interface value.
+	if isptr {
+		// The interface is of pointer type, thus it is a direct interface.
+		// The data word is the pointer data itself. We take its address.
+		return pointer{p: unsafe.Pointer(uintptr(unsafe.Pointer(i)) + ptrSize)}
+	}
+	// The interface is not of pointer type. The data word is the pointer
+	// to the data.
+	return pointer{p: (*[2]unsafe.Pointer)(unsafe.Pointer(i))[1]}
+}
+
+// valToPointer converts v to a pointer. v must be of pointer type.
+func valToPointer(v reflect.Value) pointer {
+	return pointer{p: unsafe.Pointer(v.Pointer())}
+}
+
+// offset converts from a pointer to a structure to a pointer to
+// one of its fields.
+func (p pointer) offset(f field) pointer {
+	// For safety, we should panic if !f.IsValid, however calling panic causes
+	// this to no longer be inlineable, which is a serious performance cost.
+	/*
+		if !f.IsValid() {
+			panic("invalid field")
+		}
+	*/
+	return pointer{p: unsafe.Pointer(uintptr(p.p) + uintptr(f))}
+}
+
+func (p pointer) isNil() bool {
+	return p.p == nil
+}
+
+func (p pointer) toInt64() *int64 {
+	return (*int64)(p.p)
+}
+func (p pointer) toInt64Ptr() **int64 {
+	return (**int64)(p.p)
+}
+func (p pointer) toInt64Slice() *[]int64 {
+	return (*[]int64)(p.p)
+}
+func (p pointer) toInt32() *int32 {
+	return (*int32)(p.p)
+}
+
+// See pointer_reflect.go for why toInt32Ptr/Slice doesn't exist.
+/*
+	func (p pointer) toInt32Ptr() **int32 {
+		return (**int32)(p.p)
+	}
+	func (p pointer) toInt32Slice() *[]int32 {
+		return (*[]int32)(p.p)
+	}
+*/
+func (p pointer) getInt32Ptr() *int32 {
+	return *(**int32)(p.p)
+}
+func (p pointer) setInt32Ptr(v int32) {
+	*(**int32)(p.p) = &v
+}
+
+// getInt32Slice loads a []int32 from p.
+// The value returned is aliased with the original slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) getInt32Slice() []int32 {
+	return *(*[]int32)(p.p)
+}
+
+// setInt32Slice stores a []int32 to p.
+// The value set is aliased with the input slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) setInt32Slice(v []int32) {
+	*(*[]int32)(p.p) = v
+}
+
+// TODO: Can we get rid of appendInt32Slice and use setInt32Slice instead?
+func (p pointer) appendInt32Slice(v int32) {
+	s := (*[]int32)(p.p)
+	*s = append(*s, v)
+}
+
+func (p pointer) toUint64() *uint64 {
+	return (*uint64)(p.p)
+}
+func (p pointer) toUint64Ptr() **uint64 {
+	return (**uint64)(p.p)
+}
+func (p pointer) toUint64Slice() *[]uint64 {
+	return (*[]uint64)(p.p)
+}
+func (p pointer) toUint32() *uint32 {
+	return (*uint32)(p.p)
+}
+func (p pointer) toUint32Ptr() **uint32 {
+	return (**uint32)(p.p)
+}
+func (p pointer) toUint32Slice() *[]uint32 {
+	return (*[]uint32)(p.p)
+}
+func (p pointer) toBool() *bool {
+	return (*bool)(p.p)
+}
+func (p pointer) toBoolPtr() **bool {
+	return (**bool)(p.p)
+}
+func (p pointer) toBoolSlice() *[]bool {
+	return (*[]bool)(p.p)
+}
+func (p pointer) toFloat64() *float64 {
+	return (*float64)(p.p)
+}
+func (p pointer) toFloat64Ptr() **float64 {
+	return (**float64)(p.p)
+}
+func (p pointer) toFloat64Slice() *[]float64 {
+	return (*[]float64)(p.p)
+}
+func (p pointer) toFloat32() *float32 {
+	return (*float32)(p.p)
+}
+func (p pointer) toFloat32Ptr() **float32 {
+	return (**float32)(p.p)
+}
+func (p pointer) toFloat32Slice() *[]float32 {
+	return (*[]float32)(p.p)
+}
+func (p pointer) toString() *string {
+	return (*string)(p.p)
+}
+func (p pointer) toStringPtr() **string {
+	return (**string)(p.p)
+}
+func (p pointer) toStringSlice() *[]string {
+	return (*[]string)(p.p)
+}
+func (p pointer) toBytes() *[]byte {
+	return (*[]byte)(p.p)
+}
+func (p pointer) toBytesSlice() *[][]byte {
+	return (*[][]byte)(p.p)
+}
+func (p pointer) toExtensions() *XXX_InternalExtensions {
+	return (*XXX_InternalExtensions)(p.p)
+}
+func (p pointer) toOldExtensions() *map[int32]Extension {
+	return (*map[int32]Extension)(p.p)
+}
+
+// getPointerSlice loads []*T from p as a []pointer.
+// The value returned is aliased with the original slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) getPointerSlice() []pointer {
+	// Super-tricky - p should point to a []*T where T is a
+	// message type. We load it as []pointer.
+	return *(*[]pointer)(p.p)
+}
+
+// setPointerSlice stores []pointer into p as a []*T.
+// The value set is aliased with the input slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) setPointerSlice(v []pointer) {
+	// Super-tricky - p should point to a []*T where T is a
+	// message type. We store it as []pointer.
+	*(*[]pointer)(p.p) = v
+}
+
+// getPointer loads the pointer at p and returns it.
+func (p pointer) getPointer() pointer {
+	return pointer{p: *(*unsafe.Pointer)(p.p)}
+}
+
+// setPointer stores the pointer q at p.
+func (p pointer) setPointer(q pointer) {
+	*(*unsafe.Pointer)(p.p) = q.p
+}
+
+// append q to the slice pointed to by p.
+func (p pointer) appendPointer(q pointer) {
+	s := (*[]unsafe.Pointer)(p.p)
+	*s = append(*s, q.p)
+}
+
+// getInterfacePointer returns a pointer that points to the
+// interface data of the interface pointed by p.
+func (p pointer) getInterfacePointer() pointer {
+	// Super-tricky - read pointer out of data word of interface value.
+	return pointer{p: (*(*[2]unsafe.Pointer)(p.p))[1]}
+}
+
+// asPointerTo returns a reflect.Value that is a pointer to an
+// object of type t stored at p.
+func (p pointer) asPointerTo(t reflect.Type) reflect.Value {
+	return reflect.NewAt(t, p.p)
+}
+
+func atomicLoadUnmarshalInfo(p **unmarshalInfo) *unmarshalInfo {
+	return (*unmarshalInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreUnmarshalInfo(p **unmarshalInfo, v *unmarshalInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadMarshalInfo(p **marshalInfo) *marshalInfo {
+	return (*marshalInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreMarshalInfo(p **marshalInfo, v *marshalInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadMergeInfo(p **mergeInfo) *mergeInfo {
+	return (*mergeInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreMergeInfo(p **mergeInfo, v *mergeInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadDiscardInfo(p **discardInfo) *discardInfo {
+	return (*discardInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreDiscardInfo(p **discardInfo, v *discardInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/pointer_unsafe_gogo.go b/vendor/github.com/gogo/protobuf/proto/pointer_unsafe_gogo.go
new file mode 100644
index 0000000..aca8eed
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/pointer_unsafe_gogo.go
@@ -0,0 +1,56 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build !purego,!appengine,!js
+
+// This file contains the implementation of the proto field accesses using package unsafe.
+
+package proto
+
+import (
+	"reflect"
+	"unsafe"
+)
+
+func (p pointer) getRef() pointer {
+	return pointer{p: (unsafe.Pointer)(&p.p)}
+}
+
+func (p pointer) appendRef(v pointer, typ reflect.Type) {
+	slice := p.getSlice(typ)
+	elem := v.asPointerTo(typ).Elem()
+	newSlice := reflect.Append(slice, elem)
+	slice.Set(newSlice)
+}
+
+func (p pointer) getSlice(typ reflect.Type) reflect.Value {
+	sliceTyp := reflect.SliceOf(typ)
+	slice := p.asPointerTo(sliceTyp)
+	slice = slice.Elem()
+	return slice
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/properties.go b/vendor/github.com/gogo/protobuf/proto/properties.go
new file mode 100644
index 0000000..c9e5fa0
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/properties.go
@@ -0,0 +1,599 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for encoding data into the wire format for protocol buffers.
+ */
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"reflect"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+)
+
+const debug bool = false
+
+// Constants that identify the encoding of a value on the wire.
+const (
+	WireVarint     = 0
+	WireFixed64    = 1
+	WireBytes      = 2
+	WireStartGroup = 3
+	WireEndGroup   = 4
+	WireFixed32    = 5
+)
+
+// tagMap is an optimization over map[int]int for typical protocol buffer
+// use-cases. Encoded protocol buffers are often in tag order with small tag
+// numbers.
+type tagMap struct {
+	fastTags []int
+	slowTags map[int]int
+}
+
+// tagMapFastLimit is the upper bound on the tag number that will be stored in
+// the tagMap slice rather than its map.
+const tagMapFastLimit = 1024
+
+func (p *tagMap) get(t int) (int, bool) {
+	if t > 0 && t < tagMapFastLimit {
+		if t >= len(p.fastTags) {
+			return 0, false
+		}
+		fi := p.fastTags[t]
+		return fi, fi >= 0
+	}
+	fi, ok := p.slowTags[t]
+	return fi, ok
+}
+
+func (p *tagMap) put(t int, fi int) {
+	if t > 0 && t < tagMapFastLimit {
+		for len(p.fastTags) < t+1 {
+			p.fastTags = append(p.fastTags, -1)
+		}
+		p.fastTags[t] = fi
+		return
+	}
+	if p.slowTags == nil {
+		p.slowTags = make(map[int]int)
+	}
+	p.slowTags[t] = fi
+}
+
+// StructProperties represents properties for all the fields of a struct.
+// decoderTags and decoderOrigNames should only be used by the decoder.
+type StructProperties struct {
+	Prop             []*Properties  // properties for each field
+	reqCount         int            // required count
+	decoderTags      tagMap         // map from proto tag to struct field number
+	decoderOrigNames map[string]int // map from original name to struct field number
+	order            []int          // list of struct field numbers in tag order
+
+	// OneofTypes contains information about the oneof fields in this message.
+	// It is keyed by the original name of a field.
+	OneofTypes map[string]*OneofProperties
+}
+
+// OneofProperties represents information about a specific field in a oneof.
+type OneofProperties struct {
+	Type  reflect.Type // pointer to generated struct type for this oneof field
+	Field int          // struct field number of the containing oneof in the message
+	Prop  *Properties
+}
+
+// Implement the sorting interface so we can sort the fields in tag order, as recommended by the spec.
+// See encode.go, (*Buffer).enc_struct.
+
+func (sp *StructProperties) Len() int { return len(sp.order) }
+func (sp *StructProperties) Less(i, j int) bool {
+	return sp.Prop[sp.order[i]].Tag < sp.Prop[sp.order[j]].Tag
+}
+func (sp *StructProperties) Swap(i, j int) { sp.order[i], sp.order[j] = sp.order[j], sp.order[i] }
+
+// Properties represents the protocol-specific behavior of a single struct field.
+type Properties struct {
+	Name     string // name of the field, for error messages
+	OrigName string // original name before protocol compiler (always set)
+	JSONName string // name to use for JSON; determined by protoc
+	Wire     string
+	WireType int
+	Tag      int
+	Required bool
+	Optional bool
+	Repeated bool
+	Packed   bool   // relevant for repeated primitives only
+	Enum     string // set for enum types only
+	proto3   bool   // whether this is known to be a proto3 field
+	oneof    bool   // whether this is a oneof field
+
+	Default     string // default value
+	HasDefault  bool   // whether an explicit default was provided
+	CustomType  string
+	CastType    string
+	StdTime     bool
+	StdDuration bool
+	WktPointer  bool
+
+	stype reflect.Type      // set for struct types only
+	ctype reflect.Type      // set for custom types only
+	sprop *StructProperties // set for struct types only
+
+	mtype      reflect.Type // set for map types only
+	MapKeyProp *Properties  // set for map types only
+	MapValProp *Properties  // set for map types only
+}
+
+// String formats the properties in the protobuf struct field tag style.
+func (p *Properties) String() string {
+	s := p.Wire
+	s += ","
+	s += strconv.Itoa(p.Tag)
+	if p.Required {
+		s += ",req"
+	}
+	if p.Optional {
+		s += ",opt"
+	}
+	if p.Repeated {
+		s += ",rep"
+	}
+	if p.Packed {
+		s += ",packed"
+	}
+	s += ",name=" + p.OrigName
+	if p.JSONName != p.OrigName {
+		s += ",json=" + p.JSONName
+	}
+	if p.proto3 {
+		s += ",proto3"
+	}
+	if p.oneof {
+		s += ",oneof"
+	}
+	if len(p.Enum) > 0 {
+		s += ",enum=" + p.Enum
+	}
+	if p.HasDefault {
+		s += ",def=" + p.Default
+	}
+	return s
+}
+
+// Parse populates p by parsing a string in the protobuf struct field tag style.
+func (p *Properties) Parse(s string) {
+	// "bytes,49,opt,name=foo,def=hello!"
+	fields := strings.Split(s, ",") // breaks def=, but handled below.
+	if len(fields) < 2 {
+		fmt.Fprintf(os.Stderr, "proto: tag has too few fields: %q\n", s)
+		return
+	}
+
+	p.Wire = fields[0]
+	switch p.Wire {
+	case "varint":
+		p.WireType = WireVarint
+	case "fixed32":
+		p.WireType = WireFixed32
+	case "fixed64":
+		p.WireType = WireFixed64
+	case "zigzag32":
+		p.WireType = WireVarint
+	case "zigzag64":
+		p.WireType = WireVarint
+	case "bytes", "group":
+		p.WireType = WireBytes
+		// no numeric converter for non-numeric types
+	default:
+		fmt.Fprintf(os.Stderr, "proto: tag has unknown wire type: %q\n", s)
+		return
+	}
+
+	var err error
+	p.Tag, err = strconv.Atoi(fields[1])
+	if err != nil {
+		return
+	}
+
+outer:
+	for i := 2; i < len(fields); i++ {
+		f := fields[i]
+		switch {
+		case f == "req":
+			p.Required = true
+		case f == "opt":
+			p.Optional = true
+		case f == "rep":
+			p.Repeated = true
+		case f == "packed":
+			p.Packed = true
+		case strings.HasPrefix(f, "name="):
+			p.OrigName = f[5:]
+		case strings.HasPrefix(f, "json="):
+			p.JSONName = f[5:]
+		case strings.HasPrefix(f, "enum="):
+			p.Enum = f[5:]
+		case f == "proto3":
+			p.proto3 = true
+		case f == "oneof":
+			p.oneof = true
+		case strings.HasPrefix(f, "def="):
+			p.HasDefault = true
+			p.Default = f[4:] // rest of string
+			if i+1 < len(fields) {
+				// Commas aren't escaped, and def is always last.
+				p.Default += "," + strings.Join(fields[i+1:], ",")
+				break outer
+			}
+		case strings.HasPrefix(f, "embedded="):
+			p.OrigName = strings.Split(f, "=")[1]
+		case strings.HasPrefix(f, "customtype="):
+			p.CustomType = strings.Split(f, "=")[1]
+		case strings.HasPrefix(f, "casttype="):
+			p.CastType = strings.Split(f, "=")[1]
+		case f == "stdtime":
+			p.StdTime = true
+		case f == "stdduration":
+			p.StdDuration = true
+		case f == "wktptr":
+			p.WktPointer = true
+		}
+	}
+}
+
+var protoMessageType = reflect.TypeOf((*Message)(nil)).Elem()
+
+// setFieldProps initializes the field properties for submessages and maps.
+func (p *Properties) setFieldProps(typ reflect.Type, f *reflect.StructField, lockGetProp bool) {
+	isMap := typ.Kind() == reflect.Map
+	if len(p.CustomType) > 0 && !isMap {
+		p.ctype = typ
+		p.setTag(lockGetProp)
+		return
+	}
+	if p.StdTime && !isMap {
+		p.setTag(lockGetProp)
+		return
+	}
+	if p.StdDuration && !isMap {
+		p.setTag(lockGetProp)
+		return
+	}
+	if p.WktPointer && !isMap {
+		p.setTag(lockGetProp)
+		return
+	}
+	switch t1 := typ; t1.Kind() {
+	case reflect.Struct:
+		p.stype = typ
+	case reflect.Ptr:
+		if t1.Elem().Kind() == reflect.Struct {
+			p.stype = t1.Elem()
+		}
+	case reflect.Slice:
+		switch t2 := t1.Elem(); t2.Kind() {
+		case reflect.Ptr:
+			switch t3 := t2.Elem(); t3.Kind() {
+			case reflect.Struct:
+				p.stype = t3
+			}
+		case reflect.Struct:
+			p.stype = t2
+		}
+
+	case reflect.Map:
+
+		p.mtype = t1
+		p.MapKeyProp = &Properties{}
+		p.MapKeyProp.init(reflect.PtrTo(p.mtype.Key()), "Key", f.Tag.Get("protobuf_key"), nil, lockGetProp)
+		p.MapValProp = &Properties{}
+		vtype := p.mtype.Elem()
+		if vtype.Kind() != reflect.Ptr && vtype.Kind() != reflect.Slice {
+			// The value type is not a message (*T) or bytes ([]byte),
+			// so we need encoders for the pointer to this type.
+			vtype = reflect.PtrTo(vtype)
+		}
+
+		p.MapValProp.CustomType = p.CustomType
+		p.MapValProp.StdDuration = p.StdDuration
+		p.MapValProp.StdTime = p.StdTime
+		p.MapValProp.WktPointer = p.WktPointer
+		p.MapValProp.init(vtype, "Value", f.Tag.Get("protobuf_val"), nil, lockGetProp)
+	}
+	p.setTag(lockGetProp)
+}
+
+func (p *Properties) setTag(lockGetProp bool) {
+	if p.stype != nil {
+		if lockGetProp {
+			p.sprop = GetProperties(p.stype)
+		} else {
+			p.sprop = getPropertiesLocked(p.stype)
+		}
+	}
+}
+
+var (
+	marshalerType = reflect.TypeOf((*Marshaler)(nil)).Elem()
+)
+
+// Init populates the properties from a protocol buffer struct tag.
+func (p *Properties) Init(typ reflect.Type, name, tag string, f *reflect.StructField) {
+	p.init(typ, name, tag, f, true)
+}
+
+func (p *Properties) init(typ reflect.Type, name, tag string, f *reflect.StructField, lockGetProp bool) {
+	// "bytes,49,opt,def=hello!"
+	p.Name = name
+	p.OrigName = name
+	if tag == "" {
+		return
+	}
+	p.Parse(tag)
+	p.setFieldProps(typ, f, lockGetProp)
+}
+
+var (
+	propertiesMu  sync.RWMutex
+	propertiesMap = make(map[reflect.Type]*StructProperties)
+)
+
+// GetProperties returns the list of properties for the type represented by t.
+// t must represent a generated struct type of a protocol message.
+func GetProperties(t reflect.Type) *StructProperties {
+	if t.Kind() != reflect.Struct {
+		panic("proto: type must have kind struct")
+	}
+
+	// Most calls to GetProperties in a long-running program will be
+	// retrieving details for types we have seen before.
+	propertiesMu.RLock()
+	sprop, ok := propertiesMap[t]
+	propertiesMu.RUnlock()
+	if ok {
+		return sprop
+	}
+
+	propertiesMu.Lock()
+	sprop = getPropertiesLocked(t)
+	propertiesMu.Unlock()
+	return sprop
+}
+
+// getPropertiesLocked requires that propertiesMu is held.
+func getPropertiesLocked(t reflect.Type) *StructProperties {
+	if prop, ok := propertiesMap[t]; ok {
+		return prop
+	}
+
+	prop := new(StructProperties)
+	// in case of recursive protos, fill this in now.
+	propertiesMap[t] = prop
+
+	// build properties
+	prop.Prop = make([]*Properties, t.NumField())
+	prop.order = make([]int, t.NumField())
+
+	isOneofMessage := false
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		p := new(Properties)
+		name := f.Name
+		p.init(f.Type, name, f.Tag.Get("protobuf"), &f, false)
+
+		oneof := f.Tag.Get("protobuf_oneof") // special case
+		if oneof != "" {
+			isOneofMessage = true
+			// Oneof fields don't use the traditional protobuf tag.
+			p.OrigName = oneof
+		}
+		prop.Prop[i] = p
+		prop.order[i] = i
+		if debug {
+			print(i, " ", f.Name, " ", t.String(), " ")
+			if p.Tag > 0 {
+				print(p.String())
+			}
+			print("\n")
+		}
+	}
+
+	// Re-order prop.order.
+	sort.Sort(prop)
+
+	type oneofMessage interface {
+		XXX_OneofFuncs() (func(Message, *Buffer) error, func(Message, int, int, *Buffer) (bool, error), func(Message) int, []interface{})
+	}
+	if om, ok := reflect.Zero(reflect.PtrTo(t)).Interface().(oneofMessage); isOneofMessage && ok {
+		var oots []interface{}
+		_, _, _, oots = om.XXX_OneofFuncs()
+
+		// Interpret oneof metadata.
+		prop.OneofTypes = make(map[string]*OneofProperties)
+		for _, oot := range oots {
+			oop := &OneofProperties{
+				Type: reflect.ValueOf(oot).Type(), // *T
+				Prop: new(Properties),
+			}
+			sft := oop.Type.Elem().Field(0)
+			oop.Prop.Name = sft.Name
+			oop.Prop.Parse(sft.Tag.Get("protobuf"))
+			// There will be exactly one interface field that
+			// this new value is assignable to.
+			for i := 0; i < t.NumField(); i++ {
+				f := t.Field(i)
+				if f.Type.Kind() != reflect.Interface {
+					continue
+				}
+				if !oop.Type.AssignableTo(f.Type) {
+					continue
+				}
+				oop.Field = i
+				break
+			}
+			prop.OneofTypes[oop.Prop.OrigName] = oop
+		}
+	}
+
+	// build required counts
+	// build tags
+	reqCount := 0
+	prop.decoderOrigNames = make(map[string]int)
+	for i, p := range prop.Prop {
+		if strings.HasPrefix(p.Name, "XXX_") {
+			// Internal fields should not appear in tags/origNames maps.
+			// They are handled specially when encoding and decoding.
+			continue
+		}
+		if p.Required {
+			reqCount++
+		}
+		prop.decoderTags.put(p.Tag, i)
+		prop.decoderOrigNames[p.OrigName] = i
+	}
+	prop.reqCount = reqCount
+
+	return prop
+}
+
+// A global registry of enum types.
+// The generated code will register the generated maps by calling RegisterEnum.
+
+var enumValueMaps = make(map[string]map[string]int32)
+var enumStringMaps = make(map[string]map[int32]string)
+
+// RegisterEnum is called from the generated code to install the enum descriptor
+// maps into the global table to aid parsing text format protocol buffers.
+func RegisterEnum(typeName string, unusedNameMap map[int32]string, valueMap map[string]int32) {
+	if _, ok := enumValueMaps[typeName]; ok {
+		panic("proto: duplicate enum registered: " + typeName)
+	}
+	enumValueMaps[typeName] = valueMap
+	if _, ok := enumStringMaps[typeName]; ok {
+		panic("proto: duplicate enum registered: " + typeName)
+	}
+	enumStringMaps[typeName] = unusedNameMap
+}
+
+// EnumValueMap returns the mapping from names to integers of the
+// enum type enumType, or a nil if not found.
+func EnumValueMap(enumType string) map[string]int32 {
+	return enumValueMaps[enumType]
+}
+
+// A registry of all linked message types.
+// The string is a fully-qualified proto name ("pkg.Message").
+var (
+	protoTypedNils = make(map[string]Message)      // a map from proto names to typed nil pointers
+	protoMapTypes  = make(map[string]reflect.Type) // a map from proto names to map types
+	revProtoTypes  = make(map[reflect.Type]string)
+)
+
+// RegisterType is called from generated code and maps from the fully qualified
+// proto name to the type (pointer to struct) of the protocol buffer.
+func RegisterType(x Message, name string) {
+	if _, ok := protoTypedNils[name]; ok {
+		// TODO: Some day, make this a panic.
+		log.Printf("proto: duplicate proto type registered: %s", name)
+		return
+	}
+	t := reflect.TypeOf(x)
+	if v := reflect.ValueOf(x); v.Kind() == reflect.Ptr && v.Pointer() == 0 {
+		// Generated code always calls RegisterType with nil x.
+		// This check is just for extra safety.
+		protoTypedNils[name] = x
+	} else {
+		protoTypedNils[name] = reflect.Zero(t).Interface().(Message)
+	}
+	revProtoTypes[t] = name
+}
+
+// RegisterMapType is called from generated code and maps from the fully qualified
+// proto name to the native map type of the proto map definition.
+func RegisterMapType(x interface{}, name string) {
+	if reflect.TypeOf(x).Kind() != reflect.Map {
+		panic(fmt.Sprintf("RegisterMapType(%T, %q); want map", x, name))
+	}
+	if _, ok := protoMapTypes[name]; ok {
+		log.Printf("proto: duplicate proto type registered: %s", name)
+		return
+	}
+	t := reflect.TypeOf(x)
+	protoMapTypes[name] = t
+	revProtoTypes[t] = name
+}
+
+// MessageName returns the fully-qualified proto name for the given message type.
+func MessageName(x Message) string {
+	type xname interface {
+		XXX_MessageName() string
+	}
+	if m, ok := x.(xname); ok {
+		return m.XXX_MessageName()
+	}
+	return revProtoTypes[reflect.TypeOf(x)]
+}
+
+// MessageType returns the message type (pointer to struct) for a named message.
+// The type is not guaranteed to implement proto.Message if the name refers to a
+// map entry.
+func MessageType(name string) reflect.Type {
+	if t, ok := protoTypedNils[name]; ok {
+		return reflect.TypeOf(t)
+	}
+	return protoMapTypes[name]
+}
+
+// A registry of all linked proto files.
+var (
+	protoFiles = make(map[string][]byte) // file name => fileDescriptor
+)
+
+// RegisterFile is called from generated code and maps from the
+// full file name of a .proto file to its compressed FileDescriptorProto.
+func RegisterFile(filename string, fileDescriptor []byte) {
+	protoFiles[filename] = fileDescriptor
+}
+
+// FileDescriptor returns the compressed FileDescriptorProto for a .proto file.
+func FileDescriptor(filename string) []byte { return protoFiles[filename] }
diff --git a/vendor/github.com/gogo/protobuf/proto/properties_gogo.go b/vendor/github.com/gogo/protobuf/proto/properties_gogo.go
new file mode 100644
index 0000000..40ea3dd
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/properties_gogo.go
@@ -0,0 +1,36 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"reflect"
+)
+
+var sizerType = reflect.TypeOf((*Sizer)(nil)).Elem()
+var protosizerType = reflect.TypeOf((*ProtoSizer)(nil)).Elem()
diff --git a/vendor/github.com/gogo/protobuf/proto/skip_gogo.go b/vendor/github.com/gogo/protobuf/proto/skip_gogo.go
new file mode 100644
index 0000000..5a5fd93
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/skip_gogo.go
@@ -0,0 +1,119 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"fmt"
+	"io"
+)
+
+func Skip(data []byte) (n int, err error) {
+	l := len(data)
+	index := 0
+	for index < l {
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if index >= l {
+				return 0, io.ErrUnexpectedEOF
+			}
+			b := data[index]
+			index++
+			wire |= (uint64(b) & 0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		wireType := int(wire & 0x7)
+		switch wireType {
+		case 0:
+			for {
+				if index >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				index++
+				if data[index-1] < 0x80 {
+					break
+				}
+			}
+			return index, nil
+		case 1:
+			index += 8
+			return index, nil
+		case 2:
+			var length int
+			for shift := uint(0); ; shift += 7 {
+				if index >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				b := data[index]
+				index++
+				length |= (int(b) & 0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			index += length
+			return index, nil
+		case 3:
+			for {
+				var innerWire uint64
+				var start int = index
+				for shift := uint(0); ; shift += 7 {
+					if index >= l {
+						return 0, io.ErrUnexpectedEOF
+					}
+					b := data[index]
+					index++
+					innerWire |= (uint64(b) & 0x7F) << shift
+					if b < 0x80 {
+						break
+					}
+				}
+				innerWireType := int(innerWire & 0x7)
+				if innerWireType == 4 {
+					break
+				}
+				next, err := Skip(data[start:])
+				if err != nil {
+					return 0, err
+				}
+				index = start + next
+			}
+			return index, nil
+		case 4:
+			return index, nil
+		case 5:
+			index += 4
+			return index, nil
+		default:
+			return 0, fmt.Errorf("proto: illegal wireType %d", wireType)
+		}
+	}
+	panic("unreachable")
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/table_marshal.go b/vendor/github.com/gogo/protobuf/proto/table_marshal.go
new file mode 100644
index 0000000..9b1538d
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/table_marshal.go
@@ -0,0 +1,3006 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"reflect"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"unicode/utf8"
+)
+
+// a sizer takes a pointer to a field and the size of its tag, computes the size of
+// the encoded data.
+type sizer func(pointer, int) int
+
+// a marshaler takes a byte slice, a pointer to a field, and its tag (in wire format),
+// marshals the field to the end of the slice, returns the slice and error (if any).
+type marshaler func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error)
+
+// marshalInfo is the information used for marshaling a message.
+type marshalInfo struct {
+	typ          reflect.Type
+	fields       []*marshalFieldInfo
+	unrecognized field                      // offset of XXX_unrecognized
+	extensions   field                      // offset of XXX_InternalExtensions
+	v1extensions field                      // offset of XXX_extensions
+	sizecache    field                      // offset of XXX_sizecache
+	initialized  int32                      // 0 -- only typ is set, 1 -- fully initialized
+	messageset   bool                       // uses message set wire format
+	hasmarshaler bool                       // has custom marshaler
+	sync.RWMutex                            // protect extElems map, also for initialization
+	extElems     map[int32]*marshalElemInfo // info of extension elements
+
+	hassizer      bool // has custom sizer
+	hasprotosizer bool // has custom protosizer
+
+	bytesExtensions field // offset of XXX_extensions where the field type is []byte
+}
+
+// marshalFieldInfo is the information used for marshaling a field of a message.
+type marshalFieldInfo struct {
+	field      field
+	wiretag    uint64 // tag in wire format
+	tagsize    int    // size of tag in wire format
+	sizer      sizer
+	marshaler  marshaler
+	isPointer  bool
+	required   bool                              // field is required
+	name       string                            // name of the field, for error reporting
+	oneofElems map[reflect.Type]*marshalElemInfo // info of oneof elements
+}
+
+// marshalElemInfo is the information used for marshaling an extension or oneof element.
+type marshalElemInfo struct {
+	wiretag   uint64 // tag in wire format
+	tagsize   int    // size of tag in wire format
+	sizer     sizer
+	marshaler marshaler
+	isptr     bool // elem is pointer typed, thus interface of this type is a direct interface (extension only)
+}
+
+var (
+	marshalInfoMap  = map[reflect.Type]*marshalInfo{}
+	marshalInfoLock sync.Mutex
+
+	uint8SliceType = reflect.TypeOf(([]uint8)(nil)).Kind()
+)
+
+// getMarshalInfo returns the information to marshal a given type of message.
+// The info it returns may not necessarily initialized.
+// t is the type of the message (NOT the pointer to it).
+func getMarshalInfo(t reflect.Type) *marshalInfo {
+	marshalInfoLock.Lock()
+	u, ok := marshalInfoMap[t]
+	if !ok {
+		u = &marshalInfo{typ: t}
+		marshalInfoMap[t] = u
+	}
+	marshalInfoLock.Unlock()
+	return u
+}
+
+// Size is the entry point from generated code,
+// and should be ONLY called by generated code.
+// It computes the size of encoded data of msg.
+// a is a pointer to a place to store cached marshal info.
+func (a *InternalMessageInfo) Size(msg Message) int {
+	u := getMessageMarshalInfo(msg, a)
+	ptr := toPointer(&msg)
+	if ptr.isNil() {
+		// We get here if msg is a typed nil ((*SomeMessage)(nil)),
+		// so it satisfies the interface, and msg == nil wouldn't
+		// catch it. We don't want crash in this case.
+		return 0
+	}
+	return u.size(ptr)
+}
+
+// Marshal is the entry point from generated code,
+// and should be ONLY called by generated code.
+// It marshals msg to the end of b.
+// a is a pointer to a place to store cached marshal info.
+func (a *InternalMessageInfo) Marshal(b []byte, msg Message, deterministic bool) ([]byte, error) {
+	u := getMessageMarshalInfo(msg, a)
+	ptr := toPointer(&msg)
+	if ptr.isNil() {
+		// We get here if msg is a typed nil ((*SomeMessage)(nil)),
+		// so it satisfies the interface, and msg == nil wouldn't
+		// catch it. We don't want crash in this case.
+		return b, ErrNil
+	}
+	return u.marshal(b, ptr, deterministic)
+}
+
+func getMessageMarshalInfo(msg interface{}, a *InternalMessageInfo) *marshalInfo {
+	// u := a.marshal, but atomically.
+	// We use an atomic here to ensure memory consistency.
+	u := atomicLoadMarshalInfo(&a.marshal)
+	if u == nil {
+		// Get marshal information from type of message.
+		t := reflect.ValueOf(msg).Type()
+		if t.Kind() != reflect.Ptr {
+			panic(fmt.Sprintf("cannot handle non-pointer message type %v", t))
+		}
+		u = getMarshalInfo(t.Elem())
+		// Store it in the cache for later users.
+		// a.marshal = u, but atomically.
+		atomicStoreMarshalInfo(&a.marshal, u)
+	}
+	return u
+}
+
+// size is the main function to compute the size of the encoded data of a message.
+// ptr is the pointer to the message.
+func (u *marshalInfo) size(ptr pointer) int {
+	if atomic.LoadInt32(&u.initialized) == 0 {
+		u.computeMarshalInfo()
+	}
+
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if u.hasmarshaler {
+		// Uses the message's Size method if available
+		if u.hassizer {
+			s := ptr.asPointerTo(u.typ).Interface().(Sizer)
+			return s.Size()
+		}
+		// Uses the message's ProtoSize method if available
+		if u.hasprotosizer {
+			s := ptr.asPointerTo(u.typ).Interface().(ProtoSizer)
+			return s.ProtoSize()
+		}
+
+		m := ptr.asPointerTo(u.typ).Interface().(Marshaler)
+		b, _ := m.Marshal()
+		return len(b)
+	}
+
+	n := 0
+	for _, f := range u.fields {
+		if f.isPointer && ptr.offset(f.field).getPointer().isNil() {
+			// nil pointer always marshals to nothing
+			continue
+		}
+		n += f.sizer(ptr.offset(f.field), f.tagsize)
+	}
+	if u.extensions.IsValid() {
+		e := ptr.offset(u.extensions).toExtensions()
+		if u.messageset {
+			n += u.sizeMessageSet(e)
+		} else {
+			n += u.sizeExtensions(e)
+		}
+	}
+	if u.v1extensions.IsValid() {
+		m := *ptr.offset(u.v1extensions).toOldExtensions()
+		n += u.sizeV1Extensions(m)
+	}
+	if u.bytesExtensions.IsValid() {
+		s := *ptr.offset(u.bytesExtensions).toBytes()
+		n += len(s)
+	}
+	if u.unrecognized.IsValid() {
+		s := *ptr.offset(u.unrecognized).toBytes()
+		n += len(s)
+	}
+
+	// cache the result for use in marshal
+	if u.sizecache.IsValid() {
+		atomic.StoreInt32(ptr.offset(u.sizecache).toInt32(), int32(n))
+	}
+	return n
+}
+
+// cachedsize gets the size from cache. If there is no cache (i.e. message is not generated),
+// fall back to compute the size.
+func (u *marshalInfo) cachedsize(ptr pointer) int {
+	if u.sizecache.IsValid() {
+		return int(atomic.LoadInt32(ptr.offset(u.sizecache).toInt32()))
+	}
+	return u.size(ptr)
+}
+
+// marshal is the main function to marshal a message. It takes a byte slice and appends
+// the encoded data to the end of the slice, returns the slice and error (if any).
+// ptr is the pointer to the message.
+// If deterministic is true, map is marshaled in deterministic order.
+func (u *marshalInfo) marshal(b []byte, ptr pointer, deterministic bool) ([]byte, error) {
+	if atomic.LoadInt32(&u.initialized) == 0 {
+		u.computeMarshalInfo()
+	}
+
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if u.hasmarshaler {
+		m := ptr.asPointerTo(u.typ).Interface().(Marshaler)
+		b1, err := m.Marshal()
+		b = append(b, b1...)
+		return b, err
+	}
+
+	var err, errLater error
+	// The old marshaler encodes extensions at beginning.
+	if u.extensions.IsValid() {
+		e := ptr.offset(u.extensions).toExtensions()
+		if u.messageset {
+			b, err = u.appendMessageSet(b, e, deterministic)
+		} else {
+			b, err = u.appendExtensions(b, e, deterministic)
+		}
+		if err != nil {
+			return b, err
+		}
+	}
+	if u.v1extensions.IsValid() {
+		m := *ptr.offset(u.v1extensions).toOldExtensions()
+		b, err = u.appendV1Extensions(b, m, deterministic)
+		if err != nil {
+			return b, err
+		}
+	}
+	if u.bytesExtensions.IsValid() {
+		s := *ptr.offset(u.bytesExtensions).toBytes()
+		b = append(b, s...)
+	}
+	for _, f := range u.fields {
+		if f.required {
+			if f.isPointer && ptr.offset(f.field).getPointer().isNil() {
+				// Required field is not set.
+				// We record the error but keep going, to give a complete marshaling.
+				if errLater == nil {
+					errLater = &RequiredNotSetError{f.name}
+				}
+				continue
+			}
+		}
+		if f.isPointer && ptr.offset(f.field).getPointer().isNil() {
+			// nil pointer always marshals to nothing
+			continue
+		}
+		b, err = f.marshaler(b, ptr.offset(f.field), f.wiretag, deterministic)
+		if err != nil {
+			if err1, ok := err.(*RequiredNotSetError); ok {
+				// Required field in submessage is not set.
+				// We record the error but keep going, to give a complete marshaling.
+				if errLater == nil {
+					errLater = &RequiredNotSetError{f.name + "." + err1.field}
+				}
+				continue
+			}
+			if err == errRepeatedHasNil {
+				err = errors.New("proto: repeated field " + f.name + " has nil element")
+			}
+			if err == errInvalidUTF8 {
+				if errLater == nil {
+					fullName := revProtoTypes[reflect.PtrTo(u.typ)] + "." + f.name
+					errLater = &invalidUTF8Error{fullName}
+				}
+				continue
+			}
+			return b, err
+		}
+	}
+	if u.unrecognized.IsValid() {
+		s := *ptr.offset(u.unrecognized).toBytes()
+		b = append(b, s...)
+	}
+	return b, errLater
+}
+
+// computeMarshalInfo initializes the marshal info.
+func (u *marshalInfo) computeMarshalInfo() {
+	u.Lock()
+	defer u.Unlock()
+	if u.initialized != 0 { // non-atomic read is ok as it is protected by the lock
+		return
+	}
+
+	t := u.typ
+	u.unrecognized = invalidField
+	u.extensions = invalidField
+	u.v1extensions = invalidField
+	u.bytesExtensions = invalidField
+	u.sizecache = invalidField
+	isOneofMessage := false
+
+	if reflect.PtrTo(t).Implements(sizerType) {
+		u.hassizer = true
+	}
+	if reflect.PtrTo(t).Implements(protosizerType) {
+		u.hasprotosizer = true
+	}
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if reflect.PtrTo(t).Implements(marshalerType) {
+		u.hasmarshaler = true
+		atomic.StoreInt32(&u.initialized, 1)
+		return
+	}
+
+	n := t.NumField()
+
+	// deal with XXX fields first
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		if f.Tag.Get("protobuf_oneof") != "" {
+			isOneofMessage = true
+		}
+		if !strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		switch f.Name {
+		case "XXX_sizecache":
+			u.sizecache = toField(&f)
+		case "XXX_unrecognized":
+			u.unrecognized = toField(&f)
+		case "XXX_InternalExtensions":
+			u.extensions = toField(&f)
+			u.messageset = f.Tag.Get("protobuf_messageset") == "1"
+		case "XXX_extensions":
+			if f.Type.Kind() == reflect.Map {
+				u.v1extensions = toField(&f)
+			} else {
+				u.bytesExtensions = toField(&f)
+			}
+		case "XXX_NoUnkeyedLiteral":
+			// nothing to do
+		default:
+			panic("unknown XXX field: " + f.Name)
+		}
+		n--
+	}
+
+	// get oneof implementers
+	var oneofImplementers []interface{}
+	// gogo: isOneofMessage is needed for embedded oneof messages, without a marshaler and unmarshaler
+	if m, ok := reflect.Zero(reflect.PtrTo(t)).Interface().(oneofMessage); ok && isOneofMessage {
+		_, _, _, oneofImplementers = m.XXX_OneofFuncs()
+	}
+
+	// normal fields
+	fields := make([]marshalFieldInfo, n) // batch allocation
+	u.fields = make([]*marshalFieldInfo, 0, n)
+	for i, j := 0, 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		field := &fields[j]
+		j++
+		field.name = f.Name
+		u.fields = append(u.fields, field)
+		if f.Tag.Get("protobuf_oneof") != "" {
+			field.computeOneofFieldInfo(&f, oneofImplementers)
+			continue
+		}
+		if f.Tag.Get("protobuf") == "" {
+			// field has no tag (not in generated message), ignore it
+			u.fields = u.fields[:len(u.fields)-1]
+			j--
+			continue
+		}
+		field.computeMarshalFieldInfo(&f)
+	}
+
+	// fields are marshaled in tag order on the wire.
+	sort.Sort(byTag(u.fields))
+
+	atomic.StoreInt32(&u.initialized, 1)
+}
+
+// helper for sorting fields by tag
+type byTag []*marshalFieldInfo
+
+func (a byTag) Len() int           { return len(a) }
+func (a byTag) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
+func (a byTag) Less(i, j int) bool { return a[i].wiretag < a[j].wiretag }
+
+// getExtElemInfo returns the information to marshal an extension element.
+// The info it returns is initialized.
+func (u *marshalInfo) getExtElemInfo(desc *ExtensionDesc) *marshalElemInfo {
+	// get from cache first
+	u.RLock()
+	e, ok := u.extElems[desc.Field]
+	u.RUnlock()
+	if ok {
+		return e
+	}
+
+	t := reflect.TypeOf(desc.ExtensionType) // pointer or slice to basic type or struct
+	tags := strings.Split(desc.Tag, ",")
+	tag, err := strconv.Atoi(tags[1])
+	if err != nil {
+		panic("tag is not an integer")
+	}
+	wt := wiretype(tags[0])
+	sizr, marshalr := typeMarshaler(t, tags, false, false)
+	e = &marshalElemInfo{
+		wiretag:   uint64(tag)<<3 | wt,
+		tagsize:   SizeVarint(uint64(tag) << 3),
+		sizer:     sizr,
+		marshaler: marshalr,
+		isptr:     t.Kind() == reflect.Ptr,
+	}
+
+	// update cache
+	u.Lock()
+	if u.extElems == nil {
+		u.extElems = make(map[int32]*marshalElemInfo)
+	}
+	u.extElems[desc.Field] = e
+	u.Unlock()
+	return e
+}
+
+// computeMarshalFieldInfo fills up the information to marshal a field.
+func (fi *marshalFieldInfo) computeMarshalFieldInfo(f *reflect.StructField) {
+	// parse protobuf tag of the field.
+	// tag has format of "bytes,49,opt,name=foo,def=hello!"
+	tags := strings.Split(f.Tag.Get("protobuf"), ",")
+	if tags[0] == "" {
+		return
+	}
+	tag, err := strconv.Atoi(tags[1])
+	if err != nil {
+		panic("tag is not an integer")
+	}
+	wt := wiretype(tags[0])
+	if tags[2] == "req" {
+		fi.required = true
+	}
+	fi.setTag(f, tag, wt)
+	fi.setMarshaler(f, tags)
+}
+
+func (fi *marshalFieldInfo) computeOneofFieldInfo(f *reflect.StructField, oneofImplementers []interface{}) {
+	fi.field = toField(f)
+	fi.wiretag = math.MaxInt32 // Use a large tag number, make oneofs sorted at the end. This tag will not appear on the wire.
+	fi.isPointer = true
+	fi.sizer, fi.marshaler = makeOneOfMarshaler(fi, f)
+	fi.oneofElems = make(map[reflect.Type]*marshalElemInfo)
+
+	ityp := f.Type // interface type
+	for _, o := range oneofImplementers {
+		t := reflect.TypeOf(o)
+		if !t.Implements(ityp) {
+			continue
+		}
+		sf := t.Elem().Field(0) // oneof implementer is a struct with a single field
+		tags := strings.Split(sf.Tag.Get("protobuf"), ",")
+		tag, err := strconv.Atoi(tags[1])
+		if err != nil {
+			panic("tag is not an integer")
+		}
+		wt := wiretype(tags[0])
+		sizr, marshalr := typeMarshaler(sf.Type, tags, false, true) // oneof should not omit any zero value
+		fi.oneofElems[t.Elem()] = &marshalElemInfo{
+			wiretag:   uint64(tag)<<3 | wt,
+			tagsize:   SizeVarint(uint64(tag) << 3),
+			sizer:     sizr,
+			marshaler: marshalr,
+		}
+	}
+}
+
+type oneofMessage interface {
+	XXX_OneofFuncs() (func(Message, *Buffer) error, func(Message, int, int, *Buffer) (bool, error), func(Message) int, []interface{})
+}
+
+// wiretype returns the wire encoding of the type.
+func wiretype(encoding string) uint64 {
+	switch encoding {
+	case "fixed32":
+		return WireFixed32
+	case "fixed64":
+		return WireFixed64
+	case "varint", "zigzag32", "zigzag64":
+		return WireVarint
+	case "bytes":
+		return WireBytes
+	case "group":
+		return WireStartGroup
+	}
+	panic("unknown wire type " + encoding)
+}
+
+// setTag fills up the tag (in wire format) and its size in the info of a field.
+func (fi *marshalFieldInfo) setTag(f *reflect.StructField, tag int, wt uint64) {
+	fi.field = toField(f)
+	fi.wiretag = uint64(tag)<<3 | wt
+	fi.tagsize = SizeVarint(uint64(tag) << 3)
+}
+
+// setMarshaler fills up the sizer and marshaler in the info of a field.
+func (fi *marshalFieldInfo) setMarshaler(f *reflect.StructField, tags []string) {
+	switch f.Type.Kind() {
+	case reflect.Map:
+		// map field
+		fi.isPointer = true
+		fi.sizer, fi.marshaler = makeMapMarshaler(f)
+		return
+	case reflect.Ptr, reflect.Slice:
+		fi.isPointer = true
+	}
+	fi.sizer, fi.marshaler = typeMarshaler(f.Type, tags, true, false)
+}
+
+// typeMarshaler returns the sizer and marshaler of a given field.
+// t is the type of the field.
+// tags is the generated "protobuf" tag of the field.
+// If nozero is true, zero value is not marshaled to the wire.
+// If oneof is true, it is a oneof field.
+func typeMarshaler(t reflect.Type, tags []string, nozero, oneof bool) (sizer, marshaler) {
+	encoding := tags[0]
+
+	pointer := false
+	slice := false
+	if t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8 {
+		slice = true
+		t = t.Elem()
+	}
+	if t.Kind() == reflect.Ptr {
+		pointer = true
+		t = t.Elem()
+	}
+
+	packed := false
+	proto3 := false
+	ctype := false
+	isTime := false
+	isDuration := false
+	isWktPointer := false
+	validateUTF8 := true
+	for i := 2; i < len(tags); i++ {
+		if tags[i] == "packed" {
+			packed = true
+		}
+		if tags[i] == "proto3" {
+			proto3 = true
+		}
+		if strings.HasPrefix(tags[i], "customtype=") {
+			ctype = true
+		}
+		if tags[i] == "stdtime" {
+			isTime = true
+		}
+		if tags[i] == "stdduration" {
+			isDuration = true
+		}
+		if tags[i] == "wktptr" {
+			isWktPointer = true
+		}
+	}
+	validateUTF8 = validateUTF8 && proto3
+	if !proto3 && !pointer && !slice {
+		nozero = false
+	}
+
+	if ctype {
+		if reflect.PtrTo(t).Implements(customType) {
+			if slice {
+				return makeMessageRefSliceMarshaler(getMarshalInfo(t))
+			}
+			if pointer {
+				return makeCustomPtrMarshaler(getMarshalInfo(t))
+			}
+			return makeCustomMarshaler(getMarshalInfo(t))
+		} else {
+			panic(fmt.Sprintf("custom type: type: %v, does not implement the proto.custom interface", t))
+		}
+	}
+
+	if isTime {
+		if pointer {
+			if slice {
+				return makeTimePtrSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeTimePtrMarshaler(getMarshalInfo(t))
+		}
+		if slice {
+			return makeTimeSliceMarshaler(getMarshalInfo(t))
+		}
+		return makeTimeMarshaler(getMarshalInfo(t))
+	}
+
+	if isDuration {
+		if pointer {
+			if slice {
+				return makeDurationPtrSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeDurationPtrMarshaler(getMarshalInfo(t))
+		}
+		if slice {
+			return makeDurationSliceMarshaler(getMarshalInfo(t))
+		}
+		return makeDurationMarshaler(getMarshalInfo(t))
+	}
+
+	if isWktPointer {
+		switch t.Kind() {
+		case reflect.Float64:
+			if pointer {
+				if slice {
+					return makeStdDoubleValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdDoubleValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdDoubleValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdDoubleValueMarshaler(getMarshalInfo(t))
+		case reflect.Float32:
+			if pointer {
+				if slice {
+					return makeStdFloatValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdFloatValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdFloatValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdFloatValueMarshaler(getMarshalInfo(t))
+		case reflect.Int64:
+			if pointer {
+				if slice {
+					return makeStdInt64ValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdInt64ValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdInt64ValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdInt64ValueMarshaler(getMarshalInfo(t))
+		case reflect.Uint64:
+			if pointer {
+				if slice {
+					return makeStdUInt64ValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdUInt64ValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdUInt64ValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdUInt64ValueMarshaler(getMarshalInfo(t))
+		case reflect.Int32:
+			if pointer {
+				if slice {
+					return makeStdInt32ValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdInt32ValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdInt32ValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdInt32ValueMarshaler(getMarshalInfo(t))
+		case reflect.Uint32:
+			if pointer {
+				if slice {
+					return makeStdUInt32ValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdUInt32ValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdUInt32ValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdUInt32ValueMarshaler(getMarshalInfo(t))
+		case reflect.Bool:
+			if pointer {
+				if slice {
+					return makeStdBoolValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdBoolValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdBoolValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdBoolValueMarshaler(getMarshalInfo(t))
+		case reflect.String:
+			if pointer {
+				if slice {
+					return makeStdStringValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdStringValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdStringValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdStringValueMarshaler(getMarshalInfo(t))
+		case uint8SliceType:
+			if pointer {
+				if slice {
+					return makeStdBytesValuePtrSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeStdBytesValuePtrMarshaler(getMarshalInfo(t))
+			}
+			if slice {
+				return makeStdBytesValueSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeStdBytesValueMarshaler(getMarshalInfo(t))
+		default:
+			panic(fmt.Sprintf("unknown wktpointer type %#v", t))
+		}
+	}
+
+	switch t.Kind() {
+	case reflect.Bool:
+		if pointer {
+			return sizeBoolPtr, appendBoolPtr
+		}
+		if slice {
+			if packed {
+				return sizeBoolPackedSlice, appendBoolPackedSlice
+			}
+			return sizeBoolSlice, appendBoolSlice
+		}
+		if nozero {
+			return sizeBoolValueNoZero, appendBoolValueNoZero
+		}
+		return sizeBoolValue, appendBoolValue
+	case reflect.Uint32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return sizeFixed32Ptr, appendFixed32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixed32PackedSlice, appendFixed32PackedSlice
+				}
+				return sizeFixed32Slice, appendFixed32Slice
+			}
+			if nozero {
+				return sizeFixed32ValueNoZero, appendFixed32ValueNoZero
+			}
+			return sizeFixed32Value, appendFixed32Value
+		case "varint":
+			if pointer {
+				return sizeVarint32Ptr, appendVarint32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarint32PackedSlice, appendVarint32PackedSlice
+				}
+				return sizeVarint32Slice, appendVarint32Slice
+			}
+			if nozero {
+				return sizeVarint32ValueNoZero, appendVarint32ValueNoZero
+			}
+			return sizeVarint32Value, appendVarint32Value
+		}
+	case reflect.Int32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return sizeFixedS32Ptr, appendFixedS32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixedS32PackedSlice, appendFixedS32PackedSlice
+				}
+				return sizeFixedS32Slice, appendFixedS32Slice
+			}
+			if nozero {
+				return sizeFixedS32ValueNoZero, appendFixedS32ValueNoZero
+			}
+			return sizeFixedS32Value, appendFixedS32Value
+		case "varint":
+			if pointer {
+				return sizeVarintS32Ptr, appendVarintS32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarintS32PackedSlice, appendVarintS32PackedSlice
+				}
+				return sizeVarintS32Slice, appendVarintS32Slice
+			}
+			if nozero {
+				return sizeVarintS32ValueNoZero, appendVarintS32ValueNoZero
+			}
+			return sizeVarintS32Value, appendVarintS32Value
+		case "zigzag32":
+			if pointer {
+				return sizeZigzag32Ptr, appendZigzag32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeZigzag32PackedSlice, appendZigzag32PackedSlice
+				}
+				return sizeZigzag32Slice, appendZigzag32Slice
+			}
+			if nozero {
+				return sizeZigzag32ValueNoZero, appendZigzag32ValueNoZero
+			}
+			return sizeZigzag32Value, appendZigzag32Value
+		}
+	case reflect.Uint64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return sizeFixed64Ptr, appendFixed64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixed64PackedSlice, appendFixed64PackedSlice
+				}
+				return sizeFixed64Slice, appendFixed64Slice
+			}
+			if nozero {
+				return sizeFixed64ValueNoZero, appendFixed64ValueNoZero
+			}
+			return sizeFixed64Value, appendFixed64Value
+		case "varint":
+			if pointer {
+				return sizeVarint64Ptr, appendVarint64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarint64PackedSlice, appendVarint64PackedSlice
+				}
+				return sizeVarint64Slice, appendVarint64Slice
+			}
+			if nozero {
+				return sizeVarint64ValueNoZero, appendVarint64ValueNoZero
+			}
+			return sizeVarint64Value, appendVarint64Value
+		}
+	case reflect.Int64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return sizeFixedS64Ptr, appendFixedS64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixedS64PackedSlice, appendFixedS64PackedSlice
+				}
+				return sizeFixedS64Slice, appendFixedS64Slice
+			}
+			if nozero {
+				return sizeFixedS64ValueNoZero, appendFixedS64ValueNoZero
+			}
+			return sizeFixedS64Value, appendFixedS64Value
+		case "varint":
+			if pointer {
+				return sizeVarintS64Ptr, appendVarintS64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarintS64PackedSlice, appendVarintS64PackedSlice
+				}
+				return sizeVarintS64Slice, appendVarintS64Slice
+			}
+			if nozero {
+				return sizeVarintS64ValueNoZero, appendVarintS64ValueNoZero
+			}
+			return sizeVarintS64Value, appendVarintS64Value
+		case "zigzag64":
+			if pointer {
+				return sizeZigzag64Ptr, appendZigzag64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeZigzag64PackedSlice, appendZigzag64PackedSlice
+				}
+				return sizeZigzag64Slice, appendZigzag64Slice
+			}
+			if nozero {
+				return sizeZigzag64ValueNoZero, appendZigzag64ValueNoZero
+			}
+			return sizeZigzag64Value, appendZigzag64Value
+		}
+	case reflect.Float32:
+		if pointer {
+			return sizeFloat32Ptr, appendFloat32Ptr
+		}
+		if slice {
+			if packed {
+				return sizeFloat32PackedSlice, appendFloat32PackedSlice
+			}
+			return sizeFloat32Slice, appendFloat32Slice
+		}
+		if nozero {
+			return sizeFloat32ValueNoZero, appendFloat32ValueNoZero
+		}
+		return sizeFloat32Value, appendFloat32Value
+	case reflect.Float64:
+		if pointer {
+			return sizeFloat64Ptr, appendFloat64Ptr
+		}
+		if slice {
+			if packed {
+				return sizeFloat64PackedSlice, appendFloat64PackedSlice
+			}
+			return sizeFloat64Slice, appendFloat64Slice
+		}
+		if nozero {
+			return sizeFloat64ValueNoZero, appendFloat64ValueNoZero
+		}
+		return sizeFloat64Value, appendFloat64Value
+	case reflect.String:
+		if validateUTF8 {
+			if pointer {
+				return sizeStringPtr, appendUTF8StringPtr
+			}
+			if slice {
+				return sizeStringSlice, appendUTF8StringSlice
+			}
+			if nozero {
+				return sizeStringValueNoZero, appendUTF8StringValueNoZero
+			}
+			return sizeStringValue, appendUTF8StringValue
+		}
+		if pointer {
+			return sizeStringPtr, appendStringPtr
+		}
+		if slice {
+			return sizeStringSlice, appendStringSlice
+		}
+		if nozero {
+			return sizeStringValueNoZero, appendStringValueNoZero
+		}
+		return sizeStringValue, appendStringValue
+	case reflect.Slice:
+		if slice {
+			return sizeBytesSlice, appendBytesSlice
+		}
+		if oneof {
+			// Oneof bytes field may also have "proto3" tag.
+			// We want to marshal it as a oneof field. Do this
+			// check before the proto3 check.
+			return sizeBytesOneof, appendBytesOneof
+		}
+		if proto3 {
+			return sizeBytes3, appendBytes3
+		}
+		return sizeBytes, appendBytes
+	case reflect.Struct:
+		switch encoding {
+		case "group":
+			if slice {
+				return makeGroupSliceMarshaler(getMarshalInfo(t))
+			}
+			return makeGroupMarshaler(getMarshalInfo(t))
+		case "bytes":
+			if pointer {
+				if slice {
+					return makeMessageSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeMessageMarshaler(getMarshalInfo(t))
+			} else {
+				if slice {
+					return makeMessageRefSliceMarshaler(getMarshalInfo(t))
+				}
+				return makeMessageRefMarshaler(getMarshalInfo(t))
+			}
+		}
+	}
+	panic(fmt.Sprintf("unknown or mismatched type: type: %v, wire type: %v", t, encoding))
+}
+
+// Below are functions to size/marshal a specific type of a field.
+// They are stored in the field's info, and called by function pointers.
+// They have type sizer or marshaler.
+
+func sizeFixed32Value(_ pointer, tagsize int) int {
+	return 4 + tagsize
+}
+func sizeFixed32ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toUint32()
+	if v == 0 {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFixed32Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toUint32Ptr()
+	if p == nil {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFixed32Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint32Slice()
+	return (4 + tagsize) * len(s)
+}
+func sizeFixed32PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 4*len(s) + SizeVarint(uint64(4*len(s))) + tagsize
+}
+func sizeFixedS32Value(_ pointer, tagsize int) int {
+	return 4 + tagsize
+}
+func sizeFixedS32ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFixedS32Ptr(ptr pointer, tagsize int) int {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFixedS32Slice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	return (4 + tagsize) * len(s)
+}
+func sizeFixedS32PackedSlice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 4*len(s) + SizeVarint(uint64(4*len(s))) + tagsize
+}
+func sizeFloat32Value(_ pointer, tagsize int) int {
+	return 4 + tagsize
+}
+func sizeFloat32ValueNoZero(ptr pointer, tagsize int) int {
+	v := math.Float32bits(*ptr.toFloat32())
+	if v == 0 {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFloat32Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toFloat32Ptr()
+	if p == nil {
+		return 0
+	}
+	return 4 + tagsize
+}
+func sizeFloat32Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toFloat32Slice()
+	return (4 + tagsize) * len(s)
+}
+func sizeFloat32PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toFloat32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 4*len(s) + SizeVarint(uint64(4*len(s))) + tagsize
+}
+func sizeFixed64Value(_ pointer, tagsize int) int {
+	return 8 + tagsize
+}
+func sizeFixed64ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toUint64()
+	if v == 0 {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFixed64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toUint64Ptr()
+	if p == nil {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFixed64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint64Slice()
+	return (8 + tagsize) * len(s)
+}
+func sizeFixed64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 8*len(s) + SizeVarint(uint64(8*len(s))) + tagsize
+}
+func sizeFixedS64Value(_ pointer, tagsize int) int {
+	return 8 + tagsize
+}
+func sizeFixedS64ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFixedS64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFixedS64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	return (8 + tagsize) * len(s)
+}
+func sizeFixedS64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 8*len(s) + SizeVarint(uint64(8*len(s))) + tagsize
+}
+func sizeFloat64Value(_ pointer, tagsize int) int {
+	return 8 + tagsize
+}
+func sizeFloat64ValueNoZero(ptr pointer, tagsize int) int {
+	v := math.Float64bits(*ptr.toFloat64())
+	if v == 0 {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFloat64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toFloat64Ptr()
+	if p == nil {
+		return 0
+	}
+	return 8 + tagsize
+}
+func sizeFloat64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toFloat64Slice()
+	return (8 + tagsize) * len(s)
+}
+func sizeFloat64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toFloat64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	return 8*len(s) + SizeVarint(uint64(8*len(s))) + tagsize
+}
+func sizeVarint32Value(ptr pointer, tagsize int) int {
+	v := *ptr.toUint32()
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarint32ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toUint32()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarint32Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toUint32Ptr()
+	if p == nil {
+		return 0
+	}
+	return SizeVarint(uint64(*p)) + tagsize
+}
+func sizeVarint32Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint32Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v)) + tagsize
+	}
+	return n
+}
+func sizeVarint32PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeVarintS32Value(ptr pointer, tagsize int) int {
+	v := *ptr.toInt32()
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarintS32ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarintS32Ptr(ptr pointer, tagsize int) int {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return 0
+	}
+	return SizeVarint(uint64(*p)) + tagsize
+}
+func sizeVarintS32Slice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v)) + tagsize
+	}
+	return n
+}
+func sizeVarintS32PackedSlice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeVarint64Value(ptr pointer, tagsize int) int {
+	v := *ptr.toUint64()
+	return SizeVarint(v) + tagsize
+}
+func sizeVarint64ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toUint64()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(v) + tagsize
+}
+func sizeVarint64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toUint64Ptr()
+	if p == nil {
+		return 0
+	}
+	return SizeVarint(*p) + tagsize
+}
+func sizeVarint64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint64Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(v) + tagsize
+	}
+	return n
+}
+func sizeVarint64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toUint64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(v)
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeVarintS64Value(ptr pointer, tagsize int) int {
+	v := *ptr.toInt64()
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarintS64ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(uint64(v)) + tagsize
+}
+func sizeVarintS64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return 0
+	}
+	return SizeVarint(uint64(*p)) + tagsize
+}
+func sizeVarintS64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v)) + tagsize
+	}
+	return n
+}
+func sizeVarintS64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeZigzag32Value(ptr pointer, tagsize int) int {
+	v := *ptr.toInt32()
+	return SizeVarint(uint64((uint32(v)<<1)^uint32((int32(v)>>31)))) + tagsize
+}
+func sizeZigzag32ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(uint64((uint32(v)<<1)^uint32((int32(v)>>31)))) + tagsize
+}
+func sizeZigzag32Ptr(ptr pointer, tagsize int) int {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return 0
+	}
+	v := *p
+	return SizeVarint(uint64((uint32(v)<<1)^uint32((int32(v)>>31)))) + tagsize
+}
+func sizeZigzag32Slice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64((uint32(v)<<1)^uint32((int32(v)>>31)))) + tagsize
+	}
+	return n
+}
+func sizeZigzag32PackedSlice(ptr pointer, tagsize int) int {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64((uint32(v) << 1) ^ uint32((int32(v) >> 31))))
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeZigzag64Value(ptr pointer, tagsize int) int {
+	v := *ptr.toInt64()
+	return SizeVarint(uint64(v<<1)^uint64((int64(v)>>63))) + tagsize
+}
+func sizeZigzag64ValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return 0
+	}
+	return SizeVarint(uint64(v<<1)^uint64((int64(v)>>63))) + tagsize
+}
+func sizeZigzag64Ptr(ptr pointer, tagsize int) int {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return 0
+	}
+	v := *p
+	return SizeVarint(uint64(v<<1)^uint64((int64(v)>>63))) + tagsize
+}
+func sizeZigzag64Slice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v<<1)^uint64((int64(v)>>63))) + tagsize
+	}
+	return n
+}
+func sizeZigzag64PackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return 0
+	}
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v<<1) ^ uint64((int64(v) >> 63)))
+	}
+	return n + SizeVarint(uint64(n)) + tagsize
+}
+func sizeBoolValue(_ pointer, tagsize int) int {
+	return 1 + tagsize
+}
+func sizeBoolValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toBool()
+	if !v {
+		return 0
+	}
+	return 1 + tagsize
+}
+func sizeBoolPtr(ptr pointer, tagsize int) int {
+	p := *ptr.toBoolPtr()
+	if p == nil {
+		return 0
+	}
+	return 1 + tagsize
+}
+func sizeBoolSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toBoolSlice()
+	return (1 + tagsize) * len(s)
+}
+func sizeBoolPackedSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toBoolSlice()
+	if len(s) == 0 {
+		return 0
+	}
+	return len(s) + SizeVarint(uint64(len(s))) + tagsize
+}
+func sizeStringValue(ptr pointer, tagsize int) int {
+	v := *ptr.toString()
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeStringValueNoZero(ptr pointer, tagsize int) int {
+	v := *ptr.toString()
+	if v == "" {
+		return 0
+	}
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeStringPtr(ptr pointer, tagsize int) int {
+	p := *ptr.toStringPtr()
+	if p == nil {
+		return 0
+	}
+	v := *p
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeStringSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toStringSlice()
+	n := 0
+	for _, v := range s {
+		n += len(v) + SizeVarint(uint64(len(v))) + tagsize
+	}
+	return n
+}
+func sizeBytes(ptr pointer, tagsize int) int {
+	v := *ptr.toBytes()
+	if v == nil {
+		return 0
+	}
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeBytes3(ptr pointer, tagsize int) int {
+	v := *ptr.toBytes()
+	if len(v) == 0 {
+		return 0
+	}
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeBytesOneof(ptr pointer, tagsize int) int {
+	v := *ptr.toBytes()
+	return len(v) + SizeVarint(uint64(len(v))) + tagsize
+}
+func sizeBytesSlice(ptr pointer, tagsize int) int {
+	s := *ptr.toBytesSlice()
+	n := 0
+	for _, v := range s {
+		n += len(v) + SizeVarint(uint64(len(v))) + tagsize
+	}
+	return n
+}
+
+// appendFixed32 appends an encoded fixed32 to b.
+func appendFixed32(b []byte, v uint32) []byte {
+	b = append(b,
+		byte(v),
+		byte(v>>8),
+		byte(v>>16),
+		byte(v>>24))
+	return b
+}
+
+// appendFixed64 appends an encoded fixed64 to b.
+func appendFixed64(b []byte, v uint64) []byte {
+	b = append(b,
+		byte(v),
+		byte(v>>8),
+		byte(v>>16),
+		byte(v>>24),
+		byte(v>>32),
+		byte(v>>40),
+		byte(v>>48),
+		byte(v>>56))
+	return b
+}
+
+// appendVarint appends an encoded varint to b.
+func appendVarint(b []byte, v uint64) []byte {
+	// TODO: make 1-byte (maybe 2-byte) case inline-able, once we
+	// have non-leaf inliner.
+	switch {
+	case v < 1<<7:
+		b = append(b, byte(v))
+	case v < 1<<14:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte(v>>7))
+	case v < 1<<21:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte(v>>14))
+	case v < 1<<28:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte(v>>21))
+	case v < 1<<35:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte(v>>28))
+	case v < 1<<42:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte((v>>28)&0x7f|0x80),
+			byte(v>>35))
+	case v < 1<<49:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte((v>>28)&0x7f|0x80),
+			byte((v>>35)&0x7f|0x80),
+			byte(v>>42))
+	case v < 1<<56:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte((v>>28)&0x7f|0x80),
+			byte((v>>35)&0x7f|0x80),
+			byte((v>>42)&0x7f|0x80),
+			byte(v>>49))
+	case v < 1<<63:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte((v>>28)&0x7f|0x80),
+			byte((v>>35)&0x7f|0x80),
+			byte((v>>42)&0x7f|0x80),
+			byte((v>>49)&0x7f|0x80),
+			byte(v>>56))
+	default:
+		b = append(b,
+			byte(v&0x7f|0x80),
+			byte((v>>7)&0x7f|0x80),
+			byte((v>>14)&0x7f|0x80),
+			byte((v>>21)&0x7f|0x80),
+			byte((v>>28)&0x7f|0x80),
+			byte((v>>35)&0x7f|0x80),
+			byte((v>>42)&0x7f|0x80),
+			byte((v>>49)&0x7f|0x80),
+			byte((v>>56)&0x7f|0x80),
+			1)
+	}
+	return b
+}
+
+func appendFixed32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint32()
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, v)
+	return b, nil
+}
+func appendFixed32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint32()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, v)
+	return b, nil
+}
+func appendFixed32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toUint32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, *p)
+	return b, nil
+}
+func appendFixed32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed32(b, v)
+	}
+	return b, nil
+}
+func appendFixed32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(4*len(s)))
+	for _, v := range s {
+		b = appendFixed32(b, v)
+	}
+	return b, nil
+}
+func appendFixedS32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, uint32(v))
+	return b, nil
+}
+func appendFixedS32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, uint32(v))
+	return b, nil
+}
+func appendFixedS32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, uint32(*p))
+	return b, nil
+}
+func appendFixedS32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed32(b, uint32(v))
+	}
+	return b, nil
+}
+func appendFixedS32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(4*len(s)))
+	for _, v := range s {
+		b = appendFixed32(b, uint32(v))
+	}
+	return b, nil
+}
+func appendFloat32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := math.Float32bits(*ptr.toFloat32())
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, v)
+	return b, nil
+}
+func appendFloat32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := math.Float32bits(*ptr.toFloat32())
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, v)
+	return b, nil
+}
+func appendFloat32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toFloat32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed32(b, math.Float32bits(*p))
+	return b, nil
+}
+func appendFloat32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toFloat32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed32(b, math.Float32bits(v))
+	}
+	return b, nil
+}
+func appendFloat32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toFloat32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(4*len(s)))
+	for _, v := range s {
+		b = appendFixed32(b, math.Float32bits(v))
+	}
+	return b, nil
+}
+func appendFixed64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint64()
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, v)
+	return b, nil
+}
+func appendFixed64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint64()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, v)
+	return b, nil
+}
+func appendFixed64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toUint64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, *p)
+	return b, nil
+}
+func appendFixed64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed64(b, v)
+	}
+	return b, nil
+}
+func appendFixed64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(8*len(s)))
+	for _, v := range s {
+		b = appendFixed64(b, v)
+	}
+	return b, nil
+}
+func appendFixedS64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, uint64(v))
+	return b, nil
+}
+func appendFixedS64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, uint64(v))
+	return b, nil
+}
+func appendFixedS64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, uint64(*p))
+	return b, nil
+}
+func appendFixedS64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed64(b, uint64(v))
+	}
+	return b, nil
+}
+func appendFixedS64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(8*len(s)))
+	for _, v := range s {
+		b = appendFixed64(b, uint64(v))
+	}
+	return b, nil
+}
+func appendFloat64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := math.Float64bits(*ptr.toFloat64())
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, v)
+	return b, nil
+}
+func appendFloat64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := math.Float64bits(*ptr.toFloat64())
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, v)
+	return b, nil
+}
+func appendFloat64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toFloat64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendFixed64(b, math.Float64bits(*p))
+	return b, nil
+}
+func appendFloat64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toFloat64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendFixed64(b, math.Float64bits(v))
+	}
+	return b, nil
+}
+func appendFloat64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toFloat64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(8*len(s)))
+	for _, v := range s {
+		b = appendFixed64(b, math.Float64bits(v))
+	}
+	return b, nil
+}
+func appendVarint32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint32()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarint32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint32()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarint32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toUint32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(*p))
+	return b, nil
+}
+func appendVarint32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendVarint32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendVarintS32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarintS32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarintS32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(*p))
+	return b, nil
+}
+func appendVarintS32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendVarintS32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendVarint64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint64()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, v)
+	return b, nil
+}
+func appendVarint64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toUint64()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, v)
+	return b, nil
+}
+func appendVarint64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toUint64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, *p)
+	return b, nil
+}
+func appendVarint64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, v)
+	}
+	return b, nil
+}
+func appendVarint64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toUint64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(v)
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, v)
+	}
+	return b, nil
+}
+func appendVarintS64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarintS64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v))
+	return b, nil
+}
+func appendVarintS64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(*p))
+	return b, nil
+}
+func appendVarintS64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendVarintS64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v))
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, uint64(v))
+	}
+	return b, nil
+}
+func appendZigzag32Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64((uint32(v)<<1)^uint32((int32(v)>>31))))
+	return b, nil
+}
+func appendZigzag32ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt32()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64((uint32(v)<<1)^uint32((int32(v)>>31))))
+	return b, nil
+}
+func appendZigzag32Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := ptr.getInt32Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	v := *p
+	b = appendVarint(b, uint64((uint32(v)<<1)^uint32((int32(v)>>31))))
+	return b, nil
+}
+func appendZigzag32Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64((uint32(v)<<1)^uint32((int32(v)>>31))))
+	}
+	return b, nil
+}
+func appendZigzag32PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := ptr.getInt32Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64((uint32(v) << 1) ^ uint32((int32(v) >> 31))))
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, uint64((uint32(v)<<1)^uint32((int32(v)>>31))))
+	}
+	return b, nil
+}
+func appendZigzag64Value(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v<<1)^uint64((int64(v)>>63)))
+	return b, nil
+}
+func appendZigzag64ValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toInt64()
+	if v == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(v<<1)^uint64((int64(v)>>63)))
+	return b, nil
+}
+func appendZigzag64Ptr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toInt64Ptr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	v := *p
+	b = appendVarint(b, uint64(v<<1)^uint64((int64(v)>>63)))
+	return b, nil
+}
+func appendZigzag64Slice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(v<<1)^uint64((int64(v)>>63)))
+	}
+	return b, nil
+}
+func appendZigzag64PackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toInt64Slice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	// compute size
+	n := 0
+	for _, v := range s {
+		n += SizeVarint(uint64(v<<1) ^ uint64((int64(v) >> 63)))
+	}
+	b = appendVarint(b, uint64(n))
+	for _, v := range s {
+		b = appendVarint(b, uint64(v<<1)^uint64((int64(v)>>63)))
+	}
+	return b, nil
+}
+func appendBoolValue(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toBool()
+	b = appendVarint(b, wiretag)
+	if v {
+		b = append(b, 1)
+	} else {
+		b = append(b, 0)
+	}
+	return b, nil
+}
+func appendBoolValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toBool()
+	if !v {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = append(b, 1)
+	return b, nil
+}
+
+func appendBoolPtr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toBoolPtr()
+	if p == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	if *p {
+		b = append(b, 1)
+	} else {
+		b = append(b, 0)
+	}
+	return b, nil
+}
+func appendBoolSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toBoolSlice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		if v {
+			b = append(b, 1)
+		} else {
+			b = append(b, 0)
+		}
+	}
+	return b, nil
+}
+func appendBoolPackedSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toBoolSlice()
+	if len(s) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag&^7|WireBytes)
+	b = appendVarint(b, uint64(len(s)))
+	for _, v := range s {
+		if v {
+			b = append(b, 1)
+		} else {
+			b = append(b, 0)
+		}
+	}
+	return b, nil
+}
+func appendStringValue(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toString()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendStringValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toString()
+	if v == "" {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendStringPtr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	p := *ptr.toStringPtr()
+	if p == nil {
+		return b, nil
+	}
+	v := *p
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendStringSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toStringSlice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(len(v)))
+		b = append(b, v...)
+	}
+	return b, nil
+}
+func appendUTF8StringValue(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	var invalidUTF8 bool
+	v := *ptr.toString()
+	if !utf8.ValidString(v) {
+		invalidUTF8 = true
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	if invalidUTF8 {
+		return b, errInvalidUTF8
+	}
+	return b, nil
+}
+func appendUTF8StringValueNoZero(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	var invalidUTF8 bool
+	v := *ptr.toString()
+	if v == "" {
+		return b, nil
+	}
+	if !utf8.ValidString(v) {
+		invalidUTF8 = true
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	if invalidUTF8 {
+		return b, errInvalidUTF8
+	}
+	return b, nil
+}
+func appendUTF8StringPtr(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	var invalidUTF8 bool
+	p := *ptr.toStringPtr()
+	if p == nil {
+		return b, nil
+	}
+	v := *p
+	if !utf8.ValidString(v) {
+		invalidUTF8 = true
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	if invalidUTF8 {
+		return b, errInvalidUTF8
+	}
+	return b, nil
+}
+func appendUTF8StringSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	var invalidUTF8 bool
+	s := *ptr.toStringSlice()
+	for _, v := range s {
+		if !utf8.ValidString(v) {
+			invalidUTF8 = true
+		}
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(len(v)))
+		b = append(b, v...)
+	}
+	if invalidUTF8 {
+		return b, errInvalidUTF8
+	}
+	return b, nil
+}
+func appendBytes(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toBytes()
+	if v == nil {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendBytes3(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toBytes()
+	if len(v) == 0 {
+		return b, nil
+	}
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendBytesOneof(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	v := *ptr.toBytes()
+	b = appendVarint(b, wiretag)
+	b = appendVarint(b, uint64(len(v)))
+	b = append(b, v...)
+	return b, nil
+}
+func appendBytesSlice(b []byte, ptr pointer, wiretag uint64, _ bool) ([]byte, error) {
+	s := *ptr.toBytesSlice()
+	for _, v := range s {
+		b = appendVarint(b, wiretag)
+		b = appendVarint(b, uint64(len(v)))
+		b = append(b, v...)
+	}
+	return b, nil
+}
+
+// makeGroupMarshaler returns the sizer and marshaler for a group.
+// u is the marshal info of the underlying message.
+func makeGroupMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			p := ptr.getPointer()
+			if p.isNil() {
+				return 0
+			}
+			return u.size(p) + 2*tagsize
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			p := ptr.getPointer()
+			if p.isNil() {
+				return b, nil
+			}
+			var err error
+			b = appendVarint(b, wiretag) // start group
+			b, err = u.marshal(b, p, deterministic)
+			b = appendVarint(b, wiretag+(WireEndGroup-WireStartGroup)) // end group
+			return b, err
+		}
+}
+
+// makeGroupSliceMarshaler returns the sizer and marshaler for a group slice.
+// u is the marshal info of the underlying message.
+func makeGroupSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getPointerSlice()
+			n := 0
+			for _, v := range s {
+				if v.isNil() {
+					continue
+				}
+				n += u.size(v) + 2*tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getPointerSlice()
+			var err error
+			var nerr nonFatal
+			for _, v := range s {
+				if v.isNil() {
+					return b, errRepeatedHasNil
+				}
+				b = appendVarint(b, wiretag) // start group
+				b, err = u.marshal(b, v, deterministic)
+				b = appendVarint(b, wiretag+(WireEndGroup-WireStartGroup)) // end group
+				if !nerr.Merge(err) {
+					if err == ErrNil {
+						err = errRepeatedHasNil
+					}
+					return b, err
+				}
+			}
+			return b, nerr.E
+		}
+}
+
+// makeMessageMarshaler returns the sizer and marshaler for a message field.
+// u is the marshal info of the message.
+func makeMessageMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			p := ptr.getPointer()
+			if p.isNil() {
+				return 0
+			}
+			siz := u.size(p)
+			return siz + SizeVarint(uint64(siz)) + tagsize
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			p := ptr.getPointer()
+			if p.isNil() {
+				return b, nil
+			}
+			b = appendVarint(b, wiretag)
+			siz := u.cachedsize(p)
+			b = appendVarint(b, uint64(siz))
+			return u.marshal(b, p, deterministic)
+		}
+}
+
+// makeMessageSliceMarshaler returns the sizer and marshaler for a message slice.
+// u is the marshal info of the message.
+func makeMessageSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getPointerSlice()
+			n := 0
+			for _, v := range s {
+				if v.isNil() {
+					continue
+				}
+				siz := u.size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getPointerSlice()
+			var err error
+			var nerr nonFatal
+			for _, v := range s {
+				if v.isNil() {
+					return b, errRepeatedHasNil
+				}
+				b = appendVarint(b, wiretag)
+				siz := u.cachedsize(v)
+				b = appendVarint(b, uint64(siz))
+				b, err = u.marshal(b, v, deterministic)
+
+				if !nerr.Merge(err) {
+					if err == ErrNil {
+						err = errRepeatedHasNil
+					}
+					return b, err
+				}
+			}
+			return b, nerr.E
+		}
+}
+
+// makeMapMarshaler returns the sizer and marshaler for a map field.
+// f is the pointer to the reflect data structure of the field.
+func makeMapMarshaler(f *reflect.StructField) (sizer, marshaler) {
+	// figure out key and value type
+	t := f.Type
+	keyType := t.Key()
+	valType := t.Elem()
+	tags := strings.Split(f.Tag.Get("protobuf"), ",")
+	keyTags := strings.Split(f.Tag.Get("protobuf_key"), ",")
+	valTags := strings.Split(f.Tag.Get("protobuf_val"), ",")
+	stdOptions := false
+	for _, t := range tags {
+		if strings.HasPrefix(t, "customtype=") {
+			valTags = append(valTags, t)
+		}
+		if t == "stdtime" {
+			valTags = append(valTags, t)
+			stdOptions = true
+		}
+		if t == "stdduration" {
+			valTags = append(valTags, t)
+			stdOptions = true
+		}
+		if t == "wktptr" {
+			valTags = append(valTags, t)
+		}
+	}
+	keySizer, keyMarshaler := typeMarshaler(keyType, keyTags, false, false) // don't omit zero value in map
+	valSizer, valMarshaler := typeMarshaler(valType, valTags, false, false) // don't omit zero value in map
+	keyWireTag := 1<<3 | wiretype(keyTags[0])
+	valWireTag := 2<<3 | wiretype(valTags[0])
+
+	// We create an interface to get the addresses of the map key and value.
+	// If value is pointer-typed, the interface is a direct interface, the
+	// idata itself is the value. Otherwise, the idata is the pointer to the
+	// value.
+	// Key cannot be pointer-typed.
+	valIsPtr := valType.Kind() == reflect.Ptr
+
+	// If value is a message with nested maps, calling
+	// valSizer in marshal may be quadratic. We should use
+	// cached version in marshal (but not in size).
+	// If value is not message type, we don't have size cache,
+	// but it cannot be nested either. Just use valSizer.
+	valCachedSizer := valSizer
+	if valIsPtr && !stdOptions && valType.Elem().Kind() == reflect.Struct {
+		u := getMarshalInfo(valType.Elem())
+		valCachedSizer = func(ptr pointer, tagsize int) int {
+			// Same as message sizer, but use cache.
+			p := ptr.getPointer()
+			if p.isNil() {
+				return 0
+			}
+			siz := u.cachedsize(p)
+			return siz + SizeVarint(uint64(siz)) + tagsize
+		}
+	}
+	return func(ptr pointer, tagsize int) int {
+			m := ptr.asPointerTo(t).Elem() // the map
+			n := 0
+			for _, k := range m.MapKeys() {
+				ki := k.Interface()
+				vi := m.MapIndex(k).Interface()
+				kaddr := toAddrPointer(&ki, false)             // pointer to key
+				vaddr := toAddrPointer(&vi, valIsPtr)          // pointer to value
+				siz := keySizer(kaddr, 1) + valSizer(vaddr, 1) // tag of key = 1 (size=1), tag of val = 2 (size=1)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, tag uint64, deterministic bool) ([]byte, error) {
+			m := ptr.asPointerTo(t).Elem() // the map
+			var err error
+			keys := m.MapKeys()
+			if len(keys) > 1 && deterministic {
+				sort.Sort(mapKeys(keys))
+			}
+
+			var nerr nonFatal
+			for _, k := range keys {
+				ki := k.Interface()
+				vi := m.MapIndex(k).Interface()
+				kaddr := toAddrPointer(&ki, false)    // pointer to key
+				vaddr := toAddrPointer(&vi, valIsPtr) // pointer to value
+				b = appendVarint(b, tag)
+				siz := keySizer(kaddr, 1) + valCachedSizer(vaddr, 1) // tag of key = 1 (size=1), tag of val = 2 (size=1)
+				b = appendVarint(b, uint64(siz))
+				b, err = keyMarshaler(b, kaddr, keyWireTag, deterministic)
+				if !nerr.Merge(err) {
+					return b, err
+				}
+				b, err = valMarshaler(b, vaddr, valWireTag, deterministic)
+				if err != ErrNil && !nerr.Merge(err) { // allow nil value in map
+					return b, err
+				}
+			}
+			return b, nerr.E
+		}
+}
+
+// makeOneOfMarshaler returns the sizer and marshaler for a oneof field.
+// fi is the marshal info of the field.
+// f is the pointer to the reflect data structure of the field.
+func makeOneOfMarshaler(fi *marshalFieldInfo, f *reflect.StructField) (sizer, marshaler) {
+	// Oneof field is an interface. We need to get the actual data type on the fly.
+	t := f.Type
+	return func(ptr pointer, _ int) int {
+			p := ptr.getInterfacePointer()
+			if p.isNil() {
+				return 0
+			}
+			v := ptr.asPointerTo(t).Elem().Elem().Elem() // *interface -> interface -> *struct -> struct
+			telem := v.Type()
+			e := fi.oneofElems[telem]
+			return e.sizer(p, e.tagsize)
+		},
+		func(b []byte, ptr pointer, _ uint64, deterministic bool) ([]byte, error) {
+			p := ptr.getInterfacePointer()
+			if p.isNil() {
+				return b, nil
+			}
+			v := ptr.asPointerTo(t).Elem().Elem().Elem() // *interface -> interface -> *struct -> struct
+			telem := v.Type()
+			if telem.Field(0).Type.Kind() == reflect.Ptr && p.getPointer().isNil() {
+				return b, errOneofHasNil
+			}
+			e := fi.oneofElems[telem]
+			return e.marshaler(b, p, e.wiretag, deterministic)
+		}
+}
+
+// sizeExtensions computes the size of encoded data for a XXX_InternalExtensions field.
+func (u *marshalInfo) sizeExtensions(ext *XXX_InternalExtensions) int {
+	m, mu := ext.extensionsRead()
+	if m == nil {
+		return 0
+	}
+	mu.Lock()
+
+	n := 0
+	for _, e := range m {
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			n += len(e.enc)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		n += ei.sizer(p, ei.tagsize)
+	}
+	mu.Unlock()
+	return n
+}
+
+// appendExtensions marshals a XXX_InternalExtensions field to the end of byte slice b.
+func (u *marshalInfo) appendExtensions(b []byte, ext *XXX_InternalExtensions, deterministic bool) ([]byte, error) {
+	m, mu := ext.extensionsRead()
+	if m == nil {
+		return b, nil
+	}
+	mu.Lock()
+	defer mu.Unlock()
+
+	var err error
+	var nerr nonFatal
+
+	// Fast-path for common cases: zero or one extensions.
+	// Don't bother sorting the keys.
+	if len(m) <= 1 {
+		for _, e := range m {
+			if e.value == nil || e.desc == nil {
+				// Extension is only in its encoded form.
+				b = append(b, e.enc...)
+				continue
+			}
+
+			// We don't skip extensions that have an encoded form set,
+			// because the extension value may have been mutated after
+			// the last time this function was called.
+
+			ei := u.getExtElemInfo(e.desc)
+			v := e.value
+			p := toAddrPointer(&v, ei.isptr)
+			b, err = ei.marshaler(b, p, ei.wiretag, deterministic)
+			if !nerr.Merge(err) {
+				return b, err
+			}
+		}
+		return b, nerr.E
+	}
+
+	// Sort the keys to provide a deterministic encoding.
+	// Not sure this is required, but the old code does it.
+	keys := make([]int, 0, len(m))
+	for k := range m {
+		keys = append(keys, int(k))
+	}
+	sort.Ints(keys)
+
+	for _, k := range keys {
+		e := m[int32(k)]
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			b = append(b, e.enc...)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		b, err = ei.marshaler(b, p, ei.wiretag, deterministic)
+		if !nerr.Merge(err) {
+			return b, err
+		}
+	}
+	return b, nerr.E
+}
+
+// message set format is:
+//   message MessageSet {
+//     repeated group Item = 1 {
+//       required int32 type_id = 2;
+//       required string message = 3;
+//     };
+//   }
+
+// sizeMessageSet computes the size of encoded data for a XXX_InternalExtensions field
+// in message set format (above).
+func (u *marshalInfo) sizeMessageSet(ext *XXX_InternalExtensions) int {
+	m, mu := ext.extensionsRead()
+	if m == nil {
+		return 0
+	}
+	mu.Lock()
+
+	n := 0
+	for id, e := range m {
+		n += 2                          // start group, end group. tag = 1 (size=1)
+		n += SizeVarint(uint64(id)) + 1 // type_id, tag = 2 (size=1)
+
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			msgWithLen := skipVarint(e.enc) // skip old tag, but leave the length varint
+			siz := len(msgWithLen)
+			n += siz + 1 // message, tag = 3 (size=1)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		n += ei.sizer(p, 1) // message, tag = 3 (size=1)
+	}
+	mu.Unlock()
+	return n
+}
+
+// appendMessageSet marshals a XXX_InternalExtensions field in message set format (above)
+// to the end of byte slice b.
+func (u *marshalInfo) appendMessageSet(b []byte, ext *XXX_InternalExtensions, deterministic bool) ([]byte, error) {
+	m, mu := ext.extensionsRead()
+	if m == nil {
+		return b, nil
+	}
+	mu.Lock()
+	defer mu.Unlock()
+
+	var err error
+	var nerr nonFatal
+
+	// Fast-path for common cases: zero or one extensions.
+	// Don't bother sorting the keys.
+	if len(m) <= 1 {
+		for id, e := range m {
+			b = append(b, 1<<3|WireStartGroup)
+			b = append(b, 2<<3|WireVarint)
+			b = appendVarint(b, uint64(id))
+
+			if e.value == nil || e.desc == nil {
+				// Extension is only in its encoded form.
+				msgWithLen := skipVarint(e.enc) // skip old tag, but leave the length varint
+				b = append(b, 3<<3|WireBytes)
+				b = append(b, msgWithLen...)
+				b = append(b, 1<<3|WireEndGroup)
+				continue
+			}
+
+			// We don't skip extensions that have an encoded form set,
+			// because the extension value may have been mutated after
+			// the last time this function was called.
+
+			ei := u.getExtElemInfo(e.desc)
+			v := e.value
+			p := toAddrPointer(&v, ei.isptr)
+			b, err = ei.marshaler(b, p, 3<<3|WireBytes, deterministic)
+			if !nerr.Merge(err) {
+				return b, err
+			}
+			b = append(b, 1<<3|WireEndGroup)
+		}
+		return b, nerr.E
+	}
+
+	// Sort the keys to provide a deterministic encoding.
+	keys := make([]int, 0, len(m))
+	for k := range m {
+		keys = append(keys, int(k))
+	}
+	sort.Ints(keys)
+
+	for _, id := range keys {
+		e := m[int32(id)]
+		b = append(b, 1<<3|WireStartGroup)
+		b = append(b, 2<<3|WireVarint)
+		b = appendVarint(b, uint64(id))
+
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			msgWithLen := skipVarint(e.enc) // skip old tag, but leave the length varint
+			b = append(b, 3<<3|WireBytes)
+			b = append(b, msgWithLen...)
+			b = append(b, 1<<3|WireEndGroup)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		b, err = ei.marshaler(b, p, 3<<3|WireBytes, deterministic)
+		b = append(b, 1<<3|WireEndGroup)
+		if !nerr.Merge(err) {
+			return b, err
+		}
+	}
+	return b, nerr.E
+}
+
+// sizeV1Extensions computes the size of encoded data for a V1-API extension field.
+func (u *marshalInfo) sizeV1Extensions(m map[int32]Extension) int {
+	if m == nil {
+		return 0
+	}
+
+	n := 0
+	for _, e := range m {
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			n += len(e.enc)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		n += ei.sizer(p, ei.tagsize)
+	}
+	return n
+}
+
+// appendV1Extensions marshals a V1-API extension field to the end of byte slice b.
+func (u *marshalInfo) appendV1Extensions(b []byte, m map[int32]Extension, deterministic bool) ([]byte, error) {
+	if m == nil {
+		return b, nil
+	}
+
+	// Sort the keys to provide a deterministic encoding.
+	keys := make([]int, 0, len(m))
+	for k := range m {
+		keys = append(keys, int(k))
+	}
+	sort.Ints(keys)
+
+	var err error
+	var nerr nonFatal
+	for _, k := range keys {
+		e := m[int32(k)]
+		if e.value == nil || e.desc == nil {
+			// Extension is only in its encoded form.
+			b = append(b, e.enc...)
+			continue
+		}
+
+		// We don't skip extensions that have an encoded form set,
+		// because the extension value may have been mutated after
+		// the last time this function was called.
+
+		ei := u.getExtElemInfo(e.desc)
+		v := e.value
+		p := toAddrPointer(&v, ei.isptr)
+		b, err = ei.marshaler(b, p, ei.wiretag, deterministic)
+		if !nerr.Merge(err) {
+			return b, err
+		}
+	}
+	return b, nerr.E
+}
+
+// newMarshaler is the interface representing objects that can marshal themselves.
+//
+// This exists to support protoc-gen-go generated messages.
+// The proto package will stop type-asserting to this interface in the future.
+//
+// DO NOT DEPEND ON THIS.
+type newMarshaler interface {
+	XXX_Size() int
+	XXX_Marshal(b []byte, deterministic bool) ([]byte, error)
+}
+
+// Size returns the encoded size of a protocol buffer message.
+// This is the main entry point.
+func Size(pb Message) int {
+	if m, ok := pb.(newMarshaler); ok {
+		return m.XXX_Size()
+	}
+	if m, ok := pb.(Marshaler); ok {
+		// If the message can marshal itself, let it do it, for compatibility.
+		// NOTE: This is not efficient.
+		b, _ := m.Marshal()
+		return len(b)
+	}
+	// in case somehow we didn't generate the wrapper
+	if pb == nil {
+		return 0
+	}
+	var info InternalMessageInfo
+	return info.Size(pb)
+}
+
+// Marshal takes a protocol buffer message
+// and encodes it into the wire format, returning the data.
+// This is the main entry point.
+func Marshal(pb Message) ([]byte, error) {
+	if m, ok := pb.(newMarshaler); ok {
+		siz := m.XXX_Size()
+		b := make([]byte, 0, siz)
+		return m.XXX_Marshal(b, false)
+	}
+	if m, ok := pb.(Marshaler); ok {
+		// If the message can marshal itself, let it do it, for compatibility.
+		// NOTE: This is not efficient.
+		return m.Marshal()
+	}
+	// in case somehow we didn't generate the wrapper
+	if pb == nil {
+		return nil, ErrNil
+	}
+	var info InternalMessageInfo
+	siz := info.Size(pb)
+	b := make([]byte, 0, siz)
+	return info.Marshal(b, pb, false)
+}
+
+// Marshal takes a protocol buffer message
+// and encodes it into the wire format, writing the result to the
+// Buffer.
+// This is an alternative entry point. It is not necessary to use
+// a Buffer for most applications.
+func (p *Buffer) Marshal(pb Message) error {
+	var err error
+	if p.deterministic {
+		if _, ok := pb.(Marshaler); ok {
+			return fmt.Errorf("proto: deterministic not supported by the Marshal method of %T", pb)
+		}
+	}
+	if m, ok := pb.(newMarshaler); ok {
+		siz := m.XXX_Size()
+		p.grow(siz) // make sure buf has enough capacity
+		p.buf, err = m.XXX_Marshal(p.buf, p.deterministic)
+		return err
+	}
+	if m, ok := pb.(Marshaler); ok {
+		// If the message can marshal itself, let it do it, for compatibility.
+		// NOTE: This is not efficient.
+		var b []byte
+		b, err = m.Marshal()
+		p.buf = append(p.buf, b...)
+		return err
+	}
+	// in case somehow we didn't generate the wrapper
+	if pb == nil {
+		return ErrNil
+	}
+	var info InternalMessageInfo
+	siz := info.Size(pb)
+	p.grow(siz) // make sure buf has enough capacity
+	p.buf, err = info.Marshal(p.buf, pb, p.deterministic)
+	return err
+}
+
+// grow grows the buffer's capacity, if necessary, to guarantee space for
+// another n bytes. After grow(n), at least n bytes can be written to the
+// buffer without another allocation.
+func (p *Buffer) grow(n int) {
+	need := len(p.buf) + n
+	if need <= cap(p.buf) {
+		return
+	}
+	newCap := len(p.buf) * 2
+	if newCap < need {
+		newCap = need
+	}
+	p.buf = append(make([]byte, 0, newCap), p.buf...)
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/table_marshal_gogo.go b/vendor/github.com/gogo/protobuf/proto/table_marshal_gogo.go
new file mode 100644
index 0000000..997f57c
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/table_marshal_gogo.go
@@ -0,0 +1,388 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"reflect"
+	"time"
+)
+
+// makeMessageRefMarshaler differs a bit from makeMessageMarshaler
+// It marshal a message T instead of a *T
+func makeMessageRefMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			siz := u.size(ptr)
+			return siz + SizeVarint(uint64(siz)) + tagsize
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			b = appendVarint(b, wiretag)
+			siz := u.cachedsize(ptr)
+			b = appendVarint(b, uint64(siz))
+			return u.marshal(b, ptr, deterministic)
+		}
+}
+
+// makeMessageRefSliceMarshaler differs quite a lot from makeMessageSliceMarshaler
+// It marshals a slice of messages []T instead of []*T
+func makeMessageRefSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				e := elem.Interface()
+				v := toAddrPointer(&e, false)
+				siz := u.size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			var err, errreq error
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				e := elem.Interface()
+				v := toAddrPointer(&e, false)
+				b = appendVarint(b, wiretag)
+				siz := u.size(v)
+				b = appendVarint(b, uint64(siz))
+				b, err = u.marshal(b, v, deterministic)
+
+				if err != nil {
+					if _, ok := err.(*RequiredNotSetError); ok {
+						// Required field in submessage is not set.
+						// We record the error but keep going, to give a complete marshaling.
+						if errreq == nil {
+							errreq = err
+						}
+						continue
+					}
+					if err == ErrNil {
+						err = errRepeatedHasNil
+					}
+					return b, err
+				}
+			}
+
+			return b, errreq
+		}
+}
+
+func makeCustomPtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			m := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(custom)
+			siz := m.Size()
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			m := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(custom)
+			siz := m.Size()
+			buf, err := m.Marshal()
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(siz))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeCustomMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			m := ptr.asPointerTo(u.typ).Interface().(custom)
+			siz := m.Size()
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			m := ptr.asPointerTo(u.typ).Interface().(custom)
+			siz := m.Size()
+			buf, err := m.Marshal()
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(siz))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeTimeMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*time.Time)
+			ts, err := timestampProto(*t)
+			if err != nil {
+				return 0
+			}
+			siz := Size(ts)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*time.Time)
+			ts, err := timestampProto(*t)
+			if err != nil {
+				return nil, err
+			}
+			buf, err := Marshal(ts)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeTimePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*time.Time)
+			ts, err := timestampProto(*t)
+			if err != nil {
+				return 0
+			}
+			siz := Size(ts)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*time.Time)
+			ts, err := timestampProto(*t)
+			if err != nil {
+				return nil, err
+			}
+			buf, err := Marshal(ts)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeTimeSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(time.Time)
+				ts, err := timestampProto(t)
+				if err != nil {
+					return 0
+				}
+				siz := Size(ts)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(time.Time)
+				ts, err := timestampProto(t)
+				if err != nil {
+					return nil, err
+				}
+				siz := Size(ts)
+				buf, err := Marshal(ts)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeTimePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*time.Time)
+				ts, err := timestampProto(*t)
+				if err != nil {
+					return 0
+				}
+				siz := Size(ts)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*time.Time)
+				ts, err := timestampProto(*t)
+				if err != nil {
+					return nil, err
+				}
+				siz := Size(ts)
+				buf, err := Marshal(ts)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeDurationMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			d := ptr.asPointerTo(u.typ).Interface().(*time.Duration)
+			dur := durationProto(*d)
+			siz := Size(dur)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			d := ptr.asPointerTo(u.typ).Interface().(*time.Duration)
+			dur := durationProto(*d)
+			buf, err := Marshal(dur)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeDurationPtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			d := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*time.Duration)
+			dur := durationProto(*d)
+			siz := Size(dur)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			d := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*time.Duration)
+			dur := durationProto(*d)
+			buf, err := Marshal(dur)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeDurationSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				d := elem.Interface().(time.Duration)
+				dur := durationProto(d)
+				siz := Size(dur)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				d := elem.Interface().(time.Duration)
+				dur := durationProto(d)
+				siz := Size(dur)
+				buf, err := Marshal(dur)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeDurationPtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				d := elem.Interface().(*time.Duration)
+				dur := durationProto(*d)
+				siz := Size(dur)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				d := elem.Interface().(*time.Duration)
+				dur := durationProto(*d)
+				siz := Size(dur)
+				buf, err := Marshal(dur)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/table_merge.go b/vendor/github.com/gogo/protobuf/proto/table_merge.go
new file mode 100644
index 0000000..f520106
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/table_merge.go
@@ -0,0 +1,657 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+	"sync"
+	"sync/atomic"
+)
+
+// Merge merges the src message into dst.
+// This assumes that dst and src of the same type and are non-nil.
+func (a *InternalMessageInfo) Merge(dst, src Message) {
+	mi := atomicLoadMergeInfo(&a.merge)
+	if mi == nil {
+		mi = getMergeInfo(reflect.TypeOf(dst).Elem())
+		atomicStoreMergeInfo(&a.merge, mi)
+	}
+	mi.merge(toPointer(&dst), toPointer(&src))
+}
+
+type mergeInfo struct {
+	typ reflect.Type
+
+	initialized int32 // 0: only typ is valid, 1: everything is valid
+	lock        sync.Mutex
+
+	fields       []mergeFieldInfo
+	unrecognized field // Offset of XXX_unrecognized
+}
+
+type mergeFieldInfo struct {
+	field field // Offset of field, guaranteed to be valid
+
+	// isPointer reports whether the value in the field is a pointer.
+	// This is true for the following situations:
+	//	* Pointer to struct
+	//	* Pointer to basic type (proto2 only)
+	//	* Slice (first value in slice header is a pointer)
+	//	* String (first value in string header is a pointer)
+	isPointer bool
+
+	// basicWidth reports the width of the field assuming that it is directly
+	// embedded in the struct (as is the case for basic types in proto3).
+	// The possible values are:
+	// 	0: invalid
+	//	1: bool
+	//	4: int32, uint32, float32
+	//	8: int64, uint64, float64
+	basicWidth int
+
+	// Where dst and src are pointers to the types being merged.
+	merge func(dst, src pointer)
+}
+
+var (
+	mergeInfoMap  = map[reflect.Type]*mergeInfo{}
+	mergeInfoLock sync.Mutex
+)
+
+func getMergeInfo(t reflect.Type) *mergeInfo {
+	mergeInfoLock.Lock()
+	defer mergeInfoLock.Unlock()
+	mi := mergeInfoMap[t]
+	if mi == nil {
+		mi = &mergeInfo{typ: t}
+		mergeInfoMap[t] = mi
+	}
+	return mi
+}
+
+// merge merges src into dst assuming they are both of type *mi.typ.
+func (mi *mergeInfo) merge(dst, src pointer) {
+	if dst.isNil() {
+		panic("proto: nil destination")
+	}
+	if src.isNil() {
+		return // Nothing to do.
+	}
+
+	if atomic.LoadInt32(&mi.initialized) == 0 {
+		mi.computeMergeInfo()
+	}
+
+	for _, fi := range mi.fields {
+		sfp := src.offset(fi.field)
+
+		// As an optimization, we can avoid the merge function call cost
+		// if we know for sure that the source will have no effect
+		// by checking if it is the zero value.
+		if unsafeAllowed {
+			if fi.isPointer && sfp.getPointer().isNil() { // Could be slice or string
+				continue
+			}
+			if fi.basicWidth > 0 {
+				switch {
+				case fi.basicWidth == 1 && !*sfp.toBool():
+					continue
+				case fi.basicWidth == 4 && *sfp.toUint32() == 0:
+					continue
+				case fi.basicWidth == 8 && *sfp.toUint64() == 0:
+					continue
+				}
+			}
+		}
+
+		dfp := dst.offset(fi.field)
+		fi.merge(dfp, sfp)
+	}
+
+	// TODO: Make this faster?
+	out := dst.asPointerTo(mi.typ).Elem()
+	in := src.asPointerTo(mi.typ).Elem()
+	if emIn, err := extendable(in.Addr().Interface()); err == nil {
+		emOut, _ := extendable(out.Addr().Interface())
+		mIn, muIn := emIn.extensionsRead()
+		if mIn != nil {
+			mOut := emOut.extensionsWrite()
+			muIn.Lock()
+			mergeExtension(mOut, mIn)
+			muIn.Unlock()
+		}
+	}
+
+	if mi.unrecognized.IsValid() {
+		if b := *src.offset(mi.unrecognized).toBytes(); len(b) > 0 {
+			*dst.offset(mi.unrecognized).toBytes() = append([]byte(nil), b...)
+		}
+	}
+}
+
+func (mi *mergeInfo) computeMergeInfo() {
+	mi.lock.Lock()
+	defer mi.lock.Unlock()
+	if mi.initialized != 0 {
+		return
+	}
+	t := mi.typ
+	n := t.NumField()
+
+	props := GetProperties(t)
+	for i := 0; i < n; i++ {
+		f := t.Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+
+		mfi := mergeFieldInfo{field: toField(&f)}
+		tf := f.Type
+
+		// As an optimization, we can avoid the merge function call cost
+		// if we know for sure that the source will have no effect
+		// by checking if it is the zero value.
+		if unsafeAllowed {
+			switch tf.Kind() {
+			case reflect.Ptr, reflect.Slice, reflect.String:
+				// As a special case, we assume slices and strings are pointers
+				// since we know that the first field in the SliceSlice or
+				// StringHeader is a data pointer.
+				mfi.isPointer = true
+			case reflect.Bool:
+				mfi.basicWidth = 1
+			case reflect.Int32, reflect.Uint32, reflect.Float32:
+				mfi.basicWidth = 4
+			case reflect.Int64, reflect.Uint64, reflect.Float64:
+				mfi.basicWidth = 8
+			}
+		}
+
+		// Unwrap tf to get at its most basic type.
+		var isPointer, isSlice bool
+		if tf.Kind() == reflect.Slice && tf.Elem().Kind() != reflect.Uint8 {
+			isSlice = true
+			tf = tf.Elem()
+		}
+		if tf.Kind() == reflect.Ptr {
+			isPointer = true
+			tf = tf.Elem()
+		}
+		if isPointer && isSlice && tf.Kind() != reflect.Struct {
+			panic("both pointer and slice for basic type in " + tf.Name())
+		}
+
+		switch tf.Kind() {
+		case reflect.Int32:
+			switch {
+			case isSlice: // E.g., []int32
+				mfi.merge = func(dst, src pointer) {
+					// NOTE: toInt32Slice is not defined (see pointer_reflect.go).
+					/*
+						sfsp := src.toInt32Slice()
+						if *sfsp != nil {
+							dfsp := dst.toInt32Slice()
+							*dfsp = append(*dfsp, *sfsp...)
+							if *dfsp == nil {
+								*dfsp = []int64{}
+							}
+						}
+					*/
+					sfs := src.getInt32Slice()
+					if sfs != nil {
+						dfs := dst.getInt32Slice()
+						dfs = append(dfs, sfs...)
+						if dfs == nil {
+							dfs = []int32{}
+						}
+						dst.setInt32Slice(dfs)
+					}
+				}
+			case isPointer: // E.g., *int32
+				mfi.merge = func(dst, src pointer) {
+					// NOTE: toInt32Ptr is not defined (see pointer_reflect.go).
+					/*
+						sfpp := src.toInt32Ptr()
+						if *sfpp != nil {
+							dfpp := dst.toInt32Ptr()
+							if *dfpp == nil {
+								*dfpp = Int32(**sfpp)
+							} else {
+								**dfpp = **sfpp
+							}
+						}
+					*/
+					sfp := src.getInt32Ptr()
+					if sfp != nil {
+						dfp := dst.getInt32Ptr()
+						if dfp == nil {
+							dst.setInt32Ptr(*sfp)
+						} else {
+							*dfp = *sfp
+						}
+					}
+				}
+			default: // E.g., int32
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toInt32(); v != 0 {
+						*dst.toInt32() = v
+					}
+				}
+			}
+		case reflect.Int64:
+			switch {
+			case isSlice: // E.g., []int64
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toInt64Slice()
+					if *sfsp != nil {
+						dfsp := dst.toInt64Slice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []int64{}
+						}
+					}
+				}
+			case isPointer: // E.g., *int64
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toInt64Ptr()
+					if *sfpp != nil {
+						dfpp := dst.toInt64Ptr()
+						if *dfpp == nil {
+							*dfpp = Int64(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., int64
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toInt64(); v != 0 {
+						*dst.toInt64() = v
+					}
+				}
+			}
+		case reflect.Uint32:
+			switch {
+			case isSlice: // E.g., []uint32
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toUint32Slice()
+					if *sfsp != nil {
+						dfsp := dst.toUint32Slice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []uint32{}
+						}
+					}
+				}
+			case isPointer: // E.g., *uint32
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toUint32Ptr()
+					if *sfpp != nil {
+						dfpp := dst.toUint32Ptr()
+						if *dfpp == nil {
+							*dfpp = Uint32(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., uint32
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toUint32(); v != 0 {
+						*dst.toUint32() = v
+					}
+				}
+			}
+		case reflect.Uint64:
+			switch {
+			case isSlice: // E.g., []uint64
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toUint64Slice()
+					if *sfsp != nil {
+						dfsp := dst.toUint64Slice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []uint64{}
+						}
+					}
+				}
+			case isPointer: // E.g., *uint64
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toUint64Ptr()
+					if *sfpp != nil {
+						dfpp := dst.toUint64Ptr()
+						if *dfpp == nil {
+							*dfpp = Uint64(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., uint64
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toUint64(); v != 0 {
+						*dst.toUint64() = v
+					}
+				}
+			}
+		case reflect.Float32:
+			switch {
+			case isSlice: // E.g., []float32
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toFloat32Slice()
+					if *sfsp != nil {
+						dfsp := dst.toFloat32Slice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []float32{}
+						}
+					}
+				}
+			case isPointer: // E.g., *float32
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toFloat32Ptr()
+					if *sfpp != nil {
+						dfpp := dst.toFloat32Ptr()
+						if *dfpp == nil {
+							*dfpp = Float32(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., float32
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toFloat32(); v != 0 {
+						*dst.toFloat32() = v
+					}
+				}
+			}
+		case reflect.Float64:
+			switch {
+			case isSlice: // E.g., []float64
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toFloat64Slice()
+					if *sfsp != nil {
+						dfsp := dst.toFloat64Slice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []float64{}
+						}
+					}
+				}
+			case isPointer: // E.g., *float64
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toFloat64Ptr()
+					if *sfpp != nil {
+						dfpp := dst.toFloat64Ptr()
+						if *dfpp == nil {
+							*dfpp = Float64(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., float64
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toFloat64(); v != 0 {
+						*dst.toFloat64() = v
+					}
+				}
+			}
+		case reflect.Bool:
+			switch {
+			case isSlice: // E.g., []bool
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toBoolSlice()
+					if *sfsp != nil {
+						dfsp := dst.toBoolSlice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []bool{}
+						}
+					}
+				}
+			case isPointer: // E.g., *bool
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toBoolPtr()
+					if *sfpp != nil {
+						dfpp := dst.toBoolPtr()
+						if *dfpp == nil {
+							*dfpp = Bool(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., bool
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toBool(); v {
+						*dst.toBool() = v
+					}
+				}
+			}
+		case reflect.String:
+			switch {
+			case isSlice: // E.g., []string
+				mfi.merge = func(dst, src pointer) {
+					sfsp := src.toStringSlice()
+					if *sfsp != nil {
+						dfsp := dst.toStringSlice()
+						*dfsp = append(*dfsp, *sfsp...)
+						if *dfsp == nil {
+							*dfsp = []string{}
+						}
+					}
+				}
+			case isPointer: // E.g., *string
+				mfi.merge = func(dst, src pointer) {
+					sfpp := src.toStringPtr()
+					if *sfpp != nil {
+						dfpp := dst.toStringPtr()
+						if *dfpp == nil {
+							*dfpp = String(**sfpp)
+						} else {
+							**dfpp = **sfpp
+						}
+					}
+				}
+			default: // E.g., string
+				mfi.merge = func(dst, src pointer) {
+					if v := *src.toString(); v != "" {
+						*dst.toString() = v
+					}
+				}
+			}
+		case reflect.Slice:
+			isProto3 := props.Prop[i].proto3
+			switch {
+			case isPointer:
+				panic("bad pointer in byte slice case in " + tf.Name())
+			case tf.Elem().Kind() != reflect.Uint8:
+				panic("bad element kind in byte slice case in " + tf.Name())
+			case isSlice: // E.g., [][]byte
+				mfi.merge = func(dst, src pointer) {
+					sbsp := src.toBytesSlice()
+					if *sbsp != nil {
+						dbsp := dst.toBytesSlice()
+						for _, sb := range *sbsp {
+							if sb == nil {
+								*dbsp = append(*dbsp, nil)
+							} else {
+								*dbsp = append(*dbsp, append([]byte{}, sb...))
+							}
+						}
+						if *dbsp == nil {
+							*dbsp = [][]byte{}
+						}
+					}
+				}
+			default: // E.g., []byte
+				mfi.merge = func(dst, src pointer) {
+					sbp := src.toBytes()
+					if *sbp != nil {
+						dbp := dst.toBytes()
+						if !isProto3 || len(*sbp) > 0 {
+							*dbp = append([]byte{}, *sbp...)
+						}
+					}
+				}
+			}
+		case reflect.Struct:
+			switch {
+			case !isPointer:
+				mergeInfo := getMergeInfo(tf)
+				mfi.merge = func(dst, src pointer) {
+					mergeInfo.merge(dst, src)
+				}
+			case isSlice: // E.g., []*pb.T
+				mergeInfo := getMergeInfo(tf)
+				mfi.merge = func(dst, src pointer) {
+					sps := src.getPointerSlice()
+					if sps != nil {
+						dps := dst.getPointerSlice()
+						for _, sp := range sps {
+							var dp pointer
+							if !sp.isNil() {
+								dp = valToPointer(reflect.New(tf))
+								mergeInfo.merge(dp, sp)
+							}
+							dps = append(dps, dp)
+						}
+						if dps == nil {
+							dps = []pointer{}
+						}
+						dst.setPointerSlice(dps)
+					}
+				}
+			default: // E.g., *pb.T
+				mergeInfo := getMergeInfo(tf)
+				mfi.merge = func(dst, src pointer) {
+					sp := src.getPointer()
+					if !sp.isNil() {
+						dp := dst.getPointer()
+						if dp.isNil() {
+							dp = valToPointer(reflect.New(tf))
+							dst.setPointer(dp)
+						}
+						mergeInfo.merge(dp, sp)
+					}
+				}
+			}
+		case reflect.Map:
+			switch {
+			case isPointer || isSlice:
+				panic("bad pointer or slice in map case in " + tf.Name())
+			default: // E.g., map[K]V
+				mfi.merge = func(dst, src pointer) {
+					sm := src.asPointerTo(tf).Elem()
+					if sm.Len() == 0 {
+						return
+					}
+					dm := dst.asPointerTo(tf).Elem()
+					if dm.IsNil() {
+						dm.Set(reflect.MakeMap(tf))
+					}
+
+					switch tf.Elem().Kind() {
+					case reflect.Ptr: // Proto struct (e.g., *T)
+						for _, key := range sm.MapKeys() {
+							val := sm.MapIndex(key)
+							val = reflect.ValueOf(Clone(val.Interface().(Message)))
+							dm.SetMapIndex(key, val)
+						}
+					case reflect.Slice: // E.g. Bytes type (e.g., []byte)
+						for _, key := range sm.MapKeys() {
+							val := sm.MapIndex(key)
+							val = reflect.ValueOf(append([]byte{}, val.Bytes()...))
+							dm.SetMapIndex(key, val)
+						}
+					default: // Basic type (e.g., string)
+						for _, key := range sm.MapKeys() {
+							val := sm.MapIndex(key)
+							dm.SetMapIndex(key, val)
+						}
+					}
+				}
+			}
+		case reflect.Interface:
+			// Must be oneof field.
+			switch {
+			case isPointer || isSlice:
+				panic("bad pointer or slice in interface case in " + tf.Name())
+			default: // E.g., interface{}
+				// TODO: Make this faster?
+				mfi.merge = func(dst, src pointer) {
+					su := src.asPointerTo(tf).Elem()
+					if !su.IsNil() {
+						du := dst.asPointerTo(tf).Elem()
+						typ := su.Elem().Type()
+						if du.IsNil() || du.Elem().Type() != typ {
+							du.Set(reflect.New(typ.Elem())) // Initialize interface if empty
+						}
+						sv := su.Elem().Elem().Field(0)
+						if sv.Kind() == reflect.Ptr && sv.IsNil() {
+							return
+						}
+						dv := du.Elem().Elem().Field(0)
+						if dv.Kind() == reflect.Ptr && dv.IsNil() {
+							dv.Set(reflect.New(sv.Type().Elem())) // Initialize proto message if empty
+						}
+						switch sv.Type().Kind() {
+						case reflect.Ptr: // Proto struct (e.g., *T)
+							Merge(dv.Interface().(Message), sv.Interface().(Message))
+						case reflect.Slice: // E.g. Bytes type (e.g., []byte)
+							dv.Set(reflect.ValueOf(append([]byte{}, sv.Bytes()...)))
+						default: // Basic type (e.g., string)
+							dv.Set(sv)
+						}
+					}
+				}
+			}
+		default:
+			panic(fmt.Sprintf("merger not found for type:%s", tf))
+		}
+		mi.fields = append(mi.fields, mfi)
+	}
+
+	mi.unrecognized = invalidField
+	if f, ok := t.FieldByName("XXX_unrecognized"); ok {
+		if f.Type != reflect.TypeOf([]byte{}) {
+			panic("expected XXX_unrecognized to be of type []byte")
+		}
+		mi.unrecognized = toField(&f)
+	}
+
+	atomic.StoreInt32(&mi.initialized, 1)
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/table_unmarshal.go b/vendor/github.com/gogo/protobuf/proto/table_unmarshal.go
new file mode 100644
index 0000000..bb2622f
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/table_unmarshal.go
@@ -0,0 +1,2245 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"reflect"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"unicode/utf8"
+)
+
+// Unmarshal is the entry point from the generated .pb.go files.
+// This function is not intended to be used by non-generated code.
+// This function is not subject to any compatibility guarantee.
+// msg contains a pointer to a protocol buffer struct.
+// b is the data to be unmarshaled into the protocol buffer.
+// a is a pointer to a place to store cached unmarshal information.
+func (a *InternalMessageInfo) Unmarshal(msg Message, b []byte) error {
+	// Load the unmarshal information for this message type.
+	// The atomic load ensures memory consistency.
+	u := atomicLoadUnmarshalInfo(&a.unmarshal)
+	if u == nil {
+		// Slow path: find unmarshal info for msg, update a with it.
+		u = getUnmarshalInfo(reflect.TypeOf(msg).Elem())
+		atomicStoreUnmarshalInfo(&a.unmarshal, u)
+	}
+	// Then do the unmarshaling.
+	err := u.unmarshal(toPointer(&msg), b)
+	return err
+}
+
+type unmarshalInfo struct {
+	typ reflect.Type // type of the protobuf struct
+
+	// 0 = only typ field is initialized
+	// 1 = completely initialized
+	initialized     int32
+	lock            sync.Mutex                    // prevents double initialization
+	dense           []unmarshalFieldInfo          // fields indexed by tag #
+	sparse          map[uint64]unmarshalFieldInfo // fields indexed by tag #
+	reqFields       []string                      // names of required fields
+	reqMask         uint64                        // 1<<len(reqFields)-1
+	unrecognized    field                         // offset of []byte to put unrecognized data (or invalidField if we should throw it away)
+	extensions      field                         // offset of extensions field (of type proto.XXX_InternalExtensions), or invalidField if it does not exist
+	oldExtensions   field                         // offset of old-form extensions field (of type map[int]Extension)
+	extensionRanges []ExtensionRange              // if non-nil, implies extensions field is valid
+	isMessageSet    bool                          // if true, implies extensions field is valid
+
+	bytesExtensions field // offset of XXX_extensions with type []byte
+}
+
+// An unmarshaler takes a stream of bytes and a pointer to a field of a message.
+// It decodes the field, stores it at f, and returns the unused bytes.
+// w is the wire encoding.
+// b is the data after the tag and wire encoding have been read.
+type unmarshaler func(b []byte, f pointer, w int) ([]byte, error)
+
+type unmarshalFieldInfo struct {
+	// location of the field in the proto message structure.
+	field field
+
+	// function to unmarshal the data for the field.
+	unmarshal unmarshaler
+
+	// if a required field, contains a single set bit at this field's index in the required field list.
+	reqMask uint64
+
+	name string // name of the field, for error reporting
+}
+
+var (
+	unmarshalInfoMap  = map[reflect.Type]*unmarshalInfo{}
+	unmarshalInfoLock sync.Mutex
+)
+
+// getUnmarshalInfo returns the data structure which can be
+// subsequently used to unmarshal a message of the given type.
+// t is the type of the message (note: not pointer to message).
+func getUnmarshalInfo(t reflect.Type) *unmarshalInfo {
+	// It would be correct to return a new unmarshalInfo
+	// unconditionally. We would end up allocating one
+	// per occurrence of that type as a message or submessage.
+	// We use a cache here just to reduce memory usage.
+	unmarshalInfoLock.Lock()
+	defer unmarshalInfoLock.Unlock()
+	u := unmarshalInfoMap[t]
+	if u == nil {
+		u = &unmarshalInfo{typ: t}
+		// Note: we just set the type here. The rest of the fields
+		// will be initialized on first use.
+		unmarshalInfoMap[t] = u
+	}
+	return u
+}
+
+// unmarshal does the main work of unmarshaling a message.
+// u provides type information used to unmarshal the message.
+// m is a pointer to a protocol buffer message.
+// b is a byte stream to unmarshal into m.
+// This is top routine used when recursively unmarshaling submessages.
+func (u *unmarshalInfo) unmarshal(m pointer, b []byte) error {
+	if atomic.LoadInt32(&u.initialized) == 0 {
+		u.computeUnmarshalInfo()
+	}
+	if u.isMessageSet {
+		return unmarshalMessageSet(b, m.offset(u.extensions).toExtensions())
+	}
+	var reqMask uint64 // bitmask of required fields we've seen.
+	var errLater error
+	for len(b) > 0 {
+		// Read tag and wire type.
+		// Special case 1 and 2 byte varints.
+		var x uint64
+		if b[0] < 128 {
+			x = uint64(b[0])
+			b = b[1:]
+		} else if len(b) >= 2 && b[1] < 128 {
+			x = uint64(b[0]&0x7f) + uint64(b[1])<<7
+			b = b[2:]
+		} else {
+			var n int
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+		}
+		tag := x >> 3
+		wire := int(x) & 7
+
+		// Dispatch on the tag to one of the unmarshal* functions below.
+		var f unmarshalFieldInfo
+		if tag < uint64(len(u.dense)) {
+			f = u.dense[tag]
+		} else {
+			f = u.sparse[tag]
+		}
+		if fn := f.unmarshal; fn != nil {
+			var err error
+			b, err = fn(b, m.offset(f.field), wire)
+			if err == nil {
+				reqMask |= f.reqMask
+				continue
+			}
+			if r, ok := err.(*RequiredNotSetError); ok {
+				// Remember this error, but keep parsing. We need to produce
+				// a full parse even if a required field is missing.
+				if errLater == nil {
+					errLater = r
+				}
+				reqMask |= f.reqMask
+				continue
+			}
+			if err != errInternalBadWireType {
+				if err == errInvalidUTF8 {
+					if errLater == nil {
+						fullName := revProtoTypes[reflect.PtrTo(u.typ)] + "." + f.name
+						errLater = &invalidUTF8Error{fullName}
+					}
+					continue
+				}
+				return err
+			}
+			// Fragments with bad wire type are treated as unknown fields.
+		}
+
+		// Unknown tag.
+		if !u.unrecognized.IsValid() {
+			// Don't keep unrecognized data; just skip it.
+			var err error
+			b, err = skipField(b, wire)
+			if err != nil {
+				return err
+			}
+			continue
+		}
+		// Keep unrecognized data around.
+		// maybe in extensions, maybe in the unrecognized field.
+		z := m.offset(u.unrecognized).toBytes()
+		var emap map[int32]Extension
+		var e Extension
+		for _, r := range u.extensionRanges {
+			if uint64(r.Start) <= tag && tag <= uint64(r.End) {
+				if u.extensions.IsValid() {
+					mp := m.offset(u.extensions).toExtensions()
+					emap = mp.extensionsWrite()
+					e = emap[int32(tag)]
+					z = &e.enc
+					break
+				}
+				if u.oldExtensions.IsValid() {
+					p := m.offset(u.oldExtensions).toOldExtensions()
+					emap = *p
+					if emap == nil {
+						emap = map[int32]Extension{}
+						*p = emap
+					}
+					e = emap[int32(tag)]
+					z = &e.enc
+					break
+				}
+				if u.bytesExtensions.IsValid() {
+					z = m.offset(u.bytesExtensions).toBytes()
+					break
+				}
+				panic("no extensions field available")
+			}
+		}
+		// Use wire type to skip data.
+		var err error
+		b0 := b
+		b, err = skipField(b, wire)
+		if err != nil {
+			return err
+		}
+		*z = encodeVarint(*z, tag<<3|uint64(wire))
+		*z = append(*z, b0[:len(b0)-len(b)]...)
+
+		if emap != nil {
+			emap[int32(tag)] = e
+		}
+	}
+	if reqMask != u.reqMask && errLater == nil {
+		// A required field of this message is missing.
+		for _, n := range u.reqFields {
+			if reqMask&1 == 0 {
+				errLater = &RequiredNotSetError{n}
+			}
+			reqMask >>= 1
+		}
+	}
+	return errLater
+}
+
+// computeUnmarshalInfo fills in u with information for use
+// in unmarshaling protocol buffers of type u.typ.
+func (u *unmarshalInfo) computeUnmarshalInfo() {
+	u.lock.Lock()
+	defer u.lock.Unlock()
+	if u.initialized != 0 {
+		return
+	}
+	t := u.typ
+	n := t.NumField()
+
+	// Set up the "not found" value for the unrecognized byte buffer.
+	// This is the default for proto3.
+	u.unrecognized = invalidField
+	u.extensions = invalidField
+	u.oldExtensions = invalidField
+	u.bytesExtensions = invalidField
+
+	// List of the generated type and offset for each oneof field.
+	type oneofField struct {
+		ityp  reflect.Type // interface type of oneof field
+		field field        // offset in containing message
+	}
+	var oneofFields []oneofField
+
+	for i := 0; i < n; i++ {
+		f := t.Field(i)
+		if f.Name == "XXX_unrecognized" {
+			// The byte slice used to hold unrecognized input is special.
+			if f.Type != reflect.TypeOf(([]byte)(nil)) {
+				panic("bad type for XXX_unrecognized field: " + f.Type.Name())
+			}
+			u.unrecognized = toField(&f)
+			continue
+		}
+		if f.Name == "XXX_InternalExtensions" {
+			// Ditto here.
+			if f.Type != reflect.TypeOf(XXX_InternalExtensions{}) {
+				panic("bad type for XXX_InternalExtensions field: " + f.Type.Name())
+			}
+			u.extensions = toField(&f)
+			if f.Tag.Get("protobuf_messageset") == "1" {
+				u.isMessageSet = true
+			}
+			continue
+		}
+		if f.Name == "XXX_extensions" {
+			// An older form of the extensions field.
+			if f.Type == reflect.TypeOf((map[int32]Extension)(nil)) {
+				u.oldExtensions = toField(&f)
+				continue
+			} else if f.Type == reflect.TypeOf(([]byte)(nil)) {
+				u.bytesExtensions = toField(&f)
+				continue
+			}
+			panic("bad type for XXX_extensions field: " + f.Type.Name())
+		}
+		if f.Name == "XXX_NoUnkeyedLiteral" || f.Name == "XXX_sizecache" {
+			continue
+		}
+
+		oneof := f.Tag.Get("protobuf_oneof")
+		if oneof != "" {
+			oneofFields = append(oneofFields, oneofField{f.Type, toField(&f)})
+			// The rest of oneof processing happens below.
+			continue
+		}
+
+		tags := f.Tag.Get("protobuf")
+		tagArray := strings.Split(tags, ",")
+		if len(tagArray) < 2 {
+			panic("protobuf tag not enough fields in " + t.Name() + "." + f.Name + ": " + tags)
+		}
+		tag, err := strconv.Atoi(tagArray[1])
+		if err != nil {
+			panic("protobuf tag field not an integer: " + tagArray[1])
+		}
+
+		name := ""
+		for _, tag := range tagArray[3:] {
+			if strings.HasPrefix(tag, "name=") {
+				name = tag[5:]
+			}
+		}
+
+		// Extract unmarshaling function from the field (its type and tags).
+		unmarshal := fieldUnmarshaler(&f)
+
+		// Required field?
+		var reqMask uint64
+		if tagArray[2] == "req" {
+			bit := len(u.reqFields)
+			u.reqFields = append(u.reqFields, name)
+			reqMask = uint64(1) << uint(bit)
+			// TODO: if we have more than 64 required fields, we end up
+			// not verifying that all required fields are present.
+			// Fix this, perhaps using a count of required fields?
+		}
+
+		// Store the info in the correct slot in the message.
+		u.setTag(tag, toField(&f), unmarshal, reqMask, name)
+	}
+
+	// Find any types associated with oneof fields.
+	// TODO: XXX_OneofFuncs returns more info than we need.  Get rid of some of it?
+	fn := reflect.Zero(reflect.PtrTo(t)).MethodByName("XXX_OneofFuncs")
+	// gogo: len(oneofFields) > 0 is needed for embedded oneof messages, without a marshaler and unmarshaler
+	if fn.IsValid() && len(oneofFields) > 0 {
+		res := fn.Call(nil)[3] // last return value from XXX_OneofFuncs: []interface{}
+		for i := res.Len() - 1; i >= 0; i-- {
+			v := res.Index(i)                             // interface{}
+			tptr := reflect.ValueOf(v.Interface()).Type() // *Msg_X
+			typ := tptr.Elem()                            // Msg_X
+
+			f := typ.Field(0) // oneof implementers have one field
+			baseUnmarshal := fieldUnmarshaler(&f)
+			tags := strings.Split(f.Tag.Get("protobuf"), ",")
+			fieldNum, err := strconv.Atoi(tags[1])
+			if err != nil {
+				panic("protobuf tag field not an integer: " + tags[1])
+			}
+			var name string
+			for _, tag := range tags {
+				if strings.HasPrefix(tag, "name=") {
+					name = strings.TrimPrefix(tag, "name=")
+					break
+				}
+			}
+
+			// Find the oneof field that this struct implements.
+			// Might take O(n^2) to process all of the oneofs, but who cares.
+			for _, of := range oneofFields {
+				if tptr.Implements(of.ityp) {
+					// We have found the corresponding interface for this struct.
+					// That lets us know where this struct should be stored
+					// when we encounter it during unmarshaling.
+					unmarshal := makeUnmarshalOneof(typ, of.ityp, baseUnmarshal)
+					u.setTag(fieldNum, of.field, unmarshal, 0, name)
+				}
+			}
+		}
+	}
+
+	// Get extension ranges, if any.
+	fn = reflect.Zero(reflect.PtrTo(t)).MethodByName("ExtensionRangeArray")
+	if fn.IsValid() {
+		if !u.extensions.IsValid() && !u.oldExtensions.IsValid() && !u.bytesExtensions.IsValid() {
+			panic("a message with extensions, but no extensions field in " + t.Name())
+		}
+		u.extensionRanges = fn.Call(nil)[0].Interface().([]ExtensionRange)
+	}
+
+	// Explicitly disallow tag 0. This will ensure we flag an error
+	// when decoding a buffer of all zeros. Without this code, we
+	// would decode and skip an all-zero buffer of even length.
+	// [0 0] is [tag=0/wiretype=varint varint-encoded-0].
+	u.setTag(0, zeroField, func(b []byte, f pointer, w int) ([]byte, error) {
+		return nil, fmt.Errorf("proto: %s: illegal tag 0 (wire type %d)", t, w)
+	}, 0, "")
+
+	// Set mask for required field check.
+	u.reqMask = uint64(1)<<uint(len(u.reqFields)) - 1
+
+	atomic.StoreInt32(&u.initialized, 1)
+}
+
+// setTag stores the unmarshal information for the given tag.
+// tag = tag # for field
+// field/unmarshal = unmarshal info for that field.
+// reqMask = if required, bitmask for field position in required field list. 0 otherwise.
+// name = short name of the field.
+func (u *unmarshalInfo) setTag(tag int, field field, unmarshal unmarshaler, reqMask uint64, name string) {
+	i := unmarshalFieldInfo{field: field, unmarshal: unmarshal, reqMask: reqMask, name: name}
+	n := u.typ.NumField()
+	if tag >= 0 && (tag < 16 || tag < 2*n) { // TODO: what are the right numbers here?
+		for len(u.dense) <= tag {
+			u.dense = append(u.dense, unmarshalFieldInfo{})
+		}
+		u.dense[tag] = i
+		return
+	}
+	if u.sparse == nil {
+		u.sparse = map[uint64]unmarshalFieldInfo{}
+	}
+	u.sparse[uint64(tag)] = i
+}
+
+// fieldUnmarshaler returns an unmarshaler for the given field.
+func fieldUnmarshaler(f *reflect.StructField) unmarshaler {
+	if f.Type.Kind() == reflect.Map {
+		return makeUnmarshalMap(f)
+	}
+	return typeUnmarshaler(f.Type, f.Tag.Get("protobuf"))
+}
+
+// typeUnmarshaler returns an unmarshaler for the given field type / field tag pair.
+func typeUnmarshaler(t reflect.Type, tags string) unmarshaler {
+	tagArray := strings.Split(tags, ",")
+	encoding := tagArray[0]
+	name := "unknown"
+	ctype := false
+	isTime := false
+	isDuration := false
+	isWktPointer := false
+	proto3 := false
+	validateUTF8 := true
+	for _, tag := range tagArray[3:] {
+		if strings.HasPrefix(tag, "name=") {
+			name = tag[5:]
+		}
+		if tag == "proto3" {
+			proto3 = true
+		}
+		if strings.HasPrefix(tag, "customtype=") {
+			ctype = true
+		}
+		if tag == "stdtime" {
+			isTime = true
+		}
+		if tag == "stdduration" {
+			isDuration = true
+		}
+		if tag == "wktptr" {
+			isWktPointer = true
+		}
+	}
+	validateUTF8 = validateUTF8 && proto3
+
+	// Figure out packaging (pointer, slice, or both)
+	slice := false
+	pointer := false
+	if t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8 {
+		slice = true
+		t = t.Elem()
+	}
+	if t.Kind() == reflect.Ptr {
+		pointer = true
+		t = t.Elem()
+	}
+
+	if ctype {
+		if reflect.PtrTo(t).Implements(customType) {
+			if slice {
+				return makeUnmarshalCustomSlice(getUnmarshalInfo(t), name)
+			}
+			if pointer {
+				return makeUnmarshalCustomPtr(getUnmarshalInfo(t), name)
+			}
+			return makeUnmarshalCustom(getUnmarshalInfo(t), name)
+		} else {
+			panic(fmt.Sprintf("custom type: type: %v, does not implement the proto.custom interface", t))
+		}
+	}
+
+	if isTime {
+		if pointer {
+			if slice {
+				return makeUnmarshalTimePtrSlice(getUnmarshalInfo(t), name)
+			}
+			return makeUnmarshalTimePtr(getUnmarshalInfo(t), name)
+		}
+		if slice {
+			return makeUnmarshalTimeSlice(getUnmarshalInfo(t), name)
+		}
+		return makeUnmarshalTime(getUnmarshalInfo(t), name)
+	}
+
+	if isDuration {
+		if pointer {
+			if slice {
+				return makeUnmarshalDurationPtrSlice(getUnmarshalInfo(t), name)
+			}
+			return makeUnmarshalDurationPtr(getUnmarshalInfo(t), name)
+		}
+		if slice {
+			return makeUnmarshalDurationSlice(getUnmarshalInfo(t), name)
+		}
+		return makeUnmarshalDuration(getUnmarshalInfo(t), name)
+	}
+
+	if isWktPointer {
+		switch t.Kind() {
+		case reflect.Float64:
+			if pointer {
+				if slice {
+					return makeStdDoubleValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdDoubleValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdDoubleValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdDoubleValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Float32:
+			if pointer {
+				if slice {
+					return makeStdFloatValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdFloatValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdFloatValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdFloatValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Int64:
+			if pointer {
+				if slice {
+					return makeStdInt64ValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdInt64ValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdInt64ValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdInt64ValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Uint64:
+			if pointer {
+				if slice {
+					return makeStdUInt64ValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdUInt64ValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdUInt64ValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdUInt64ValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Int32:
+			if pointer {
+				if slice {
+					return makeStdInt32ValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdInt32ValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdInt32ValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdInt32ValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Uint32:
+			if pointer {
+				if slice {
+					return makeStdUInt32ValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdUInt32ValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdUInt32ValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdUInt32ValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.Bool:
+			if pointer {
+				if slice {
+					return makeStdBoolValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdBoolValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdBoolValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdBoolValueUnmarshaler(getUnmarshalInfo(t), name)
+		case reflect.String:
+			if pointer {
+				if slice {
+					return makeStdStringValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdStringValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdStringValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdStringValueUnmarshaler(getUnmarshalInfo(t), name)
+		case uint8SliceType:
+			if pointer {
+				if slice {
+					return makeStdBytesValuePtrSliceUnmarshaler(getUnmarshalInfo(t), name)
+				}
+				return makeStdBytesValuePtrUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			if slice {
+				return makeStdBytesValueSliceUnmarshaler(getUnmarshalInfo(t), name)
+			}
+			return makeStdBytesValueUnmarshaler(getUnmarshalInfo(t), name)
+		default:
+			panic(fmt.Sprintf("unknown wktpointer type %#v", t))
+		}
+	}
+
+	// We'll never have both pointer and slice for basic types.
+	if pointer && slice && t.Kind() != reflect.Struct {
+		panic("both pointer and slice for basic type in " + t.Name())
+	}
+
+	switch t.Kind() {
+	case reflect.Bool:
+		if pointer {
+			return unmarshalBoolPtr
+		}
+		if slice {
+			return unmarshalBoolSlice
+		}
+		return unmarshalBoolValue
+	case reflect.Int32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return unmarshalFixedS32Ptr
+			}
+			if slice {
+				return unmarshalFixedS32Slice
+			}
+			return unmarshalFixedS32Value
+		case "varint":
+			// this could be int32 or enum
+			if pointer {
+				return unmarshalInt32Ptr
+			}
+			if slice {
+				return unmarshalInt32Slice
+			}
+			return unmarshalInt32Value
+		case "zigzag32":
+			if pointer {
+				return unmarshalSint32Ptr
+			}
+			if slice {
+				return unmarshalSint32Slice
+			}
+			return unmarshalSint32Value
+		}
+	case reflect.Int64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return unmarshalFixedS64Ptr
+			}
+			if slice {
+				return unmarshalFixedS64Slice
+			}
+			return unmarshalFixedS64Value
+		case "varint":
+			if pointer {
+				return unmarshalInt64Ptr
+			}
+			if slice {
+				return unmarshalInt64Slice
+			}
+			return unmarshalInt64Value
+		case "zigzag64":
+			if pointer {
+				return unmarshalSint64Ptr
+			}
+			if slice {
+				return unmarshalSint64Slice
+			}
+			return unmarshalSint64Value
+		}
+	case reflect.Uint32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return unmarshalFixed32Ptr
+			}
+			if slice {
+				return unmarshalFixed32Slice
+			}
+			return unmarshalFixed32Value
+		case "varint":
+			if pointer {
+				return unmarshalUint32Ptr
+			}
+			if slice {
+				return unmarshalUint32Slice
+			}
+			return unmarshalUint32Value
+		}
+	case reflect.Uint64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return unmarshalFixed64Ptr
+			}
+			if slice {
+				return unmarshalFixed64Slice
+			}
+			return unmarshalFixed64Value
+		case "varint":
+			if pointer {
+				return unmarshalUint64Ptr
+			}
+			if slice {
+				return unmarshalUint64Slice
+			}
+			return unmarshalUint64Value
+		}
+	case reflect.Float32:
+		if pointer {
+			return unmarshalFloat32Ptr
+		}
+		if slice {
+			return unmarshalFloat32Slice
+		}
+		return unmarshalFloat32Value
+	case reflect.Float64:
+		if pointer {
+			return unmarshalFloat64Ptr
+		}
+		if slice {
+			return unmarshalFloat64Slice
+		}
+		return unmarshalFloat64Value
+	case reflect.Map:
+		panic("map type in typeUnmarshaler in " + t.Name())
+	case reflect.Slice:
+		if pointer {
+			panic("bad pointer in slice case in " + t.Name())
+		}
+		if slice {
+			return unmarshalBytesSlice
+		}
+		return unmarshalBytesValue
+	case reflect.String:
+		if validateUTF8 {
+			if pointer {
+				return unmarshalUTF8StringPtr
+			}
+			if slice {
+				return unmarshalUTF8StringSlice
+			}
+			return unmarshalUTF8StringValue
+		}
+		if pointer {
+			return unmarshalStringPtr
+		}
+		if slice {
+			return unmarshalStringSlice
+		}
+		return unmarshalStringValue
+	case reflect.Struct:
+		// message or group field
+		if !pointer {
+			switch encoding {
+			case "bytes":
+				if slice {
+					return makeUnmarshalMessageSlice(getUnmarshalInfo(t), name)
+				}
+				return makeUnmarshalMessage(getUnmarshalInfo(t), name)
+			}
+		}
+		switch encoding {
+		case "bytes":
+			if slice {
+				return makeUnmarshalMessageSlicePtr(getUnmarshalInfo(t), name)
+			}
+			return makeUnmarshalMessagePtr(getUnmarshalInfo(t), name)
+		case "group":
+			if slice {
+				return makeUnmarshalGroupSlicePtr(getUnmarshalInfo(t), name)
+			}
+			return makeUnmarshalGroupPtr(getUnmarshalInfo(t), name)
+		}
+	}
+	panic(fmt.Sprintf("unmarshaler not found type:%s encoding:%s", t, encoding))
+}
+
+// Below are all the unmarshalers for individual fields of various types.
+
+func unmarshalInt64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x)
+	*f.toInt64() = v
+	return b, nil
+}
+
+func unmarshalInt64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x)
+	*f.toInt64Ptr() = &v
+	return b, nil
+}
+
+func unmarshalInt64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := int64(x)
+			s := f.toInt64Slice()
+			*s = append(*s, v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x)
+	s := f.toInt64Slice()
+	*s = append(*s, v)
+	return b, nil
+}
+
+func unmarshalSint64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x>>1) ^ int64(x)<<63>>63
+	*f.toInt64() = v
+	return b, nil
+}
+
+func unmarshalSint64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x>>1) ^ int64(x)<<63>>63
+	*f.toInt64Ptr() = &v
+	return b, nil
+}
+
+func unmarshalSint64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := int64(x>>1) ^ int64(x)<<63>>63
+			s := f.toInt64Slice()
+			*s = append(*s, v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int64(x>>1) ^ int64(x)<<63>>63
+	s := f.toInt64Slice()
+	*s = append(*s, v)
+	return b, nil
+}
+
+func unmarshalUint64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint64(x)
+	*f.toUint64() = v
+	return b, nil
+}
+
+func unmarshalUint64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint64(x)
+	*f.toUint64Ptr() = &v
+	return b, nil
+}
+
+func unmarshalUint64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := uint64(x)
+			s := f.toUint64Slice()
+			*s = append(*s, v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint64(x)
+	s := f.toUint64Slice()
+	*s = append(*s, v)
+	return b, nil
+}
+
+func unmarshalInt32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x)
+	*f.toInt32() = v
+	return b, nil
+}
+
+func unmarshalInt32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x)
+	f.setInt32Ptr(v)
+	return b, nil
+}
+
+func unmarshalInt32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := int32(x)
+			f.appendInt32Slice(v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x)
+	f.appendInt32Slice(v)
+	return b, nil
+}
+
+func unmarshalSint32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x>>1) ^ int32(x)<<31>>31
+	*f.toInt32() = v
+	return b, nil
+}
+
+func unmarshalSint32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x>>1) ^ int32(x)<<31>>31
+	f.setInt32Ptr(v)
+	return b, nil
+}
+
+func unmarshalSint32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := int32(x>>1) ^ int32(x)<<31>>31
+			f.appendInt32Slice(v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := int32(x>>1) ^ int32(x)<<31>>31
+	f.appendInt32Slice(v)
+	return b, nil
+}
+
+func unmarshalUint32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint32(x)
+	*f.toUint32() = v
+	return b, nil
+}
+
+func unmarshalUint32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint32(x)
+	*f.toUint32Ptr() = &v
+	return b, nil
+}
+
+func unmarshalUint32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			b = b[n:]
+			v := uint32(x)
+			s := f.toUint32Slice()
+			*s = append(*s, v)
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	v := uint32(x)
+	s := f.toUint32Slice()
+	*s = append(*s, v)
+	return b, nil
+}
+
+func unmarshalFixed64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+	*f.toUint64() = v
+	return b[8:], nil
+}
+
+func unmarshalFixed64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+	*f.toUint64Ptr() = &v
+	return b[8:], nil
+}
+
+func unmarshalFixed64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 8 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+			s := f.toUint64Slice()
+			*s = append(*s, v)
+			b = b[8:]
+		}
+		return res, nil
+	}
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
+	s := f.toUint64Slice()
+	*s = append(*s, v)
+	return b[8:], nil
+}
+
+func unmarshalFixedS64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int64(b[0]) | int64(b[1])<<8 | int64(b[2])<<16 | int64(b[3])<<24 | int64(b[4])<<32 | int64(b[5])<<40 | int64(b[6])<<48 | int64(b[7])<<56
+	*f.toInt64() = v
+	return b[8:], nil
+}
+
+func unmarshalFixedS64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int64(b[0]) | int64(b[1])<<8 | int64(b[2])<<16 | int64(b[3])<<24 | int64(b[4])<<32 | int64(b[5])<<40 | int64(b[6])<<48 | int64(b[7])<<56
+	*f.toInt64Ptr() = &v
+	return b[8:], nil
+}
+
+func unmarshalFixedS64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 8 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := int64(b[0]) | int64(b[1])<<8 | int64(b[2])<<16 | int64(b[3])<<24 | int64(b[4])<<32 | int64(b[5])<<40 | int64(b[6])<<48 | int64(b[7])<<56
+			s := f.toInt64Slice()
+			*s = append(*s, v)
+			b = b[8:]
+		}
+		return res, nil
+	}
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int64(b[0]) | int64(b[1])<<8 | int64(b[2])<<16 | int64(b[3])<<24 | int64(b[4])<<32 | int64(b[5])<<40 | int64(b[6])<<48 | int64(b[7])<<56
+	s := f.toInt64Slice()
+	*s = append(*s, v)
+	return b[8:], nil
+}
+
+func unmarshalFixed32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+	*f.toUint32() = v
+	return b[4:], nil
+}
+
+func unmarshalFixed32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+	*f.toUint32Ptr() = &v
+	return b[4:], nil
+}
+
+func unmarshalFixed32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 4 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+			s := f.toUint32Slice()
+			*s = append(*s, v)
+			b = b[4:]
+		}
+		return res, nil
+	}
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
+	s := f.toUint32Slice()
+	*s = append(*s, v)
+	return b[4:], nil
+}
+
+func unmarshalFixedS32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int32(b[0]) | int32(b[1])<<8 | int32(b[2])<<16 | int32(b[3])<<24
+	*f.toInt32() = v
+	return b[4:], nil
+}
+
+func unmarshalFixedS32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int32(b[0]) | int32(b[1])<<8 | int32(b[2])<<16 | int32(b[3])<<24
+	f.setInt32Ptr(v)
+	return b[4:], nil
+}
+
+func unmarshalFixedS32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 4 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := int32(b[0]) | int32(b[1])<<8 | int32(b[2])<<16 | int32(b[3])<<24
+			f.appendInt32Slice(v)
+			b = b[4:]
+		}
+		return res, nil
+	}
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := int32(b[0]) | int32(b[1])<<8 | int32(b[2])<<16 | int32(b[3])<<24
+	f.appendInt32Slice(v)
+	return b[4:], nil
+}
+
+func unmarshalBoolValue(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	// Note: any length varint is allowed, even though any sane
+	// encoder will use one byte.
+	// See https://github.com/golang/protobuf/issues/76
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	// TODO: check if x>1? Tests seem to indicate no.
+	v := x != 0
+	*f.toBool() = v
+	return b[n:], nil
+}
+
+func unmarshalBoolPtr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := x != 0
+	*f.toBoolPtr() = &v
+	return b[n:], nil
+}
+
+func unmarshalBoolSlice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			x, n = decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := x != 0
+			s := f.toBoolSlice()
+			*s = append(*s, v)
+			b = b[n:]
+		}
+		return res, nil
+	}
+	if w != WireVarint {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := x != 0
+	s := f.toBoolSlice()
+	*s = append(*s, v)
+	return b[n:], nil
+}
+
+func unmarshalFloat64Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float64frombits(uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56)
+	*f.toFloat64() = v
+	return b[8:], nil
+}
+
+func unmarshalFloat64Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float64frombits(uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56)
+	*f.toFloat64Ptr() = &v
+	return b[8:], nil
+}
+
+func unmarshalFloat64Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 8 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := math.Float64frombits(uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56)
+			s := f.toFloat64Slice()
+			*s = append(*s, v)
+			b = b[8:]
+		}
+		return res, nil
+	}
+	if w != WireFixed64 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 8 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float64frombits(uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56)
+	s := f.toFloat64Slice()
+	*s = append(*s, v)
+	return b[8:], nil
+}
+
+func unmarshalFloat32Value(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float32frombits(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24)
+	*f.toFloat32() = v
+	return b[4:], nil
+}
+
+func unmarshalFloat32Ptr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float32frombits(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24)
+	*f.toFloat32Ptr() = &v
+	return b[4:], nil
+}
+
+func unmarshalFloat32Slice(b []byte, f pointer, w int) ([]byte, error) {
+	if w == WireBytes { // packed
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		res := b[x:]
+		b = b[:x]
+		for len(b) > 0 {
+			if len(b) < 4 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			v := math.Float32frombits(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24)
+			s := f.toFloat32Slice()
+			*s = append(*s, v)
+			b = b[4:]
+		}
+		return res, nil
+	}
+	if w != WireFixed32 {
+		return b, errInternalBadWireType
+	}
+	if len(b) < 4 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := math.Float32frombits(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24)
+	s := f.toFloat32Slice()
+	*s = append(*s, v)
+	return b[4:], nil
+}
+
+func unmarshalStringValue(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	*f.toString() = v
+	return b[x:], nil
+}
+
+func unmarshalStringPtr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	*f.toStringPtr() = &v
+	return b[x:], nil
+}
+
+func unmarshalStringSlice(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	s := f.toStringSlice()
+	*s = append(*s, v)
+	return b[x:], nil
+}
+
+func unmarshalUTF8StringValue(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	*f.toString() = v
+	if !utf8.ValidString(v) {
+		return b[x:], errInvalidUTF8
+	}
+	return b[x:], nil
+}
+
+func unmarshalUTF8StringPtr(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	*f.toStringPtr() = &v
+	if !utf8.ValidString(v) {
+		return b[x:], errInvalidUTF8
+	}
+	return b[x:], nil
+}
+
+func unmarshalUTF8StringSlice(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := string(b[:x])
+	s := f.toStringSlice()
+	*s = append(*s, v)
+	if !utf8.ValidString(v) {
+		return b[x:], errInvalidUTF8
+	}
+	return b[x:], nil
+}
+
+var emptyBuf [0]byte
+
+func unmarshalBytesValue(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	// The use of append here is a trick which avoids the zeroing
+	// that would be required if we used a make/copy pair.
+	// We append to emptyBuf instead of nil because we want
+	// a non-nil result even when the length is 0.
+	v := append(emptyBuf[:], b[:x]...)
+	*f.toBytes() = v
+	return b[x:], nil
+}
+
+func unmarshalBytesSlice(b []byte, f pointer, w int) ([]byte, error) {
+	if w != WireBytes {
+		return b, errInternalBadWireType
+	}
+	x, n := decodeVarint(b)
+	if n == 0 {
+		return nil, io.ErrUnexpectedEOF
+	}
+	b = b[n:]
+	if x > uint64(len(b)) {
+		return nil, io.ErrUnexpectedEOF
+	}
+	v := append(emptyBuf[:], b[:x]...)
+	s := f.toBytesSlice()
+	*s = append(*s, v)
+	return b[x:], nil
+}
+
+func makeUnmarshalMessagePtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return b, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		// First read the message field to see if something is there.
+		// The semantics of multiple submessages are weird.  Instead of
+		// the last one winning (as it is for all other fields), multiple
+		// submessages are merged.
+		v := f.getPointer()
+		if v.isNil() {
+			v = valToPointer(reflect.New(sub.typ))
+			f.setPointer(v)
+		}
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		return b[x:], err
+	}
+}
+
+func makeUnmarshalMessageSlicePtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return b, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		v := valToPointer(reflect.New(sub.typ))
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		f.appendPointer(v)
+		return b[x:], err
+	}
+}
+
+func makeUnmarshalGroupPtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireStartGroup {
+			return b, errInternalBadWireType
+		}
+		x, y := findEndGroup(b)
+		if x < 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		v := f.getPointer()
+		if v.isNil() {
+			v = valToPointer(reflect.New(sub.typ))
+			f.setPointer(v)
+		}
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		return b[y:], err
+	}
+}
+
+func makeUnmarshalGroupSlicePtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireStartGroup {
+			return b, errInternalBadWireType
+		}
+		x, y := findEndGroup(b)
+		if x < 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		v := valToPointer(reflect.New(sub.typ))
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		f.appendPointer(v)
+		return b[y:], err
+	}
+}
+
+func makeUnmarshalMap(f *reflect.StructField) unmarshaler {
+	t := f.Type
+	kt := t.Key()
+	vt := t.Elem()
+	tagArray := strings.Split(f.Tag.Get("protobuf"), ",")
+	valTags := strings.Split(f.Tag.Get("protobuf_val"), ",")
+	for _, t := range tagArray {
+		if strings.HasPrefix(t, "customtype=") {
+			valTags = append(valTags, t)
+		}
+		if t == "stdtime" {
+			valTags = append(valTags, t)
+		}
+		if t == "stdduration" {
+			valTags = append(valTags, t)
+		}
+		if t == "wktptr" {
+			valTags = append(valTags, t)
+		}
+	}
+	unmarshalKey := typeUnmarshaler(kt, f.Tag.Get("protobuf_key"))
+	unmarshalVal := typeUnmarshaler(vt, strings.Join(valTags, ","))
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		// The map entry is a submessage. Figure out how big it is.
+		if w != WireBytes {
+			return nil, fmt.Errorf("proto: bad wiretype for map field: got %d want %d", w, WireBytes)
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		r := b[x:] // unused data to return
+		b = b[:x]  // data for map entry
+
+		// Note: we could use #keys * #values ~= 200 functions
+		// to do map decoding without reflection. Probably not worth it.
+		// Maps will be somewhat slow. Oh well.
+
+		// Read key and value from data.
+		var nerr nonFatal
+		k := reflect.New(kt)
+		v := reflect.New(vt)
+		for len(b) > 0 {
+			x, n := decodeVarint(b)
+			if n == 0 {
+				return nil, io.ErrUnexpectedEOF
+			}
+			wire := int(x) & 7
+			b = b[n:]
+
+			var err error
+			switch x >> 3 {
+			case 1:
+				b, err = unmarshalKey(b, valToPointer(k), wire)
+			case 2:
+				b, err = unmarshalVal(b, valToPointer(v), wire)
+			default:
+				err = errInternalBadWireType // skip unknown tag
+			}
+
+			if nerr.Merge(err) {
+				continue
+			}
+			if err != errInternalBadWireType {
+				return nil, err
+			}
+
+			// Skip past unknown fields.
+			b, err = skipField(b, wire)
+			if err != nil {
+				return nil, err
+			}
+		}
+
+		// Get map, allocate if needed.
+		m := f.asPointerTo(t).Elem() // an addressable map[K]T
+		if m.IsNil() {
+			m.Set(reflect.MakeMap(t))
+		}
+
+		// Insert into map.
+		m.SetMapIndex(k.Elem(), v.Elem())
+
+		return r, nerr.E
+	}
+}
+
+// makeUnmarshalOneof makes an unmarshaler for oneof fields.
+// for:
+// message Msg {
+//   oneof F {
+//     int64 X = 1;
+//     float64 Y = 2;
+//   }
+// }
+// typ is the type of the concrete entry for a oneof case (e.g. Msg_X).
+// ityp is the interface type of the oneof field (e.g. isMsg_F).
+// unmarshal is the unmarshaler for the base type of the oneof case (e.g. int64).
+// Note that this function will be called once for each case in the oneof.
+func makeUnmarshalOneof(typ, ityp reflect.Type, unmarshal unmarshaler) unmarshaler {
+	sf := typ.Field(0)
+	field0 := toField(&sf)
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		// Allocate holder for value.
+		v := reflect.New(typ)
+
+		// Unmarshal data into holder.
+		// We unmarshal into the first field of the holder object.
+		var err error
+		var nerr nonFatal
+		b, err = unmarshal(b, valToPointer(v).offset(field0), w)
+		if !nerr.Merge(err) {
+			return nil, err
+		}
+
+		// Write pointer to holder into target field.
+		f.asPointerTo(ityp).Elem().Set(v)
+
+		return b, nerr.E
+	}
+}
+
+// Error used by decode internally.
+var errInternalBadWireType = errors.New("proto: internal error: bad wiretype")
+
+// skipField skips past a field of type wire and returns the remaining bytes.
+func skipField(b []byte, wire int) ([]byte, error) {
+	switch wire {
+	case WireVarint:
+		_, k := decodeVarint(b)
+		if k == 0 {
+			return b, io.ErrUnexpectedEOF
+		}
+		b = b[k:]
+	case WireFixed32:
+		if len(b) < 4 {
+			return b, io.ErrUnexpectedEOF
+		}
+		b = b[4:]
+	case WireFixed64:
+		if len(b) < 8 {
+			return b, io.ErrUnexpectedEOF
+		}
+		b = b[8:]
+	case WireBytes:
+		m, k := decodeVarint(b)
+		if k == 0 || uint64(len(b)-k) < m {
+			return b, io.ErrUnexpectedEOF
+		}
+		b = b[uint64(k)+m:]
+	case WireStartGroup:
+		_, i := findEndGroup(b)
+		if i == -1 {
+			return b, io.ErrUnexpectedEOF
+		}
+		b = b[i:]
+	default:
+		return b, fmt.Errorf("proto: can't skip unknown wire type %d", wire)
+	}
+	return b, nil
+}
+
+// findEndGroup finds the index of the next EndGroup tag.
+// Groups may be nested, so the "next" EndGroup tag is the first
+// unpaired EndGroup.
+// findEndGroup returns the indexes of the start and end of the EndGroup tag.
+// Returns (-1,-1) if it can't find one.
+func findEndGroup(b []byte) (int, int) {
+	depth := 1
+	i := 0
+	for {
+		x, n := decodeVarint(b[i:])
+		if n == 0 {
+			return -1, -1
+		}
+		j := i
+		i += n
+		switch x & 7 {
+		case WireVarint:
+			_, k := decodeVarint(b[i:])
+			if k == 0 {
+				return -1, -1
+			}
+			i += k
+		case WireFixed32:
+			if len(b)-4 < i {
+				return -1, -1
+			}
+			i += 4
+		case WireFixed64:
+			if len(b)-8 < i {
+				return -1, -1
+			}
+			i += 8
+		case WireBytes:
+			m, k := decodeVarint(b[i:])
+			if k == 0 {
+				return -1, -1
+			}
+			i += k
+			if uint64(len(b)-i) < m {
+				return -1, -1
+			}
+			i += int(m)
+		case WireStartGroup:
+			depth++
+		case WireEndGroup:
+			depth--
+			if depth == 0 {
+				return j, i
+			}
+		default:
+			return -1, -1
+		}
+	}
+}
+
+// encodeVarint appends a varint-encoded integer to b and returns the result.
+func encodeVarint(b []byte, x uint64) []byte {
+	for x >= 1<<7 {
+		b = append(b, byte(x&0x7f|0x80))
+		x >>= 7
+	}
+	return append(b, byte(x))
+}
+
+// decodeVarint reads a varint-encoded integer from b.
+// Returns the decoded integer and the number of bytes read.
+// If there is an error, it returns 0,0.
+func decodeVarint(b []byte) (uint64, int) {
+	var x, y uint64
+	if len(b) == 0 {
+		goto bad
+	}
+	x = uint64(b[0])
+	if x < 0x80 {
+		return x, 1
+	}
+	x -= 0x80
+
+	if len(b) <= 1 {
+		goto bad
+	}
+	y = uint64(b[1])
+	x += y << 7
+	if y < 0x80 {
+		return x, 2
+	}
+	x -= 0x80 << 7
+
+	if len(b) <= 2 {
+		goto bad
+	}
+	y = uint64(b[2])
+	x += y << 14
+	if y < 0x80 {
+		return x, 3
+	}
+	x -= 0x80 << 14
+
+	if len(b) <= 3 {
+		goto bad
+	}
+	y = uint64(b[3])
+	x += y << 21
+	if y < 0x80 {
+		return x, 4
+	}
+	x -= 0x80 << 21
+
+	if len(b) <= 4 {
+		goto bad
+	}
+	y = uint64(b[4])
+	x += y << 28
+	if y < 0x80 {
+		return x, 5
+	}
+	x -= 0x80 << 28
+
+	if len(b) <= 5 {
+		goto bad
+	}
+	y = uint64(b[5])
+	x += y << 35
+	if y < 0x80 {
+		return x, 6
+	}
+	x -= 0x80 << 35
+
+	if len(b) <= 6 {
+		goto bad
+	}
+	y = uint64(b[6])
+	x += y << 42
+	if y < 0x80 {
+		return x, 7
+	}
+	x -= 0x80 << 42
+
+	if len(b) <= 7 {
+		goto bad
+	}
+	y = uint64(b[7])
+	x += y << 49
+	if y < 0x80 {
+		return x, 8
+	}
+	x -= 0x80 << 49
+
+	if len(b) <= 8 {
+		goto bad
+	}
+	y = uint64(b[8])
+	x += y << 56
+	if y < 0x80 {
+		return x, 9
+	}
+	x -= 0x80 << 56
+
+	if len(b) <= 9 {
+		goto bad
+	}
+	y = uint64(b[9])
+	x += y << 63
+	if y < 2 {
+		return x, 10
+	}
+
+bad:
+	return 0, 0
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/table_unmarshal_gogo.go b/vendor/github.com/gogo/protobuf/proto/table_unmarshal_gogo.go
new file mode 100644
index 0000000..00d6c7a
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/table_unmarshal_gogo.go
@@ -0,0 +1,385 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"io"
+	"reflect"
+)
+
+func makeUnmarshalMessage(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		// First read the message field to see if something is there.
+		// The semantics of multiple submessages are weird.  Instead of
+		// the last one winning (as it is for all other fields), multiple
+		// submessages are merged.
+		v := f // gogo: changed from v := f.getPointer()
+		if v.isNil() {
+			v = valToPointer(reflect.New(sub.typ))
+			f.setPointer(v)
+		}
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		return b[x:], err
+	}
+}
+
+func makeUnmarshalMessageSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		v := valToPointer(reflect.New(sub.typ))
+		err := sub.unmarshal(v, b[:x])
+		if err != nil {
+			if r, ok := err.(*RequiredNotSetError); ok {
+				r.field = name + "." + r.field
+			} else {
+				return nil, err
+			}
+		}
+		f.appendRef(v, sub.typ) // gogo: changed from f.appendPointer(v)
+		return b[x:], err
+	}
+}
+
+func makeUnmarshalCustomPtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.New(sub.typ))
+		m := s.Interface().(custom)
+		if err := m.Unmarshal(b[:x]); err != nil {
+			return nil, err
+		}
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalCustomSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := reflect.New(sub.typ)
+		c := m.Interface().(custom)
+		if err := c.Unmarshal(b[:x]); err != nil {
+			return nil, err
+		}
+		v := valToPointer(m)
+		f.appendRef(v, sub.typ)
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalCustom(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+
+		m := f.asPointerTo(sub.typ).Interface().(custom)
+		if err := m.Unmarshal(b[:x]); err != nil {
+			return nil, err
+		}
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalTime(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &timestamp{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		t, err := timestampFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(t))
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalTimePtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &timestamp{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		t, err := timestampFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&t))
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalTimePtrSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &timestamp{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		t, err := timestampFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&t))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalTimeSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &timestamp{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		t, err := timestampFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(t))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalDurationPtr(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &duration{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		d, err := durationFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&d))
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalDuration(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &duration{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		d, err := durationFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(d))
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalDurationPtrSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &duration{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		d, err := durationFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&d))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeUnmarshalDurationSlice(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &duration{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		d, err := durationFromProto(m)
+		if err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(d))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/text.go b/vendor/github.com/gogo/protobuf/proto/text.go
new file mode 100644
index 0000000..0407ba8
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/text.go
@@ -0,0 +1,928 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+// Functions for writing the text protocol buffer format.
+
+import (
+	"bufio"
+	"bytes"
+	"encoding"
+	"errors"
+	"fmt"
+	"io"
+	"log"
+	"math"
+	"reflect"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+var (
+	newline         = []byte("\n")
+	spaces          = []byte("                                        ")
+	endBraceNewline = []byte("}\n")
+	backslashN      = []byte{'\\', 'n'}
+	backslashR      = []byte{'\\', 'r'}
+	backslashT      = []byte{'\\', 't'}
+	backslashDQ     = []byte{'\\', '"'}
+	backslashBS     = []byte{'\\', '\\'}
+	posInf          = []byte("inf")
+	negInf          = []byte("-inf")
+	nan             = []byte("nan")
+)
+
+type writer interface {
+	io.Writer
+	WriteByte(byte) error
+}
+
+// textWriter is an io.Writer that tracks its indentation level.
+type textWriter struct {
+	ind      int
+	complete bool // if the current position is a complete line
+	compact  bool // whether to write out as a one-liner
+	w        writer
+}
+
+func (w *textWriter) WriteString(s string) (n int, err error) {
+	if !strings.Contains(s, "\n") {
+		if !w.compact && w.complete {
+			w.writeIndent()
+		}
+		w.complete = false
+		return io.WriteString(w.w, s)
+	}
+	// WriteString is typically called without newlines, so this
+	// codepath and its copy are rare.  We copy to avoid
+	// duplicating all of Write's logic here.
+	return w.Write([]byte(s))
+}
+
+func (w *textWriter) Write(p []byte) (n int, err error) {
+	newlines := bytes.Count(p, newline)
+	if newlines == 0 {
+		if !w.compact && w.complete {
+			w.writeIndent()
+		}
+		n, err = w.w.Write(p)
+		w.complete = false
+		return n, err
+	}
+
+	frags := bytes.SplitN(p, newline, newlines+1)
+	if w.compact {
+		for i, frag := range frags {
+			if i > 0 {
+				if err := w.w.WriteByte(' '); err != nil {
+					return n, err
+				}
+				n++
+			}
+			nn, err := w.w.Write(frag)
+			n += nn
+			if err != nil {
+				return n, err
+			}
+		}
+		return n, nil
+	}
+
+	for i, frag := range frags {
+		if w.complete {
+			w.writeIndent()
+		}
+		nn, err := w.w.Write(frag)
+		n += nn
+		if err != nil {
+			return n, err
+		}
+		if i+1 < len(frags) {
+			if err := w.w.WriteByte('\n'); err != nil {
+				return n, err
+			}
+			n++
+		}
+	}
+	w.complete = len(frags[len(frags)-1]) == 0
+	return n, nil
+}
+
+func (w *textWriter) WriteByte(c byte) error {
+	if w.compact && c == '\n' {
+		c = ' '
+	}
+	if !w.compact && w.complete {
+		w.writeIndent()
+	}
+	err := w.w.WriteByte(c)
+	w.complete = c == '\n'
+	return err
+}
+
+func (w *textWriter) indent() { w.ind++ }
+
+func (w *textWriter) unindent() {
+	if w.ind == 0 {
+		log.Print("proto: textWriter unindented too far")
+		return
+	}
+	w.ind--
+}
+
+func writeName(w *textWriter, props *Properties) error {
+	if _, err := w.WriteString(props.OrigName); err != nil {
+		return err
+	}
+	if props.Wire != "group" {
+		return w.WriteByte(':')
+	}
+	return nil
+}
+
+func requiresQuotes(u string) bool {
+	// When type URL contains any characters except [0-9A-Za-z./\-]*, it must be quoted.
+	for _, ch := range u {
+		switch {
+		case ch == '.' || ch == '/' || ch == '_':
+			continue
+		case '0' <= ch && ch <= '9':
+			continue
+		case 'A' <= ch && ch <= 'Z':
+			continue
+		case 'a' <= ch && ch <= 'z':
+			continue
+		default:
+			return true
+		}
+	}
+	return false
+}
+
+// isAny reports whether sv is a google.protobuf.Any message
+func isAny(sv reflect.Value) bool {
+	type wkt interface {
+		XXX_WellKnownType() string
+	}
+	t, ok := sv.Addr().Interface().(wkt)
+	return ok && t.XXX_WellKnownType() == "Any"
+}
+
+// writeProto3Any writes an expanded google.protobuf.Any message.
+//
+// It returns (false, nil) if sv value can't be unmarshaled (e.g. because
+// required messages are not linked in).
+//
+// It returns (true, error) when sv was written in expanded format or an error
+// was encountered.
+func (tm *TextMarshaler) writeProto3Any(w *textWriter, sv reflect.Value) (bool, error) {
+	turl := sv.FieldByName("TypeUrl")
+	val := sv.FieldByName("Value")
+	if !turl.IsValid() || !val.IsValid() {
+		return true, errors.New("proto: invalid google.protobuf.Any message")
+	}
+
+	b, ok := val.Interface().([]byte)
+	if !ok {
+		return true, errors.New("proto: invalid google.protobuf.Any message")
+	}
+
+	parts := strings.Split(turl.String(), "/")
+	mt := MessageType(parts[len(parts)-1])
+	if mt == nil {
+		return false, nil
+	}
+	m := reflect.New(mt.Elem())
+	if err := Unmarshal(b, m.Interface().(Message)); err != nil {
+		return false, nil
+	}
+	w.Write([]byte("["))
+	u := turl.String()
+	if requiresQuotes(u) {
+		writeString(w, u)
+	} else {
+		w.Write([]byte(u))
+	}
+	if w.compact {
+		w.Write([]byte("]:<"))
+	} else {
+		w.Write([]byte("]: <\n"))
+		w.ind++
+	}
+	if err := tm.writeStruct(w, m.Elem()); err != nil {
+		return true, err
+	}
+	if w.compact {
+		w.Write([]byte("> "))
+	} else {
+		w.ind--
+		w.Write([]byte(">\n"))
+	}
+	return true, nil
+}
+
+func (tm *TextMarshaler) writeStruct(w *textWriter, sv reflect.Value) error {
+	if tm.ExpandAny && isAny(sv) {
+		if canExpand, err := tm.writeProto3Any(w, sv); canExpand {
+			return err
+		}
+	}
+	st := sv.Type()
+	sprops := GetProperties(st)
+	for i := 0; i < sv.NumField(); i++ {
+		fv := sv.Field(i)
+		props := sprops.Prop[i]
+		name := st.Field(i).Name
+
+		if name == "XXX_NoUnkeyedLiteral" {
+			continue
+		}
+
+		if strings.HasPrefix(name, "XXX_") {
+			// There are two XXX_ fields:
+			//   XXX_unrecognized []byte
+			//   XXX_extensions   map[int32]proto.Extension
+			// The first is handled here;
+			// the second is handled at the bottom of this function.
+			if name == "XXX_unrecognized" && !fv.IsNil() {
+				if err := writeUnknownStruct(w, fv.Interface().([]byte)); err != nil {
+					return err
+				}
+			}
+			continue
+		}
+		if fv.Kind() == reflect.Ptr && fv.IsNil() {
+			// Field not filled in. This could be an optional field or
+			// a required field that wasn't filled in. Either way, there
+			// isn't anything we can show for it.
+			continue
+		}
+		if fv.Kind() == reflect.Slice && fv.IsNil() {
+			// Repeated field that is empty, or a bytes field that is unused.
+			continue
+		}
+
+		if props.Repeated && fv.Kind() == reflect.Slice {
+			// Repeated field.
+			for j := 0; j < fv.Len(); j++ {
+				if err := writeName(w, props); err != nil {
+					return err
+				}
+				if !w.compact {
+					if err := w.WriteByte(' '); err != nil {
+						return err
+					}
+				}
+				v := fv.Index(j)
+				if v.Kind() == reflect.Ptr && v.IsNil() {
+					// A nil message in a repeated field is not valid,
+					// but we can handle that more gracefully than panicking.
+					if _, err := w.Write([]byte("<nil>\n")); err != nil {
+						return err
+					}
+					continue
+				}
+				if len(props.Enum) > 0 {
+					if err := tm.writeEnum(w, v, props); err != nil {
+						return err
+					}
+				} else if err := tm.writeAny(w, v, props); err != nil {
+					return err
+				}
+				if err := w.WriteByte('\n'); err != nil {
+					return err
+				}
+			}
+			continue
+		}
+		if fv.Kind() == reflect.Map {
+			// Map fields are rendered as a repeated struct with key/value fields.
+			keys := fv.MapKeys()
+			sort.Sort(mapKeys(keys))
+			for _, key := range keys {
+				val := fv.MapIndex(key)
+				if err := writeName(w, props); err != nil {
+					return err
+				}
+				if !w.compact {
+					if err := w.WriteByte(' '); err != nil {
+						return err
+					}
+				}
+				// open struct
+				if err := w.WriteByte('<'); err != nil {
+					return err
+				}
+				if !w.compact {
+					if err := w.WriteByte('\n'); err != nil {
+						return err
+					}
+				}
+				w.indent()
+				// key
+				if _, err := w.WriteString("key:"); err != nil {
+					return err
+				}
+				if !w.compact {
+					if err := w.WriteByte(' '); err != nil {
+						return err
+					}
+				}
+				if err := tm.writeAny(w, key, props.MapKeyProp); err != nil {
+					return err
+				}
+				if err := w.WriteByte('\n'); err != nil {
+					return err
+				}
+				// nil values aren't legal, but we can avoid panicking because of them.
+				if val.Kind() != reflect.Ptr || !val.IsNil() {
+					// value
+					if _, err := w.WriteString("value:"); err != nil {
+						return err
+					}
+					if !w.compact {
+						if err := w.WriteByte(' '); err != nil {
+							return err
+						}
+					}
+					if err := tm.writeAny(w, val, props.MapValProp); err != nil {
+						return err
+					}
+					if err := w.WriteByte('\n'); err != nil {
+						return err
+					}
+				}
+				// close struct
+				w.unindent()
+				if err := w.WriteByte('>'); err != nil {
+					return err
+				}
+				if err := w.WriteByte('\n'); err != nil {
+					return err
+				}
+			}
+			continue
+		}
+		if props.proto3 && fv.Kind() == reflect.Slice && fv.Len() == 0 {
+			// empty bytes field
+			continue
+		}
+		if props.proto3 && fv.Kind() != reflect.Ptr && fv.Kind() != reflect.Slice {
+			// proto3 non-repeated scalar field; skip if zero value
+			if isProto3Zero(fv) {
+				continue
+			}
+		}
+
+		if fv.Kind() == reflect.Interface {
+			// Check if it is a oneof.
+			if st.Field(i).Tag.Get("protobuf_oneof") != "" {
+				// fv is nil, or holds a pointer to generated struct.
+				// That generated struct has exactly one field,
+				// which has a protobuf struct tag.
+				if fv.IsNil() {
+					continue
+				}
+				inner := fv.Elem().Elem() // interface -> *T -> T
+				tag := inner.Type().Field(0).Tag.Get("protobuf")
+				props = new(Properties) // Overwrite the outer props var, but not its pointee.
+				props.Parse(tag)
+				// Write the value in the oneof, not the oneof itself.
+				fv = inner.Field(0)
+
+				// Special case to cope with malformed messages gracefully:
+				// If the value in the oneof is a nil pointer, don't panic
+				// in writeAny.
+				if fv.Kind() == reflect.Ptr && fv.IsNil() {
+					// Use errors.New so writeAny won't render quotes.
+					msg := errors.New("/* nil */")
+					fv = reflect.ValueOf(&msg).Elem()
+				}
+			}
+		}
+
+		if err := writeName(w, props); err != nil {
+			return err
+		}
+		if !w.compact {
+			if err := w.WriteByte(' '); err != nil {
+				return err
+			}
+		}
+
+		if len(props.Enum) > 0 {
+			if err := tm.writeEnum(w, fv, props); err != nil {
+				return err
+			}
+		} else if err := tm.writeAny(w, fv, props); err != nil {
+			return err
+		}
+
+		if err := w.WriteByte('\n'); err != nil {
+			return err
+		}
+	}
+
+	// Extensions (the XXX_extensions field).
+	pv := sv
+	if pv.CanAddr() {
+		pv = sv.Addr()
+	} else {
+		pv = reflect.New(sv.Type())
+		pv.Elem().Set(sv)
+	}
+	if _, err := extendable(pv.Interface()); err == nil {
+		if err := tm.writeExtensions(w, pv); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// writeAny writes an arbitrary field.
+func (tm *TextMarshaler) writeAny(w *textWriter, v reflect.Value, props *Properties) error {
+	v = reflect.Indirect(v)
+
+	if props != nil {
+		if len(props.CustomType) > 0 {
+			custom, ok := v.Interface().(Marshaler)
+			if ok {
+				data, err := custom.Marshal()
+				if err != nil {
+					return err
+				}
+				if err := writeString(w, string(data)); err != nil {
+					return err
+				}
+				return nil
+			}
+		} else if len(props.CastType) > 0 {
+			if _, ok := v.Interface().(interface {
+				String() string
+			}); ok {
+				switch v.Kind() {
+				case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64,
+					reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64:
+					_, err := fmt.Fprintf(w, "%d", v.Interface())
+					return err
+				}
+			}
+		} else if props.StdTime {
+			t, ok := v.Interface().(time.Time)
+			if !ok {
+				return fmt.Errorf("stdtime is not time.Time, but %T", v.Interface())
+			}
+			tproto, err := timestampProto(t)
+			if err != nil {
+				return err
+			}
+			propsCopy := *props // Make a copy so that this is goroutine-safe
+			propsCopy.StdTime = false
+			err = tm.writeAny(w, reflect.ValueOf(tproto), &propsCopy)
+			return err
+		} else if props.StdDuration {
+			d, ok := v.Interface().(time.Duration)
+			if !ok {
+				return fmt.Errorf("stdtime is not time.Duration, but %T", v.Interface())
+			}
+			dproto := durationProto(d)
+			propsCopy := *props // Make a copy so that this is goroutine-safe
+			propsCopy.StdDuration = false
+			err := tm.writeAny(w, reflect.ValueOf(dproto), &propsCopy)
+			return err
+		}
+	}
+
+	// Floats have special cases.
+	if v.Kind() == reflect.Float32 || v.Kind() == reflect.Float64 {
+		x := v.Float()
+		var b []byte
+		switch {
+		case math.IsInf(x, 1):
+			b = posInf
+		case math.IsInf(x, -1):
+			b = negInf
+		case math.IsNaN(x):
+			b = nan
+		}
+		if b != nil {
+			_, err := w.Write(b)
+			return err
+		}
+		// Other values are handled below.
+	}
+
+	// We don't attempt to serialise every possible value type; only those
+	// that can occur in protocol buffers.
+	switch v.Kind() {
+	case reflect.Slice:
+		// Should only be a []byte; repeated fields are handled in writeStruct.
+		if err := writeString(w, string(v.Bytes())); err != nil {
+			return err
+		}
+	case reflect.String:
+		if err := writeString(w, v.String()); err != nil {
+			return err
+		}
+	case reflect.Struct:
+		// Required/optional group/message.
+		var bra, ket byte = '<', '>'
+		if props != nil && props.Wire == "group" {
+			bra, ket = '{', '}'
+		}
+		if err := w.WriteByte(bra); err != nil {
+			return err
+		}
+		if !w.compact {
+			if err := w.WriteByte('\n'); err != nil {
+				return err
+			}
+		}
+		w.indent()
+		if v.CanAddr() {
+			// Calling v.Interface on a struct causes the reflect package to
+			// copy the entire struct. This is racy with the new Marshaler
+			// since we atomically update the XXX_sizecache.
+			//
+			// Thus, we retrieve a pointer to the struct if possible to avoid
+			// a race since v.Interface on the pointer doesn't copy the struct.
+			//
+			// If v is not addressable, then we are not worried about a race
+			// since it implies that the binary Marshaler cannot possibly be
+			// mutating this value.
+			v = v.Addr()
+		}
+		if etm, ok := v.Interface().(encoding.TextMarshaler); ok {
+			text, err := etm.MarshalText()
+			if err != nil {
+				return err
+			}
+			if _, err = w.Write(text); err != nil {
+				return err
+			}
+		} else {
+			if v.Kind() == reflect.Ptr {
+				v = v.Elem()
+			}
+			if err := tm.writeStruct(w, v); err != nil {
+				return err
+			}
+		}
+		w.unindent()
+		if err := w.WriteByte(ket); err != nil {
+			return err
+		}
+	default:
+		_, err := fmt.Fprint(w, v.Interface())
+		return err
+	}
+	return nil
+}
+
+// equivalent to C's isprint.
+func isprint(c byte) bool {
+	return c >= 0x20 && c < 0x7f
+}
+
+// writeString writes a string in the protocol buffer text format.
+// It is similar to strconv.Quote except we don't use Go escape sequences,
+// we treat the string as a byte sequence, and we use octal escapes.
+// These differences are to maintain interoperability with the other
+// languages' implementations of the text format.
+func writeString(w *textWriter, s string) error {
+	// use WriteByte here to get any needed indent
+	if err := w.WriteByte('"'); err != nil {
+		return err
+	}
+	// Loop over the bytes, not the runes.
+	for i := 0; i < len(s); i++ {
+		var err error
+		// Divergence from C++: we don't escape apostrophes.
+		// There's no need to escape them, and the C++ parser
+		// copes with a naked apostrophe.
+		switch c := s[i]; c {
+		case '\n':
+			_, err = w.w.Write(backslashN)
+		case '\r':
+			_, err = w.w.Write(backslashR)
+		case '\t':
+			_, err = w.w.Write(backslashT)
+		case '"':
+			_, err = w.w.Write(backslashDQ)
+		case '\\':
+			_, err = w.w.Write(backslashBS)
+		default:
+			if isprint(c) {
+				err = w.w.WriteByte(c)
+			} else {
+				_, err = fmt.Fprintf(w.w, "\\%03o", c)
+			}
+		}
+		if err != nil {
+			return err
+		}
+	}
+	return w.WriteByte('"')
+}
+
+func writeUnknownStruct(w *textWriter, data []byte) (err error) {
+	if !w.compact {
+		if _, err := fmt.Fprintf(w, "/* %d unknown bytes */\n", len(data)); err != nil {
+			return err
+		}
+	}
+	b := NewBuffer(data)
+	for b.index < len(b.buf) {
+		x, err := b.DecodeVarint()
+		if err != nil {
+			_, ferr := fmt.Fprintf(w, "/* %v */\n", err)
+			return ferr
+		}
+		wire, tag := x&7, x>>3
+		if wire == WireEndGroup {
+			w.unindent()
+			if _, werr := w.Write(endBraceNewline); werr != nil {
+				return werr
+			}
+			continue
+		}
+		if _, ferr := fmt.Fprint(w, tag); ferr != nil {
+			return ferr
+		}
+		if wire != WireStartGroup {
+			if err = w.WriteByte(':'); err != nil {
+				return err
+			}
+		}
+		if !w.compact || wire == WireStartGroup {
+			if err = w.WriteByte(' '); err != nil {
+				return err
+			}
+		}
+		switch wire {
+		case WireBytes:
+			buf, e := b.DecodeRawBytes(false)
+			if e == nil {
+				_, err = fmt.Fprintf(w, "%q", buf)
+			} else {
+				_, err = fmt.Fprintf(w, "/* %v */", e)
+			}
+		case WireFixed32:
+			x, err = b.DecodeFixed32()
+			err = writeUnknownInt(w, x, err)
+		case WireFixed64:
+			x, err = b.DecodeFixed64()
+			err = writeUnknownInt(w, x, err)
+		case WireStartGroup:
+			err = w.WriteByte('{')
+			w.indent()
+		case WireVarint:
+			x, err = b.DecodeVarint()
+			err = writeUnknownInt(w, x, err)
+		default:
+			_, err = fmt.Fprintf(w, "/* unknown wire type %d */", wire)
+		}
+		if err != nil {
+			return err
+		}
+		if err := w.WriteByte('\n'); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func writeUnknownInt(w *textWriter, x uint64, err error) error {
+	if err == nil {
+		_, err = fmt.Fprint(w, x)
+	} else {
+		_, err = fmt.Fprintf(w, "/* %v */", err)
+	}
+	return err
+}
+
+type int32Slice []int32
+
+func (s int32Slice) Len() int           { return len(s) }
+func (s int32Slice) Less(i, j int) bool { return s[i] < s[j] }
+func (s int32Slice) Swap(i, j int)      { s[i], s[j] = s[j], s[i] }
+
+// writeExtensions writes all the extensions in pv.
+// pv is assumed to be a pointer to a protocol message struct that is extendable.
+func (tm *TextMarshaler) writeExtensions(w *textWriter, pv reflect.Value) error {
+	emap := extensionMaps[pv.Type().Elem()]
+	e := pv.Interface().(Message)
+
+	var m map[int32]Extension
+	var mu sync.Locker
+	if em, ok := e.(extensionsBytes); ok {
+		eb := em.GetExtensions()
+		var err error
+		m, err = BytesToExtensionsMap(*eb)
+		if err != nil {
+			return err
+		}
+		mu = notLocker{}
+	} else if _, ok := e.(extendableProto); ok {
+		ep, _ := extendable(e)
+		m, mu = ep.extensionsRead()
+		if m == nil {
+			return nil
+		}
+	}
+
+	// Order the extensions by ID.
+	// This isn't strictly necessary, but it will give us
+	// canonical output, which will also make testing easier.
+
+	mu.Lock()
+	ids := make([]int32, 0, len(m))
+	for id := range m {
+		ids = append(ids, id)
+	}
+	sort.Sort(int32Slice(ids))
+	mu.Unlock()
+
+	for _, extNum := range ids {
+		ext := m[extNum]
+		var desc *ExtensionDesc
+		if emap != nil {
+			desc = emap[extNum]
+		}
+		if desc == nil {
+			// Unknown extension.
+			if err := writeUnknownStruct(w, ext.enc); err != nil {
+				return err
+			}
+			continue
+		}
+
+		pb, err := GetExtension(e, desc)
+		if err != nil {
+			return fmt.Errorf("failed getting extension: %v", err)
+		}
+
+		// Repeated extensions will appear as a slice.
+		if !desc.repeated() {
+			if err := tm.writeExtension(w, desc.Name, pb); err != nil {
+				return err
+			}
+		} else {
+			v := reflect.ValueOf(pb)
+			for i := 0; i < v.Len(); i++ {
+				if err := tm.writeExtension(w, desc.Name, v.Index(i).Interface()); err != nil {
+					return err
+				}
+			}
+		}
+	}
+	return nil
+}
+
+func (tm *TextMarshaler) writeExtension(w *textWriter, name string, pb interface{}) error {
+	if _, err := fmt.Fprintf(w, "[%s]:", name); err != nil {
+		return err
+	}
+	if !w.compact {
+		if err := w.WriteByte(' '); err != nil {
+			return err
+		}
+	}
+	if err := tm.writeAny(w, reflect.ValueOf(pb), nil); err != nil {
+		return err
+	}
+	if err := w.WriteByte('\n'); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (w *textWriter) writeIndent() {
+	if !w.complete {
+		return
+	}
+	remain := w.ind * 2
+	for remain > 0 {
+		n := remain
+		if n > len(spaces) {
+			n = len(spaces)
+		}
+		w.w.Write(spaces[:n])
+		remain -= n
+	}
+	w.complete = false
+}
+
+// TextMarshaler is a configurable text format marshaler.
+type TextMarshaler struct {
+	Compact   bool // use compact text format (one line).
+	ExpandAny bool // expand google.protobuf.Any messages of known types
+}
+
+// Marshal writes a given protocol buffer in text format.
+// The only errors returned are from w.
+func (tm *TextMarshaler) Marshal(w io.Writer, pb Message) error {
+	val := reflect.ValueOf(pb)
+	if pb == nil || val.IsNil() {
+		w.Write([]byte("<nil>"))
+		return nil
+	}
+	var bw *bufio.Writer
+	ww, ok := w.(writer)
+	if !ok {
+		bw = bufio.NewWriter(w)
+		ww = bw
+	}
+	aw := &textWriter{
+		w:        ww,
+		complete: true,
+		compact:  tm.Compact,
+	}
+
+	if etm, ok := pb.(encoding.TextMarshaler); ok {
+		text, err := etm.MarshalText()
+		if err != nil {
+			return err
+		}
+		if _, err = aw.Write(text); err != nil {
+			return err
+		}
+		if bw != nil {
+			return bw.Flush()
+		}
+		return nil
+	}
+	// Dereference the received pointer so we don't have outer < and >.
+	v := reflect.Indirect(val)
+	if err := tm.writeStruct(aw, v); err != nil {
+		return err
+	}
+	if bw != nil {
+		return bw.Flush()
+	}
+	return nil
+}
+
+// Text is the same as Marshal, but returns the string directly.
+func (tm *TextMarshaler) Text(pb Message) string {
+	var buf bytes.Buffer
+	tm.Marshal(&buf, pb)
+	return buf.String()
+}
+
+var (
+	defaultTextMarshaler = TextMarshaler{}
+	compactTextMarshaler = TextMarshaler{Compact: true}
+)
+
+// TODO: consider removing some of the Marshal functions below.
+
+// MarshalText writes a given protocol buffer in text format.
+// The only errors returned are from w.
+func MarshalText(w io.Writer, pb Message) error { return defaultTextMarshaler.Marshal(w, pb) }
+
+// MarshalTextString is the same as MarshalText, but returns the string directly.
+func MarshalTextString(pb Message) string { return defaultTextMarshaler.Text(pb) }
+
+// CompactText writes a given protocol buffer in compact text format (one line).
+func CompactText(w io.Writer, pb Message) error { return compactTextMarshaler.Marshal(w, pb) }
+
+// CompactTextString is the same as CompactText, but returns the string directly.
+func CompactTextString(pb Message) string { return compactTextMarshaler.Text(pb) }
diff --git a/vendor/github.com/gogo/protobuf/proto/text_gogo.go b/vendor/github.com/gogo/protobuf/proto/text_gogo.go
new file mode 100644
index 0000000..1d6c6aa
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/text_gogo.go
@@ -0,0 +1,57 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"fmt"
+	"reflect"
+)
+
+func (tm *TextMarshaler) writeEnum(w *textWriter, v reflect.Value, props *Properties) error {
+	m, ok := enumStringMaps[props.Enum]
+	if !ok {
+		if err := tm.writeAny(w, v, props); err != nil {
+			return err
+		}
+	}
+	key := int32(0)
+	if v.Kind() == reflect.Ptr {
+		key = int32(v.Elem().Int())
+	} else {
+		key = int32(v.Int())
+	}
+	s, ok := m[key]
+	if !ok {
+		if err := tm.writeAny(w, v, props); err != nil {
+			return err
+		}
+	}
+	_, err := fmt.Fprint(w, s)
+	return err
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/text_parser.go b/vendor/github.com/gogo/protobuf/proto/text_parser.go
new file mode 100644
index 0000000..1ce0be2
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/text_parser.go
@@ -0,0 +1,1018 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+// Functions for parsing the Text protocol buffer format.
+// TODO: message sets.
+
+import (
+	"encoding"
+	"errors"
+	"fmt"
+	"reflect"
+	"strconv"
+	"strings"
+	"time"
+	"unicode/utf8"
+)
+
+// Error string emitted when deserializing Any and fields are already set
+const anyRepeatedlyUnpacked = "Any message unpacked multiple times, or %q already set"
+
+type ParseError struct {
+	Message string
+	Line    int // 1-based line number
+	Offset  int // 0-based byte offset from start of input
+}
+
+func (p *ParseError) Error() string {
+	if p.Line == 1 {
+		// show offset only for first line
+		return fmt.Sprintf("line 1.%d: %v", p.Offset, p.Message)
+	}
+	return fmt.Sprintf("line %d: %v", p.Line, p.Message)
+}
+
+type token struct {
+	value    string
+	err      *ParseError
+	line     int    // line number
+	offset   int    // byte number from start of input, not start of line
+	unquoted string // the unquoted version of value, if it was a quoted string
+}
+
+func (t *token) String() string {
+	if t.err == nil {
+		return fmt.Sprintf("%q (line=%d, offset=%d)", t.value, t.line, t.offset)
+	}
+	return fmt.Sprintf("parse error: %v", t.err)
+}
+
+type textParser struct {
+	s            string // remaining input
+	done         bool   // whether the parsing is finished (success or error)
+	backed       bool   // whether back() was called
+	offset, line int
+	cur          token
+}
+
+func newTextParser(s string) *textParser {
+	p := new(textParser)
+	p.s = s
+	p.line = 1
+	p.cur.line = 1
+	return p
+}
+
+func (p *textParser) errorf(format string, a ...interface{}) *ParseError {
+	pe := &ParseError{fmt.Sprintf(format, a...), p.cur.line, p.cur.offset}
+	p.cur.err = pe
+	p.done = true
+	return pe
+}
+
+// Numbers and identifiers are matched by [-+._A-Za-z0-9]
+func isIdentOrNumberChar(c byte) bool {
+	switch {
+	case 'A' <= c && c <= 'Z', 'a' <= c && c <= 'z':
+		return true
+	case '0' <= c && c <= '9':
+		return true
+	}
+	switch c {
+	case '-', '+', '.', '_':
+		return true
+	}
+	return false
+}
+
+func isWhitespace(c byte) bool {
+	switch c {
+	case ' ', '\t', '\n', '\r':
+		return true
+	}
+	return false
+}
+
+func isQuote(c byte) bool {
+	switch c {
+	case '"', '\'':
+		return true
+	}
+	return false
+}
+
+func (p *textParser) skipWhitespace() {
+	i := 0
+	for i < len(p.s) && (isWhitespace(p.s[i]) || p.s[i] == '#') {
+		if p.s[i] == '#' {
+			// comment; skip to end of line or input
+			for i < len(p.s) && p.s[i] != '\n' {
+				i++
+			}
+			if i == len(p.s) {
+				break
+			}
+		}
+		if p.s[i] == '\n' {
+			p.line++
+		}
+		i++
+	}
+	p.offset += i
+	p.s = p.s[i:len(p.s)]
+	if len(p.s) == 0 {
+		p.done = true
+	}
+}
+
+func (p *textParser) advance() {
+	// Skip whitespace
+	p.skipWhitespace()
+	if p.done {
+		return
+	}
+
+	// Start of non-whitespace
+	p.cur.err = nil
+	p.cur.offset, p.cur.line = p.offset, p.line
+	p.cur.unquoted = ""
+	switch p.s[0] {
+	case '<', '>', '{', '}', ':', '[', ']', ';', ',', '/':
+		// Single symbol
+		p.cur.value, p.s = p.s[0:1], p.s[1:len(p.s)]
+	case '"', '\'':
+		// Quoted string
+		i := 1
+		for i < len(p.s) && p.s[i] != p.s[0] && p.s[i] != '\n' {
+			if p.s[i] == '\\' && i+1 < len(p.s) {
+				// skip escaped char
+				i++
+			}
+			i++
+		}
+		if i >= len(p.s) || p.s[i] != p.s[0] {
+			p.errorf("unmatched quote")
+			return
+		}
+		unq, err := unquoteC(p.s[1:i], rune(p.s[0]))
+		if err != nil {
+			p.errorf("invalid quoted string %s: %v", p.s[0:i+1], err)
+			return
+		}
+		p.cur.value, p.s = p.s[0:i+1], p.s[i+1:len(p.s)]
+		p.cur.unquoted = unq
+	default:
+		i := 0
+		for i < len(p.s) && isIdentOrNumberChar(p.s[i]) {
+			i++
+		}
+		if i == 0 {
+			p.errorf("unexpected byte %#x", p.s[0])
+			return
+		}
+		p.cur.value, p.s = p.s[0:i], p.s[i:len(p.s)]
+	}
+	p.offset += len(p.cur.value)
+}
+
+var (
+	errBadUTF8 = errors.New("proto: bad UTF-8")
+)
+
+func unquoteC(s string, quote rune) (string, error) {
+	// This is based on C++'s tokenizer.cc.
+	// Despite its name, this is *not* parsing C syntax.
+	// For instance, "\0" is an invalid quoted string.
+
+	// Avoid allocation in trivial cases.
+	simple := true
+	for _, r := range s {
+		if r == '\\' || r == quote {
+			simple = false
+			break
+		}
+	}
+	if simple {
+		return s, nil
+	}
+
+	buf := make([]byte, 0, 3*len(s)/2)
+	for len(s) > 0 {
+		r, n := utf8.DecodeRuneInString(s)
+		if r == utf8.RuneError && n == 1 {
+			return "", errBadUTF8
+		}
+		s = s[n:]
+		if r != '\\' {
+			if r < utf8.RuneSelf {
+				buf = append(buf, byte(r))
+			} else {
+				buf = append(buf, string(r)...)
+			}
+			continue
+		}
+
+		ch, tail, err := unescape(s)
+		if err != nil {
+			return "", err
+		}
+		buf = append(buf, ch...)
+		s = tail
+	}
+	return string(buf), nil
+}
+
+func unescape(s string) (ch string, tail string, err error) {
+	r, n := utf8.DecodeRuneInString(s)
+	if r == utf8.RuneError && n == 1 {
+		return "", "", errBadUTF8
+	}
+	s = s[n:]
+	switch r {
+	case 'a':
+		return "\a", s, nil
+	case 'b':
+		return "\b", s, nil
+	case 'f':
+		return "\f", s, nil
+	case 'n':
+		return "\n", s, nil
+	case 'r':
+		return "\r", s, nil
+	case 't':
+		return "\t", s, nil
+	case 'v':
+		return "\v", s, nil
+	case '?':
+		return "?", s, nil // trigraph workaround
+	case '\'', '"', '\\':
+		return string(r), s, nil
+	case '0', '1', '2', '3', '4', '5', '6', '7':
+		if len(s) < 2 {
+			return "", "", fmt.Errorf(`\%c requires 2 following digits`, r)
+		}
+		ss := string(r) + s[:2]
+		s = s[2:]
+		i, err := strconv.ParseUint(ss, 8, 8)
+		if err != nil {
+			return "", "", fmt.Errorf(`\%s contains non-octal digits`, ss)
+		}
+		return string([]byte{byte(i)}), s, nil
+	case 'x', 'X', 'u', 'U':
+		var n int
+		switch r {
+		case 'x', 'X':
+			n = 2
+		case 'u':
+			n = 4
+		case 'U':
+			n = 8
+		}
+		if len(s) < n {
+			return "", "", fmt.Errorf(`\%c requires %d following digits`, r, n)
+		}
+		ss := s[:n]
+		s = s[n:]
+		i, err := strconv.ParseUint(ss, 16, 64)
+		if err != nil {
+			return "", "", fmt.Errorf(`\%c%s contains non-hexadecimal digits`, r, ss)
+		}
+		if r == 'x' || r == 'X' {
+			return string([]byte{byte(i)}), s, nil
+		}
+		if i > utf8.MaxRune {
+			return "", "", fmt.Errorf(`\%c%s is not a valid Unicode code point`, r, ss)
+		}
+		return string(i), s, nil
+	}
+	return "", "", fmt.Errorf(`unknown escape \%c`, r)
+}
+
+// Back off the parser by one token. Can only be done between calls to next().
+// It makes the next advance() a no-op.
+func (p *textParser) back() { p.backed = true }
+
+// Advances the parser and returns the new current token.
+func (p *textParser) next() *token {
+	if p.backed || p.done {
+		p.backed = false
+		return &p.cur
+	}
+	p.advance()
+	if p.done {
+		p.cur.value = ""
+	} else if len(p.cur.value) > 0 && isQuote(p.cur.value[0]) {
+		// Look for multiple quoted strings separated by whitespace,
+		// and concatenate them.
+		cat := p.cur
+		for {
+			p.skipWhitespace()
+			if p.done || !isQuote(p.s[0]) {
+				break
+			}
+			p.advance()
+			if p.cur.err != nil {
+				return &p.cur
+			}
+			cat.value += " " + p.cur.value
+			cat.unquoted += p.cur.unquoted
+		}
+		p.done = false // parser may have seen EOF, but we want to return cat
+		p.cur = cat
+	}
+	return &p.cur
+}
+
+func (p *textParser) consumeToken(s string) error {
+	tok := p.next()
+	if tok.err != nil {
+		return tok.err
+	}
+	if tok.value != s {
+		p.back()
+		return p.errorf("expected %q, found %q", s, tok.value)
+	}
+	return nil
+}
+
+// Return a RequiredNotSetError indicating which required field was not set.
+func (p *textParser) missingRequiredFieldError(sv reflect.Value) *RequiredNotSetError {
+	st := sv.Type()
+	sprops := GetProperties(st)
+	for i := 0; i < st.NumField(); i++ {
+		if !isNil(sv.Field(i)) {
+			continue
+		}
+
+		props := sprops.Prop[i]
+		if props.Required {
+			return &RequiredNotSetError{fmt.Sprintf("%v.%v", st, props.OrigName)}
+		}
+	}
+	return &RequiredNotSetError{fmt.Sprintf("%v.<unknown field name>", st)} // should not happen
+}
+
+// Returns the index in the struct for the named field, as well as the parsed tag properties.
+func structFieldByName(sprops *StructProperties, name string) (int, *Properties, bool) {
+	i, ok := sprops.decoderOrigNames[name]
+	if ok {
+		return i, sprops.Prop[i], true
+	}
+	return -1, nil, false
+}
+
+// Consume a ':' from the input stream (if the next token is a colon),
+// returning an error if a colon is needed but not present.
+func (p *textParser) checkForColon(props *Properties, typ reflect.Type) *ParseError {
+	tok := p.next()
+	if tok.err != nil {
+		return tok.err
+	}
+	if tok.value != ":" {
+		// Colon is optional when the field is a group or message.
+		needColon := true
+		switch props.Wire {
+		case "group":
+			needColon = false
+		case "bytes":
+			// A "bytes" field is either a message, a string, or a repeated field;
+			// those three become *T, *string and []T respectively, so we can check for
+			// this field being a pointer to a non-string.
+			if typ.Kind() == reflect.Ptr {
+				// *T or *string
+				if typ.Elem().Kind() == reflect.String {
+					break
+				}
+			} else if typ.Kind() == reflect.Slice {
+				// []T or []*T
+				if typ.Elem().Kind() != reflect.Ptr {
+					break
+				}
+			} else if typ.Kind() == reflect.String {
+				// The proto3 exception is for a string field,
+				// which requires a colon.
+				break
+			}
+			needColon = false
+		}
+		if needColon {
+			return p.errorf("expected ':', found %q", tok.value)
+		}
+		p.back()
+	}
+	return nil
+}
+
+func (p *textParser) readStruct(sv reflect.Value, terminator string) error {
+	st := sv.Type()
+	sprops := GetProperties(st)
+	reqCount := sprops.reqCount
+	var reqFieldErr error
+	fieldSet := make(map[string]bool)
+	// A struct is a sequence of "name: value", terminated by one of
+	// '>' or '}', or the end of the input.  A name may also be
+	// "[extension]" or "[type/url]".
+	//
+	// The whole struct can also be an expanded Any message, like:
+	// [type/url] < ... struct contents ... >
+	for {
+		tok := p.next()
+		if tok.err != nil {
+			return tok.err
+		}
+		if tok.value == terminator {
+			break
+		}
+		if tok.value == "[" {
+			// Looks like an extension or an Any.
+			//
+			// TODO: Check whether we need to handle
+			// namespace rooted names (e.g. ".something.Foo").
+			extName, err := p.consumeExtName()
+			if err != nil {
+				return err
+			}
+
+			if s := strings.LastIndex(extName, "/"); s >= 0 {
+				// If it contains a slash, it's an Any type URL.
+				messageName := extName[s+1:]
+				mt := MessageType(messageName)
+				if mt == nil {
+					return p.errorf("unrecognized message %q in google.protobuf.Any", messageName)
+				}
+				tok = p.next()
+				if tok.err != nil {
+					return tok.err
+				}
+				// consume an optional colon
+				if tok.value == ":" {
+					tok = p.next()
+					if tok.err != nil {
+						return tok.err
+					}
+				}
+				var terminator string
+				switch tok.value {
+				case "<":
+					terminator = ">"
+				case "{":
+					terminator = "}"
+				default:
+					return p.errorf("expected '{' or '<', found %q", tok.value)
+				}
+				v := reflect.New(mt.Elem())
+				if pe := p.readStruct(v.Elem(), terminator); pe != nil {
+					return pe
+				}
+				b, err := Marshal(v.Interface().(Message))
+				if err != nil {
+					return p.errorf("failed to marshal message of type %q: %v", messageName, err)
+				}
+				if fieldSet["type_url"] {
+					return p.errorf(anyRepeatedlyUnpacked, "type_url")
+				}
+				if fieldSet["value"] {
+					return p.errorf(anyRepeatedlyUnpacked, "value")
+				}
+				sv.FieldByName("TypeUrl").SetString(extName)
+				sv.FieldByName("Value").SetBytes(b)
+				fieldSet["type_url"] = true
+				fieldSet["value"] = true
+				continue
+			}
+
+			var desc *ExtensionDesc
+			// This could be faster, but it's functional.
+			// TODO: Do something smarter than a linear scan.
+			for _, d := range RegisteredExtensions(reflect.New(st).Interface().(Message)) {
+				if d.Name == extName {
+					desc = d
+					break
+				}
+			}
+			if desc == nil {
+				return p.errorf("unrecognized extension %q", extName)
+			}
+
+			props := &Properties{}
+			props.Parse(desc.Tag)
+
+			typ := reflect.TypeOf(desc.ExtensionType)
+			if err := p.checkForColon(props, typ); err != nil {
+				return err
+			}
+
+			rep := desc.repeated()
+
+			// Read the extension structure, and set it in
+			// the value we're constructing.
+			var ext reflect.Value
+			if !rep {
+				ext = reflect.New(typ).Elem()
+			} else {
+				ext = reflect.New(typ.Elem()).Elem()
+			}
+			if err := p.readAny(ext, props); err != nil {
+				if _, ok := err.(*RequiredNotSetError); !ok {
+					return err
+				}
+				reqFieldErr = err
+			}
+			ep := sv.Addr().Interface().(Message)
+			if !rep {
+				SetExtension(ep, desc, ext.Interface())
+			} else {
+				old, err := GetExtension(ep, desc)
+				var sl reflect.Value
+				if err == nil {
+					sl = reflect.ValueOf(old) // existing slice
+				} else {
+					sl = reflect.MakeSlice(typ, 0, 1)
+				}
+				sl = reflect.Append(sl, ext)
+				SetExtension(ep, desc, sl.Interface())
+			}
+			if err := p.consumeOptionalSeparator(); err != nil {
+				return err
+			}
+			continue
+		}
+
+		// This is a normal, non-extension field.
+		name := tok.value
+		var dst reflect.Value
+		fi, props, ok := structFieldByName(sprops, name)
+		if ok {
+			dst = sv.Field(fi)
+		} else if oop, ok := sprops.OneofTypes[name]; ok {
+			// It is a oneof.
+			props = oop.Prop
+			nv := reflect.New(oop.Type.Elem())
+			dst = nv.Elem().Field(0)
+			field := sv.Field(oop.Field)
+			if !field.IsNil() {
+				return p.errorf("field '%s' would overwrite already parsed oneof '%s'", name, sv.Type().Field(oop.Field).Name)
+			}
+			field.Set(nv)
+		}
+		if !dst.IsValid() {
+			return p.errorf("unknown field name %q in %v", name, st)
+		}
+
+		if dst.Kind() == reflect.Map {
+			// Consume any colon.
+			if err := p.checkForColon(props, dst.Type()); err != nil {
+				return err
+			}
+
+			// Construct the map if it doesn't already exist.
+			if dst.IsNil() {
+				dst.Set(reflect.MakeMap(dst.Type()))
+			}
+			key := reflect.New(dst.Type().Key()).Elem()
+			val := reflect.New(dst.Type().Elem()).Elem()
+
+			// The map entry should be this sequence of tokens:
+			//	< key : KEY value : VALUE >
+			// However, implementations may omit key or value, and technically
+			// we should support them in any order.  See b/28924776 for a time
+			// this went wrong.
+
+			tok := p.next()
+			var terminator string
+			switch tok.value {
+			case "<":
+				terminator = ">"
+			case "{":
+				terminator = "}"
+			default:
+				return p.errorf("expected '{' or '<', found %q", tok.value)
+			}
+			for {
+				tok := p.next()
+				if tok.err != nil {
+					return tok.err
+				}
+				if tok.value == terminator {
+					break
+				}
+				switch tok.value {
+				case "key":
+					if err := p.consumeToken(":"); err != nil {
+						return err
+					}
+					if err := p.readAny(key, props.MapKeyProp); err != nil {
+						return err
+					}
+					if err := p.consumeOptionalSeparator(); err != nil {
+						return err
+					}
+				case "value":
+					if err := p.checkForColon(props.MapValProp, dst.Type().Elem()); err != nil {
+						return err
+					}
+					if err := p.readAny(val, props.MapValProp); err != nil {
+						return err
+					}
+					if err := p.consumeOptionalSeparator(); err != nil {
+						return err
+					}
+				default:
+					p.back()
+					return p.errorf(`expected "key", "value", or %q, found %q`, terminator, tok.value)
+				}
+			}
+
+			dst.SetMapIndex(key, val)
+			continue
+		}
+
+		// Check that it's not already set if it's not a repeated field.
+		if !props.Repeated && fieldSet[name] {
+			return p.errorf("non-repeated field %q was repeated", name)
+		}
+
+		if err := p.checkForColon(props, dst.Type()); err != nil {
+			return err
+		}
+
+		// Parse into the field.
+		fieldSet[name] = true
+		if err := p.readAny(dst, props); err != nil {
+			if _, ok := err.(*RequiredNotSetError); !ok {
+				return err
+			}
+			reqFieldErr = err
+		}
+		if props.Required {
+			reqCount--
+		}
+
+		if err := p.consumeOptionalSeparator(); err != nil {
+			return err
+		}
+
+	}
+
+	if reqCount > 0 {
+		return p.missingRequiredFieldError(sv)
+	}
+	return reqFieldErr
+}
+
+// consumeExtName consumes extension name or expanded Any type URL and the
+// following ']'. It returns the name or URL consumed.
+func (p *textParser) consumeExtName() (string, error) {
+	tok := p.next()
+	if tok.err != nil {
+		return "", tok.err
+	}
+
+	// If extension name or type url is quoted, it's a single token.
+	if len(tok.value) > 2 && isQuote(tok.value[0]) && tok.value[len(tok.value)-1] == tok.value[0] {
+		name, err := unquoteC(tok.value[1:len(tok.value)-1], rune(tok.value[0]))
+		if err != nil {
+			return "", err
+		}
+		return name, p.consumeToken("]")
+	}
+
+	// Consume everything up to "]"
+	var parts []string
+	for tok.value != "]" {
+		parts = append(parts, tok.value)
+		tok = p.next()
+		if tok.err != nil {
+			return "", p.errorf("unrecognized type_url or extension name: %s", tok.err)
+		}
+		if p.done && tok.value != "]" {
+			return "", p.errorf("unclosed type_url or extension name")
+		}
+	}
+	return strings.Join(parts, ""), nil
+}
+
+// consumeOptionalSeparator consumes an optional semicolon or comma.
+// It is used in readStruct to provide backward compatibility.
+func (p *textParser) consumeOptionalSeparator() error {
+	tok := p.next()
+	if tok.err != nil {
+		return tok.err
+	}
+	if tok.value != ";" && tok.value != "," {
+		p.back()
+	}
+	return nil
+}
+
+func (p *textParser) readAny(v reflect.Value, props *Properties) error {
+	tok := p.next()
+	if tok.err != nil {
+		return tok.err
+	}
+	if tok.value == "" {
+		return p.errorf("unexpected EOF")
+	}
+	if len(props.CustomType) > 0 {
+		if props.Repeated {
+			t := reflect.TypeOf(v.Interface())
+			if t.Kind() == reflect.Slice {
+				tc := reflect.TypeOf(new(Marshaler))
+				ok := t.Elem().Implements(tc.Elem())
+				if ok {
+					fv := v
+					flen := fv.Len()
+					if flen == fv.Cap() {
+						nav := reflect.MakeSlice(v.Type(), flen, 2*flen+1)
+						reflect.Copy(nav, fv)
+						fv.Set(nav)
+					}
+					fv.SetLen(flen + 1)
+
+					// Read one.
+					p.back()
+					return p.readAny(fv.Index(flen), props)
+				}
+			}
+		}
+		if reflect.TypeOf(v.Interface()).Kind() == reflect.Ptr {
+			custom := reflect.New(props.ctype.Elem()).Interface().(Unmarshaler)
+			err := custom.Unmarshal([]byte(tok.unquoted))
+			if err != nil {
+				return p.errorf("%v %v: %v", err, v.Type(), tok.value)
+			}
+			v.Set(reflect.ValueOf(custom))
+		} else {
+			custom := reflect.New(reflect.TypeOf(v.Interface())).Interface().(Unmarshaler)
+			err := custom.Unmarshal([]byte(tok.unquoted))
+			if err != nil {
+				return p.errorf("%v %v: %v", err, v.Type(), tok.value)
+			}
+			v.Set(reflect.Indirect(reflect.ValueOf(custom)))
+		}
+		return nil
+	}
+	if props.StdTime {
+		fv := v
+		p.back()
+		props.StdTime = false
+		tproto := &timestamp{}
+		err := p.readAny(reflect.ValueOf(tproto).Elem(), props)
+		props.StdTime = true
+		if err != nil {
+			return err
+		}
+		tim, err := timestampFromProto(tproto)
+		if err != nil {
+			return err
+		}
+		if props.Repeated {
+			t := reflect.TypeOf(v.Interface())
+			if t.Kind() == reflect.Slice {
+				if t.Elem().Kind() == reflect.Ptr {
+					ts := fv.Interface().([]*time.Time)
+					ts = append(ts, &tim)
+					fv.Set(reflect.ValueOf(ts))
+					return nil
+				} else {
+					ts := fv.Interface().([]time.Time)
+					ts = append(ts, tim)
+					fv.Set(reflect.ValueOf(ts))
+					return nil
+				}
+			}
+		}
+		if reflect.TypeOf(v.Interface()).Kind() == reflect.Ptr {
+			v.Set(reflect.ValueOf(&tim))
+		} else {
+			v.Set(reflect.Indirect(reflect.ValueOf(&tim)))
+		}
+		return nil
+	}
+	if props.StdDuration {
+		fv := v
+		p.back()
+		props.StdDuration = false
+		dproto := &duration{}
+		err := p.readAny(reflect.ValueOf(dproto).Elem(), props)
+		props.StdDuration = true
+		if err != nil {
+			return err
+		}
+		dur, err := durationFromProto(dproto)
+		if err != nil {
+			return err
+		}
+		if props.Repeated {
+			t := reflect.TypeOf(v.Interface())
+			if t.Kind() == reflect.Slice {
+				if t.Elem().Kind() == reflect.Ptr {
+					ds := fv.Interface().([]*time.Duration)
+					ds = append(ds, &dur)
+					fv.Set(reflect.ValueOf(ds))
+					return nil
+				} else {
+					ds := fv.Interface().([]time.Duration)
+					ds = append(ds, dur)
+					fv.Set(reflect.ValueOf(ds))
+					return nil
+				}
+			}
+		}
+		if reflect.TypeOf(v.Interface()).Kind() == reflect.Ptr {
+			v.Set(reflect.ValueOf(&dur))
+		} else {
+			v.Set(reflect.Indirect(reflect.ValueOf(&dur)))
+		}
+		return nil
+	}
+	switch fv := v; fv.Kind() {
+	case reflect.Slice:
+		at := v.Type()
+		if at.Elem().Kind() == reflect.Uint8 {
+			// Special case for []byte
+			if tok.value[0] != '"' && tok.value[0] != '\'' {
+				// Deliberately written out here, as the error after
+				// this switch statement would write "invalid []byte: ...",
+				// which is not as user-friendly.
+				return p.errorf("invalid string: %v", tok.value)
+			}
+			bytes := []byte(tok.unquoted)
+			fv.Set(reflect.ValueOf(bytes))
+			return nil
+		}
+		// Repeated field.
+		if tok.value == "[" {
+			// Repeated field with list notation, like [1,2,3].
+			for {
+				fv.Set(reflect.Append(fv, reflect.New(at.Elem()).Elem()))
+				err := p.readAny(fv.Index(fv.Len()-1), props)
+				if err != nil {
+					return err
+				}
+				ntok := p.next()
+				if ntok.err != nil {
+					return ntok.err
+				}
+				if ntok.value == "]" {
+					break
+				}
+				if ntok.value != "," {
+					return p.errorf("Expected ']' or ',' found %q", ntok.value)
+				}
+			}
+			return nil
+		}
+		// One value of the repeated field.
+		p.back()
+		fv.Set(reflect.Append(fv, reflect.New(at.Elem()).Elem()))
+		return p.readAny(fv.Index(fv.Len()-1), props)
+	case reflect.Bool:
+		// true/1/t/True or false/f/0/False.
+		switch tok.value {
+		case "true", "1", "t", "True":
+			fv.SetBool(true)
+			return nil
+		case "false", "0", "f", "False":
+			fv.SetBool(false)
+			return nil
+		}
+	case reflect.Float32, reflect.Float64:
+		v := tok.value
+		// Ignore 'f' for compatibility with output generated by C++, but don't
+		// remove 'f' when the value is "-inf" or "inf".
+		if strings.HasSuffix(v, "f") && tok.value != "-inf" && tok.value != "inf" {
+			v = v[:len(v)-1]
+		}
+		if f, err := strconv.ParseFloat(v, fv.Type().Bits()); err == nil {
+			fv.SetFloat(f)
+			return nil
+		}
+	case reflect.Int8:
+		if x, err := strconv.ParseInt(tok.value, 0, 8); err == nil {
+			fv.SetInt(x)
+			return nil
+		}
+	case reflect.Int16:
+		if x, err := strconv.ParseInt(tok.value, 0, 16); err == nil {
+			fv.SetInt(x)
+			return nil
+		}
+	case reflect.Int32:
+		if x, err := strconv.ParseInt(tok.value, 0, 32); err == nil {
+			fv.SetInt(x)
+			return nil
+		}
+
+		if len(props.Enum) == 0 {
+			break
+		}
+		m, ok := enumValueMaps[props.Enum]
+		if !ok {
+			break
+		}
+		x, ok := m[tok.value]
+		if !ok {
+			break
+		}
+		fv.SetInt(int64(x))
+		return nil
+	case reflect.Int64:
+		if x, err := strconv.ParseInt(tok.value, 0, 64); err == nil {
+			fv.SetInt(x)
+			return nil
+		}
+
+	case reflect.Ptr:
+		// A basic field (indirected through pointer), or a repeated message/group
+		p.back()
+		fv.Set(reflect.New(fv.Type().Elem()))
+		return p.readAny(fv.Elem(), props)
+	case reflect.String:
+		if tok.value[0] == '"' || tok.value[0] == '\'' {
+			fv.SetString(tok.unquoted)
+			return nil
+		}
+	case reflect.Struct:
+		var terminator string
+		switch tok.value {
+		case "{":
+			terminator = "}"
+		case "<":
+			terminator = ">"
+		default:
+			return p.errorf("expected '{' or '<', found %q", tok.value)
+		}
+		// TODO: Handle nested messages which implement encoding.TextUnmarshaler.
+		return p.readStruct(fv, terminator)
+	case reflect.Uint8:
+		if x, err := strconv.ParseUint(tok.value, 0, 8); err == nil {
+			fv.SetUint(x)
+			return nil
+		}
+	case reflect.Uint16:
+		if x, err := strconv.ParseUint(tok.value, 0, 16); err == nil {
+			fv.SetUint(x)
+			return nil
+		}
+	case reflect.Uint32:
+		if x, err := strconv.ParseUint(tok.value, 0, 32); err == nil {
+			fv.SetUint(uint64(x))
+			return nil
+		}
+	case reflect.Uint64:
+		if x, err := strconv.ParseUint(tok.value, 0, 64); err == nil {
+			fv.SetUint(x)
+			return nil
+		}
+	}
+	return p.errorf("invalid %v: %v", v.Type(), tok.value)
+}
+
+// UnmarshalText reads a protocol buffer in Text format. UnmarshalText resets pb
+// before starting to unmarshal, so any existing data in pb is always removed.
+// If a required field is not set and no other error occurs,
+// UnmarshalText returns *RequiredNotSetError.
+func UnmarshalText(s string, pb Message) error {
+	if um, ok := pb.(encoding.TextUnmarshaler); ok {
+		return um.UnmarshalText([]byte(s))
+	}
+	pb.Reset()
+	v := reflect.ValueOf(pb)
+	return newTextParser(s).readStruct(v.Elem(), "")
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/timestamp.go b/vendor/github.com/gogo/protobuf/proto/timestamp.go
new file mode 100644
index 0000000..9324f65
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/timestamp.go
@@ -0,0 +1,113 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+// This file implements operations on google.protobuf.Timestamp.
+
+import (
+	"errors"
+	"fmt"
+	"time"
+)
+
+const (
+	// Seconds field of the earliest valid Timestamp.
+	// This is time.Date(1, 1, 1, 0, 0, 0, 0, time.UTC).Unix().
+	minValidSeconds = -62135596800
+	// Seconds field just after the latest valid Timestamp.
+	// This is time.Date(10000, 1, 1, 0, 0, 0, 0, time.UTC).Unix().
+	maxValidSeconds = 253402300800
+)
+
+// validateTimestamp determines whether a Timestamp is valid.
+// A valid timestamp represents a time in the range
+// [0001-01-01, 10000-01-01) and has a Nanos field
+// in the range [0, 1e9).
+//
+// If the Timestamp is valid, validateTimestamp returns nil.
+// Otherwise, it returns an error that describes
+// the problem.
+//
+// Every valid Timestamp can be represented by a time.Time, but the converse is not true.
+func validateTimestamp(ts *timestamp) error {
+	if ts == nil {
+		return errors.New("timestamp: nil Timestamp")
+	}
+	if ts.Seconds < minValidSeconds {
+		return fmt.Errorf("timestamp: %#v before 0001-01-01", ts)
+	}
+	if ts.Seconds >= maxValidSeconds {
+		return fmt.Errorf("timestamp: %#v after 10000-01-01", ts)
+	}
+	if ts.Nanos < 0 || ts.Nanos >= 1e9 {
+		return fmt.Errorf("timestamp: %#v: nanos not in range [0, 1e9)", ts)
+	}
+	return nil
+}
+
+// TimestampFromProto converts a google.protobuf.Timestamp proto to a time.Time.
+// It returns an error if the argument is invalid.
+//
+// Unlike most Go functions, if Timestamp returns an error, the first return value
+// is not the zero time.Time. Instead, it is the value obtained from the
+// time.Unix function when passed the contents of the Timestamp, in the UTC
+// locale. This may or may not be a meaningful time; many invalid Timestamps
+// do map to valid time.Times.
+//
+// A nil Timestamp returns an error. The first return value in that case is
+// undefined.
+func timestampFromProto(ts *timestamp) (time.Time, error) {
+	// Don't return the zero value on error, because corresponds to a valid
+	// timestamp. Instead return whatever time.Unix gives us.
+	var t time.Time
+	if ts == nil {
+		t = time.Unix(0, 0).UTC() // treat nil like the empty Timestamp
+	} else {
+		t = time.Unix(ts.Seconds, int64(ts.Nanos)).UTC()
+	}
+	return t, validateTimestamp(ts)
+}
+
+// TimestampProto converts the time.Time to a google.protobuf.Timestamp proto.
+// It returns an error if the resulting Timestamp is invalid.
+func timestampProto(t time.Time) (*timestamp, error) {
+	seconds := t.Unix()
+	nanos := int32(t.Sub(time.Unix(seconds, 0)))
+	ts := &timestamp{
+		Seconds: seconds,
+		Nanos:   nanos,
+	}
+	if err := validateTimestamp(ts); err != nil {
+		return nil, err
+	}
+	return ts, nil
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/timestamp_gogo.go b/vendor/github.com/gogo/protobuf/proto/timestamp_gogo.go
new file mode 100644
index 0000000..38439fa
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/timestamp_gogo.go
@@ -0,0 +1,49 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2016, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"reflect"
+	"time"
+)
+
+var timeType = reflect.TypeOf((*time.Time)(nil)).Elem()
+
+type timestamp struct {
+	Seconds int64 `protobuf:"varint,1,opt,name=seconds,proto3" json:"seconds,omitempty"`
+	Nanos   int32 `protobuf:"varint,2,opt,name=nanos,proto3" json:"nanos,omitempty"`
+}
+
+func (m *timestamp) Reset()       { *m = timestamp{} }
+func (*timestamp) ProtoMessage()  {}
+func (*timestamp) String() string { return "timestamp<string>" }
+
+func init() {
+	RegisterType((*timestamp)(nil), "gogo.protobuf.proto.timestamp")
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/wrappers.go b/vendor/github.com/gogo/protobuf/proto/wrappers.go
new file mode 100644
index 0000000..b175d1b
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/wrappers.go
@@ -0,0 +1,1888 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"io"
+	"reflect"
+)
+
+func makeStdDoubleValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*float64)
+			v := &float64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*float64)
+			v := &float64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdDoubleValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*float64)
+			v := &float64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*float64)
+			v := &float64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdDoubleValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(float64)
+				v := &float64Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(float64)
+				v := &float64Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdDoubleValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*float64)
+				v := &float64Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*float64)
+				v := &float64Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdDoubleValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdDoubleValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdDoubleValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdDoubleValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdFloatValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*float32)
+			v := &float32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*float32)
+			v := &float32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdFloatValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*float32)
+			v := &float32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*float32)
+			v := &float32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdFloatValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(float32)
+				v := &float32Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(float32)
+				v := &float32Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdFloatValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*float32)
+				v := &float32Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*float32)
+				v := &float32Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdFloatValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdFloatValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdFloatValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdFloatValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &float32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdInt64ValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*int64)
+			v := &int64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*int64)
+			v := &int64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdInt64ValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*int64)
+			v := &int64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*int64)
+			v := &int64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdInt64ValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(int64)
+				v := &int64Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(int64)
+				v := &int64Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdInt64ValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*int64)
+				v := &int64Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*int64)
+				v := &int64Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdInt64ValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdInt64ValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdInt64ValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdInt64ValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt64ValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*uint64)
+			v := &uint64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*uint64)
+			v := &uint64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdUInt64ValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*uint64)
+			v := &uint64Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*uint64)
+			v := &uint64Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdUInt64ValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(uint64)
+				v := &uint64Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(uint64)
+				v := &uint64Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdUInt64ValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*uint64)
+				v := &uint64Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*uint64)
+				v := &uint64Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdUInt64ValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt64ValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt64ValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt64ValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint64Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdInt32ValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*int32)
+			v := &int32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*int32)
+			v := &int32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdInt32ValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*int32)
+			v := &int32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*int32)
+			v := &int32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdInt32ValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(int32)
+				v := &int32Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(int32)
+				v := &int32Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdInt32ValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*int32)
+				v := &int32Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*int32)
+				v := &int32Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdInt32ValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdInt32ValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdInt32ValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdInt32ValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &int32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt32ValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*uint32)
+			v := &uint32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*uint32)
+			v := &uint32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdUInt32ValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*uint32)
+			v := &uint32Value{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*uint32)
+			v := &uint32Value{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdUInt32ValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(uint32)
+				v := &uint32Value{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(uint32)
+				v := &uint32Value{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdUInt32ValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*uint32)
+				v := &uint32Value{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*uint32)
+				v := &uint32Value{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdUInt32ValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt32ValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt32ValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdUInt32ValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &uint32Value{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdBoolValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*bool)
+			v := &boolValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*bool)
+			v := &boolValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdBoolValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*bool)
+			v := &boolValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*bool)
+			v := &boolValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdBoolValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(bool)
+				v := &boolValue{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(bool)
+				v := &boolValue{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdBoolValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*bool)
+				v := &boolValue{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*bool)
+				v := &boolValue{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdBoolValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &boolValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdBoolValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &boolValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdBoolValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &boolValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdBoolValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &boolValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdStringValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*string)
+			v := &stringValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*string)
+			v := &stringValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdStringValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*string)
+			v := &stringValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*string)
+			v := &stringValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdStringValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(string)
+				v := &stringValue{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(string)
+				v := &stringValue{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdStringValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*string)
+				v := &stringValue{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*string)
+				v := &stringValue{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdStringValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &stringValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdStringValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &stringValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdStringValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &stringValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdStringValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &stringValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdBytesValueMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			t := ptr.asPointerTo(u.typ).Interface().(*[]byte)
+			v := &bytesValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			t := ptr.asPointerTo(u.typ).Interface().(*[]byte)
+			v := &bytesValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdBytesValuePtrMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			if ptr.isNil() {
+				return 0
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*[]byte)
+			v := &bytesValue{*t}
+			siz := Size(v)
+			return tagsize + SizeVarint(uint64(siz)) + siz
+		}, func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			if ptr.isNil() {
+				return b, nil
+			}
+			t := ptr.asPointerTo(reflect.PtrTo(u.typ)).Elem().Interface().(*[]byte)
+			v := &bytesValue{*t}
+			buf, err := Marshal(v)
+			if err != nil {
+				return nil, err
+			}
+			b = appendVarint(b, wiretag)
+			b = appendVarint(b, uint64(len(buf)))
+			b = append(b, buf...)
+			return b, nil
+		}
+}
+
+func makeStdBytesValueSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(u.typ)
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().([]byte)
+				v := &bytesValue{t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(u.typ)
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().([]byte)
+				v := &bytesValue{t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdBytesValuePtrSliceMarshaler(u *marshalInfo) (sizer, marshaler) {
+	return func(ptr pointer, tagsize int) int {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			n := 0
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*[]byte)
+				v := &bytesValue{*t}
+				siz := Size(v)
+				n += siz + SizeVarint(uint64(siz)) + tagsize
+			}
+			return n
+		},
+		func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error) {
+			s := ptr.getSlice(reflect.PtrTo(u.typ))
+			for i := 0; i < s.Len(); i++ {
+				elem := s.Index(i)
+				t := elem.Interface().(*[]byte)
+				v := &bytesValue{*t}
+				siz := Size(v)
+				buf, err := Marshal(v)
+				if err != nil {
+					return nil, err
+				}
+				b = appendVarint(b, wiretag)
+				b = appendVarint(b, uint64(siz))
+				b = append(b, buf...)
+			}
+
+			return b, nil
+		}
+}
+
+func makeStdBytesValueUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &bytesValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(sub.typ).Elem()
+		s.Set(reflect.ValueOf(m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdBytesValuePtrUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &bytesValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		s := f.asPointerTo(reflect.PtrTo(sub.typ)).Elem()
+		s.Set(reflect.ValueOf(&m.Value))
+		return b[x:], nil
+	}
+}
+
+func makeStdBytesValuePtrSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &bytesValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(reflect.PtrTo(sub.typ))
+		newSlice := reflect.Append(slice, reflect.ValueOf(&m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
+
+func makeStdBytesValueSliceUnmarshaler(sub *unmarshalInfo, name string) unmarshaler {
+	return func(b []byte, f pointer, w int) ([]byte, error) {
+		if w != WireBytes {
+			return nil, errInternalBadWireType
+		}
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		if x > uint64(len(b)) {
+			return nil, io.ErrUnexpectedEOF
+		}
+		m := &bytesValue{}
+		if err := Unmarshal(b[:x], m); err != nil {
+			return nil, err
+		}
+		slice := f.getSlice(sub.typ)
+		newSlice := reflect.Append(slice, reflect.ValueOf(m.Value))
+		slice.Set(newSlice)
+		return b[x:], nil
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/proto/wrappers_gogo.go b/vendor/github.com/gogo/protobuf/proto/wrappers_gogo.go
new file mode 100644
index 0000000..c1cf7bf
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/proto/wrappers_gogo.go
@@ -0,0 +1,113 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2018, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+type float64Value struct {
+	Value float64 `protobuf:"fixed64,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *float64Value) Reset()       { *m = float64Value{} }
+func (*float64Value) ProtoMessage()  {}
+func (*float64Value) String() string { return "float64<string>" }
+
+type float32Value struct {
+	Value float32 `protobuf:"fixed32,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *float32Value) Reset()       { *m = float32Value{} }
+func (*float32Value) ProtoMessage()  {}
+func (*float32Value) String() string { return "float32<string>" }
+
+type int64Value struct {
+	Value int64 `protobuf:"varint,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *int64Value) Reset()       { *m = int64Value{} }
+func (*int64Value) ProtoMessage()  {}
+func (*int64Value) String() string { return "int64<string>" }
+
+type uint64Value struct {
+	Value uint64 `protobuf:"varint,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *uint64Value) Reset()       { *m = uint64Value{} }
+func (*uint64Value) ProtoMessage()  {}
+func (*uint64Value) String() string { return "uint64<string>" }
+
+type int32Value struct {
+	Value int32 `protobuf:"varint,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *int32Value) Reset()       { *m = int32Value{} }
+func (*int32Value) ProtoMessage()  {}
+func (*int32Value) String() string { return "int32<string>" }
+
+type uint32Value struct {
+	Value uint32 `protobuf:"varint,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *uint32Value) Reset()       { *m = uint32Value{} }
+func (*uint32Value) ProtoMessage()  {}
+func (*uint32Value) String() string { return "uint32<string>" }
+
+type boolValue struct {
+	Value bool `protobuf:"varint,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *boolValue) Reset()       { *m = boolValue{} }
+func (*boolValue) ProtoMessage()  {}
+func (*boolValue) String() string { return "bool<string>" }
+
+type stringValue struct {
+	Value string `protobuf:"bytes,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *stringValue) Reset()       { *m = stringValue{} }
+func (*stringValue) ProtoMessage()  {}
+func (*stringValue) String() string { return "string<string>" }
+
+type bytesValue struct {
+	Value []byte `protobuf:"bytes,1,opt,name=value,proto3" json:"value,omitempty"`
+}
+
+func (m *bytesValue) Reset()       { *m = bytesValue{} }
+func (*bytesValue) ProtoMessage()  {}
+func (*bytesValue) String() string { return "[]byte<string>" }
+
+func init() {
+	RegisterType((*float64Value)(nil), "gogo.protobuf.proto.DoubleValue")
+	RegisterType((*float32Value)(nil), "gogo.protobuf.proto.FloatValue")
+	RegisterType((*int64Value)(nil), "gogo.protobuf.proto.Int64Value")
+	RegisterType((*uint64Value)(nil), "gogo.protobuf.proto.UInt64Value")
+	RegisterType((*int32Value)(nil), "gogo.protobuf.proto.Int32Value")
+	RegisterType((*uint32Value)(nil), "gogo.protobuf.proto.UInt32Value")
+	RegisterType((*boolValue)(nil), "gogo.protobuf.proto.BoolValue")
+	RegisterType((*stringValue)(nil), "gogo.protobuf.proto.StringValue")
+	RegisterType((*bytesValue)(nil), "gogo.protobuf.proto.BytesValue")
+}
diff --git a/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/Makefile b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/Makefile
new file mode 100644
index 0000000..3496dc9
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/Makefile
@@ -0,0 +1,36 @@
+# Go support for Protocol Buffers - Google's data interchange format
+#
+# Copyright 2010 The Go Authors.  All rights reserved.
+# https://github.com/golang/protobuf
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+regenerate:
+	go install github.com/gogo/protobuf/protoc-gen-gogo
+	go install github.com/gogo/protobuf/protoc-gen-gostring
+	protoc --gogo_out=. -I=../../protobuf/google/protobuf ../../protobuf/google/protobuf/descriptor.proto
+	protoc --gostring_out=. -I=../../protobuf/google/protobuf ../../protobuf/google/protobuf/descriptor.proto
diff --git a/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.go b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.go
new file mode 100644
index 0000000..a85bf19
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.go
@@ -0,0 +1,118 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Package descriptor provides functions for obtaining protocol buffer
+// descriptors for generated Go types.
+//
+// These functions cannot go in package proto because they depend on the
+// generated protobuf descriptor messages, which themselves depend on proto.
+package descriptor
+
+import (
+	"bytes"
+	"compress/gzip"
+	"fmt"
+	"io/ioutil"
+
+	"github.com/gogo/protobuf/proto"
+)
+
+// extractFile extracts a FileDescriptorProto from a gzip'd buffer.
+func extractFile(gz []byte) (*FileDescriptorProto, error) {
+	r, err := gzip.NewReader(bytes.NewReader(gz))
+	if err != nil {
+		return nil, fmt.Errorf("failed to open gzip reader: %v", err)
+	}
+	defer r.Close()
+
+	b, err := ioutil.ReadAll(r)
+	if err != nil {
+		return nil, fmt.Errorf("failed to uncompress descriptor: %v", err)
+	}
+
+	fd := new(FileDescriptorProto)
+	if err := proto.Unmarshal(b, fd); err != nil {
+		return nil, fmt.Errorf("malformed FileDescriptorProto: %v", err)
+	}
+
+	return fd, nil
+}
+
+// Message is a proto.Message with a method to return its descriptor.
+//
+// Message types generated by the protocol compiler always satisfy
+// the Message interface.
+type Message interface {
+	proto.Message
+	Descriptor() ([]byte, []int)
+}
+
+// ForMessage returns a FileDescriptorProto and a DescriptorProto from within it
+// describing the given message.
+func ForMessage(msg Message) (fd *FileDescriptorProto, md *DescriptorProto) {
+	gz, path := msg.Descriptor()
+	fd, err := extractFile(gz)
+	if err != nil {
+		panic(fmt.Sprintf("invalid FileDescriptorProto for %T: %v", msg, err))
+	}
+
+	md = fd.MessageType[path[0]]
+	for _, i := range path[1:] {
+		md = md.NestedType[i]
+	}
+	return fd, md
+}
+
+// Is this field a scalar numeric type?
+func (field *FieldDescriptorProto) IsScalar() bool {
+	if field.Type == nil {
+		return false
+	}
+	switch *field.Type {
+	case FieldDescriptorProto_TYPE_DOUBLE,
+		FieldDescriptorProto_TYPE_FLOAT,
+		FieldDescriptorProto_TYPE_INT64,
+		FieldDescriptorProto_TYPE_UINT64,
+		FieldDescriptorProto_TYPE_INT32,
+		FieldDescriptorProto_TYPE_FIXED64,
+		FieldDescriptorProto_TYPE_FIXED32,
+		FieldDescriptorProto_TYPE_BOOL,
+		FieldDescriptorProto_TYPE_UINT32,
+		FieldDescriptorProto_TYPE_ENUM,
+		FieldDescriptorProto_TYPE_SFIXED32,
+		FieldDescriptorProto_TYPE_SFIXED64,
+		FieldDescriptorProto_TYPE_SINT32,
+		FieldDescriptorProto_TYPE_SINT64:
+		return true
+	default:
+		return false
+	}
+}
diff --git a/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.pb.go b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.pb.go
new file mode 100644
index 0000000..cacfa39
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor.pb.go
@@ -0,0 +1,2865 @@
+// Code generated by protoc-gen-gogo. DO NOT EDIT.
+// source: descriptor.proto
+
+package descriptor
+
+import (
+	fmt "fmt"
+	proto "github.com/gogo/protobuf/proto"
+	math "math"
+)
+
+// Reference imports to suppress errors if they are not otherwise used.
+var _ = proto.Marshal
+var _ = fmt.Errorf
+var _ = math.Inf
+
+// This is a compile-time assertion to ensure that this generated file
+// is compatible with the proto package it is being compiled against.
+// A compilation error at this line likely means your copy of the
+// proto package needs to be updated.
+const _ = proto.GoGoProtoPackageIsVersion2 // please upgrade the proto package
+
+type FieldDescriptorProto_Type int32
+
+const (
+	// 0 is reserved for errors.
+	// Order is weird for historical reasons.
+	FieldDescriptorProto_TYPE_DOUBLE FieldDescriptorProto_Type = 1
+	FieldDescriptorProto_TYPE_FLOAT  FieldDescriptorProto_Type = 2
+	// Not ZigZag encoded.  Negative numbers take 10 bytes.  Use TYPE_SINT64 if
+	// negative values are likely.
+	FieldDescriptorProto_TYPE_INT64  FieldDescriptorProto_Type = 3
+	FieldDescriptorProto_TYPE_UINT64 FieldDescriptorProto_Type = 4
+	// Not ZigZag encoded.  Negative numbers take 10 bytes.  Use TYPE_SINT32 if
+	// negative values are likely.
+	FieldDescriptorProto_TYPE_INT32   FieldDescriptorProto_Type = 5
+	FieldDescriptorProto_TYPE_FIXED64 FieldDescriptorProto_Type = 6
+	FieldDescriptorProto_TYPE_FIXED32 FieldDescriptorProto_Type = 7
+	FieldDescriptorProto_TYPE_BOOL    FieldDescriptorProto_Type = 8
+	FieldDescriptorProto_TYPE_STRING  FieldDescriptorProto_Type = 9
+	// Tag-delimited aggregate.
+	// Group type is deprecated and not supported in proto3. However, Proto3
+	// implementations should still be able to parse the group wire format and
+	// treat group fields as unknown fields.
+	FieldDescriptorProto_TYPE_GROUP   FieldDescriptorProto_Type = 10
+	FieldDescriptorProto_TYPE_MESSAGE FieldDescriptorProto_Type = 11
+	// New in version 2.
+	FieldDescriptorProto_TYPE_BYTES    FieldDescriptorProto_Type = 12
+	FieldDescriptorProto_TYPE_UINT32   FieldDescriptorProto_Type = 13
+	FieldDescriptorProto_TYPE_ENUM     FieldDescriptorProto_Type = 14
+	FieldDescriptorProto_TYPE_SFIXED32 FieldDescriptorProto_Type = 15
+	FieldDescriptorProto_TYPE_SFIXED64 FieldDescriptorProto_Type = 16
+	FieldDescriptorProto_TYPE_SINT32   FieldDescriptorProto_Type = 17
+	FieldDescriptorProto_TYPE_SINT64   FieldDescriptorProto_Type = 18
+)
+
+var FieldDescriptorProto_Type_name = map[int32]string{
+	1:  "TYPE_DOUBLE",
+	2:  "TYPE_FLOAT",
+	3:  "TYPE_INT64",
+	4:  "TYPE_UINT64",
+	5:  "TYPE_INT32",
+	6:  "TYPE_FIXED64",
+	7:  "TYPE_FIXED32",
+	8:  "TYPE_BOOL",
+	9:  "TYPE_STRING",
+	10: "TYPE_GROUP",
+	11: "TYPE_MESSAGE",
+	12: "TYPE_BYTES",
+	13: "TYPE_UINT32",
+	14: "TYPE_ENUM",
+	15: "TYPE_SFIXED32",
+	16: "TYPE_SFIXED64",
+	17: "TYPE_SINT32",
+	18: "TYPE_SINT64",
+}
+
+var FieldDescriptorProto_Type_value = map[string]int32{
+	"TYPE_DOUBLE":   1,
+	"TYPE_FLOAT":    2,
+	"TYPE_INT64":    3,
+	"TYPE_UINT64":   4,
+	"TYPE_INT32":    5,
+	"TYPE_FIXED64":  6,
+	"TYPE_FIXED32":  7,
+	"TYPE_BOOL":     8,
+	"TYPE_STRING":   9,
+	"TYPE_GROUP":    10,
+	"TYPE_MESSAGE":  11,
+	"TYPE_BYTES":    12,
+	"TYPE_UINT32":   13,
+	"TYPE_ENUM":     14,
+	"TYPE_SFIXED32": 15,
+	"TYPE_SFIXED64": 16,
+	"TYPE_SINT32":   17,
+	"TYPE_SINT64":   18,
+}
+
+func (x FieldDescriptorProto_Type) Enum() *FieldDescriptorProto_Type {
+	p := new(FieldDescriptorProto_Type)
+	*p = x
+	return p
+}
+
+func (x FieldDescriptorProto_Type) String() string {
+	return proto.EnumName(FieldDescriptorProto_Type_name, int32(x))
+}
+
+func (x *FieldDescriptorProto_Type) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(FieldDescriptorProto_Type_value, data, "FieldDescriptorProto_Type")
+	if err != nil {
+		return err
+	}
+	*x = FieldDescriptorProto_Type(value)
+	return nil
+}
+
+func (FieldDescriptorProto_Type) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{4, 0}
+}
+
+type FieldDescriptorProto_Label int32
+
+const (
+	// 0 is reserved for errors
+	FieldDescriptorProto_LABEL_OPTIONAL FieldDescriptorProto_Label = 1
+	FieldDescriptorProto_LABEL_REQUIRED FieldDescriptorProto_Label = 2
+	FieldDescriptorProto_LABEL_REPEATED FieldDescriptorProto_Label = 3
+)
+
+var FieldDescriptorProto_Label_name = map[int32]string{
+	1: "LABEL_OPTIONAL",
+	2: "LABEL_REQUIRED",
+	3: "LABEL_REPEATED",
+}
+
+var FieldDescriptorProto_Label_value = map[string]int32{
+	"LABEL_OPTIONAL": 1,
+	"LABEL_REQUIRED": 2,
+	"LABEL_REPEATED": 3,
+}
+
+func (x FieldDescriptorProto_Label) Enum() *FieldDescriptorProto_Label {
+	p := new(FieldDescriptorProto_Label)
+	*p = x
+	return p
+}
+
+func (x FieldDescriptorProto_Label) String() string {
+	return proto.EnumName(FieldDescriptorProto_Label_name, int32(x))
+}
+
+func (x *FieldDescriptorProto_Label) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(FieldDescriptorProto_Label_value, data, "FieldDescriptorProto_Label")
+	if err != nil {
+		return err
+	}
+	*x = FieldDescriptorProto_Label(value)
+	return nil
+}
+
+func (FieldDescriptorProto_Label) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{4, 1}
+}
+
+// Generated classes can be optimized for speed or code size.
+type FileOptions_OptimizeMode int32
+
+const (
+	FileOptions_SPEED FileOptions_OptimizeMode = 1
+	// etc.
+	FileOptions_CODE_SIZE    FileOptions_OptimizeMode = 2
+	FileOptions_LITE_RUNTIME FileOptions_OptimizeMode = 3
+)
+
+var FileOptions_OptimizeMode_name = map[int32]string{
+	1: "SPEED",
+	2: "CODE_SIZE",
+	3: "LITE_RUNTIME",
+}
+
+var FileOptions_OptimizeMode_value = map[string]int32{
+	"SPEED":        1,
+	"CODE_SIZE":    2,
+	"LITE_RUNTIME": 3,
+}
+
+func (x FileOptions_OptimizeMode) Enum() *FileOptions_OptimizeMode {
+	p := new(FileOptions_OptimizeMode)
+	*p = x
+	return p
+}
+
+func (x FileOptions_OptimizeMode) String() string {
+	return proto.EnumName(FileOptions_OptimizeMode_name, int32(x))
+}
+
+func (x *FileOptions_OptimizeMode) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(FileOptions_OptimizeMode_value, data, "FileOptions_OptimizeMode")
+	if err != nil {
+		return err
+	}
+	*x = FileOptions_OptimizeMode(value)
+	return nil
+}
+
+func (FileOptions_OptimizeMode) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{10, 0}
+}
+
+type FieldOptions_CType int32
+
+const (
+	// Default mode.
+	FieldOptions_STRING       FieldOptions_CType = 0
+	FieldOptions_CORD         FieldOptions_CType = 1
+	FieldOptions_STRING_PIECE FieldOptions_CType = 2
+)
+
+var FieldOptions_CType_name = map[int32]string{
+	0: "STRING",
+	1: "CORD",
+	2: "STRING_PIECE",
+}
+
+var FieldOptions_CType_value = map[string]int32{
+	"STRING":       0,
+	"CORD":         1,
+	"STRING_PIECE": 2,
+}
+
+func (x FieldOptions_CType) Enum() *FieldOptions_CType {
+	p := new(FieldOptions_CType)
+	*p = x
+	return p
+}
+
+func (x FieldOptions_CType) String() string {
+	return proto.EnumName(FieldOptions_CType_name, int32(x))
+}
+
+func (x *FieldOptions_CType) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(FieldOptions_CType_value, data, "FieldOptions_CType")
+	if err != nil {
+		return err
+	}
+	*x = FieldOptions_CType(value)
+	return nil
+}
+
+func (FieldOptions_CType) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{12, 0}
+}
+
+type FieldOptions_JSType int32
+
+const (
+	// Use the default type.
+	FieldOptions_JS_NORMAL FieldOptions_JSType = 0
+	// Use JavaScript strings.
+	FieldOptions_JS_STRING FieldOptions_JSType = 1
+	// Use JavaScript numbers.
+	FieldOptions_JS_NUMBER FieldOptions_JSType = 2
+)
+
+var FieldOptions_JSType_name = map[int32]string{
+	0: "JS_NORMAL",
+	1: "JS_STRING",
+	2: "JS_NUMBER",
+}
+
+var FieldOptions_JSType_value = map[string]int32{
+	"JS_NORMAL": 0,
+	"JS_STRING": 1,
+	"JS_NUMBER": 2,
+}
+
+func (x FieldOptions_JSType) Enum() *FieldOptions_JSType {
+	p := new(FieldOptions_JSType)
+	*p = x
+	return p
+}
+
+func (x FieldOptions_JSType) String() string {
+	return proto.EnumName(FieldOptions_JSType_name, int32(x))
+}
+
+func (x *FieldOptions_JSType) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(FieldOptions_JSType_value, data, "FieldOptions_JSType")
+	if err != nil {
+		return err
+	}
+	*x = FieldOptions_JSType(value)
+	return nil
+}
+
+func (FieldOptions_JSType) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{12, 1}
+}
+
+// Is this method side-effect-free (or safe in HTTP parlance), or idempotent,
+// or neither? HTTP based RPC implementation may choose GET verb for safe
+// methods, and PUT verb for idempotent methods instead of the default POST.
+type MethodOptions_IdempotencyLevel int32
+
+const (
+	MethodOptions_IDEMPOTENCY_UNKNOWN MethodOptions_IdempotencyLevel = 0
+	MethodOptions_NO_SIDE_EFFECTS     MethodOptions_IdempotencyLevel = 1
+	MethodOptions_IDEMPOTENT          MethodOptions_IdempotencyLevel = 2
+)
+
+var MethodOptions_IdempotencyLevel_name = map[int32]string{
+	0: "IDEMPOTENCY_UNKNOWN",
+	1: "NO_SIDE_EFFECTS",
+	2: "IDEMPOTENT",
+}
+
+var MethodOptions_IdempotencyLevel_value = map[string]int32{
+	"IDEMPOTENCY_UNKNOWN": 0,
+	"NO_SIDE_EFFECTS":     1,
+	"IDEMPOTENT":          2,
+}
+
+func (x MethodOptions_IdempotencyLevel) Enum() *MethodOptions_IdempotencyLevel {
+	p := new(MethodOptions_IdempotencyLevel)
+	*p = x
+	return p
+}
+
+func (x MethodOptions_IdempotencyLevel) String() string {
+	return proto.EnumName(MethodOptions_IdempotencyLevel_name, int32(x))
+}
+
+func (x *MethodOptions_IdempotencyLevel) UnmarshalJSON(data []byte) error {
+	value, err := proto.UnmarshalJSONEnum(MethodOptions_IdempotencyLevel_value, data, "MethodOptions_IdempotencyLevel")
+	if err != nil {
+		return err
+	}
+	*x = MethodOptions_IdempotencyLevel(value)
+	return nil
+}
+
+func (MethodOptions_IdempotencyLevel) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{17, 0}
+}
+
+// The protocol compiler can output a FileDescriptorSet containing the .proto
+// files it parses.
+type FileDescriptorSet struct {
+	File                 []*FileDescriptorProto `protobuf:"bytes,1,rep,name=file" json:"file,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}               `json:"-"`
+	XXX_unrecognized     []byte                 `json:"-"`
+	XXX_sizecache        int32                  `json:"-"`
+}
+
+func (m *FileDescriptorSet) Reset()         { *m = FileDescriptorSet{} }
+func (m *FileDescriptorSet) String() string { return proto.CompactTextString(m) }
+func (*FileDescriptorSet) ProtoMessage()    {}
+func (*FileDescriptorSet) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{0}
+}
+func (m *FileDescriptorSet) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_FileDescriptorSet.Unmarshal(m, b)
+}
+func (m *FileDescriptorSet) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_FileDescriptorSet.Marshal(b, m, deterministic)
+}
+func (m *FileDescriptorSet) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_FileDescriptorSet.Merge(m, src)
+}
+func (m *FileDescriptorSet) XXX_Size() int {
+	return xxx_messageInfo_FileDescriptorSet.Size(m)
+}
+func (m *FileDescriptorSet) XXX_DiscardUnknown() {
+	xxx_messageInfo_FileDescriptorSet.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_FileDescriptorSet proto.InternalMessageInfo
+
+func (m *FileDescriptorSet) GetFile() []*FileDescriptorProto {
+	if m != nil {
+		return m.File
+	}
+	return nil
+}
+
+// Describes a complete .proto file.
+type FileDescriptorProto struct {
+	Name    *string `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Package *string `protobuf:"bytes,2,opt,name=package" json:"package,omitempty"`
+	// Names of files imported by this file.
+	Dependency []string `protobuf:"bytes,3,rep,name=dependency" json:"dependency,omitempty"`
+	// Indexes of the public imported files in the dependency list above.
+	PublicDependency []int32 `protobuf:"varint,10,rep,name=public_dependency,json=publicDependency" json:"public_dependency,omitempty"`
+	// Indexes of the weak imported files in the dependency list.
+	// For Google-internal migration only. Do not use.
+	WeakDependency []int32 `protobuf:"varint,11,rep,name=weak_dependency,json=weakDependency" json:"weak_dependency,omitempty"`
+	// All top-level definitions in this file.
+	MessageType []*DescriptorProto        `protobuf:"bytes,4,rep,name=message_type,json=messageType" json:"message_type,omitempty"`
+	EnumType    []*EnumDescriptorProto    `protobuf:"bytes,5,rep,name=enum_type,json=enumType" json:"enum_type,omitempty"`
+	Service     []*ServiceDescriptorProto `protobuf:"bytes,6,rep,name=service" json:"service,omitempty"`
+	Extension   []*FieldDescriptorProto   `protobuf:"bytes,7,rep,name=extension" json:"extension,omitempty"`
+	Options     *FileOptions              `protobuf:"bytes,8,opt,name=options" json:"options,omitempty"`
+	// This field contains optional information about the original source code.
+	// You may safely remove this entire field without harming runtime
+	// functionality of the descriptors -- the information is needed only by
+	// development tools.
+	SourceCodeInfo *SourceCodeInfo `protobuf:"bytes,9,opt,name=source_code_info,json=sourceCodeInfo" json:"source_code_info,omitempty"`
+	// The syntax of the proto file.
+	// The supported values are "proto2" and "proto3".
+	Syntax               *string  `protobuf:"bytes,12,opt,name=syntax" json:"syntax,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *FileDescriptorProto) Reset()         { *m = FileDescriptorProto{} }
+func (m *FileDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*FileDescriptorProto) ProtoMessage()    {}
+func (*FileDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{1}
+}
+func (m *FileDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_FileDescriptorProto.Unmarshal(m, b)
+}
+func (m *FileDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_FileDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *FileDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_FileDescriptorProto.Merge(m, src)
+}
+func (m *FileDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_FileDescriptorProto.Size(m)
+}
+func (m *FileDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_FileDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_FileDescriptorProto proto.InternalMessageInfo
+
+func (m *FileDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *FileDescriptorProto) GetPackage() string {
+	if m != nil && m.Package != nil {
+		return *m.Package
+	}
+	return ""
+}
+
+func (m *FileDescriptorProto) GetDependency() []string {
+	if m != nil {
+		return m.Dependency
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetPublicDependency() []int32 {
+	if m != nil {
+		return m.PublicDependency
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetWeakDependency() []int32 {
+	if m != nil {
+		return m.WeakDependency
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetMessageType() []*DescriptorProto {
+	if m != nil {
+		return m.MessageType
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetEnumType() []*EnumDescriptorProto {
+	if m != nil {
+		return m.EnumType
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetService() []*ServiceDescriptorProto {
+	if m != nil {
+		return m.Service
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetExtension() []*FieldDescriptorProto {
+	if m != nil {
+		return m.Extension
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetOptions() *FileOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetSourceCodeInfo() *SourceCodeInfo {
+	if m != nil {
+		return m.SourceCodeInfo
+	}
+	return nil
+}
+
+func (m *FileDescriptorProto) GetSyntax() string {
+	if m != nil && m.Syntax != nil {
+		return *m.Syntax
+	}
+	return ""
+}
+
+// Describes a message type.
+type DescriptorProto struct {
+	Name           *string                           `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Field          []*FieldDescriptorProto           `protobuf:"bytes,2,rep,name=field" json:"field,omitempty"`
+	Extension      []*FieldDescriptorProto           `protobuf:"bytes,6,rep,name=extension" json:"extension,omitempty"`
+	NestedType     []*DescriptorProto                `protobuf:"bytes,3,rep,name=nested_type,json=nestedType" json:"nested_type,omitempty"`
+	EnumType       []*EnumDescriptorProto            `protobuf:"bytes,4,rep,name=enum_type,json=enumType" json:"enum_type,omitempty"`
+	ExtensionRange []*DescriptorProto_ExtensionRange `protobuf:"bytes,5,rep,name=extension_range,json=extensionRange" json:"extension_range,omitempty"`
+	OneofDecl      []*OneofDescriptorProto           `protobuf:"bytes,8,rep,name=oneof_decl,json=oneofDecl" json:"oneof_decl,omitempty"`
+	Options        *MessageOptions                   `protobuf:"bytes,7,opt,name=options" json:"options,omitempty"`
+	ReservedRange  []*DescriptorProto_ReservedRange  `protobuf:"bytes,9,rep,name=reserved_range,json=reservedRange" json:"reserved_range,omitempty"`
+	// Reserved field names, which may not be used by fields in the same message.
+	// A given name may only be reserved once.
+	ReservedName         []string `protobuf:"bytes,10,rep,name=reserved_name,json=reservedName" json:"reserved_name,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *DescriptorProto) Reset()         { *m = DescriptorProto{} }
+func (m *DescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*DescriptorProto) ProtoMessage()    {}
+func (*DescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{2}
+}
+func (m *DescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_DescriptorProto.Unmarshal(m, b)
+}
+func (m *DescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_DescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *DescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_DescriptorProto.Merge(m, src)
+}
+func (m *DescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_DescriptorProto.Size(m)
+}
+func (m *DescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_DescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_DescriptorProto proto.InternalMessageInfo
+
+func (m *DescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *DescriptorProto) GetField() []*FieldDescriptorProto {
+	if m != nil {
+		return m.Field
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetExtension() []*FieldDescriptorProto {
+	if m != nil {
+		return m.Extension
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetNestedType() []*DescriptorProto {
+	if m != nil {
+		return m.NestedType
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetEnumType() []*EnumDescriptorProto {
+	if m != nil {
+		return m.EnumType
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetExtensionRange() []*DescriptorProto_ExtensionRange {
+	if m != nil {
+		return m.ExtensionRange
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetOneofDecl() []*OneofDescriptorProto {
+	if m != nil {
+		return m.OneofDecl
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetOptions() *MessageOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetReservedRange() []*DescriptorProto_ReservedRange {
+	if m != nil {
+		return m.ReservedRange
+	}
+	return nil
+}
+
+func (m *DescriptorProto) GetReservedName() []string {
+	if m != nil {
+		return m.ReservedName
+	}
+	return nil
+}
+
+type DescriptorProto_ExtensionRange struct {
+	Start                *int32                 `protobuf:"varint,1,opt,name=start" json:"start,omitempty"`
+	End                  *int32                 `protobuf:"varint,2,opt,name=end" json:"end,omitempty"`
+	Options              *ExtensionRangeOptions `protobuf:"bytes,3,opt,name=options" json:"options,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}               `json:"-"`
+	XXX_unrecognized     []byte                 `json:"-"`
+	XXX_sizecache        int32                  `json:"-"`
+}
+
+func (m *DescriptorProto_ExtensionRange) Reset()         { *m = DescriptorProto_ExtensionRange{} }
+func (m *DescriptorProto_ExtensionRange) String() string { return proto.CompactTextString(m) }
+func (*DescriptorProto_ExtensionRange) ProtoMessage()    {}
+func (*DescriptorProto_ExtensionRange) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{2, 0}
+}
+func (m *DescriptorProto_ExtensionRange) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_DescriptorProto_ExtensionRange.Unmarshal(m, b)
+}
+func (m *DescriptorProto_ExtensionRange) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_DescriptorProto_ExtensionRange.Marshal(b, m, deterministic)
+}
+func (m *DescriptorProto_ExtensionRange) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_DescriptorProto_ExtensionRange.Merge(m, src)
+}
+func (m *DescriptorProto_ExtensionRange) XXX_Size() int {
+	return xxx_messageInfo_DescriptorProto_ExtensionRange.Size(m)
+}
+func (m *DescriptorProto_ExtensionRange) XXX_DiscardUnknown() {
+	xxx_messageInfo_DescriptorProto_ExtensionRange.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_DescriptorProto_ExtensionRange proto.InternalMessageInfo
+
+func (m *DescriptorProto_ExtensionRange) GetStart() int32 {
+	if m != nil && m.Start != nil {
+		return *m.Start
+	}
+	return 0
+}
+
+func (m *DescriptorProto_ExtensionRange) GetEnd() int32 {
+	if m != nil && m.End != nil {
+		return *m.End
+	}
+	return 0
+}
+
+func (m *DescriptorProto_ExtensionRange) GetOptions() *ExtensionRangeOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+// Range of reserved tag numbers. Reserved tag numbers may not be used by
+// fields or extension ranges in the same message. Reserved ranges may
+// not overlap.
+type DescriptorProto_ReservedRange struct {
+	Start                *int32   `protobuf:"varint,1,opt,name=start" json:"start,omitempty"`
+	End                  *int32   `protobuf:"varint,2,opt,name=end" json:"end,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *DescriptorProto_ReservedRange) Reset()         { *m = DescriptorProto_ReservedRange{} }
+func (m *DescriptorProto_ReservedRange) String() string { return proto.CompactTextString(m) }
+func (*DescriptorProto_ReservedRange) ProtoMessage()    {}
+func (*DescriptorProto_ReservedRange) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{2, 1}
+}
+func (m *DescriptorProto_ReservedRange) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_DescriptorProto_ReservedRange.Unmarshal(m, b)
+}
+func (m *DescriptorProto_ReservedRange) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_DescriptorProto_ReservedRange.Marshal(b, m, deterministic)
+}
+func (m *DescriptorProto_ReservedRange) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_DescriptorProto_ReservedRange.Merge(m, src)
+}
+func (m *DescriptorProto_ReservedRange) XXX_Size() int {
+	return xxx_messageInfo_DescriptorProto_ReservedRange.Size(m)
+}
+func (m *DescriptorProto_ReservedRange) XXX_DiscardUnknown() {
+	xxx_messageInfo_DescriptorProto_ReservedRange.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_DescriptorProto_ReservedRange proto.InternalMessageInfo
+
+func (m *DescriptorProto_ReservedRange) GetStart() int32 {
+	if m != nil && m.Start != nil {
+		return *m.Start
+	}
+	return 0
+}
+
+func (m *DescriptorProto_ReservedRange) GetEnd() int32 {
+	if m != nil && m.End != nil {
+		return *m.End
+	}
+	return 0
+}
+
+type ExtensionRangeOptions struct {
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *ExtensionRangeOptions) Reset()         { *m = ExtensionRangeOptions{} }
+func (m *ExtensionRangeOptions) String() string { return proto.CompactTextString(m) }
+func (*ExtensionRangeOptions) ProtoMessage()    {}
+func (*ExtensionRangeOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{3}
+}
+
+var extRange_ExtensionRangeOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*ExtensionRangeOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_ExtensionRangeOptions
+}
+
+func (m *ExtensionRangeOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_ExtensionRangeOptions.Unmarshal(m, b)
+}
+func (m *ExtensionRangeOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_ExtensionRangeOptions.Marshal(b, m, deterministic)
+}
+func (m *ExtensionRangeOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ExtensionRangeOptions.Merge(m, src)
+}
+func (m *ExtensionRangeOptions) XXX_Size() int {
+	return xxx_messageInfo_ExtensionRangeOptions.Size(m)
+}
+func (m *ExtensionRangeOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_ExtensionRangeOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ExtensionRangeOptions proto.InternalMessageInfo
+
+func (m *ExtensionRangeOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+// Describes a field within a message.
+type FieldDescriptorProto struct {
+	Name   *string                     `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Number *int32                      `protobuf:"varint,3,opt,name=number" json:"number,omitempty"`
+	Label  *FieldDescriptorProto_Label `protobuf:"varint,4,opt,name=label,enum=google.protobuf.FieldDescriptorProto_Label" json:"label,omitempty"`
+	// If type_name is set, this need not be set.  If both this and type_name
+	// are set, this must be one of TYPE_ENUM, TYPE_MESSAGE or TYPE_GROUP.
+	Type *FieldDescriptorProto_Type `protobuf:"varint,5,opt,name=type,enum=google.protobuf.FieldDescriptorProto_Type" json:"type,omitempty"`
+	// For message and enum types, this is the name of the type.  If the name
+	// starts with a '.', it is fully-qualified.  Otherwise, C++-like scoping
+	// rules are used to find the type (i.e. first the nested types within this
+	// message are searched, then within the parent, on up to the root
+	// namespace).
+	TypeName *string `protobuf:"bytes,6,opt,name=type_name,json=typeName" json:"type_name,omitempty"`
+	// For extensions, this is the name of the type being extended.  It is
+	// resolved in the same manner as type_name.
+	Extendee *string `protobuf:"bytes,2,opt,name=extendee" json:"extendee,omitempty"`
+	// For numeric types, contains the original text representation of the value.
+	// For booleans, "true" or "false".
+	// For strings, contains the default text contents (not escaped in any way).
+	// For bytes, contains the C escaped value.  All bytes >= 128 are escaped.
+	// TODO(kenton):  Base-64 encode?
+	DefaultValue *string `protobuf:"bytes,7,opt,name=default_value,json=defaultValue" json:"default_value,omitempty"`
+	// If set, gives the index of a oneof in the containing type's oneof_decl
+	// list.  This field is a member of that oneof.
+	OneofIndex *int32 `protobuf:"varint,9,opt,name=oneof_index,json=oneofIndex" json:"oneof_index,omitempty"`
+	// JSON name of this field. The value is set by protocol compiler. If the
+	// user has set a "json_name" option on this field, that option's value
+	// will be used. Otherwise, it's deduced from the field's name by converting
+	// it to camelCase.
+	JsonName             *string       `protobuf:"bytes,10,opt,name=json_name,json=jsonName" json:"json_name,omitempty"`
+	Options              *FieldOptions `protobuf:"bytes,8,opt,name=options" json:"options,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}      `json:"-"`
+	XXX_unrecognized     []byte        `json:"-"`
+	XXX_sizecache        int32         `json:"-"`
+}
+
+func (m *FieldDescriptorProto) Reset()         { *m = FieldDescriptorProto{} }
+func (m *FieldDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*FieldDescriptorProto) ProtoMessage()    {}
+func (*FieldDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{4}
+}
+func (m *FieldDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_FieldDescriptorProto.Unmarshal(m, b)
+}
+func (m *FieldDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_FieldDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *FieldDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_FieldDescriptorProto.Merge(m, src)
+}
+func (m *FieldDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_FieldDescriptorProto.Size(m)
+}
+func (m *FieldDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_FieldDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_FieldDescriptorProto proto.InternalMessageInfo
+
+func (m *FieldDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *FieldDescriptorProto) GetNumber() int32 {
+	if m != nil && m.Number != nil {
+		return *m.Number
+	}
+	return 0
+}
+
+func (m *FieldDescriptorProto) GetLabel() FieldDescriptorProto_Label {
+	if m != nil && m.Label != nil {
+		return *m.Label
+	}
+	return FieldDescriptorProto_LABEL_OPTIONAL
+}
+
+func (m *FieldDescriptorProto) GetType() FieldDescriptorProto_Type {
+	if m != nil && m.Type != nil {
+		return *m.Type
+	}
+	return FieldDescriptorProto_TYPE_DOUBLE
+}
+
+func (m *FieldDescriptorProto) GetTypeName() string {
+	if m != nil && m.TypeName != nil {
+		return *m.TypeName
+	}
+	return ""
+}
+
+func (m *FieldDescriptorProto) GetExtendee() string {
+	if m != nil && m.Extendee != nil {
+		return *m.Extendee
+	}
+	return ""
+}
+
+func (m *FieldDescriptorProto) GetDefaultValue() string {
+	if m != nil && m.DefaultValue != nil {
+		return *m.DefaultValue
+	}
+	return ""
+}
+
+func (m *FieldDescriptorProto) GetOneofIndex() int32 {
+	if m != nil && m.OneofIndex != nil {
+		return *m.OneofIndex
+	}
+	return 0
+}
+
+func (m *FieldDescriptorProto) GetJsonName() string {
+	if m != nil && m.JsonName != nil {
+		return *m.JsonName
+	}
+	return ""
+}
+
+func (m *FieldDescriptorProto) GetOptions() *FieldOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+// Describes a oneof.
+type OneofDescriptorProto struct {
+	Name                 *string       `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Options              *OneofOptions `protobuf:"bytes,2,opt,name=options" json:"options,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}      `json:"-"`
+	XXX_unrecognized     []byte        `json:"-"`
+	XXX_sizecache        int32         `json:"-"`
+}
+
+func (m *OneofDescriptorProto) Reset()         { *m = OneofDescriptorProto{} }
+func (m *OneofDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*OneofDescriptorProto) ProtoMessage()    {}
+func (*OneofDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{5}
+}
+func (m *OneofDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_OneofDescriptorProto.Unmarshal(m, b)
+}
+func (m *OneofDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_OneofDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *OneofDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_OneofDescriptorProto.Merge(m, src)
+}
+func (m *OneofDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_OneofDescriptorProto.Size(m)
+}
+func (m *OneofDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_OneofDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_OneofDescriptorProto proto.InternalMessageInfo
+
+func (m *OneofDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *OneofDescriptorProto) GetOptions() *OneofOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+// Describes an enum type.
+type EnumDescriptorProto struct {
+	Name    *string                     `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Value   []*EnumValueDescriptorProto `protobuf:"bytes,2,rep,name=value" json:"value,omitempty"`
+	Options *EnumOptions                `protobuf:"bytes,3,opt,name=options" json:"options,omitempty"`
+	// Range of reserved numeric values. Reserved numeric values may not be used
+	// by enum values in the same enum declaration. Reserved ranges may not
+	// overlap.
+	ReservedRange []*EnumDescriptorProto_EnumReservedRange `protobuf:"bytes,4,rep,name=reserved_range,json=reservedRange" json:"reserved_range,omitempty"`
+	// Reserved enum value names, which may not be reused. A given name may only
+	// be reserved once.
+	ReservedName         []string `protobuf:"bytes,5,rep,name=reserved_name,json=reservedName" json:"reserved_name,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *EnumDescriptorProto) Reset()         { *m = EnumDescriptorProto{} }
+func (m *EnumDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*EnumDescriptorProto) ProtoMessage()    {}
+func (*EnumDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{6}
+}
+func (m *EnumDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_EnumDescriptorProto.Unmarshal(m, b)
+}
+func (m *EnumDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_EnumDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *EnumDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_EnumDescriptorProto.Merge(m, src)
+}
+func (m *EnumDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_EnumDescriptorProto.Size(m)
+}
+func (m *EnumDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_EnumDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_EnumDescriptorProto proto.InternalMessageInfo
+
+func (m *EnumDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *EnumDescriptorProto) GetValue() []*EnumValueDescriptorProto {
+	if m != nil {
+		return m.Value
+	}
+	return nil
+}
+
+func (m *EnumDescriptorProto) GetOptions() *EnumOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+func (m *EnumDescriptorProto) GetReservedRange() []*EnumDescriptorProto_EnumReservedRange {
+	if m != nil {
+		return m.ReservedRange
+	}
+	return nil
+}
+
+func (m *EnumDescriptorProto) GetReservedName() []string {
+	if m != nil {
+		return m.ReservedName
+	}
+	return nil
+}
+
+// Range of reserved numeric values. Reserved values may not be used by
+// entries in the same enum. Reserved ranges may not overlap.
+//
+// Note that this is distinct from DescriptorProto.ReservedRange in that it
+// is inclusive such that it can appropriately represent the entire int32
+// domain.
+type EnumDescriptorProto_EnumReservedRange struct {
+	Start                *int32   `protobuf:"varint,1,opt,name=start" json:"start,omitempty"`
+	End                  *int32   `protobuf:"varint,2,opt,name=end" json:"end,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *EnumDescriptorProto_EnumReservedRange) Reset()         { *m = EnumDescriptorProto_EnumReservedRange{} }
+func (m *EnumDescriptorProto_EnumReservedRange) String() string { return proto.CompactTextString(m) }
+func (*EnumDescriptorProto_EnumReservedRange) ProtoMessage()    {}
+func (*EnumDescriptorProto_EnumReservedRange) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{6, 0}
+}
+func (m *EnumDescriptorProto_EnumReservedRange) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_EnumDescriptorProto_EnumReservedRange.Unmarshal(m, b)
+}
+func (m *EnumDescriptorProto_EnumReservedRange) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_EnumDescriptorProto_EnumReservedRange.Marshal(b, m, deterministic)
+}
+func (m *EnumDescriptorProto_EnumReservedRange) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_EnumDescriptorProto_EnumReservedRange.Merge(m, src)
+}
+func (m *EnumDescriptorProto_EnumReservedRange) XXX_Size() int {
+	return xxx_messageInfo_EnumDescriptorProto_EnumReservedRange.Size(m)
+}
+func (m *EnumDescriptorProto_EnumReservedRange) XXX_DiscardUnknown() {
+	xxx_messageInfo_EnumDescriptorProto_EnumReservedRange.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_EnumDescriptorProto_EnumReservedRange proto.InternalMessageInfo
+
+func (m *EnumDescriptorProto_EnumReservedRange) GetStart() int32 {
+	if m != nil && m.Start != nil {
+		return *m.Start
+	}
+	return 0
+}
+
+func (m *EnumDescriptorProto_EnumReservedRange) GetEnd() int32 {
+	if m != nil && m.End != nil {
+		return *m.End
+	}
+	return 0
+}
+
+// Describes a value within an enum.
+type EnumValueDescriptorProto struct {
+	Name                 *string           `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Number               *int32            `protobuf:"varint,2,opt,name=number" json:"number,omitempty"`
+	Options              *EnumValueOptions `protobuf:"bytes,3,opt,name=options" json:"options,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}          `json:"-"`
+	XXX_unrecognized     []byte            `json:"-"`
+	XXX_sizecache        int32             `json:"-"`
+}
+
+func (m *EnumValueDescriptorProto) Reset()         { *m = EnumValueDescriptorProto{} }
+func (m *EnumValueDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*EnumValueDescriptorProto) ProtoMessage()    {}
+func (*EnumValueDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{7}
+}
+func (m *EnumValueDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_EnumValueDescriptorProto.Unmarshal(m, b)
+}
+func (m *EnumValueDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_EnumValueDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *EnumValueDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_EnumValueDescriptorProto.Merge(m, src)
+}
+func (m *EnumValueDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_EnumValueDescriptorProto.Size(m)
+}
+func (m *EnumValueDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_EnumValueDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_EnumValueDescriptorProto proto.InternalMessageInfo
+
+func (m *EnumValueDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *EnumValueDescriptorProto) GetNumber() int32 {
+	if m != nil && m.Number != nil {
+		return *m.Number
+	}
+	return 0
+}
+
+func (m *EnumValueDescriptorProto) GetOptions() *EnumValueOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+// Describes a service.
+type ServiceDescriptorProto struct {
+	Name                 *string                  `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	Method               []*MethodDescriptorProto `protobuf:"bytes,2,rep,name=method" json:"method,omitempty"`
+	Options              *ServiceOptions          `protobuf:"bytes,3,opt,name=options" json:"options,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}                 `json:"-"`
+	XXX_unrecognized     []byte                   `json:"-"`
+	XXX_sizecache        int32                    `json:"-"`
+}
+
+func (m *ServiceDescriptorProto) Reset()         { *m = ServiceDescriptorProto{} }
+func (m *ServiceDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*ServiceDescriptorProto) ProtoMessage()    {}
+func (*ServiceDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{8}
+}
+func (m *ServiceDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_ServiceDescriptorProto.Unmarshal(m, b)
+}
+func (m *ServiceDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_ServiceDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *ServiceDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ServiceDescriptorProto.Merge(m, src)
+}
+func (m *ServiceDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_ServiceDescriptorProto.Size(m)
+}
+func (m *ServiceDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_ServiceDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ServiceDescriptorProto proto.InternalMessageInfo
+
+func (m *ServiceDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *ServiceDescriptorProto) GetMethod() []*MethodDescriptorProto {
+	if m != nil {
+		return m.Method
+	}
+	return nil
+}
+
+func (m *ServiceDescriptorProto) GetOptions() *ServiceOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+// Describes a method of a service.
+type MethodDescriptorProto struct {
+	Name *string `protobuf:"bytes,1,opt,name=name" json:"name,omitempty"`
+	// Input and output type names.  These are resolved in the same way as
+	// FieldDescriptorProto.type_name, but must refer to a message type.
+	InputType  *string        `protobuf:"bytes,2,opt,name=input_type,json=inputType" json:"input_type,omitempty"`
+	OutputType *string        `protobuf:"bytes,3,opt,name=output_type,json=outputType" json:"output_type,omitempty"`
+	Options    *MethodOptions `protobuf:"bytes,4,opt,name=options" json:"options,omitempty"`
+	// Identifies if client streams multiple client messages
+	ClientStreaming *bool `protobuf:"varint,5,opt,name=client_streaming,json=clientStreaming,def=0" json:"client_streaming,omitempty"`
+	// Identifies if server streams multiple server messages
+	ServerStreaming      *bool    `protobuf:"varint,6,opt,name=server_streaming,json=serverStreaming,def=0" json:"server_streaming,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *MethodDescriptorProto) Reset()         { *m = MethodDescriptorProto{} }
+func (m *MethodDescriptorProto) String() string { return proto.CompactTextString(m) }
+func (*MethodDescriptorProto) ProtoMessage()    {}
+func (*MethodDescriptorProto) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{9}
+}
+func (m *MethodDescriptorProto) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_MethodDescriptorProto.Unmarshal(m, b)
+}
+func (m *MethodDescriptorProto) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_MethodDescriptorProto.Marshal(b, m, deterministic)
+}
+func (m *MethodDescriptorProto) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_MethodDescriptorProto.Merge(m, src)
+}
+func (m *MethodDescriptorProto) XXX_Size() int {
+	return xxx_messageInfo_MethodDescriptorProto.Size(m)
+}
+func (m *MethodDescriptorProto) XXX_DiscardUnknown() {
+	xxx_messageInfo_MethodDescriptorProto.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_MethodDescriptorProto proto.InternalMessageInfo
+
+const Default_MethodDescriptorProto_ClientStreaming bool = false
+const Default_MethodDescriptorProto_ServerStreaming bool = false
+
+func (m *MethodDescriptorProto) GetName() string {
+	if m != nil && m.Name != nil {
+		return *m.Name
+	}
+	return ""
+}
+
+func (m *MethodDescriptorProto) GetInputType() string {
+	if m != nil && m.InputType != nil {
+		return *m.InputType
+	}
+	return ""
+}
+
+func (m *MethodDescriptorProto) GetOutputType() string {
+	if m != nil && m.OutputType != nil {
+		return *m.OutputType
+	}
+	return ""
+}
+
+func (m *MethodDescriptorProto) GetOptions() *MethodOptions {
+	if m != nil {
+		return m.Options
+	}
+	return nil
+}
+
+func (m *MethodDescriptorProto) GetClientStreaming() bool {
+	if m != nil && m.ClientStreaming != nil {
+		return *m.ClientStreaming
+	}
+	return Default_MethodDescriptorProto_ClientStreaming
+}
+
+func (m *MethodDescriptorProto) GetServerStreaming() bool {
+	if m != nil && m.ServerStreaming != nil {
+		return *m.ServerStreaming
+	}
+	return Default_MethodDescriptorProto_ServerStreaming
+}
+
+type FileOptions struct {
+	// Sets the Java package where classes generated from this .proto will be
+	// placed.  By default, the proto package is used, but this is often
+	// inappropriate because proto packages do not normally start with backwards
+	// domain names.
+	JavaPackage *string `protobuf:"bytes,1,opt,name=java_package,json=javaPackage" json:"java_package,omitempty"`
+	// If set, all the classes from the .proto file are wrapped in a single
+	// outer class with the given name.  This applies to both Proto1
+	// (equivalent to the old "--one_java_file" option) and Proto2 (where
+	// a .proto always translates to a single class, but you may want to
+	// explicitly choose the class name).
+	JavaOuterClassname *string `protobuf:"bytes,8,opt,name=java_outer_classname,json=javaOuterClassname" json:"java_outer_classname,omitempty"`
+	// If set true, then the Java code generator will generate a separate .java
+	// file for each top-level message, enum, and service defined in the .proto
+	// file.  Thus, these types will *not* be nested inside the outer class
+	// named by java_outer_classname.  However, the outer class will still be
+	// generated to contain the file's getDescriptor() method as well as any
+	// top-level extensions defined in the file.
+	JavaMultipleFiles *bool `protobuf:"varint,10,opt,name=java_multiple_files,json=javaMultipleFiles,def=0" json:"java_multiple_files,omitempty"`
+	// This option does nothing.
+	JavaGenerateEqualsAndHash *bool `protobuf:"varint,20,opt,name=java_generate_equals_and_hash,json=javaGenerateEqualsAndHash" json:"java_generate_equals_and_hash,omitempty"` // Deprecated: Do not use.
+	// If set true, then the Java2 code generator will generate code that
+	// throws an exception whenever an attempt is made to assign a non-UTF-8
+	// byte sequence to a string field.
+	// Message reflection will do the same.
+	// However, an extension field still accepts non-UTF-8 byte sequences.
+	// This option has no effect on when used with the lite runtime.
+	JavaStringCheckUtf8 *bool                     `protobuf:"varint,27,opt,name=java_string_check_utf8,json=javaStringCheckUtf8,def=0" json:"java_string_check_utf8,omitempty"`
+	OptimizeFor         *FileOptions_OptimizeMode `protobuf:"varint,9,opt,name=optimize_for,json=optimizeFor,enum=google.protobuf.FileOptions_OptimizeMode,def=1" json:"optimize_for,omitempty"`
+	// Sets the Go package where structs generated from this .proto will be
+	// placed. If omitted, the Go package will be derived from the following:
+	//   - The basename of the package import path, if provided.
+	//   - Otherwise, the package statement in the .proto file, if present.
+	//   - Otherwise, the basename of the .proto file, without extension.
+	GoPackage *string `protobuf:"bytes,11,opt,name=go_package,json=goPackage" json:"go_package,omitempty"`
+	// Should generic services be generated in each language?  "Generic" services
+	// are not specific to any particular RPC system.  They are generated by the
+	// main code generators in each language (without additional plugins).
+	// Generic services were the only kind of service generation supported by
+	// early versions of google.protobuf.
+	//
+	// Generic services are now considered deprecated in favor of using plugins
+	// that generate code specific to your particular RPC system.  Therefore,
+	// these default to false.  Old code which depends on generic services should
+	// explicitly set them to true.
+	CcGenericServices   *bool `protobuf:"varint,16,opt,name=cc_generic_services,json=ccGenericServices,def=0" json:"cc_generic_services,omitempty"`
+	JavaGenericServices *bool `protobuf:"varint,17,opt,name=java_generic_services,json=javaGenericServices,def=0" json:"java_generic_services,omitempty"`
+	PyGenericServices   *bool `protobuf:"varint,18,opt,name=py_generic_services,json=pyGenericServices,def=0" json:"py_generic_services,omitempty"`
+	PhpGenericServices  *bool `protobuf:"varint,42,opt,name=php_generic_services,json=phpGenericServices,def=0" json:"php_generic_services,omitempty"`
+	// Is this file deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for everything in the file, or it will be completely ignored; in the very
+	// least, this is a formalization for deprecating files.
+	Deprecated *bool `protobuf:"varint,23,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// Enables the use of arenas for the proto messages in this file. This applies
+	// only to generated classes for C++.
+	CcEnableArenas *bool `protobuf:"varint,31,opt,name=cc_enable_arenas,json=ccEnableArenas,def=0" json:"cc_enable_arenas,omitempty"`
+	// Sets the objective c class prefix which is prepended to all objective c
+	// generated classes from this .proto. There is no default.
+	ObjcClassPrefix *string `protobuf:"bytes,36,opt,name=objc_class_prefix,json=objcClassPrefix" json:"objc_class_prefix,omitempty"`
+	// Namespace for generated classes; defaults to the package.
+	CsharpNamespace *string `protobuf:"bytes,37,opt,name=csharp_namespace,json=csharpNamespace" json:"csharp_namespace,omitempty"`
+	// By default Swift generators will take the proto package and CamelCase it
+	// replacing '.' with underscore and use that to prefix the types/symbols
+	// defined. When this options is provided, they will use this value instead
+	// to prefix the types/symbols defined.
+	SwiftPrefix *string `protobuf:"bytes,39,opt,name=swift_prefix,json=swiftPrefix" json:"swift_prefix,omitempty"`
+	// Sets the php class prefix which is prepended to all php generated classes
+	// from this .proto. Default is empty.
+	PhpClassPrefix *string `protobuf:"bytes,40,opt,name=php_class_prefix,json=phpClassPrefix" json:"php_class_prefix,omitempty"`
+	// Use this option to change the namespace of php generated classes. Default
+	// is empty. When this option is empty, the package name will be used for
+	// determining the namespace.
+	PhpNamespace *string `protobuf:"bytes,41,opt,name=php_namespace,json=phpNamespace" json:"php_namespace,omitempty"`
+	// Use this option to change the namespace of php generated metadata classes.
+	// Default is empty. When this option is empty, the proto file name will be used
+	// for determining the namespace.
+	PhpMetadataNamespace *string `protobuf:"bytes,44,opt,name=php_metadata_namespace,json=phpMetadataNamespace" json:"php_metadata_namespace,omitempty"`
+	// Use this option to change the package of ruby generated classes. Default
+	// is empty. When this option is not set, the package name will be used for
+	// determining the ruby package.
+	RubyPackage *string `protobuf:"bytes,45,opt,name=ruby_package,json=rubyPackage" json:"ruby_package,omitempty"`
+	// The parser stores options it doesn't recognize here.
+	// See the documentation for the "Options" section above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *FileOptions) Reset()         { *m = FileOptions{} }
+func (m *FileOptions) String() string { return proto.CompactTextString(m) }
+func (*FileOptions) ProtoMessage()    {}
+func (*FileOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{10}
+}
+
+var extRange_FileOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*FileOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_FileOptions
+}
+
+func (m *FileOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_FileOptions.Unmarshal(m, b)
+}
+func (m *FileOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_FileOptions.Marshal(b, m, deterministic)
+}
+func (m *FileOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_FileOptions.Merge(m, src)
+}
+func (m *FileOptions) XXX_Size() int {
+	return xxx_messageInfo_FileOptions.Size(m)
+}
+func (m *FileOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_FileOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_FileOptions proto.InternalMessageInfo
+
+const Default_FileOptions_JavaMultipleFiles bool = false
+const Default_FileOptions_JavaStringCheckUtf8 bool = false
+const Default_FileOptions_OptimizeFor FileOptions_OptimizeMode = FileOptions_SPEED
+const Default_FileOptions_CcGenericServices bool = false
+const Default_FileOptions_JavaGenericServices bool = false
+const Default_FileOptions_PyGenericServices bool = false
+const Default_FileOptions_PhpGenericServices bool = false
+const Default_FileOptions_Deprecated bool = false
+const Default_FileOptions_CcEnableArenas bool = false
+
+func (m *FileOptions) GetJavaPackage() string {
+	if m != nil && m.JavaPackage != nil {
+		return *m.JavaPackage
+	}
+	return ""
+}
+
+func (m *FileOptions) GetJavaOuterClassname() string {
+	if m != nil && m.JavaOuterClassname != nil {
+		return *m.JavaOuterClassname
+	}
+	return ""
+}
+
+func (m *FileOptions) GetJavaMultipleFiles() bool {
+	if m != nil && m.JavaMultipleFiles != nil {
+		return *m.JavaMultipleFiles
+	}
+	return Default_FileOptions_JavaMultipleFiles
+}
+
+// Deprecated: Do not use.
+func (m *FileOptions) GetJavaGenerateEqualsAndHash() bool {
+	if m != nil && m.JavaGenerateEqualsAndHash != nil {
+		return *m.JavaGenerateEqualsAndHash
+	}
+	return false
+}
+
+func (m *FileOptions) GetJavaStringCheckUtf8() bool {
+	if m != nil && m.JavaStringCheckUtf8 != nil {
+		return *m.JavaStringCheckUtf8
+	}
+	return Default_FileOptions_JavaStringCheckUtf8
+}
+
+func (m *FileOptions) GetOptimizeFor() FileOptions_OptimizeMode {
+	if m != nil && m.OptimizeFor != nil {
+		return *m.OptimizeFor
+	}
+	return Default_FileOptions_OptimizeFor
+}
+
+func (m *FileOptions) GetGoPackage() string {
+	if m != nil && m.GoPackage != nil {
+		return *m.GoPackage
+	}
+	return ""
+}
+
+func (m *FileOptions) GetCcGenericServices() bool {
+	if m != nil && m.CcGenericServices != nil {
+		return *m.CcGenericServices
+	}
+	return Default_FileOptions_CcGenericServices
+}
+
+func (m *FileOptions) GetJavaGenericServices() bool {
+	if m != nil && m.JavaGenericServices != nil {
+		return *m.JavaGenericServices
+	}
+	return Default_FileOptions_JavaGenericServices
+}
+
+func (m *FileOptions) GetPyGenericServices() bool {
+	if m != nil && m.PyGenericServices != nil {
+		return *m.PyGenericServices
+	}
+	return Default_FileOptions_PyGenericServices
+}
+
+func (m *FileOptions) GetPhpGenericServices() bool {
+	if m != nil && m.PhpGenericServices != nil {
+		return *m.PhpGenericServices
+	}
+	return Default_FileOptions_PhpGenericServices
+}
+
+func (m *FileOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_FileOptions_Deprecated
+}
+
+func (m *FileOptions) GetCcEnableArenas() bool {
+	if m != nil && m.CcEnableArenas != nil {
+		return *m.CcEnableArenas
+	}
+	return Default_FileOptions_CcEnableArenas
+}
+
+func (m *FileOptions) GetObjcClassPrefix() string {
+	if m != nil && m.ObjcClassPrefix != nil {
+		return *m.ObjcClassPrefix
+	}
+	return ""
+}
+
+func (m *FileOptions) GetCsharpNamespace() string {
+	if m != nil && m.CsharpNamespace != nil {
+		return *m.CsharpNamespace
+	}
+	return ""
+}
+
+func (m *FileOptions) GetSwiftPrefix() string {
+	if m != nil && m.SwiftPrefix != nil {
+		return *m.SwiftPrefix
+	}
+	return ""
+}
+
+func (m *FileOptions) GetPhpClassPrefix() string {
+	if m != nil && m.PhpClassPrefix != nil {
+		return *m.PhpClassPrefix
+	}
+	return ""
+}
+
+func (m *FileOptions) GetPhpNamespace() string {
+	if m != nil && m.PhpNamespace != nil {
+		return *m.PhpNamespace
+	}
+	return ""
+}
+
+func (m *FileOptions) GetPhpMetadataNamespace() string {
+	if m != nil && m.PhpMetadataNamespace != nil {
+		return *m.PhpMetadataNamespace
+	}
+	return ""
+}
+
+func (m *FileOptions) GetRubyPackage() string {
+	if m != nil && m.RubyPackage != nil {
+		return *m.RubyPackage
+	}
+	return ""
+}
+
+func (m *FileOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type MessageOptions struct {
+	// Set true to use the old proto1 MessageSet wire format for extensions.
+	// This is provided for backwards-compatibility with the MessageSet wire
+	// format.  You should not use this for any other reason:  It's less
+	// efficient, has fewer features, and is more complicated.
+	//
+	// The message must be defined exactly as follows:
+	//   message Foo {
+	//     option message_set_wire_format = true;
+	//     extensions 4 to max;
+	//   }
+	// Note that the message cannot have any defined fields; MessageSets only
+	// have extensions.
+	//
+	// All extensions of your type must be singular messages; e.g. they cannot
+	// be int32s, enums, or repeated messages.
+	//
+	// Because this is an option, the above two restrictions are not enforced by
+	// the protocol compiler.
+	MessageSetWireFormat *bool `protobuf:"varint,1,opt,name=message_set_wire_format,json=messageSetWireFormat,def=0" json:"message_set_wire_format,omitempty"`
+	// Disables the generation of the standard "descriptor()" accessor, which can
+	// conflict with a field of the same name.  This is meant to make migration
+	// from proto1 easier; new code should avoid fields named "descriptor".
+	NoStandardDescriptorAccessor *bool `protobuf:"varint,2,opt,name=no_standard_descriptor_accessor,json=noStandardDescriptorAccessor,def=0" json:"no_standard_descriptor_accessor,omitempty"`
+	// Is this message deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for the message, or it will be completely ignored; in the very least,
+	// this is a formalization for deprecating messages.
+	Deprecated *bool `protobuf:"varint,3,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// Whether the message is an automatically generated map entry type for the
+	// maps field.
+	//
+	// For maps fields:
+	//     map<KeyType, ValueType> map_field = 1;
+	// The parsed descriptor looks like:
+	//     message MapFieldEntry {
+	//         option map_entry = true;
+	//         optional KeyType key = 1;
+	//         optional ValueType value = 2;
+	//     }
+	//     repeated MapFieldEntry map_field = 1;
+	//
+	// Implementations may choose not to generate the map_entry=true message, but
+	// use a native map in the target language to hold the keys and values.
+	// The reflection APIs in such implementions still need to work as
+	// if the field is a repeated message field.
+	//
+	// NOTE: Do not set the option in .proto files. Always use the maps syntax
+	// instead. The option should only be implicitly set by the proto compiler
+	// parser.
+	MapEntry *bool `protobuf:"varint,7,opt,name=map_entry,json=mapEntry" json:"map_entry,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *MessageOptions) Reset()         { *m = MessageOptions{} }
+func (m *MessageOptions) String() string { return proto.CompactTextString(m) }
+func (*MessageOptions) ProtoMessage()    {}
+func (*MessageOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{11}
+}
+
+var extRange_MessageOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*MessageOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_MessageOptions
+}
+
+func (m *MessageOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_MessageOptions.Unmarshal(m, b)
+}
+func (m *MessageOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_MessageOptions.Marshal(b, m, deterministic)
+}
+func (m *MessageOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_MessageOptions.Merge(m, src)
+}
+func (m *MessageOptions) XXX_Size() int {
+	return xxx_messageInfo_MessageOptions.Size(m)
+}
+func (m *MessageOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_MessageOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_MessageOptions proto.InternalMessageInfo
+
+const Default_MessageOptions_MessageSetWireFormat bool = false
+const Default_MessageOptions_NoStandardDescriptorAccessor bool = false
+const Default_MessageOptions_Deprecated bool = false
+
+func (m *MessageOptions) GetMessageSetWireFormat() bool {
+	if m != nil && m.MessageSetWireFormat != nil {
+		return *m.MessageSetWireFormat
+	}
+	return Default_MessageOptions_MessageSetWireFormat
+}
+
+func (m *MessageOptions) GetNoStandardDescriptorAccessor() bool {
+	if m != nil && m.NoStandardDescriptorAccessor != nil {
+		return *m.NoStandardDescriptorAccessor
+	}
+	return Default_MessageOptions_NoStandardDescriptorAccessor
+}
+
+func (m *MessageOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_MessageOptions_Deprecated
+}
+
+func (m *MessageOptions) GetMapEntry() bool {
+	if m != nil && m.MapEntry != nil {
+		return *m.MapEntry
+	}
+	return false
+}
+
+func (m *MessageOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type FieldOptions struct {
+	// The ctype option instructs the C++ code generator to use a different
+	// representation of the field than it normally would.  See the specific
+	// options below.  This option is not yet implemented in the open source
+	// release -- sorry, we'll try to include it in a future version!
+	Ctype *FieldOptions_CType `protobuf:"varint,1,opt,name=ctype,enum=google.protobuf.FieldOptions_CType,def=0" json:"ctype,omitempty"`
+	// The packed option can be enabled for repeated primitive fields to enable
+	// a more efficient representation on the wire. Rather than repeatedly
+	// writing the tag and type for each element, the entire array is encoded as
+	// a single length-delimited blob. In proto3, only explicit setting it to
+	// false will avoid using packed encoding.
+	Packed *bool `protobuf:"varint,2,opt,name=packed" json:"packed,omitempty"`
+	// The jstype option determines the JavaScript type used for values of the
+	// field.  The option is permitted only for 64 bit integral and fixed types
+	// (int64, uint64, sint64, fixed64, sfixed64).  A field with jstype JS_STRING
+	// is represented as JavaScript string, which avoids loss of precision that
+	// can happen when a large value is converted to a floating point JavaScript.
+	// Specifying JS_NUMBER for the jstype causes the generated JavaScript code to
+	// use the JavaScript "number" type.  The behavior of the default option
+	// JS_NORMAL is implementation dependent.
+	//
+	// This option is an enum to permit additional types to be added, e.g.
+	// goog.math.Integer.
+	Jstype *FieldOptions_JSType `protobuf:"varint,6,opt,name=jstype,enum=google.protobuf.FieldOptions_JSType,def=0" json:"jstype,omitempty"`
+	// Should this field be parsed lazily?  Lazy applies only to message-type
+	// fields.  It means that when the outer message is initially parsed, the
+	// inner message's contents will not be parsed but instead stored in encoded
+	// form.  The inner message will actually be parsed when it is first accessed.
+	//
+	// This is only a hint.  Implementations are free to choose whether to use
+	// eager or lazy parsing regardless of the value of this option.  However,
+	// setting this option true suggests that the protocol author believes that
+	// using lazy parsing on this field is worth the additional bookkeeping
+	// overhead typically needed to implement it.
+	//
+	// This option does not affect the public interface of any generated code;
+	// all method signatures remain the same.  Furthermore, thread-safety of the
+	// interface is not affected by this option; const methods remain safe to
+	// call from multiple threads concurrently, while non-const methods continue
+	// to require exclusive access.
+	//
+	//
+	// Note that implementations may choose not to check required fields within
+	// a lazy sub-message.  That is, calling IsInitialized() on the outer message
+	// may return true even if the inner message has missing required fields.
+	// This is necessary because otherwise the inner message would have to be
+	// parsed in order to perform the check, defeating the purpose of lazy
+	// parsing.  An implementation which chooses not to check required fields
+	// must be consistent about it.  That is, for any particular sub-message, the
+	// implementation must either *always* check its required fields, or *never*
+	// check its required fields, regardless of whether or not the message has
+	// been parsed.
+	Lazy *bool `protobuf:"varint,5,opt,name=lazy,def=0" json:"lazy,omitempty"`
+	// Is this field deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for accessors, or it will be completely ignored; in the very least, this
+	// is a formalization for deprecating fields.
+	Deprecated *bool `protobuf:"varint,3,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// For Google-internal migration only. Do not use.
+	Weak *bool `protobuf:"varint,10,opt,name=weak,def=0" json:"weak,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *FieldOptions) Reset()         { *m = FieldOptions{} }
+func (m *FieldOptions) String() string { return proto.CompactTextString(m) }
+func (*FieldOptions) ProtoMessage()    {}
+func (*FieldOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{12}
+}
+
+var extRange_FieldOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*FieldOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_FieldOptions
+}
+
+func (m *FieldOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_FieldOptions.Unmarshal(m, b)
+}
+func (m *FieldOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_FieldOptions.Marshal(b, m, deterministic)
+}
+func (m *FieldOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_FieldOptions.Merge(m, src)
+}
+func (m *FieldOptions) XXX_Size() int {
+	return xxx_messageInfo_FieldOptions.Size(m)
+}
+func (m *FieldOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_FieldOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_FieldOptions proto.InternalMessageInfo
+
+const Default_FieldOptions_Ctype FieldOptions_CType = FieldOptions_STRING
+const Default_FieldOptions_Jstype FieldOptions_JSType = FieldOptions_JS_NORMAL
+const Default_FieldOptions_Lazy bool = false
+const Default_FieldOptions_Deprecated bool = false
+const Default_FieldOptions_Weak bool = false
+
+func (m *FieldOptions) GetCtype() FieldOptions_CType {
+	if m != nil && m.Ctype != nil {
+		return *m.Ctype
+	}
+	return Default_FieldOptions_Ctype
+}
+
+func (m *FieldOptions) GetPacked() bool {
+	if m != nil && m.Packed != nil {
+		return *m.Packed
+	}
+	return false
+}
+
+func (m *FieldOptions) GetJstype() FieldOptions_JSType {
+	if m != nil && m.Jstype != nil {
+		return *m.Jstype
+	}
+	return Default_FieldOptions_Jstype
+}
+
+func (m *FieldOptions) GetLazy() bool {
+	if m != nil && m.Lazy != nil {
+		return *m.Lazy
+	}
+	return Default_FieldOptions_Lazy
+}
+
+func (m *FieldOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_FieldOptions_Deprecated
+}
+
+func (m *FieldOptions) GetWeak() bool {
+	if m != nil && m.Weak != nil {
+		return *m.Weak
+	}
+	return Default_FieldOptions_Weak
+}
+
+func (m *FieldOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type OneofOptions struct {
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *OneofOptions) Reset()         { *m = OneofOptions{} }
+func (m *OneofOptions) String() string { return proto.CompactTextString(m) }
+func (*OneofOptions) ProtoMessage()    {}
+func (*OneofOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{13}
+}
+
+var extRange_OneofOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*OneofOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_OneofOptions
+}
+
+func (m *OneofOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_OneofOptions.Unmarshal(m, b)
+}
+func (m *OneofOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_OneofOptions.Marshal(b, m, deterministic)
+}
+func (m *OneofOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_OneofOptions.Merge(m, src)
+}
+func (m *OneofOptions) XXX_Size() int {
+	return xxx_messageInfo_OneofOptions.Size(m)
+}
+func (m *OneofOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_OneofOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_OneofOptions proto.InternalMessageInfo
+
+func (m *OneofOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type EnumOptions struct {
+	// Set this option to true to allow mapping different tag names to the same
+	// value.
+	AllowAlias *bool `protobuf:"varint,2,opt,name=allow_alias,json=allowAlias" json:"allow_alias,omitempty"`
+	// Is this enum deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for the enum, or it will be completely ignored; in the very least, this
+	// is a formalization for deprecating enums.
+	Deprecated *bool `protobuf:"varint,3,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *EnumOptions) Reset()         { *m = EnumOptions{} }
+func (m *EnumOptions) String() string { return proto.CompactTextString(m) }
+func (*EnumOptions) ProtoMessage()    {}
+func (*EnumOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{14}
+}
+
+var extRange_EnumOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*EnumOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_EnumOptions
+}
+
+func (m *EnumOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_EnumOptions.Unmarshal(m, b)
+}
+func (m *EnumOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_EnumOptions.Marshal(b, m, deterministic)
+}
+func (m *EnumOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_EnumOptions.Merge(m, src)
+}
+func (m *EnumOptions) XXX_Size() int {
+	return xxx_messageInfo_EnumOptions.Size(m)
+}
+func (m *EnumOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_EnumOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_EnumOptions proto.InternalMessageInfo
+
+const Default_EnumOptions_Deprecated bool = false
+
+func (m *EnumOptions) GetAllowAlias() bool {
+	if m != nil && m.AllowAlias != nil {
+		return *m.AllowAlias
+	}
+	return false
+}
+
+func (m *EnumOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_EnumOptions_Deprecated
+}
+
+func (m *EnumOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type EnumValueOptions struct {
+	// Is this enum value deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for the enum value, or it will be completely ignored; in the very least,
+	// this is a formalization for deprecating enum values.
+	Deprecated *bool `protobuf:"varint,1,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *EnumValueOptions) Reset()         { *m = EnumValueOptions{} }
+func (m *EnumValueOptions) String() string { return proto.CompactTextString(m) }
+func (*EnumValueOptions) ProtoMessage()    {}
+func (*EnumValueOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{15}
+}
+
+var extRange_EnumValueOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*EnumValueOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_EnumValueOptions
+}
+
+func (m *EnumValueOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_EnumValueOptions.Unmarshal(m, b)
+}
+func (m *EnumValueOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_EnumValueOptions.Marshal(b, m, deterministic)
+}
+func (m *EnumValueOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_EnumValueOptions.Merge(m, src)
+}
+func (m *EnumValueOptions) XXX_Size() int {
+	return xxx_messageInfo_EnumValueOptions.Size(m)
+}
+func (m *EnumValueOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_EnumValueOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_EnumValueOptions proto.InternalMessageInfo
+
+const Default_EnumValueOptions_Deprecated bool = false
+
+func (m *EnumValueOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_EnumValueOptions_Deprecated
+}
+
+func (m *EnumValueOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type ServiceOptions struct {
+	// Is this service deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for the service, or it will be completely ignored; in the very least,
+	// this is a formalization for deprecating services.
+	Deprecated *bool `protobuf:"varint,33,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *ServiceOptions) Reset()         { *m = ServiceOptions{} }
+func (m *ServiceOptions) String() string { return proto.CompactTextString(m) }
+func (*ServiceOptions) ProtoMessage()    {}
+func (*ServiceOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{16}
+}
+
+var extRange_ServiceOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*ServiceOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_ServiceOptions
+}
+
+func (m *ServiceOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_ServiceOptions.Unmarshal(m, b)
+}
+func (m *ServiceOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_ServiceOptions.Marshal(b, m, deterministic)
+}
+func (m *ServiceOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ServiceOptions.Merge(m, src)
+}
+func (m *ServiceOptions) XXX_Size() int {
+	return xxx_messageInfo_ServiceOptions.Size(m)
+}
+func (m *ServiceOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_ServiceOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ServiceOptions proto.InternalMessageInfo
+
+const Default_ServiceOptions_Deprecated bool = false
+
+func (m *ServiceOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_ServiceOptions_Deprecated
+}
+
+func (m *ServiceOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+type MethodOptions struct {
+	// Is this method deprecated?
+	// Depending on the target platform, this can emit Deprecated annotations
+	// for the method, or it will be completely ignored; in the very least,
+	// this is a formalization for deprecating methods.
+	Deprecated       *bool                           `protobuf:"varint,33,opt,name=deprecated,def=0" json:"deprecated,omitempty"`
+	IdempotencyLevel *MethodOptions_IdempotencyLevel `protobuf:"varint,34,opt,name=idempotency_level,json=idempotencyLevel,enum=google.protobuf.MethodOptions_IdempotencyLevel,def=0" json:"idempotency_level,omitempty"`
+	// The parser stores options it doesn't recognize here. See above.
+	UninterpretedOption          []*UninterpretedOption `protobuf:"bytes,999,rep,name=uninterpreted_option,json=uninterpretedOption" json:"uninterpreted_option,omitempty"`
+	XXX_NoUnkeyedLiteral         struct{}               `json:"-"`
+	proto.XXX_InternalExtensions `json:"-"`
+	XXX_unrecognized             []byte `json:"-"`
+	XXX_sizecache                int32  `json:"-"`
+}
+
+func (m *MethodOptions) Reset()         { *m = MethodOptions{} }
+func (m *MethodOptions) String() string { return proto.CompactTextString(m) }
+func (*MethodOptions) ProtoMessage()    {}
+func (*MethodOptions) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{17}
+}
+
+var extRange_MethodOptions = []proto.ExtensionRange{
+	{Start: 1000, End: 536870911},
+}
+
+func (*MethodOptions) ExtensionRangeArray() []proto.ExtensionRange {
+	return extRange_MethodOptions
+}
+
+func (m *MethodOptions) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_MethodOptions.Unmarshal(m, b)
+}
+func (m *MethodOptions) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_MethodOptions.Marshal(b, m, deterministic)
+}
+func (m *MethodOptions) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_MethodOptions.Merge(m, src)
+}
+func (m *MethodOptions) XXX_Size() int {
+	return xxx_messageInfo_MethodOptions.Size(m)
+}
+func (m *MethodOptions) XXX_DiscardUnknown() {
+	xxx_messageInfo_MethodOptions.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_MethodOptions proto.InternalMessageInfo
+
+const Default_MethodOptions_Deprecated bool = false
+const Default_MethodOptions_IdempotencyLevel MethodOptions_IdempotencyLevel = MethodOptions_IDEMPOTENCY_UNKNOWN
+
+func (m *MethodOptions) GetDeprecated() bool {
+	if m != nil && m.Deprecated != nil {
+		return *m.Deprecated
+	}
+	return Default_MethodOptions_Deprecated
+}
+
+func (m *MethodOptions) GetIdempotencyLevel() MethodOptions_IdempotencyLevel {
+	if m != nil && m.IdempotencyLevel != nil {
+		return *m.IdempotencyLevel
+	}
+	return Default_MethodOptions_IdempotencyLevel
+}
+
+func (m *MethodOptions) GetUninterpretedOption() []*UninterpretedOption {
+	if m != nil {
+		return m.UninterpretedOption
+	}
+	return nil
+}
+
+// A message representing a option the parser does not recognize. This only
+// appears in options protos created by the compiler::Parser class.
+// DescriptorPool resolves these when building Descriptor objects. Therefore,
+// options protos in descriptor objects (e.g. returned by Descriptor::options(),
+// or produced by Descriptor::CopyTo()) will never have UninterpretedOptions
+// in them.
+type UninterpretedOption struct {
+	Name []*UninterpretedOption_NamePart `protobuf:"bytes,2,rep,name=name" json:"name,omitempty"`
+	// The value of the uninterpreted option, in whatever type the tokenizer
+	// identified it as during parsing. Exactly one of these should be set.
+	IdentifierValue      *string  `protobuf:"bytes,3,opt,name=identifier_value,json=identifierValue" json:"identifier_value,omitempty"`
+	PositiveIntValue     *uint64  `protobuf:"varint,4,opt,name=positive_int_value,json=positiveIntValue" json:"positive_int_value,omitempty"`
+	NegativeIntValue     *int64   `protobuf:"varint,5,opt,name=negative_int_value,json=negativeIntValue" json:"negative_int_value,omitempty"`
+	DoubleValue          *float64 `protobuf:"fixed64,6,opt,name=double_value,json=doubleValue" json:"double_value,omitempty"`
+	StringValue          []byte   `protobuf:"bytes,7,opt,name=string_value,json=stringValue" json:"string_value,omitempty"`
+	AggregateValue       *string  `protobuf:"bytes,8,opt,name=aggregate_value,json=aggregateValue" json:"aggregate_value,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *UninterpretedOption) Reset()         { *m = UninterpretedOption{} }
+func (m *UninterpretedOption) String() string { return proto.CompactTextString(m) }
+func (*UninterpretedOption) ProtoMessage()    {}
+func (*UninterpretedOption) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{18}
+}
+func (m *UninterpretedOption) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_UninterpretedOption.Unmarshal(m, b)
+}
+func (m *UninterpretedOption) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_UninterpretedOption.Marshal(b, m, deterministic)
+}
+func (m *UninterpretedOption) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_UninterpretedOption.Merge(m, src)
+}
+func (m *UninterpretedOption) XXX_Size() int {
+	return xxx_messageInfo_UninterpretedOption.Size(m)
+}
+func (m *UninterpretedOption) XXX_DiscardUnknown() {
+	xxx_messageInfo_UninterpretedOption.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_UninterpretedOption proto.InternalMessageInfo
+
+func (m *UninterpretedOption) GetName() []*UninterpretedOption_NamePart {
+	if m != nil {
+		return m.Name
+	}
+	return nil
+}
+
+func (m *UninterpretedOption) GetIdentifierValue() string {
+	if m != nil && m.IdentifierValue != nil {
+		return *m.IdentifierValue
+	}
+	return ""
+}
+
+func (m *UninterpretedOption) GetPositiveIntValue() uint64 {
+	if m != nil && m.PositiveIntValue != nil {
+		return *m.PositiveIntValue
+	}
+	return 0
+}
+
+func (m *UninterpretedOption) GetNegativeIntValue() int64 {
+	if m != nil && m.NegativeIntValue != nil {
+		return *m.NegativeIntValue
+	}
+	return 0
+}
+
+func (m *UninterpretedOption) GetDoubleValue() float64 {
+	if m != nil && m.DoubleValue != nil {
+		return *m.DoubleValue
+	}
+	return 0
+}
+
+func (m *UninterpretedOption) GetStringValue() []byte {
+	if m != nil {
+		return m.StringValue
+	}
+	return nil
+}
+
+func (m *UninterpretedOption) GetAggregateValue() string {
+	if m != nil && m.AggregateValue != nil {
+		return *m.AggregateValue
+	}
+	return ""
+}
+
+// The name of the uninterpreted option.  Each string represents a segment in
+// a dot-separated name.  is_extension is true iff a segment represents an
+// extension (denoted with parentheses in options specs in .proto files).
+// E.g.,{ ["foo", false], ["bar.baz", true], ["qux", false] } represents
+// "foo.(bar.baz).qux".
+type UninterpretedOption_NamePart struct {
+	NamePart             *string  `protobuf:"bytes,1,req,name=name_part,json=namePart" json:"name_part,omitempty"`
+	IsExtension          *bool    `protobuf:"varint,2,req,name=is_extension,json=isExtension" json:"is_extension,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *UninterpretedOption_NamePart) Reset()         { *m = UninterpretedOption_NamePart{} }
+func (m *UninterpretedOption_NamePart) String() string { return proto.CompactTextString(m) }
+func (*UninterpretedOption_NamePart) ProtoMessage()    {}
+func (*UninterpretedOption_NamePart) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{18, 0}
+}
+func (m *UninterpretedOption_NamePart) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_UninterpretedOption_NamePart.Unmarshal(m, b)
+}
+func (m *UninterpretedOption_NamePart) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_UninterpretedOption_NamePart.Marshal(b, m, deterministic)
+}
+func (m *UninterpretedOption_NamePart) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_UninterpretedOption_NamePart.Merge(m, src)
+}
+func (m *UninterpretedOption_NamePart) XXX_Size() int {
+	return xxx_messageInfo_UninterpretedOption_NamePart.Size(m)
+}
+func (m *UninterpretedOption_NamePart) XXX_DiscardUnknown() {
+	xxx_messageInfo_UninterpretedOption_NamePart.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_UninterpretedOption_NamePart proto.InternalMessageInfo
+
+func (m *UninterpretedOption_NamePart) GetNamePart() string {
+	if m != nil && m.NamePart != nil {
+		return *m.NamePart
+	}
+	return ""
+}
+
+func (m *UninterpretedOption_NamePart) GetIsExtension() bool {
+	if m != nil && m.IsExtension != nil {
+		return *m.IsExtension
+	}
+	return false
+}
+
+// Encapsulates information about the original source file from which a
+// FileDescriptorProto was generated.
+type SourceCodeInfo struct {
+	// A Location identifies a piece of source code in a .proto file which
+	// corresponds to a particular definition.  This information is intended
+	// to be useful to IDEs, code indexers, documentation generators, and similar
+	// tools.
+	//
+	// For example, say we have a file like:
+	//   message Foo {
+	//     optional string foo = 1;
+	//   }
+	// Let's look at just the field definition:
+	//   optional string foo = 1;
+	//   ^       ^^     ^^  ^  ^^^
+	//   a       bc     de  f  ghi
+	// We have the following locations:
+	//   span   path               represents
+	//   [a,i)  [ 4, 0, 2, 0 ]     The whole field definition.
+	//   [a,b)  [ 4, 0, 2, 0, 4 ]  The label (optional).
+	//   [c,d)  [ 4, 0, 2, 0, 5 ]  The type (string).
+	//   [e,f)  [ 4, 0, 2, 0, 1 ]  The name (foo).
+	//   [g,h)  [ 4, 0, 2, 0, 3 ]  The number (1).
+	//
+	// Notes:
+	// - A location may refer to a repeated field itself (i.e. not to any
+	//   particular index within it).  This is used whenever a set of elements are
+	//   logically enclosed in a single code segment.  For example, an entire
+	//   extend block (possibly containing multiple extension definitions) will
+	//   have an outer location whose path refers to the "extensions" repeated
+	//   field without an index.
+	// - Multiple locations may have the same path.  This happens when a single
+	//   logical declaration is spread out across multiple places.  The most
+	//   obvious example is the "extend" block again -- there may be multiple
+	//   extend blocks in the same scope, each of which will have the same path.
+	// - A location's span is not always a subset of its parent's span.  For
+	//   example, the "extendee" of an extension declaration appears at the
+	//   beginning of the "extend" block and is shared by all extensions within
+	//   the block.
+	// - Just because a location's span is a subset of some other location's span
+	//   does not mean that it is a descendent.  For example, a "group" defines
+	//   both a type and a field in a single declaration.  Thus, the locations
+	//   corresponding to the type and field and their components will overlap.
+	// - Code which tries to interpret locations should probably be designed to
+	//   ignore those that it doesn't understand, as more types of locations could
+	//   be recorded in the future.
+	Location             []*SourceCodeInfo_Location `protobuf:"bytes,1,rep,name=location" json:"location,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}                   `json:"-"`
+	XXX_unrecognized     []byte                     `json:"-"`
+	XXX_sizecache        int32                      `json:"-"`
+}
+
+func (m *SourceCodeInfo) Reset()         { *m = SourceCodeInfo{} }
+func (m *SourceCodeInfo) String() string { return proto.CompactTextString(m) }
+func (*SourceCodeInfo) ProtoMessage()    {}
+func (*SourceCodeInfo) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{19}
+}
+func (m *SourceCodeInfo) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_SourceCodeInfo.Unmarshal(m, b)
+}
+func (m *SourceCodeInfo) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_SourceCodeInfo.Marshal(b, m, deterministic)
+}
+func (m *SourceCodeInfo) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_SourceCodeInfo.Merge(m, src)
+}
+func (m *SourceCodeInfo) XXX_Size() int {
+	return xxx_messageInfo_SourceCodeInfo.Size(m)
+}
+func (m *SourceCodeInfo) XXX_DiscardUnknown() {
+	xxx_messageInfo_SourceCodeInfo.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_SourceCodeInfo proto.InternalMessageInfo
+
+func (m *SourceCodeInfo) GetLocation() []*SourceCodeInfo_Location {
+	if m != nil {
+		return m.Location
+	}
+	return nil
+}
+
+type SourceCodeInfo_Location struct {
+	// Identifies which part of the FileDescriptorProto was defined at this
+	// location.
+	//
+	// Each element is a field number or an index.  They form a path from
+	// the root FileDescriptorProto to the place where the definition.  For
+	// example, this path:
+	//   [ 4, 3, 2, 7, 1 ]
+	// refers to:
+	//   file.message_type(3)  // 4, 3
+	//       .field(7)         // 2, 7
+	//       .name()           // 1
+	// This is because FileDescriptorProto.message_type has field number 4:
+	//   repeated DescriptorProto message_type = 4;
+	// and DescriptorProto.field has field number 2:
+	//   repeated FieldDescriptorProto field = 2;
+	// and FieldDescriptorProto.name has field number 1:
+	//   optional string name = 1;
+	//
+	// Thus, the above path gives the location of a field name.  If we removed
+	// the last element:
+	//   [ 4, 3, 2, 7 ]
+	// this path refers to the whole field declaration (from the beginning
+	// of the label to the terminating semicolon).
+	Path []int32 `protobuf:"varint,1,rep,packed,name=path" json:"path,omitempty"`
+	// Always has exactly three or four elements: start line, start column,
+	// end line (optional, otherwise assumed same as start line), end column.
+	// These are packed into a single field for efficiency.  Note that line
+	// and column numbers are zero-based -- typically you will want to add
+	// 1 to each before displaying to a user.
+	Span []int32 `protobuf:"varint,2,rep,packed,name=span" json:"span,omitempty"`
+	// If this SourceCodeInfo represents a complete declaration, these are any
+	// comments appearing before and after the declaration which appear to be
+	// attached to the declaration.
+	//
+	// A series of line comments appearing on consecutive lines, with no other
+	// tokens appearing on those lines, will be treated as a single comment.
+	//
+	// leading_detached_comments will keep paragraphs of comments that appear
+	// before (but not connected to) the current element. Each paragraph,
+	// separated by empty lines, will be one comment element in the repeated
+	// field.
+	//
+	// Only the comment content is provided; comment markers (e.g. //) are
+	// stripped out.  For block comments, leading whitespace and an asterisk
+	// will be stripped from the beginning of each line other than the first.
+	// Newlines are included in the output.
+	//
+	// Examples:
+	//
+	//   optional int32 foo = 1;  // Comment attached to foo.
+	//   // Comment attached to bar.
+	//   optional int32 bar = 2;
+	//
+	//   optional string baz = 3;
+	//   // Comment attached to baz.
+	//   // Another line attached to baz.
+	//
+	//   // Comment attached to qux.
+	//   //
+	//   // Another line attached to qux.
+	//   optional double qux = 4;
+	//
+	//   // Detached comment for corge. This is not leading or trailing comments
+	//   // to qux or corge because there are blank lines separating it from
+	//   // both.
+	//
+	//   // Detached comment for corge paragraph 2.
+	//
+	//   optional string corge = 5;
+	//   /* Block comment attached
+	//    * to corge.  Leading asterisks
+	//    * will be removed. */
+	//   /* Block comment attached to
+	//    * grault. */
+	//   optional int32 grault = 6;
+	//
+	//   // ignored detached comments.
+	LeadingComments         *string  `protobuf:"bytes,3,opt,name=leading_comments,json=leadingComments" json:"leading_comments,omitempty"`
+	TrailingComments        *string  `protobuf:"bytes,4,opt,name=trailing_comments,json=trailingComments" json:"trailing_comments,omitempty"`
+	LeadingDetachedComments []string `protobuf:"bytes,6,rep,name=leading_detached_comments,json=leadingDetachedComments" json:"leading_detached_comments,omitempty"`
+	XXX_NoUnkeyedLiteral    struct{} `json:"-"`
+	XXX_unrecognized        []byte   `json:"-"`
+	XXX_sizecache           int32    `json:"-"`
+}
+
+func (m *SourceCodeInfo_Location) Reset()         { *m = SourceCodeInfo_Location{} }
+func (m *SourceCodeInfo_Location) String() string { return proto.CompactTextString(m) }
+func (*SourceCodeInfo_Location) ProtoMessage()    {}
+func (*SourceCodeInfo_Location) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{19, 0}
+}
+func (m *SourceCodeInfo_Location) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_SourceCodeInfo_Location.Unmarshal(m, b)
+}
+func (m *SourceCodeInfo_Location) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_SourceCodeInfo_Location.Marshal(b, m, deterministic)
+}
+func (m *SourceCodeInfo_Location) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_SourceCodeInfo_Location.Merge(m, src)
+}
+func (m *SourceCodeInfo_Location) XXX_Size() int {
+	return xxx_messageInfo_SourceCodeInfo_Location.Size(m)
+}
+func (m *SourceCodeInfo_Location) XXX_DiscardUnknown() {
+	xxx_messageInfo_SourceCodeInfo_Location.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_SourceCodeInfo_Location proto.InternalMessageInfo
+
+func (m *SourceCodeInfo_Location) GetPath() []int32 {
+	if m != nil {
+		return m.Path
+	}
+	return nil
+}
+
+func (m *SourceCodeInfo_Location) GetSpan() []int32 {
+	if m != nil {
+		return m.Span
+	}
+	return nil
+}
+
+func (m *SourceCodeInfo_Location) GetLeadingComments() string {
+	if m != nil && m.LeadingComments != nil {
+		return *m.LeadingComments
+	}
+	return ""
+}
+
+func (m *SourceCodeInfo_Location) GetTrailingComments() string {
+	if m != nil && m.TrailingComments != nil {
+		return *m.TrailingComments
+	}
+	return ""
+}
+
+func (m *SourceCodeInfo_Location) GetLeadingDetachedComments() []string {
+	if m != nil {
+		return m.LeadingDetachedComments
+	}
+	return nil
+}
+
+// Describes the relationship between generated code and its original source
+// file. A GeneratedCodeInfo message is associated with only one generated
+// source file, but may contain references to different source .proto files.
+type GeneratedCodeInfo struct {
+	// An Annotation connects some span of text in generated code to an element
+	// of its generating .proto file.
+	Annotation           []*GeneratedCodeInfo_Annotation `protobuf:"bytes,1,rep,name=annotation" json:"annotation,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}                        `json:"-"`
+	XXX_unrecognized     []byte                          `json:"-"`
+	XXX_sizecache        int32                           `json:"-"`
+}
+
+func (m *GeneratedCodeInfo) Reset()         { *m = GeneratedCodeInfo{} }
+func (m *GeneratedCodeInfo) String() string { return proto.CompactTextString(m) }
+func (*GeneratedCodeInfo) ProtoMessage()    {}
+func (*GeneratedCodeInfo) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{20}
+}
+func (m *GeneratedCodeInfo) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_GeneratedCodeInfo.Unmarshal(m, b)
+}
+func (m *GeneratedCodeInfo) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_GeneratedCodeInfo.Marshal(b, m, deterministic)
+}
+func (m *GeneratedCodeInfo) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_GeneratedCodeInfo.Merge(m, src)
+}
+func (m *GeneratedCodeInfo) XXX_Size() int {
+	return xxx_messageInfo_GeneratedCodeInfo.Size(m)
+}
+func (m *GeneratedCodeInfo) XXX_DiscardUnknown() {
+	xxx_messageInfo_GeneratedCodeInfo.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_GeneratedCodeInfo proto.InternalMessageInfo
+
+func (m *GeneratedCodeInfo) GetAnnotation() []*GeneratedCodeInfo_Annotation {
+	if m != nil {
+		return m.Annotation
+	}
+	return nil
+}
+
+type GeneratedCodeInfo_Annotation struct {
+	// Identifies the element in the original source .proto file. This field
+	// is formatted the same as SourceCodeInfo.Location.path.
+	Path []int32 `protobuf:"varint,1,rep,packed,name=path" json:"path,omitempty"`
+	// Identifies the filesystem path to the original source .proto.
+	SourceFile *string `protobuf:"bytes,2,opt,name=source_file,json=sourceFile" json:"source_file,omitempty"`
+	// Identifies the starting offset in bytes in the generated code
+	// that relates to the identified object.
+	Begin *int32 `protobuf:"varint,3,opt,name=begin" json:"begin,omitempty"`
+	// Identifies the ending offset in bytes in the generated code that
+	// relates to the identified offset. The end offset should be one past
+	// the last relevant byte (so the length of the text = end - begin).
+	End                  *int32   `protobuf:"varint,4,opt,name=end" json:"end,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *GeneratedCodeInfo_Annotation) Reset()         { *m = GeneratedCodeInfo_Annotation{} }
+func (m *GeneratedCodeInfo_Annotation) String() string { return proto.CompactTextString(m) }
+func (*GeneratedCodeInfo_Annotation) ProtoMessage()    {}
+func (*GeneratedCodeInfo_Annotation) Descriptor() ([]byte, []int) {
+	return fileDescriptor_308767df5ffe18af, []int{20, 0}
+}
+func (m *GeneratedCodeInfo_Annotation) XXX_Unmarshal(b []byte) error {
+	return xxx_messageInfo_GeneratedCodeInfo_Annotation.Unmarshal(m, b)
+}
+func (m *GeneratedCodeInfo_Annotation) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	return xxx_messageInfo_GeneratedCodeInfo_Annotation.Marshal(b, m, deterministic)
+}
+func (m *GeneratedCodeInfo_Annotation) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_GeneratedCodeInfo_Annotation.Merge(m, src)
+}
+func (m *GeneratedCodeInfo_Annotation) XXX_Size() int {
+	return xxx_messageInfo_GeneratedCodeInfo_Annotation.Size(m)
+}
+func (m *GeneratedCodeInfo_Annotation) XXX_DiscardUnknown() {
+	xxx_messageInfo_GeneratedCodeInfo_Annotation.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_GeneratedCodeInfo_Annotation proto.InternalMessageInfo
+
+func (m *GeneratedCodeInfo_Annotation) GetPath() []int32 {
+	if m != nil {
+		return m.Path
+	}
+	return nil
+}
+
+func (m *GeneratedCodeInfo_Annotation) GetSourceFile() string {
+	if m != nil && m.SourceFile != nil {
+		return *m.SourceFile
+	}
+	return ""
+}
+
+func (m *GeneratedCodeInfo_Annotation) GetBegin() int32 {
+	if m != nil && m.Begin != nil {
+		return *m.Begin
+	}
+	return 0
+}
+
+func (m *GeneratedCodeInfo_Annotation) GetEnd() int32 {
+	if m != nil && m.End != nil {
+		return *m.End
+	}
+	return 0
+}
+
+func init() {
+	proto.RegisterEnum("google.protobuf.FieldDescriptorProto_Type", FieldDescriptorProto_Type_name, FieldDescriptorProto_Type_value)
+	proto.RegisterEnum("google.protobuf.FieldDescriptorProto_Label", FieldDescriptorProto_Label_name, FieldDescriptorProto_Label_value)
+	proto.RegisterEnum("google.protobuf.FileOptions_OptimizeMode", FileOptions_OptimizeMode_name, FileOptions_OptimizeMode_value)
+	proto.RegisterEnum("google.protobuf.FieldOptions_CType", FieldOptions_CType_name, FieldOptions_CType_value)
+	proto.RegisterEnum("google.protobuf.FieldOptions_JSType", FieldOptions_JSType_name, FieldOptions_JSType_value)
+	proto.RegisterEnum("google.protobuf.MethodOptions_IdempotencyLevel", MethodOptions_IdempotencyLevel_name, MethodOptions_IdempotencyLevel_value)
+	proto.RegisterType((*FileDescriptorSet)(nil), "google.protobuf.FileDescriptorSet")
+	proto.RegisterType((*FileDescriptorProto)(nil), "google.protobuf.FileDescriptorProto")
+	proto.RegisterType((*DescriptorProto)(nil), "google.protobuf.DescriptorProto")
+	proto.RegisterType((*DescriptorProto_ExtensionRange)(nil), "google.protobuf.DescriptorProto.ExtensionRange")
+	proto.RegisterType((*DescriptorProto_ReservedRange)(nil), "google.protobuf.DescriptorProto.ReservedRange")
+	proto.RegisterType((*ExtensionRangeOptions)(nil), "google.protobuf.ExtensionRangeOptions")
+	proto.RegisterType((*FieldDescriptorProto)(nil), "google.protobuf.FieldDescriptorProto")
+	proto.RegisterType((*OneofDescriptorProto)(nil), "google.protobuf.OneofDescriptorProto")
+	proto.RegisterType((*EnumDescriptorProto)(nil), "google.protobuf.EnumDescriptorProto")
+	proto.RegisterType((*EnumDescriptorProto_EnumReservedRange)(nil), "google.protobuf.EnumDescriptorProto.EnumReservedRange")
+	proto.RegisterType((*EnumValueDescriptorProto)(nil), "google.protobuf.EnumValueDescriptorProto")
+	proto.RegisterType((*ServiceDescriptorProto)(nil), "google.protobuf.ServiceDescriptorProto")
+	proto.RegisterType((*MethodDescriptorProto)(nil), "google.protobuf.MethodDescriptorProto")
+	proto.RegisterType((*FileOptions)(nil), "google.protobuf.FileOptions")
+	proto.RegisterType((*MessageOptions)(nil), "google.protobuf.MessageOptions")
+	proto.RegisterType((*FieldOptions)(nil), "google.protobuf.FieldOptions")
+	proto.RegisterType((*OneofOptions)(nil), "google.protobuf.OneofOptions")
+	proto.RegisterType((*EnumOptions)(nil), "google.protobuf.EnumOptions")
+	proto.RegisterType((*EnumValueOptions)(nil), "google.protobuf.EnumValueOptions")
+	proto.RegisterType((*ServiceOptions)(nil), "google.protobuf.ServiceOptions")
+	proto.RegisterType((*MethodOptions)(nil), "google.protobuf.MethodOptions")
+	proto.RegisterType((*UninterpretedOption)(nil), "google.protobuf.UninterpretedOption")
+	proto.RegisterType((*UninterpretedOption_NamePart)(nil), "google.protobuf.UninterpretedOption.NamePart")
+	proto.RegisterType((*SourceCodeInfo)(nil), "google.protobuf.SourceCodeInfo")
+	proto.RegisterType((*SourceCodeInfo_Location)(nil), "google.protobuf.SourceCodeInfo.Location")
+	proto.RegisterType((*GeneratedCodeInfo)(nil), "google.protobuf.GeneratedCodeInfo")
+	proto.RegisterType((*GeneratedCodeInfo_Annotation)(nil), "google.protobuf.GeneratedCodeInfo.Annotation")
+}
+
+func init() { proto.RegisterFile("descriptor.proto", fileDescriptor_308767df5ffe18af) }
+
+var fileDescriptor_308767df5ffe18af = []byte{
+	// 2522 bytes of a gzipped FileDescriptorProto
+	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0xc4, 0x59, 0xcd, 0x6f, 0xdb, 0xc8,
+	0x15, 0x5f, 0x7d, 0x5a, 0x7a, 0x92, 0x65, 0x7a, 0xec, 0x75, 0x18, 0xef, 0x47, 0x1c, 0xed, 0x66,
+	0xe3, 0x24, 0xbb, 0xca, 0xc2, 0x49, 0x9c, 0xac, 0x53, 0x6c, 0x2b, 0x4b, 0x8c, 0x57, 0xa9, 0xbe,
+	0x4a, 0xc9, 0xdd, 0x64, 0x8b, 0x82, 0x18, 0x93, 0x23, 0x89, 0x09, 0x45, 0x72, 0x49, 0x2a, 0x89,
+	0x83, 0x1e, 0x02, 0xf4, 0x54, 0xa0, 0x7f, 0x40, 0x51, 0x14, 0x3d, 0xf4, 0xb2, 0x40, 0xff, 0x80,
+	0x02, 0xed, 0xbd, 0xd7, 0x02, 0xbd, 0xf7, 0x50, 0xa0, 0x05, 0xda, 0x3f, 0xa1, 0xc7, 0x62, 0x66,
+	0x48, 0x8a, 0xd4, 0x47, 0xe2, 0x5d, 0x20, 0xd9, 0x93, 0x3d, 0xef, 0xfd, 0xde, 0x9b, 0x37, 0x8f,
+	0xbf, 0x79, 0xf3, 0x66, 0x04, 0x82, 0x46, 0x5c, 0xd5, 0xd1, 0x6d, 0xcf, 0x72, 0x2a, 0xb6, 0x63,
+	0x79, 0x16, 0x5a, 0x1b, 0x5a, 0xd6, 0xd0, 0x20, 0x7c, 0x74, 0x32, 0x19, 0x94, 0x5b, 0xb0, 0x7e,
+	0x4f, 0x37, 0x48, 0x3d, 0x04, 0xf6, 0x88, 0x87, 0xee, 0x40, 0x7a, 0xa0, 0x1b, 0x44, 0x4c, 0xec,
+	0xa4, 0x76, 0x0b, 0x7b, 0x1f, 0x56, 0x66, 0x8c, 0x2a, 0x71, 0x8b, 0x2e, 0x15, 0xcb, 0xcc, 0xa2,
+	0xfc, 0xef, 0x34, 0x6c, 0x2c, 0xd0, 0x22, 0x04, 0x69, 0x13, 0x8f, 0xa9, 0xc7, 0xc4, 0x6e, 0x5e,
+	0x66, 0xff, 0x23, 0x11, 0x56, 0x6c, 0xac, 0x3e, 0xc6, 0x43, 0x22, 0x26, 0x99, 0x38, 0x18, 0xa2,
+	0xf7, 0x01, 0x34, 0x62, 0x13, 0x53, 0x23, 0xa6, 0x7a, 0x2a, 0xa6, 0x76, 0x52, 0xbb, 0x79, 0x39,
+	0x22, 0x41, 0xd7, 0x60, 0xdd, 0x9e, 0x9c, 0x18, 0xba, 0xaa, 0x44, 0x60, 0xb0, 0x93, 0xda, 0xcd,
+	0xc8, 0x02, 0x57, 0xd4, 0xa7, 0xe0, 0xcb, 0xb0, 0xf6, 0x94, 0xe0, 0xc7, 0x51, 0x68, 0x81, 0x41,
+	0x4b, 0x54, 0x1c, 0x01, 0xd6, 0xa0, 0x38, 0x26, 0xae, 0x8b, 0x87, 0x44, 0xf1, 0x4e, 0x6d, 0x22,
+	0xa6, 0xd9, 0xea, 0x77, 0xe6, 0x56, 0x3f, 0xbb, 0xf2, 0x82, 0x6f, 0xd5, 0x3f, 0xb5, 0x09, 0xaa,
+	0x42, 0x9e, 0x98, 0x93, 0x31, 0xf7, 0x90, 0x59, 0x92, 0x3f, 0xc9, 0x9c, 0x8c, 0x67, 0xbd, 0xe4,
+	0xa8, 0x99, 0xef, 0x62, 0xc5, 0x25, 0xce, 0x13, 0x5d, 0x25, 0x62, 0x96, 0x39, 0xb8, 0x3c, 0xe7,
+	0xa0, 0xc7, 0xf5, 0xb3, 0x3e, 0x02, 0x3b, 0x54, 0x83, 0x3c, 0x79, 0xe6, 0x11, 0xd3, 0xd5, 0x2d,
+	0x53, 0x5c, 0x61, 0x4e, 0x2e, 0x2d, 0xf8, 0x8a, 0xc4, 0xd0, 0x66, 0x5d, 0x4c, 0xed, 0xd0, 0x3e,
+	0xac, 0x58, 0xb6, 0xa7, 0x5b, 0xa6, 0x2b, 0xe6, 0x76, 0x12, 0xbb, 0x85, 0xbd, 0x77, 0x17, 0x12,
+	0xa1, 0xc3, 0x31, 0x72, 0x00, 0x46, 0x0d, 0x10, 0x5c, 0x6b, 0xe2, 0xa8, 0x44, 0x51, 0x2d, 0x8d,
+	0x28, 0xba, 0x39, 0xb0, 0xc4, 0x3c, 0x73, 0x70, 0x61, 0x7e, 0x21, 0x0c, 0x58, 0xb3, 0x34, 0xd2,
+	0x30, 0x07, 0x96, 0x5c, 0x72, 0x63, 0x63, 0xb4, 0x05, 0x59, 0xf7, 0xd4, 0xf4, 0xf0, 0x33, 0xb1,
+	0xc8, 0x18, 0xe2, 0x8f, 0xca, 0x7f, 0xce, 0xc2, 0xda, 0x59, 0x28, 0x76, 0x17, 0x32, 0x03, 0xba,
+	0x4a, 0x31, 0xf9, 0x6d, 0x72, 0xc0, 0x6d, 0xe2, 0x49, 0xcc, 0x7e, 0xc7, 0x24, 0x56, 0xa1, 0x60,
+	0x12, 0xd7, 0x23, 0x1a, 0x67, 0x44, 0xea, 0x8c, 0x9c, 0x02, 0x6e, 0x34, 0x4f, 0xa9, 0xf4, 0x77,
+	0xa2, 0xd4, 0x03, 0x58, 0x0b, 0x43, 0x52, 0x1c, 0x6c, 0x0e, 0x03, 0x6e, 0x5e, 0x7f, 0x55, 0x24,
+	0x15, 0x29, 0xb0, 0x93, 0xa9, 0x99, 0x5c, 0x22, 0xb1, 0x31, 0xaa, 0x03, 0x58, 0x26, 0xb1, 0x06,
+	0x8a, 0x46, 0x54, 0x43, 0xcc, 0x2d, 0xc9, 0x52, 0x87, 0x42, 0xe6, 0xb2, 0x64, 0x71, 0xa9, 0x6a,
+	0xa0, 0xcf, 0xa6, 0x54, 0x5b, 0x59, 0xc2, 0x94, 0x16, 0xdf, 0x64, 0x73, 0x6c, 0x3b, 0x86, 0x92,
+	0x43, 0x28, 0xef, 0x89, 0xe6, 0xaf, 0x2c, 0xcf, 0x82, 0xa8, 0xbc, 0x72, 0x65, 0xb2, 0x6f, 0xc6,
+	0x17, 0xb6, 0xea, 0x44, 0x87, 0xe8, 0x03, 0x08, 0x05, 0x0a, 0xa3, 0x15, 0xb0, 0x2a, 0x54, 0x0c,
+	0x84, 0x6d, 0x3c, 0x26, 0xdb, 0xcf, 0xa1, 0x14, 0x4f, 0x0f, 0xda, 0x84, 0x8c, 0xeb, 0x61, 0xc7,
+	0x63, 0x2c, 0xcc, 0xc8, 0x7c, 0x80, 0x04, 0x48, 0x11, 0x53, 0x63, 0x55, 0x2e, 0x23, 0xd3, 0x7f,
+	0xd1, 0x8f, 0xa6, 0x0b, 0x4e, 0xb1, 0x05, 0x7f, 0x34, 0xff, 0x45, 0x63, 0x9e, 0x67, 0xd7, 0xbd,
+	0x7d, 0x1b, 0x56, 0x63, 0x0b, 0x38, 0xeb, 0xd4, 0xe5, 0x5f, 0xc0, 0xdb, 0x0b, 0x5d, 0xa3, 0x07,
+	0xb0, 0x39, 0x31, 0x75, 0xd3, 0x23, 0x8e, 0xed, 0x10, 0xca, 0x58, 0x3e, 0x95, 0xf8, 0x9f, 0x95,
+	0x25, 0x9c, 0x3b, 0x8e, 0xa2, 0xb9, 0x17, 0x79, 0x63, 0x32, 0x2f, 0xbc, 0x9a, 0xcf, 0xfd, 0x77,
+	0x45, 0x78, 0xf1, 0xe2, 0xc5, 0x8b, 0x64, 0xf9, 0x37, 0x59, 0xd8, 0x5c, 0xb4, 0x67, 0x16, 0x6e,
+	0xdf, 0x2d, 0xc8, 0x9a, 0x93, 0xf1, 0x09, 0x71, 0x58, 0x92, 0x32, 0xb2, 0x3f, 0x42, 0x55, 0xc8,
+	0x18, 0xf8, 0x84, 0x18, 0x62, 0x7a, 0x27, 0xb1, 0x5b, 0xda, 0xbb, 0x76, 0xa6, 0x5d, 0x59, 0x69,
+	0x52, 0x13, 0x99, 0x5b, 0xa2, 0xcf, 0x21, 0xed, 0x97, 0x68, 0xea, 0xe1, 0xea, 0xd9, 0x3c, 0xd0,
+	0xbd, 0x24, 0x33, 0x3b, 0xf4, 0x0e, 0xe4, 0xe9, 0x5f, 0xce, 0x8d, 0x2c, 0x8b, 0x39, 0x47, 0x05,
+	0x94, 0x17, 0x68, 0x1b, 0x72, 0x6c, 0x9b, 0x68, 0x24, 0x38, 0xda, 0xc2, 0x31, 0x25, 0x96, 0x46,
+	0x06, 0x78, 0x62, 0x78, 0xca, 0x13, 0x6c, 0x4c, 0x08, 0x23, 0x7c, 0x5e, 0x2e, 0xfa, 0xc2, 0x9f,
+	0x52, 0x19, 0xba, 0x00, 0x05, 0xbe, 0xab, 0x74, 0x53, 0x23, 0xcf, 0x58, 0xf5, 0xcc, 0xc8, 0x7c,
+	0xa3, 0x35, 0xa8, 0x84, 0x4e, 0xff, 0xc8, 0xb5, 0xcc, 0x80, 0x9a, 0x6c, 0x0a, 0x2a, 0x60, 0xd3,
+	0xdf, 0x9e, 0x2d, 0xdc, 0xef, 0x2d, 0x5e, 0xde, 0x2c, 0xa7, 0xca, 0x7f, 0x4a, 0x42, 0x9a, 0xd5,
+	0x8b, 0x35, 0x28, 0xf4, 0x1f, 0x76, 0x25, 0xa5, 0xde, 0x39, 0x3e, 0x6c, 0x4a, 0x42, 0x02, 0x95,
+	0x00, 0x98, 0xe0, 0x5e, 0xb3, 0x53, 0xed, 0x0b, 0xc9, 0x70, 0xdc, 0x68, 0xf7, 0xf7, 0x6f, 0x0a,
+	0xa9, 0xd0, 0xe0, 0x98, 0x0b, 0xd2, 0x51, 0xc0, 0x8d, 0x3d, 0x21, 0x83, 0x04, 0x28, 0x72, 0x07,
+	0x8d, 0x07, 0x52, 0x7d, 0xff, 0xa6, 0x90, 0x8d, 0x4b, 0x6e, 0xec, 0x09, 0x2b, 0x68, 0x15, 0xf2,
+	0x4c, 0x72, 0xd8, 0xe9, 0x34, 0x85, 0x5c, 0xe8, 0xb3, 0xd7, 0x97, 0x1b, 0xed, 0x23, 0x21, 0x1f,
+	0xfa, 0x3c, 0x92, 0x3b, 0xc7, 0x5d, 0x01, 0x42, 0x0f, 0x2d, 0xa9, 0xd7, 0xab, 0x1e, 0x49, 0x42,
+	0x21, 0x44, 0x1c, 0x3e, 0xec, 0x4b, 0x3d, 0xa1, 0x18, 0x0b, 0xeb, 0xc6, 0x9e, 0xb0, 0x1a, 0x4e,
+	0x21, 0xb5, 0x8f, 0x5b, 0x42, 0x09, 0xad, 0xc3, 0x2a, 0x9f, 0x22, 0x08, 0x62, 0x6d, 0x46, 0xb4,
+	0x7f, 0x53, 0x10, 0xa6, 0x81, 0x70, 0x2f, 0xeb, 0x31, 0xc1, 0xfe, 0x4d, 0x01, 0x95, 0x6b, 0x90,
+	0x61, 0xec, 0x42, 0x08, 0x4a, 0xcd, 0xea, 0xa1, 0xd4, 0x54, 0x3a, 0xdd, 0x7e, 0xa3, 0xd3, 0xae,
+	0x36, 0x85, 0xc4, 0x54, 0x26, 0x4b, 0x3f, 0x39, 0x6e, 0xc8, 0x52, 0x5d, 0x48, 0x46, 0x65, 0x5d,
+	0xa9, 0xda, 0x97, 0xea, 0x42, 0xaa, 0xac, 0xc2, 0xe6, 0xa2, 0x3a, 0xb9, 0x70, 0x67, 0x44, 0x3e,
+	0x71, 0x72, 0xc9, 0x27, 0x66, 0xbe, 0xe6, 0x3e, 0xf1, 0xbf, 0x92, 0xb0, 0xb1, 0xe0, 0xac, 0x58,
+	0x38, 0xc9, 0x0f, 0x21, 0xc3, 0x29, 0xca, 0x4f, 0xcf, 0x2b, 0x0b, 0x0f, 0x1d, 0x46, 0xd8, 0xb9,
+	0x13, 0x94, 0xd9, 0x45, 0x3b, 0x88, 0xd4, 0x92, 0x0e, 0x82, 0xba, 0x98, 0xab, 0xe9, 0x3f, 0x9f,
+	0xab, 0xe9, 0xfc, 0xd8, 0xdb, 0x3f, 0xcb, 0xb1, 0xc7, 0x64, 0xdf, 0xae, 0xb6, 0x67, 0x16, 0xd4,
+	0xf6, 0xbb, 0xb0, 0x3e, 0xe7, 0xe8, 0xcc, 0x35, 0xf6, 0x97, 0x09, 0x10, 0x97, 0x25, 0xe7, 0x15,
+	0x95, 0x2e, 0x19, 0xab, 0x74, 0x77, 0x67, 0x33, 0x78, 0x71, 0xf9, 0x47, 0x98, 0xfb, 0xd6, 0xdf,
+	0x24, 0x60, 0x6b, 0x71, 0xa7, 0xb8, 0x30, 0x86, 0xcf, 0x21, 0x3b, 0x26, 0xde, 0xc8, 0x0a, 0xba,
+	0xa5, 0x8f, 0x16, 0x9c, 0xc1, 0x54, 0x3d, 0xfb, 0xb1, 0x7d, 0xab, 0xe8, 0x21, 0x9e, 0x5a, 0xd6,
+	0xee, 0xf1, 0x68, 0xe6, 0x22, 0xfd, 0x55, 0x12, 0xde, 0x5e, 0xe8, 0x7c, 0x61, 0xa0, 0xef, 0x01,
+	0xe8, 0xa6, 0x3d, 0xf1, 0x78, 0x47, 0xc4, 0x0b, 0x6c, 0x9e, 0x49, 0x58, 0xf1, 0xa2, 0xc5, 0x73,
+	0xe2, 0x85, 0xfa, 0x14, 0xd3, 0x03, 0x17, 0x31, 0xc0, 0x9d, 0x69, 0xa0, 0x69, 0x16, 0xe8, 0xfb,
+	0x4b, 0x56, 0x3a, 0x47, 0xcc, 0x4f, 0x41, 0x50, 0x0d, 0x9d, 0x98, 0x9e, 0xe2, 0x7a, 0x0e, 0xc1,
+	0x63, 0xdd, 0x1c, 0xb2, 0x13, 0x24, 0x77, 0x90, 0x19, 0x60, 0xc3, 0x25, 0xf2, 0x1a, 0x57, 0xf7,
+	0x02, 0x2d, 0xb5, 0x60, 0x04, 0x72, 0x22, 0x16, 0xd9, 0x98, 0x05, 0x57, 0x87, 0x16, 0xe5, 0x5f,
+	0xe7, 0xa1, 0x10, 0xe9, 0xab, 0xd1, 0x45, 0x28, 0x3e, 0xc2, 0x4f, 0xb0, 0x12, 0xdc, 0x95, 0x78,
+	0x26, 0x0a, 0x54, 0xd6, 0xf5, 0xef, 0x4b, 0x9f, 0xc2, 0x26, 0x83, 0x58, 0x13, 0x8f, 0x38, 0x8a,
+	0x6a, 0x60, 0xd7, 0x65, 0x49, 0xcb, 0x31, 0x28, 0xa2, 0xba, 0x0e, 0x55, 0xd5, 0x02, 0x0d, 0xba,
+	0x05, 0x1b, 0xcc, 0x62, 0x3c, 0x31, 0x3c, 0xdd, 0x36, 0x88, 0x42, 0x6f, 0x6f, 0x2e, 0x3b, 0x49,
+	0xc2, 0xc8, 0xd6, 0x29, 0xa2, 0xe5, 0x03, 0x68, 0x44, 0x2e, 0xaa, 0xc3, 0x7b, 0xcc, 0x6c, 0x48,
+	0x4c, 0xe2, 0x60, 0x8f, 0x28, 0xe4, 0xeb, 0x09, 0x36, 0x5c, 0x05, 0x9b, 0x9a, 0x32, 0xc2, 0xee,
+	0x48, 0xdc, 0xa4, 0x0e, 0x0e, 0x93, 0x62, 0x42, 0x3e, 0x4f, 0x81, 0x47, 0x3e, 0x4e, 0x62, 0xb0,
+	0xaa, 0xa9, 0x7d, 0x81, 0xdd, 0x11, 0x3a, 0x80, 0x2d, 0xe6, 0xc5, 0xf5, 0x1c, 0xdd, 0x1c, 0x2a,
+	0xea, 0x88, 0xa8, 0x8f, 0x95, 0x89, 0x37, 0xb8, 0x23, 0xbe, 0x13, 0x9d, 0x9f, 0x45, 0xd8, 0x63,
+	0x98, 0x1a, 0x85, 0x1c, 0x7b, 0x83, 0x3b, 0xa8, 0x07, 0x45, 0xfa, 0x31, 0xc6, 0xfa, 0x73, 0xa2,
+	0x0c, 0x2c, 0x87, 0x1d, 0x8d, 0xa5, 0x05, 0xa5, 0x29, 0x92, 0xc1, 0x4a, 0xc7, 0x37, 0x68, 0x59,
+	0x1a, 0x39, 0xc8, 0xf4, 0xba, 0x92, 0x54, 0x97, 0x0b, 0x81, 0x97, 0x7b, 0x96, 0x43, 0x09, 0x35,
+	0xb4, 0xc2, 0x04, 0x17, 0x38, 0xa1, 0x86, 0x56, 0x90, 0xde, 0x5b, 0xb0, 0xa1, 0xaa, 0x7c, 0xcd,
+	0xba, 0xaa, 0xf8, 0x77, 0x2c, 0x57, 0x14, 0x62, 0xc9, 0x52, 0xd5, 0x23, 0x0e, 0xf0, 0x39, 0xee,
+	0xa2, 0xcf, 0xe0, 0xed, 0x69, 0xb2, 0xa2, 0x86, 0xeb, 0x73, 0xab, 0x9c, 0x35, 0xbd, 0x05, 0x1b,
+	0xf6, 0xe9, 0xbc, 0x21, 0x8a, 0xcd, 0x68, 0x9f, 0xce, 0x9a, 0xdd, 0x86, 0x4d, 0x7b, 0x64, 0xcf,
+	0xdb, 0x5d, 0x8d, 0xda, 0x21, 0x7b, 0x64, 0xcf, 0x1a, 0x5e, 0x62, 0x17, 0x6e, 0x87, 0xa8, 0xd8,
+	0x23, 0x9a, 0x78, 0x2e, 0x0a, 0x8f, 0x28, 0xd0, 0x75, 0x10, 0x54, 0x55, 0x21, 0x26, 0x3e, 0x31,
+	0x88, 0x82, 0x1d, 0x62, 0x62, 0x57, 0xbc, 0x10, 0x05, 0x97, 0x54, 0x55, 0x62, 0xda, 0x2a, 0x53,
+	0xa2, 0xab, 0xb0, 0x6e, 0x9d, 0x3c, 0x52, 0x39, 0x25, 0x15, 0xdb, 0x21, 0x03, 0xfd, 0x99, 0xf8,
+	0x21, 0xcb, 0xef, 0x1a, 0x55, 0x30, 0x42, 0x76, 0x99, 0x18, 0x5d, 0x01, 0x41, 0x75, 0x47, 0xd8,
+	0xb1, 0x59, 0x4d, 0x76, 0x6d, 0xac, 0x12, 0xf1, 0x12, 0x87, 0x72, 0x79, 0x3b, 0x10, 0xd3, 0x2d,
+	0xe1, 0x3e, 0xd5, 0x07, 0x5e, 0xe0, 0xf1, 0x32, 0xdf, 0x12, 0x4c, 0xe6, 0x7b, 0xdb, 0x05, 0x81,
+	0xa6, 0x22, 0x36, 0xf1, 0x2e, 0x83, 0x95, 0xec, 0x91, 0x1d, 0x9d, 0xf7, 0x03, 0x58, 0xa5, 0xc8,
+	0xe9, 0xa4, 0x57, 0x78, 0x43, 0x66, 0x8f, 0x22, 0x33, 0xde, 0x84, 0x2d, 0x0a, 0x1a, 0x13, 0x0f,
+	0x6b, 0xd8, 0xc3, 0x11, 0xf4, 0xc7, 0x0c, 0x4d, 0xf3, 0xde, 0xf2, 0x95, 0xb1, 0x38, 0x9d, 0xc9,
+	0xc9, 0x69, 0xc8, 0xac, 0x4f, 0x78, 0x9c, 0x54, 0x16, 0x70, 0xeb, 0xb5, 0x35, 0xdd, 0xe5, 0x03,
+	0x28, 0x46, 0x89, 0x8f, 0xf2, 0xc0, 0xa9, 0x2f, 0x24, 0x68, 0x17, 0x54, 0xeb, 0xd4, 0x69, 0xff,
+	0xf2, 0x95, 0x24, 0x24, 0x69, 0x1f, 0xd5, 0x6c, 0xf4, 0x25, 0x45, 0x3e, 0x6e, 0xf7, 0x1b, 0x2d,
+	0x49, 0x48, 0x45, 0x1b, 0xf6, 0xbf, 0x26, 0xa1, 0x14, 0xbf, 0x7b, 0xa1, 0x1f, 0xc0, 0xb9, 0xe0,
+	0xa1, 0xc4, 0x25, 0x9e, 0xf2, 0x54, 0x77, 0xd8, 0x5e, 0x1c, 0x63, 0x7e, 0x2e, 0x86, 0x6c, 0xd8,
+	0xf4, 0x51, 0x3d, 0xe2, 0x7d, 0xa9, 0x3b, 0x74, 0xa7, 0x8d, 0xb1, 0x87, 0x9a, 0x70, 0xc1, 0xb4,
+	0x14, 0xd7, 0xc3, 0xa6, 0x86, 0x1d, 0x4d, 0x99, 0x3e, 0x51, 0x29, 0x58, 0x55, 0x89, 0xeb, 0x5a,
+	0xfc, 0x0c, 0x0c, 0xbd, 0xbc, 0x6b, 0x5a, 0x3d, 0x1f, 0x3c, 0x3d, 0x1c, 0xaa, 0x3e, 0x74, 0x86,
+	0xb9, 0xa9, 0x65, 0xcc, 0x7d, 0x07, 0xf2, 0x63, 0x6c, 0x2b, 0xc4, 0xf4, 0x9c, 0x53, 0xd6, 0x71,
+	0xe7, 0xe4, 0xdc, 0x18, 0xdb, 0x12, 0x1d, 0xbf, 0x99, 0x8b, 0xcf, 0x3f, 0x52, 0x50, 0x8c, 0x76,
+	0xdd, 0xf4, 0x12, 0xa3, 0xb2, 0x03, 0x2a, 0xc1, 0x4a, 0xd8, 0x07, 0x2f, 0xed, 0xd1, 0x2b, 0x35,
+	0x7a, 0x72, 0x1d, 0x64, 0x79, 0x2f, 0x2c, 0x73, 0x4b, 0xda, 0x35, 0x50, 0x6a, 0x11, 0xde, 0x7b,
+	0xe4, 0x64, 0x7f, 0x84, 0x8e, 0x20, 0xfb, 0xc8, 0x65, 0xbe, 0xb3, 0xcc, 0xf7, 0x87, 0x2f, 0xf7,
+	0x7d, 0xbf, 0xc7, 0x9c, 0xe7, 0xef, 0xf7, 0x94, 0x76, 0x47, 0x6e, 0x55, 0x9b, 0xb2, 0x6f, 0x8e,
+	0xce, 0x43, 0xda, 0xc0, 0xcf, 0x4f, 0xe3, 0x67, 0x1c, 0x13, 0x9d, 0x35, 0xf1, 0xe7, 0x21, 0xfd,
+	0x94, 0xe0, 0xc7, 0xf1, 0x93, 0x85, 0x89, 0x5e, 0x23, 0xf5, 0xaf, 0x43, 0x86, 0xe5, 0x0b, 0x01,
+	0xf8, 0x19, 0x13, 0xde, 0x42, 0x39, 0x48, 0xd7, 0x3a, 0x32, 0xa5, 0xbf, 0x00, 0x45, 0x2e, 0x55,
+	0xba, 0x0d, 0xa9, 0x26, 0x09, 0xc9, 0xf2, 0x2d, 0xc8, 0xf2, 0x24, 0xd0, 0xad, 0x11, 0xa6, 0x41,
+	0x78, 0xcb, 0x1f, 0xfa, 0x3e, 0x12, 0x81, 0xf6, 0xb8, 0x75, 0x28, 0xc9, 0x42, 0x32, 0xfa, 0x79,
+	0x5d, 0x28, 0x46, 0x1b, 0xee, 0x37, 0xc3, 0xa9, 0xbf, 0x24, 0xa0, 0x10, 0x69, 0xa0, 0x69, 0xe7,
+	0x83, 0x0d, 0xc3, 0x7a, 0xaa, 0x60, 0x43, 0xc7, 0xae, 0x4f, 0x0a, 0x60, 0xa2, 0x2a, 0x95, 0x9c,
+	0xf5, 0xa3, 0xbd, 0x91, 0xe0, 0x7f, 0x9f, 0x00, 0x61, 0xb6, 0x77, 0x9d, 0x09, 0x30, 0xf1, 0xbd,
+	0x06, 0xf8, 0xbb, 0x04, 0x94, 0xe2, 0x0d, 0xeb, 0x4c, 0x78, 0x17, 0xbf, 0xd7, 0xf0, 0xfe, 0x99,
+	0x84, 0xd5, 0x58, 0x9b, 0x7a, 0xd6, 0xe8, 0xbe, 0x86, 0x75, 0x5d, 0x23, 0x63, 0xdb, 0xf2, 0x88,
+	0xa9, 0x9e, 0x2a, 0x06, 0x79, 0x42, 0x0c, 0xb1, 0xcc, 0x0a, 0xc5, 0xf5, 0x97, 0x37, 0xc2, 0x95,
+	0xc6, 0xd4, 0xae, 0x49, 0xcd, 0x0e, 0x36, 0x1a, 0x75, 0xa9, 0xd5, 0xed, 0xf4, 0xa5, 0x76, 0xed,
+	0xa1, 0x72, 0xdc, 0xfe, 0x71, 0xbb, 0xf3, 0x65, 0x5b, 0x16, 0xf4, 0x19, 0xd8, 0x6b, 0xdc, 0xea,
+	0x5d, 0x10, 0x66, 0x83, 0x42, 0xe7, 0x60, 0x51, 0x58, 0xc2, 0x5b, 0x68, 0x03, 0xd6, 0xda, 0x1d,
+	0xa5, 0xd7, 0xa8, 0x4b, 0x8a, 0x74, 0xef, 0x9e, 0x54, 0xeb, 0xf7, 0xf8, 0xd3, 0x46, 0x88, 0xee,
+	0xc7, 0x37, 0xf5, 0x6f, 0x53, 0xb0, 0xb1, 0x20, 0x12, 0x54, 0xf5, 0x2f, 0x25, 0xfc, 0x9e, 0xf4,
+	0xc9, 0x59, 0xa2, 0xaf, 0xd0, 0xae, 0xa0, 0x8b, 0x1d, 0xcf, 0xbf, 0xc3, 0x5c, 0x01, 0x9a, 0x25,
+	0xd3, 0xd3, 0x07, 0x3a, 0x71, 0xfc, 0x97, 0x20, 0x7e, 0x53, 0x59, 0x9b, 0xca, 0xf9, 0x63, 0xd0,
+	0xc7, 0x80, 0x6c, 0xcb, 0xd5, 0x3d, 0xfd, 0x09, 0x51, 0x74, 0x33, 0x78, 0x36, 0xa2, 0x37, 0x97,
+	0xb4, 0x2c, 0x04, 0x9a, 0x86, 0xe9, 0x85, 0x68, 0x93, 0x0c, 0xf1, 0x0c, 0x9a, 0x16, 0xf0, 0x94,
+	0x2c, 0x04, 0x9a, 0x10, 0x7d, 0x11, 0x8a, 0x9a, 0x35, 0xa1, 0xed, 0x1c, 0xc7, 0xd1, 0xf3, 0x22,
+	0x21, 0x17, 0xb8, 0x2c, 0x84, 0xf8, 0x8d, 0xfa, 0xf4, 0xbd, 0xaa, 0x28, 0x17, 0xb8, 0x8c, 0x43,
+	0x2e, 0xc3, 0x1a, 0x1e, 0x0e, 0x1d, 0xea, 0x3c, 0x70, 0xc4, 0xaf, 0x1e, 0xa5, 0x50, 0xcc, 0x80,
+	0xdb, 0xf7, 0x21, 0x17, 0xe4, 0x81, 0x1e, 0xc9, 0x34, 0x13, 0x8a, 0xcd, 0xef, 0xd3, 0xc9, 0xdd,
+	0xbc, 0x9c, 0x33, 0x03, 0xe5, 0x45, 0x28, 0xea, 0xae, 0x32, 0x7d, 0x7e, 0x4f, 0xee, 0x24, 0x77,
+	0x73, 0x72, 0x41, 0x77, 0xc3, 0xa7, 0xcb, 0xf2, 0x37, 0x49, 0x28, 0xc5, 0x7f, 0x3e, 0x40, 0x75,
+	0xc8, 0x19, 0x96, 0x8a, 0x19, 0xb5, 0xf8, 0x6f, 0x57, 0xbb, 0xaf, 0xf8, 0xc5, 0xa1, 0xd2, 0xf4,
+	0xf1, 0x72, 0x68, 0xb9, 0xfd, 0xb7, 0x04, 0xe4, 0x02, 0x31, 0xda, 0x82, 0xb4, 0x8d, 0xbd, 0x11,
+	0x73, 0x97, 0x39, 0x4c, 0x0a, 0x09, 0x99, 0x8d, 0xa9, 0xdc, 0xb5, 0xb1, 0xc9, 0x28, 0xe0, 0xcb,
+	0xe9, 0x98, 0x7e, 0x57, 0x83, 0x60, 0x8d, 0xdd, 0x6b, 0xac, 0xf1, 0x98, 0x98, 0x9e, 0x1b, 0x7c,
+	0x57, 0x5f, 0x5e, 0xf3, 0xc5, 0xe8, 0x1a, 0xac, 0x7b, 0x0e, 0xd6, 0x8d, 0x18, 0x36, 0xcd, 0xb0,
+	0x42, 0xa0, 0x08, 0xc1, 0x07, 0x70, 0x3e, 0xf0, 0xab, 0x11, 0x0f, 0xab, 0x23, 0xa2, 0x4d, 0x8d,
+	0xb2, 0xec, 0xfd, 0xe2, 0x9c, 0x0f, 0xa8, 0xfb, 0xfa, 0xc0, 0xb6, 0xfc, 0xf7, 0x04, 0xac, 0x07,
+	0x37, 0x31, 0x2d, 0x4c, 0x56, 0x0b, 0x00, 0x9b, 0xa6, 0xe5, 0x45, 0xd3, 0x35, 0x4f, 0xe5, 0x39,
+	0xbb, 0x4a, 0x35, 0x34, 0x92, 0x23, 0x0e, 0xb6, 0xc7, 0x00, 0x53, 0xcd, 0xd2, 0xb4, 0x5d, 0x80,
+	0x82, 0xff, 0xdb, 0x10, 0xfb, 0x81, 0x91, 0xdf, 0xdd, 0x81, 0x8b, 0xe8, 0x95, 0x0d, 0x6d, 0x42,
+	0xe6, 0x84, 0x0c, 0x75, 0xd3, 0x7f, 0xf1, 0xe5, 0x83, 0xe0, 0x85, 0x25, 0x1d, 0xbe, 0xb0, 0x1c,
+	0xfe, 0x0c, 0x36, 0x54, 0x6b, 0x3c, 0x1b, 0xee, 0xa1, 0x30, 0xf3, 0x7e, 0xe0, 0x7e, 0x91, 0xf8,
+	0x0a, 0xa6, 0x2d, 0xe6, 0xff, 0x12, 0x89, 0x3f, 0x24, 0x53, 0x47, 0xdd, 0xc3, 0x3f, 0x26, 0xb7,
+	0x8f, 0xb8, 0x69, 0x37, 0x58, 0xa9, 0x4c, 0x06, 0x06, 0x51, 0x69, 0xf4, 0xff, 0x0f, 0x00, 0x00,
+	0xff, 0xff, 0x88, 0x17, 0xc1, 0xbe, 0x38, 0x1d, 0x00, 0x00,
+}
diff --git a/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor_gostring.gen.go b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor_gostring.gen.go
new file mode 100644
index 0000000..165b211
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/descriptor_gostring.gen.go
@@ -0,0 +1,752 @@
+// Code generated by protoc-gen-gogo. DO NOT EDIT.
+// source: descriptor.proto
+
+package descriptor
+
+import (
+	fmt "fmt"
+	github_com_gogo_protobuf_proto "github.com/gogo/protobuf/proto"
+	proto "github.com/gogo/protobuf/proto"
+	math "math"
+	reflect "reflect"
+	sort "sort"
+	strconv "strconv"
+	strings "strings"
+)
+
+// Reference imports to suppress errors if they are not otherwise used.
+var _ = proto.Marshal
+var _ = fmt.Errorf
+var _ = math.Inf
+
+func (this *FileDescriptorSet) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 5)
+	s = append(s, "&descriptor.FileDescriptorSet{")
+	if this.File != nil {
+		s = append(s, "File: "+fmt.Sprintf("%#v", this.File)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *FileDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 16)
+	s = append(s, "&descriptor.FileDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Package != nil {
+		s = append(s, "Package: "+valueToGoStringDescriptor(this.Package, "string")+",\n")
+	}
+	if this.Dependency != nil {
+		s = append(s, "Dependency: "+fmt.Sprintf("%#v", this.Dependency)+",\n")
+	}
+	if this.PublicDependency != nil {
+		s = append(s, "PublicDependency: "+fmt.Sprintf("%#v", this.PublicDependency)+",\n")
+	}
+	if this.WeakDependency != nil {
+		s = append(s, "WeakDependency: "+fmt.Sprintf("%#v", this.WeakDependency)+",\n")
+	}
+	if this.MessageType != nil {
+		s = append(s, "MessageType: "+fmt.Sprintf("%#v", this.MessageType)+",\n")
+	}
+	if this.EnumType != nil {
+		s = append(s, "EnumType: "+fmt.Sprintf("%#v", this.EnumType)+",\n")
+	}
+	if this.Service != nil {
+		s = append(s, "Service: "+fmt.Sprintf("%#v", this.Service)+",\n")
+	}
+	if this.Extension != nil {
+		s = append(s, "Extension: "+fmt.Sprintf("%#v", this.Extension)+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.SourceCodeInfo != nil {
+		s = append(s, "SourceCodeInfo: "+fmt.Sprintf("%#v", this.SourceCodeInfo)+",\n")
+	}
+	if this.Syntax != nil {
+		s = append(s, "Syntax: "+valueToGoStringDescriptor(this.Syntax, "string")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *DescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 14)
+	s = append(s, "&descriptor.DescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Field != nil {
+		s = append(s, "Field: "+fmt.Sprintf("%#v", this.Field)+",\n")
+	}
+	if this.Extension != nil {
+		s = append(s, "Extension: "+fmt.Sprintf("%#v", this.Extension)+",\n")
+	}
+	if this.NestedType != nil {
+		s = append(s, "NestedType: "+fmt.Sprintf("%#v", this.NestedType)+",\n")
+	}
+	if this.EnumType != nil {
+		s = append(s, "EnumType: "+fmt.Sprintf("%#v", this.EnumType)+",\n")
+	}
+	if this.ExtensionRange != nil {
+		s = append(s, "ExtensionRange: "+fmt.Sprintf("%#v", this.ExtensionRange)+",\n")
+	}
+	if this.OneofDecl != nil {
+		s = append(s, "OneofDecl: "+fmt.Sprintf("%#v", this.OneofDecl)+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.ReservedRange != nil {
+		s = append(s, "ReservedRange: "+fmt.Sprintf("%#v", this.ReservedRange)+",\n")
+	}
+	if this.ReservedName != nil {
+		s = append(s, "ReservedName: "+fmt.Sprintf("%#v", this.ReservedName)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *DescriptorProto_ExtensionRange) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 7)
+	s = append(s, "&descriptor.DescriptorProto_ExtensionRange{")
+	if this.Start != nil {
+		s = append(s, "Start: "+valueToGoStringDescriptor(this.Start, "int32")+",\n")
+	}
+	if this.End != nil {
+		s = append(s, "End: "+valueToGoStringDescriptor(this.End, "int32")+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *DescriptorProto_ReservedRange) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.DescriptorProto_ReservedRange{")
+	if this.Start != nil {
+		s = append(s, "Start: "+valueToGoStringDescriptor(this.Start, "int32")+",\n")
+	}
+	if this.End != nil {
+		s = append(s, "End: "+valueToGoStringDescriptor(this.End, "int32")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *ExtensionRangeOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 5)
+	s = append(s, "&descriptor.ExtensionRangeOptions{")
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *FieldDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 14)
+	s = append(s, "&descriptor.FieldDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Number != nil {
+		s = append(s, "Number: "+valueToGoStringDescriptor(this.Number, "int32")+",\n")
+	}
+	if this.Label != nil {
+		s = append(s, "Label: "+valueToGoStringDescriptor(this.Label, "FieldDescriptorProto_Label")+",\n")
+	}
+	if this.Type != nil {
+		s = append(s, "Type: "+valueToGoStringDescriptor(this.Type, "FieldDescriptorProto_Type")+",\n")
+	}
+	if this.TypeName != nil {
+		s = append(s, "TypeName: "+valueToGoStringDescriptor(this.TypeName, "string")+",\n")
+	}
+	if this.Extendee != nil {
+		s = append(s, "Extendee: "+valueToGoStringDescriptor(this.Extendee, "string")+",\n")
+	}
+	if this.DefaultValue != nil {
+		s = append(s, "DefaultValue: "+valueToGoStringDescriptor(this.DefaultValue, "string")+",\n")
+	}
+	if this.OneofIndex != nil {
+		s = append(s, "OneofIndex: "+valueToGoStringDescriptor(this.OneofIndex, "int32")+",\n")
+	}
+	if this.JsonName != nil {
+		s = append(s, "JsonName: "+valueToGoStringDescriptor(this.JsonName, "string")+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *OneofDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.OneofDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *EnumDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 9)
+	s = append(s, "&descriptor.EnumDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Value != nil {
+		s = append(s, "Value: "+fmt.Sprintf("%#v", this.Value)+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.ReservedRange != nil {
+		s = append(s, "ReservedRange: "+fmt.Sprintf("%#v", this.ReservedRange)+",\n")
+	}
+	if this.ReservedName != nil {
+		s = append(s, "ReservedName: "+fmt.Sprintf("%#v", this.ReservedName)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *EnumDescriptorProto_EnumReservedRange) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.EnumDescriptorProto_EnumReservedRange{")
+	if this.Start != nil {
+		s = append(s, "Start: "+valueToGoStringDescriptor(this.Start, "int32")+",\n")
+	}
+	if this.End != nil {
+		s = append(s, "End: "+valueToGoStringDescriptor(this.End, "int32")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *EnumValueDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 7)
+	s = append(s, "&descriptor.EnumValueDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Number != nil {
+		s = append(s, "Number: "+valueToGoStringDescriptor(this.Number, "int32")+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *ServiceDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 7)
+	s = append(s, "&descriptor.ServiceDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.Method != nil {
+		s = append(s, "Method: "+fmt.Sprintf("%#v", this.Method)+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *MethodDescriptorProto) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 10)
+	s = append(s, "&descriptor.MethodDescriptorProto{")
+	if this.Name != nil {
+		s = append(s, "Name: "+valueToGoStringDescriptor(this.Name, "string")+",\n")
+	}
+	if this.InputType != nil {
+		s = append(s, "InputType: "+valueToGoStringDescriptor(this.InputType, "string")+",\n")
+	}
+	if this.OutputType != nil {
+		s = append(s, "OutputType: "+valueToGoStringDescriptor(this.OutputType, "string")+",\n")
+	}
+	if this.Options != nil {
+		s = append(s, "Options: "+fmt.Sprintf("%#v", this.Options)+",\n")
+	}
+	if this.ClientStreaming != nil {
+		s = append(s, "ClientStreaming: "+valueToGoStringDescriptor(this.ClientStreaming, "bool")+",\n")
+	}
+	if this.ServerStreaming != nil {
+		s = append(s, "ServerStreaming: "+valueToGoStringDescriptor(this.ServerStreaming, "bool")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *FileOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 25)
+	s = append(s, "&descriptor.FileOptions{")
+	if this.JavaPackage != nil {
+		s = append(s, "JavaPackage: "+valueToGoStringDescriptor(this.JavaPackage, "string")+",\n")
+	}
+	if this.JavaOuterClassname != nil {
+		s = append(s, "JavaOuterClassname: "+valueToGoStringDescriptor(this.JavaOuterClassname, "string")+",\n")
+	}
+	if this.JavaMultipleFiles != nil {
+		s = append(s, "JavaMultipleFiles: "+valueToGoStringDescriptor(this.JavaMultipleFiles, "bool")+",\n")
+	}
+	if this.JavaGenerateEqualsAndHash != nil {
+		s = append(s, "JavaGenerateEqualsAndHash: "+valueToGoStringDescriptor(this.JavaGenerateEqualsAndHash, "bool")+",\n")
+	}
+	if this.JavaStringCheckUtf8 != nil {
+		s = append(s, "JavaStringCheckUtf8: "+valueToGoStringDescriptor(this.JavaStringCheckUtf8, "bool")+",\n")
+	}
+	if this.OptimizeFor != nil {
+		s = append(s, "OptimizeFor: "+valueToGoStringDescriptor(this.OptimizeFor, "FileOptions_OptimizeMode")+",\n")
+	}
+	if this.GoPackage != nil {
+		s = append(s, "GoPackage: "+valueToGoStringDescriptor(this.GoPackage, "string")+",\n")
+	}
+	if this.CcGenericServices != nil {
+		s = append(s, "CcGenericServices: "+valueToGoStringDescriptor(this.CcGenericServices, "bool")+",\n")
+	}
+	if this.JavaGenericServices != nil {
+		s = append(s, "JavaGenericServices: "+valueToGoStringDescriptor(this.JavaGenericServices, "bool")+",\n")
+	}
+	if this.PyGenericServices != nil {
+		s = append(s, "PyGenericServices: "+valueToGoStringDescriptor(this.PyGenericServices, "bool")+",\n")
+	}
+	if this.PhpGenericServices != nil {
+		s = append(s, "PhpGenericServices: "+valueToGoStringDescriptor(this.PhpGenericServices, "bool")+",\n")
+	}
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.CcEnableArenas != nil {
+		s = append(s, "CcEnableArenas: "+valueToGoStringDescriptor(this.CcEnableArenas, "bool")+",\n")
+	}
+	if this.ObjcClassPrefix != nil {
+		s = append(s, "ObjcClassPrefix: "+valueToGoStringDescriptor(this.ObjcClassPrefix, "string")+",\n")
+	}
+	if this.CsharpNamespace != nil {
+		s = append(s, "CsharpNamespace: "+valueToGoStringDescriptor(this.CsharpNamespace, "string")+",\n")
+	}
+	if this.SwiftPrefix != nil {
+		s = append(s, "SwiftPrefix: "+valueToGoStringDescriptor(this.SwiftPrefix, "string")+",\n")
+	}
+	if this.PhpClassPrefix != nil {
+		s = append(s, "PhpClassPrefix: "+valueToGoStringDescriptor(this.PhpClassPrefix, "string")+",\n")
+	}
+	if this.PhpNamespace != nil {
+		s = append(s, "PhpNamespace: "+valueToGoStringDescriptor(this.PhpNamespace, "string")+",\n")
+	}
+	if this.PhpMetadataNamespace != nil {
+		s = append(s, "PhpMetadataNamespace: "+valueToGoStringDescriptor(this.PhpMetadataNamespace, "string")+",\n")
+	}
+	if this.RubyPackage != nil {
+		s = append(s, "RubyPackage: "+valueToGoStringDescriptor(this.RubyPackage, "string")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *MessageOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 9)
+	s = append(s, "&descriptor.MessageOptions{")
+	if this.MessageSetWireFormat != nil {
+		s = append(s, "MessageSetWireFormat: "+valueToGoStringDescriptor(this.MessageSetWireFormat, "bool")+",\n")
+	}
+	if this.NoStandardDescriptorAccessor != nil {
+		s = append(s, "NoStandardDescriptorAccessor: "+valueToGoStringDescriptor(this.NoStandardDescriptorAccessor, "bool")+",\n")
+	}
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.MapEntry != nil {
+		s = append(s, "MapEntry: "+valueToGoStringDescriptor(this.MapEntry, "bool")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *FieldOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 11)
+	s = append(s, "&descriptor.FieldOptions{")
+	if this.Ctype != nil {
+		s = append(s, "Ctype: "+valueToGoStringDescriptor(this.Ctype, "FieldOptions_CType")+",\n")
+	}
+	if this.Packed != nil {
+		s = append(s, "Packed: "+valueToGoStringDescriptor(this.Packed, "bool")+",\n")
+	}
+	if this.Jstype != nil {
+		s = append(s, "Jstype: "+valueToGoStringDescriptor(this.Jstype, "FieldOptions_JSType")+",\n")
+	}
+	if this.Lazy != nil {
+		s = append(s, "Lazy: "+valueToGoStringDescriptor(this.Lazy, "bool")+",\n")
+	}
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.Weak != nil {
+		s = append(s, "Weak: "+valueToGoStringDescriptor(this.Weak, "bool")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *OneofOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 5)
+	s = append(s, "&descriptor.OneofOptions{")
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *EnumOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 7)
+	s = append(s, "&descriptor.EnumOptions{")
+	if this.AllowAlias != nil {
+		s = append(s, "AllowAlias: "+valueToGoStringDescriptor(this.AllowAlias, "bool")+",\n")
+	}
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *EnumValueOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.EnumValueOptions{")
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *ServiceOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.ServiceOptions{")
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *MethodOptions) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 7)
+	s = append(s, "&descriptor.MethodOptions{")
+	if this.Deprecated != nil {
+		s = append(s, "Deprecated: "+valueToGoStringDescriptor(this.Deprecated, "bool")+",\n")
+	}
+	if this.IdempotencyLevel != nil {
+		s = append(s, "IdempotencyLevel: "+valueToGoStringDescriptor(this.IdempotencyLevel, "MethodOptions_IdempotencyLevel")+",\n")
+	}
+	if this.UninterpretedOption != nil {
+		s = append(s, "UninterpretedOption: "+fmt.Sprintf("%#v", this.UninterpretedOption)+",\n")
+	}
+	s = append(s, "XXX_InternalExtensions: "+extensionToGoStringDescriptor(this)+",\n")
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *UninterpretedOption) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 11)
+	s = append(s, "&descriptor.UninterpretedOption{")
+	if this.Name != nil {
+		s = append(s, "Name: "+fmt.Sprintf("%#v", this.Name)+",\n")
+	}
+	if this.IdentifierValue != nil {
+		s = append(s, "IdentifierValue: "+valueToGoStringDescriptor(this.IdentifierValue, "string")+",\n")
+	}
+	if this.PositiveIntValue != nil {
+		s = append(s, "PositiveIntValue: "+valueToGoStringDescriptor(this.PositiveIntValue, "uint64")+",\n")
+	}
+	if this.NegativeIntValue != nil {
+		s = append(s, "NegativeIntValue: "+valueToGoStringDescriptor(this.NegativeIntValue, "int64")+",\n")
+	}
+	if this.DoubleValue != nil {
+		s = append(s, "DoubleValue: "+valueToGoStringDescriptor(this.DoubleValue, "float64")+",\n")
+	}
+	if this.StringValue != nil {
+		s = append(s, "StringValue: "+valueToGoStringDescriptor(this.StringValue, "byte")+",\n")
+	}
+	if this.AggregateValue != nil {
+		s = append(s, "AggregateValue: "+valueToGoStringDescriptor(this.AggregateValue, "string")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *UninterpretedOption_NamePart) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 6)
+	s = append(s, "&descriptor.UninterpretedOption_NamePart{")
+	if this.NamePart != nil {
+		s = append(s, "NamePart: "+valueToGoStringDescriptor(this.NamePart, "string")+",\n")
+	}
+	if this.IsExtension != nil {
+		s = append(s, "IsExtension: "+valueToGoStringDescriptor(this.IsExtension, "bool")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *SourceCodeInfo) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 5)
+	s = append(s, "&descriptor.SourceCodeInfo{")
+	if this.Location != nil {
+		s = append(s, "Location: "+fmt.Sprintf("%#v", this.Location)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *SourceCodeInfo_Location) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 9)
+	s = append(s, "&descriptor.SourceCodeInfo_Location{")
+	if this.Path != nil {
+		s = append(s, "Path: "+fmt.Sprintf("%#v", this.Path)+",\n")
+	}
+	if this.Span != nil {
+		s = append(s, "Span: "+fmt.Sprintf("%#v", this.Span)+",\n")
+	}
+	if this.LeadingComments != nil {
+		s = append(s, "LeadingComments: "+valueToGoStringDescriptor(this.LeadingComments, "string")+",\n")
+	}
+	if this.TrailingComments != nil {
+		s = append(s, "TrailingComments: "+valueToGoStringDescriptor(this.TrailingComments, "string")+",\n")
+	}
+	if this.LeadingDetachedComments != nil {
+		s = append(s, "LeadingDetachedComments: "+fmt.Sprintf("%#v", this.LeadingDetachedComments)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *GeneratedCodeInfo) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 5)
+	s = append(s, "&descriptor.GeneratedCodeInfo{")
+	if this.Annotation != nil {
+		s = append(s, "Annotation: "+fmt.Sprintf("%#v", this.Annotation)+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func (this *GeneratedCodeInfo_Annotation) GoString() string {
+	if this == nil {
+		return "nil"
+	}
+	s := make([]string, 0, 8)
+	s = append(s, "&descriptor.GeneratedCodeInfo_Annotation{")
+	if this.Path != nil {
+		s = append(s, "Path: "+fmt.Sprintf("%#v", this.Path)+",\n")
+	}
+	if this.SourceFile != nil {
+		s = append(s, "SourceFile: "+valueToGoStringDescriptor(this.SourceFile, "string")+",\n")
+	}
+	if this.Begin != nil {
+		s = append(s, "Begin: "+valueToGoStringDescriptor(this.Begin, "int32")+",\n")
+	}
+	if this.End != nil {
+		s = append(s, "End: "+valueToGoStringDescriptor(this.End, "int32")+",\n")
+	}
+	if this.XXX_unrecognized != nil {
+		s = append(s, "XXX_unrecognized:"+fmt.Sprintf("%#v", this.XXX_unrecognized)+",\n")
+	}
+	s = append(s, "}")
+	return strings.Join(s, "")
+}
+func valueToGoStringDescriptor(v interface{}, typ string) string {
+	rv := reflect.ValueOf(v)
+	if rv.IsNil() {
+		return "nil"
+	}
+	pv := reflect.Indirect(rv).Interface()
+	return fmt.Sprintf("func(v %v) *%v { return &v } ( %#v )", typ, typ, pv)
+}
+func extensionToGoStringDescriptor(m github_com_gogo_protobuf_proto.Message) string {
+	e := github_com_gogo_protobuf_proto.GetUnsafeExtensionsMap(m)
+	if e == nil {
+		return "nil"
+	}
+	s := "proto.NewUnsafeXXX_InternalExtensions(map[int32]proto.Extension{"
+	keys := make([]int, 0, len(e))
+	for k := range e {
+		keys = append(keys, int(k))
+	}
+	sort.Ints(keys)
+	ss := []string{}
+	for _, k := range keys {
+		ss = append(ss, strconv.Itoa(k)+": "+e[int32(k)].GoString())
+	}
+	s += strings.Join(ss, ",") + "})"
+	return s
+}
diff --git a/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/helper.go b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/helper.go
new file mode 100644
index 0000000..e0846a3
--- /dev/null
+++ b/vendor/github.com/gogo/protobuf/protoc-gen-gogo/descriptor/helper.go
@@ -0,0 +1,390 @@
+// Protocol Buffers for Go with Gadgets
+//
+// Copyright (c) 2013, The GoGo Authors. All rights reserved.
+// http://github.com/gogo/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package descriptor
+
+import (
+	"strings"
+)
+
+func (msg *DescriptorProto) GetMapFields() (*FieldDescriptorProto, *FieldDescriptorProto) {
+	if !msg.GetOptions().GetMapEntry() {
+		return nil, nil
+	}
+	return msg.GetField()[0], msg.GetField()[1]
+}
+
+func dotToUnderscore(r rune) rune {
+	if r == '.' {
+		return '_'
+	}
+	return r
+}
+
+func (field *FieldDescriptorProto) WireType() (wire int) {
+	switch *field.Type {
+	case FieldDescriptorProto_TYPE_DOUBLE:
+		return 1
+	case FieldDescriptorProto_TYPE_FLOAT:
+		return 5
+	case FieldDescriptorProto_TYPE_INT64:
+		return 0
+	case FieldDescriptorProto_TYPE_UINT64:
+		return 0
+	case FieldDescriptorProto_TYPE_INT32:
+		return 0
+	case FieldDescriptorProto_TYPE_UINT32:
+		return 0
+	case FieldDescriptorProto_TYPE_FIXED64:
+		return 1
+	case FieldDescriptorProto_TYPE_FIXED32:
+		return 5
+	case FieldDescriptorProto_TYPE_BOOL:
+		return 0
+	case FieldDescriptorProto_TYPE_STRING:
+		return 2
+	case FieldDescriptorProto_TYPE_GROUP:
+		return 2
+	case FieldDescriptorProto_TYPE_MESSAGE:
+		return 2
+	case FieldDescriptorProto_TYPE_BYTES:
+		return 2
+	case FieldDescriptorProto_TYPE_ENUM:
+		return 0
+	case FieldDescriptorProto_TYPE_SFIXED32:
+		return 5
+	case FieldDescriptorProto_TYPE_SFIXED64:
+		return 1
+	case FieldDescriptorProto_TYPE_SINT32:
+		return 0
+	case FieldDescriptorProto_TYPE_SINT64:
+		return 0
+	}
+	panic("unreachable")
+}
+
+func (field *FieldDescriptorProto) GetKeyUint64() (x uint64) {
+	packed := field.IsPacked()
+	wireType := field.WireType()
+	fieldNumber := field.GetNumber()
+	if packed {
+		wireType = 2
+	}
+	x = uint64(uint32(fieldNumber)<<3 | uint32(wireType))
+	return x
+}
+
+func (field *FieldDescriptorProto) GetKey3Uint64() (x uint64) {
+	packed := field.IsPacked3()
+	wireType := field.WireType()
+	fieldNumber := field.GetNumber()
+	if packed {
+		wireType = 2
+	}
+	x = uint64(uint32(fieldNumber)<<3 | uint32(wireType))
+	return x
+}
+
+func (field *FieldDescriptorProto) GetKey() []byte {
+	x := field.GetKeyUint64()
+	i := 0
+	keybuf := make([]byte, 0)
+	for i = 0; x > 127; i++ {
+		keybuf = append(keybuf, 0x80|uint8(x&0x7F))
+		x >>= 7
+	}
+	keybuf = append(keybuf, uint8(x))
+	return keybuf
+}
+
+func (field *FieldDescriptorProto) GetKey3() []byte {
+	x := field.GetKey3Uint64()
+	i := 0
+	keybuf := make([]byte, 0)
+	for i = 0; x > 127; i++ {
+		keybuf = append(keybuf, 0x80|uint8(x&0x7F))
+		x >>= 7
+	}
+	keybuf = append(keybuf, uint8(x))
+	return keybuf
+}
+
+func (desc *FileDescriptorSet) GetField(packageName, messageName, fieldName string) *FieldDescriptorProto {
+	msg := desc.GetMessage(packageName, messageName)
+	if msg == nil {
+		return nil
+	}
+	for _, field := range msg.GetField() {
+		if field.GetName() == fieldName {
+			return field
+		}
+	}
+	return nil
+}
+
+func (file *FileDescriptorProto) GetMessage(typeName string) *DescriptorProto {
+	for _, msg := range file.GetMessageType() {
+		if msg.GetName() == typeName {
+			return msg
+		}
+		nes := file.GetNestedMessage(msg, strings.TrimPrefix(typeName, msg.GetName()+"."))
+		if nes != nil {
+			return nes
+		}
+	}
+	return nil
+}
+
+func (file *FileDescriptorProto) GetNestedMessage(msg *DescriptorProto, typeName string) *DescriptorProto {
+	for _, nes := range msg.GetNestedType() {
+		if nes.GetName() == typeName {
+			return nes
+		}
+		res := file.GetNestedMessage(nes, strings.TrimPrefix(typeName, nes.GetName()+"."))
+		if res != nil {
+			return res
+		}
+	}
+	return nil
+}
+
+func (desc *FileDescriptorSet) GetMessage(packageName string, typeName string) *DescriptorProto {
+	for _, file := range desc.GetFile() {
+		if strings.Map(dotToUnderscore, file.GetPackage()) != strings.Map(dotToUnderscore, packageName) {
+			continue
+		}
+		for _, msg := range file.GetMessageType() {
+			if msg.GetName() == typeName {
+				return msg
+			}
+		}
+		for _, msg := range file.GetMessageType() {
+			for _, nes := range msg.GetNestedType() {
+				if nes.GetName() == typeName {
+					return nes
+				}
+				if msg.GetName()+"."+nes.GetName() == typeName {
+					return nes
+				}
+			}
+		}
+	}
+	return nil
+}
+
+func (desc *FileDescriptorSet) IsProto3(packageName string, typeName string) bool {
+	for _, file := range desc.GetFile() {
+		if strings.Map(dotToUnderscore, file.GetPackage()) != strings.Map(dotToUnderscore, packageName) {
+			continue
+		}
+		for _, msg := range file.GetMessageType() {
+			if msg.GetName() == typeName {
+				return file.GetSyntax() == "proto3"
+			}
+		}
+		for _, msg := range file.GetMessageType() {
+			for _, nes := range msg.GetNestedType() {
+				if nes.GetName() == typeName {
+					return file.GetSyntax() == "proto3"
+				}
+				if msg.GetName()+"."+nes.GetName() == typeName {
+					return file.GetSyntax() == "proto3"
+				}
+			}
+		}
+	}
+	return false
+}
+
+func (msg *DescriptorProto) IsExtendable() bool {
+	return len(msg.GetExtensionRange()) > 0
+}
+
+func (desc *FileDescriptorSet) FindExtension(packageName string, typeName string, fieldName string) (extPackageName string, field *FieldDescriptorProto) {
+	parent := desc.GetMessage(packageName, typeName)
+	if parent == nil {
+		return "", nil
+	}
+	if !parent.IsExtendable() {
+		return "", nil
+	}
+	extendee := "." + packageName + "." + typeName
+	for _, file := range desc.GetFile() {
+		for _, ext := range file.GetExtension() {
+			if strings.Map(dotToUnderscore, file.GetPackage()) == strings.Map(dotToUnderscore, packageName) {
+				if !(ext.GetExtendee() == typeName || ext.GetExtendee() == extendee) {
+					continue
+				}
+			} else {
+				if ext.GetExtendee() != extendee {
+					continue
+				}
+			}
+			if ext.GetName() == fieldName {
+				return file.GetPackage(), ext
+			}
+		}
+	}
+	return "", nil
+}
+
+func (desc *FileDescriptorSet) FindExtensionByFieldNumber(packageName string, typeName string, fieldNum int32) (extPackageName string, field *FieldDescriptorProto) {
+	parent := desc.GetMessage(packageName, typeName)
+	if parent == nil {
+		return "", nil
+	}
+	if !parent.IsExtendable() {
+		return "", nil
+	}
+	extendee := "." + packageName + "." + typeName
+	for _, file := range desc.GetFile() {
+		for _, ext := range file.GetExtension() {
+			if strings.Map(dotToUnderscore, file.GetPackage()) == strings.Map(dotToUnderscore, packageName) {
+				if !(ext.GetExtendee() == typeName || ext.GetExtendee() == extendee) {
+					continue
+				}
+			} else {
+				if ext.GetExtendee() != extendee {
+					continue
+				}
+			}
+			if ext.GetNumber() == fieldNum {
+				return file.GetPackage(), ext
+			}
+		}
+	}
+	return "", nil
+}
+
+func (desc *FileDescriptorSet) FindMessage(packageName string, typeName string, fieldName string) (msgPackageName string, msgName string) {
+	parent := desc.GetMessage(packageName, typeName)
+	if parent == nil {
+		return "", ""
+	}
+	field := parent.GetFieldDescriptor(fieldName)
+	if field == nil {
+		var extPackageName string
+		extPackageName, field = desc.FindExtension(packageName, typeName, fieldName)
+		if field == nil {
+			return "", ""
+		}
+		packageName = extPackageName
+	}
+	typeNames := strings.Split(field.GetTypeName(), ".")
+	if len(typeNames) == 1 {
+		msg := desc.GetMessage(packageName, typeName)
+		if msg == nil {
+			return "", ""
+		}
+		return packageName, msg.GetName()
+	}
+	if len(typeNames) > 2 {
+		for i := 1; i < len(typeNames)-1; i++ {
+			packageName = strings.Join(typeNames[1:len(typeNames)-i], ".")
+			typeName = strings.Join(typeNames[len(typeNames)-i:], ".")
+			msg := desc.GetMessage(packageName, typeName)
+			if msg != nil {
+				typeNames := strings.Split(msg.GetName(), ".")
+				if len(typeNames) == 1 {
+					return packageName, msg.GetName()
+				}
+				return strings.Join(typeNames[1:len(typeNames)-1], "."), typeNames[len(typeNames)-1]
+			}
+		}
+	}
+	return "", ""
+}
+
+func (msg *DescriptorProto) GetFieldDescriptor(fieldName string) *FieldDescriptorProto {
+	for _, field := range msg.GetField() {
+		if field.GetName() == fieldName {
+			return field
+		}
+	}
+	return nil
+}
+
+func (desc *FileDescriptorSet) GetEnum(packageName string, typeName string) *EnumDescriptorProto {
+	for _, file := range desc.GetFile() {
+		if strings.Map(dotToUnderscore, file.GetPackage()) != strings.Map(dotToUnderscore, packageName) {
+			continue
+		}
+		for _, enum := range file.GetEnumType() {
+			if enum.GetName() == typeName {
+				return enum
+			}
+		}
+	}
+	return nil
+}
+
+func (f *FieldDescriptorProto) IsEnum() bool {
+	return *f.Type == FieldDescriptorProto_TYPE_ENUM
+}
+
+func (f *FieldDescriptorProto) IsMessage() bool {
+	return *f.Type == FieldDescriptorProto_TYPE_MESSAGE
+}
+
+func (f *FieldDescriptorProto) IsBytes() bool {
+	return *f.Type == FieldDescriptorProto_TYPE_BYTES
+}
+
+func (f *FieldDescriptorProto) IsRepeated() bool {
+	return f.Label != nil && *f.Label == FieldDescriptorProto_LABEL_REPEATED
+}
+
+func (f *FieldDescriptorProto) IsString() bool {
+	return *f.Type == FieldDescriptorProto_TYPE_STRING
+}
+
+func (f *FieldDescriptorProto) IsBool() bool {
+	return *f.Type == FieldDescriptorProto_TYPE_BOOL
+}
+
+func (f *FieldDescriptorProto) IsRequired() bool {
+	return f.Label != nil && *f.Label == FieldDescriptorProto_LABEL_REQUIRED
+}
+
+func (f *FieldDescriptorProto) IsPacked() bool {
+	return f.Options != nil && f.GetOptions().GetPacked()
+}
+
+func (f *FieldDescriptorProto) IsPacked3() bool {
+	if f.IsRepeated() && f.IsScalar() {
+		if f.Options == nil || f.GetOptions().Packed == nil {
+			return true
+		}
+		return f.Options != nil && f.GetOptions().GetPacked()
+	}
+	return false
+}
+
+func (m *DescriptorProto) HasExtension() bool {
+	return len(m.ExtensionRange) > 0
+}
diff --git a/vendor/github.com/golang/protobuf/AUTHORS b/vendor/github.com/golang/protobuf/AUTHORS
new file mode 100644
index 0000000..15167cd
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/AUTHORS
@@ -0,0 +1,3 @@
+# This source code refers to The Go Authors for copyright purposes.
+# The master list of authors is in the main Go distribution,
+# visible at http://tip.golang.org/AUTHORS.
diff --git a/vendor/github.com/golang/protobuf/CONTRIBUTORS b/vendor/github.com/golang/protobuf/CONTRIBUTORS
new file mode 100644
index 0000000..1c4577e
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/CONTRIBUTORS
@@ -0,0 +1,3 @@
+# This source code was written by the Go contributors.
+# The master list of contributors is in the main Go distribution,
+# visible at http://tip.golang.org/CONTRIBUTORS.
diff --git a/vendor/github.com/golang/protobuf/LICENSE b/vendor/github.com/golang/protobuf/LICENSE
new file mode 100644
index 0000000..0f64693
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/LICENSE
@@ -0,0 +1,28 @@
+Copyright 2010 The Go Authors.  All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/vendor/github.com/golang/protobuf/proto/clone.go b/vendor/github.com/golang/protobuf/proto/clone.go
new file mode 100644
index 0000000..3cd3249
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/clone.go
@@ -0,0 +1,253 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2011 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Protocol buffer deep copy and merge.
+// TODO: RawMessage.
+
+package proto
+
+import (
+	"fmt"
+	"log"
+	"reflect"
+	"strings"
+)
+
+// Clone returns a deep copy of a protocol buffer.
+func Clone(src Message) Message {
+	in := reflect.ValueOf(src)
+	if in.IsNil() {
+		return src
+	}
+	out := reflect.New(in.Type().Elem())
+	dst := out.Interface().(Message)
+	Merge(dst, src)
+	return dst
+}
+
+// Merger is the interface representing objects that can merge messages of the same type.
+type Merger interface {
+	// Merge merges src into this message.
+	// Required and optional fields that are set in src will be set to that value in dst.
+	// Elements of repeated fields will be appended.
+	//
+	// Merge may panic if called with a different argument type than the receiver.
+	Merge(src Message)
+}
+
+// generatedMerger is the custom merge method that generated protos will have.
+// We must add this method since a generate Merge method will conflict with
+// many existing protos that have a Merge data field already defined.
+type generatedMerger interface {
+	XXX_Merge(src Message)
+}
+
+// Merge merges src into dst.
+// Required and optional fields that are set in src will be set to that value in dst.
+// Elements of repeated fields will be appended.
+// Merge panics if src and dst are not the same type, or if dst is nil.
+func Merge(dst, src Message) {
+	if m, ok := dst.(Merger); ok {
+		m.Merge(src)
+		return
+	}
+
+	in := reflect.ValueOf(src)
+	out := reflect.ValueOf(dst)
+	if out.IsNil() {
+		panic("proto: nil destination")
+	}
+	if in.Type() != out.Type() {
+		panic(fmt.Sprintf("proto.Merge(%T, %T) type mismatch", dst, src))
+	}
+	if in.IsNil() {
+		return // Merge from nil src is a noop
+	}
+	if m, ok := dst.(generatedMerger); ok {
+		m.XXX_Merge(src)
+		return
+	}
+	mergeStruct(out.Elem(), in.Elem())
+}
+
+func mergeStruct(out, in reflect.Value) {
+	sprop := GetProperties(in.Type())
+	for i := 0; i < in.NumField(); i++ {
+		f := in.Type().Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		mergeAny(out.Field(i), in.Field(i), false, sprop.Prop[i])
+	}
+
+	if emIn, err := extendable(in.Addr().Interface()); err == nil {
+		emOut, _ := extendable(out.Addr().Interface())
+		mIn, muIn := emIn.extensionsRead()
+		if mIn != nil {
+			mOut := emOut.extensionsWrite()
+			muIn.Lock()
+			mergeExtension(mOut, mIn)
+			muIn.Unlock()
+		}
+	}
+
+	uf := in.FieldByName("XXX_unrecognized")
+	if !uf.IsValid() {
+		return
+	}
+	uin := uf.Bytes()
+	if len(uin) > 0 {
+		out.FieldByName("XXX_unrecognized").SetBytes(append([]byte(nil), uin...))
+	}
+}
+
+// mergeAny performs a merge between two values of the same type.
+// viaPtr indicates whether the values were indirected through a pointer (implying proto2).
+// prop is set if this is a struct field (it may be nil).
+func mergeAny(out, in reflect.Value, viaPtr bool, prop *Properties) {
+	if in.Type() == protoMessageType {
+		if !in.IsNil() {
+			if out.IsNil() {
+				out.Set(reflect.ValueOf(Clone(in.Interface().(Message))))
+			} else {
+				Merge(out.Interface().(Message), in.Interface().(Message))
+			}
+		}
+		return
+	}
+	switch in.Kind() {
+	case reflect.Bool, reflect.Float32, reflect.Float64, reflect.Int32, reflect.Int64,
+		reflect.String, reflect.Uint32, reflect.Uint64:
+		if !viaPtr && isProto3Zero(in) {
+			return
+		}
+		out.Set(in)
+	case reflect.Interface:
+		// Probably a oneof field; copy non-nil values.
+		if in.IsNil() {
+			return
+		}
+		// Allocate destination if it is not set, or set to a different type.
+		// Otherwise we will merge as normal.
+		if out.IsNil() || out.Elem().Type() != in.Elem().Type() {
+			out.Set(reflect.New(in.Elem().Elem().Type())) // interface -> *T -> T -> new(T)
+		}
+		mergeAny(out.Elem(), in.Elem(), false, nil)
+	case reflect.Map:
+		if in.Len() == 0 {
+			return
+		}
+		if out.IsNil() {
+			out.Set(reflect.MakeMap(in.Type()))
+		}
+		// For maps with value types of *T or []byte we need to deep copy each value.
+		elemKind := in.Type().Elem().Kind()
+		for _, key := range in.MapKeys() {
+			var val reflect.Value
+			switch elemKind {
+			case reflect.Ptr:
+				val = reflect.New(in.Type().Elem().Elem())
+				mergeAny(val, in.MapIndex(key), false, nil)
+			case reflect.Slice:
+				val = in.MapIndex(key)
+				val = reflect.ValueOf(append([]byte{}, val.Bytes()...))
+			default:
+				val = in.MapIndex(key)
+			}
+			out.SetMapIndex(key, val)
+		}
+	case reflect.Ptr:
+		if in.IsNil() {
+			return
+		}
+		if out.IsNil() {
+			out.Set(reflect.New(in.Elem().Type()))
+		}
+		mergeAny(out.Elem(), in.Elem(), true, nil)
+	case reflect.Slice:
+		if in.IsNil() {
+			return
+		}
+		if in.Type().Elem().Kind() == reflect.Uint8 {
+			// []byte is a scalar bytes field, not a repeated field.
+
+			// Edge case: if this is in a proto3 message, a zero length
+			// bytes field is considered the zero value, and should not
+			// be merged.
+			if prop != nil && prop.proto3 && in.Len() == 0 {
+				return
+			}
+
+			// Make a deep copy.
+			// Append to []byte{} instead of []byte(nil) so that we never end up
+			// with a nil result.
+			out.SetBytes(append([]byte{}, in.Bytes()...))
+			return
+		}
+		n := in.Len()
+		if out.IsNil() {
+			out.Set(reflect.MakeSlice(in.Type(), 0, n))
+		}
+		switch in.Type().Elem().Kind() {
+		case reflect.Bool, reflect.Float32, reflect.Float64, reflect.Int32, reflect.Int64,
+			reflect.String, reflect.Uint32, reflect.Uint64:
+			out.Set(reflect.AppendSlice(out, in))
+		default:
+			for i := 0; i < n; i++ {
+				x := reflect.Indirect(reflect.New(in.Type().Elem()))
+				mergeAny(x, in.Index(i), false, nil)
+				out.Set(reflect.Append(out, x))
+			}
+		}
+	case reflect.Struct:
+		mergeStruct(out, in)
+	default:
+		// unknown type, so not a protocol buffer
+		log.Printf("proto: don't know how to copy %v", in)
+	}
+}
+
+func mergeExtension(out, in map[int32]Extension) {
+	for extNum, eIn := range in {
+		eOut := Extension{desc: eIn.desc}
+		if eIn.value != nil {
+			v := reflect.New(reflect.TypeOf(eIn.value)).Elem()
+			mergeAny(v, reflect.ValueOf(eIn.value), false, nil)
+			eOut.value = v.Interface()
+		}
+		if eIn.enc != nil {
+			eOut.enc = make([]byte, len(eIn.enc))
+			copy(eOut.enc, eIn.enc)
+		}
+
+		out[extNum] = eOut
+	}
+}
diff --git a/vendor/github.com/golang/protobuf/proto/decode.go b/vendor/github.com/golang/protobuf/proto/decode.go
new file mode 100644
index 0000000..63b0f08
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/decode.go
@@ -0,0 +1,427 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for decoding protocol buffer data to construct in-memory representations.
+ */
+
+import (
+	"errors"
+	"fmt"
+	"io"
+)
+
+// errOverflow is returned when an integer is too large to be represented.
+var errOverflow = errors.New("proto: integer overflow")
+
+// ErrInternalBadWireType is returned by generated code when an incorrect
+// wire type is encountered. It does not get returned to user code.
+var ErrInternalBadWireType = errors.New("proto: internal error: bad wiretype for oneof")
+
+// DecodeVarint reads a varint-encoded integer from the slice.
+// It returns the integer and the number of bytes consumed, or
+// zero if there is not enough.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func DecodeVarint(buf []byte) (x uint64, n int) {
+	for shift := uint(0); shift < 64; shift += 7 {
+		if n >= len(buf) {
+			return 0, 0
+		}
+		b := uint64(buf[n])
+		n++
+		x |= (b & 0x7F) << shift
+		if (b & 0x80) == 0 {
+			return x, n
+		}
+	}
+
+	// The number is too large to represent in a 64-bit value.
+	return 0, 0
+}
+
+func (p *Buffer) decodeVarintSlow() (x uint64, err error) {
+	i := p.index
+	l := len(p.buf)
+
+	for shift := uint(0); shift < 64; shift += 7 {
+		if i >= l {
+			err = io.ErrUnexpectedEOF
+			return
+		}
+		b := p.buf[i]
+		i++
+		x |= (uint64(b) & 0x7F) << shift
+		if b < 0x80 {
+			p.index = i
+			return
+		}
+	}
+
+	// The number is too large to represent in a 64-bit value.
+	err = errOverflow
+	return
+}
+
+// DecodeVarint reads a varint-encoded integer from the Buffer.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func (p *Buffer) DecodeVarint() (x uint64, err error) {
+	i := p.index
+	buf := p.buf
+
+	if i >= len(buf) {
+		return 0, io.ErrUnexpectedEOF
+	} else if buf[i] < 0x80 {
+		p.index++
+		return uint64(buf[i]), nil
+	} else if len(buf)-i < 10 {
+		return p.decodeVarintSlow()
+	}
+
+	var b uint64
+	// we already checked the first byte
+	x = uint64(buf[i]) - 0x80
+	i++
+
+	b = uint64(buf[i])
+	i++
+	x += b << 7
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 7
+
+	b = uint64(buf[i])
+	i++
+	x += b << 14
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 14
+
+	b = uint64(buf[i])
+	i++
+	x += b << 21
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 21
+
+	b = uint64(buf[i])
+	i++
+	x += b << 28
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 28
+
+	b = uint64(buf[i])
+	i++
+	x += b << 35
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 35
+
+	b = uint64(buf[i])
+	i++
+	x += b << 42
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 42
+
+	b = uint64(buf[i])
+	i++
+	x += b << 49
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 49
+
+	b = uint64(buf[i])
+	i++
+	x += b << 56
+	if b&0x80 == 0 {
+		goto done
+	}
+	x -= 0x80 << 56
+
+	b = uint64(buf[i])
+	i++
+	x += b << 63
+	if b&0x80 == 0 {
+		goto done
+	}
+
+	return 0, errOverflow
+
+done:
+	p.index = i
+	return x, nil
+}
+
+// DecodeFixed64 reads a 64-bit integer from the Buffer.
+// This is the format for the
+// fixed64, sfixed64, and double protocol buffer types.
+func (p *Buffer) DecodeFixed64() (x uint64, err error) {
+	// x, err already 0
+	i := p.index + 8
+	if i < 0 || i > len(p.buf) {
+		err = io.ErrUnexpectedEOF
+		return
+	}
+	p.index = i
+
+	x = uint64(p.buf[i-8])
+	x |= uint64(p.buf[i-7]) << 8
+	x |= uint64(p.buf[i-6]) << 16
+	x |= uint64(p.buf[i-5]) << 24
+	x |= uint64(p.buf[i-4]) << 32
+	x |= uint64(p.buf[i-3]) << 40
+	x |= uint64(p.buf[i-2]) << 48
+	x |= uint64(p.buf[i-1]) << 56
+	return
+}
+
+// DecodeFixed32 reads a 32-bit integer from the Buffer.
+// This is the format for the
+// fixed32, sfixed32, and float protocol buffer types.
+func (p *Buffer) DecodeFixed32() (x uint64, err error) {
+	// x, err already 0
+	i := p.index + 4
+	if i < 0 || i > len(p.buf) {
+		err = io.ErrUnexpectedEOF
+		return
+	}
+	p.index = i
+
+	x = uint64(p.buf[i-4])
+	x |= uint64(p.buf[i-3]) << 8
+	x |= uint64(p.buf[i-2]) << 16
+	x |= uint64(p.buf[i-1]) << 24
+	return
+}
+
+// DecodeZigzag64 reads a zigzag-encoded 64-bit integer
+// from the Buffer.
+// This is the format used for the sint64 protocol buffer type.
+func (p *Buffer) DecodeZigzag64() (x uint64, err error) {
+	x, err = p.DecodeVarint()
+	if err != nil {
+		return
+	}
+	x = (x >> 1) ^ uint64((int64(x&1)<<63)>>63)
+	return
+}
+
+// DecodeZigzag32 reads a zigzag-encoded 32-bit integer
+// from  the Buffer.
+// This is the format used for the sint32 protocol buffer type.
+func (p *Buffer) DecodeZigzag32() (x uint64, err error) {
+	x, err = p.DecodeVarint()
+	if err != nil {
+		return
+	}
+	x = uint64((uint32(x) >> 1) ^ uint32((int32(x&1)<<31)>>31))
+	return
+}
+
+// DecodeRawBytes reads a count-delimited byte buffer from the Buffer.
+// This is the format used for the bytes protocol buffer
+// type and for embedded messages.
+func (p *Buffer) DecodeRawBytes(alloc bool) (buf []byte, err error) {
+	n, err := p.DecodeVarint()
+	if err != nil {
+		return nil, err
+	}
+
+	nb := int(n)
+	if nb < 0 {
+		return nil, fmt.Errorf("proto: bad byte length %d", nb)
+	}
+	end := p.index + nb
+	if end < p.index || end > len(p.buf) {
+		return nil, io.ErrUnexpectedEOF
+	}
+
+	if !alloc {
+		// todo: check if can get more uses of alloc=false
+		buf = p.buf[p.index:end]
+		p.index += nb
+		return
+	}
+
+	buf = make([]byte, nb)
+	copy(buf, p.buf[p.index:])
+	p.index += nb
+	return
+}
+
+// DecodeStringBytes reads an encoded string from the Buffer.
+// This is the format used for the proto2 string type.
+func (p *Buffer) DecodeStringBytes() (s string, err error) {
+	buf, err := p.DecodeRawBytes(false)
+	if err != nil {
+		return
+	}
+	return string(buf), nil
+}
+
+// Unmarshaler is the interface representing objects that can
+// unmarshal themselves.  The argument points to data that may be
+// overwritten, so implementations should not keep references to the
+// buffer.
+// Unmarshal implementations should not clear the receiver.
+// Any unmarshaled data should be merged into the receiver.
+// Callers of Unmarshal that do not want to retain existing data
+// should Reset the receiver before calling Unmarshal.
+type Unmarshaler interface {
+	Unmarshal([]byte) error
+}
+
+// newUnmarshaler is the interface representing objects that can
+// unmarshal themselves. The semantics are identical to Unmarshaler.
+//
+// This exists to support protoc-gen-go generated messages.
+// The proto package will stop type-asserting to this interface in the future.
+//
+// DO NOT DEPEND ON THIS.
+type newUnmarshaler interface {
+	XXX_Unmarshal([]byte) error
+}
+
+// Unmarshal parses the protocol buffer representation in buf and places the
+// decoded result in pb.  If the struct underlying pb does not match
+// the data in buf, the results can be unpredictable.
+//
+// Unmarshal resets pb before starting to unmarshal, so any
+// existing data in pb is always removed. Use UnmarshalMerge
+// to preserve and append to existing data.
+func Unmarshal(buf []byte, pb Message) error {
+	pb.Reset()
+	if u, ok := pb.(newUnmarshaler); ok {
+		return u.XXX_Unmarshal(buf)
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		return u.Unmarshal(buf)
+	}
+	return NewBuffer(buf).Unmarshal(pb)
+}
+
+// UnmarshalMerge parses the protocol buffer representation in buf and
+// writes the decoded result to pb.  If the struct underlying pb does not match
+// the data in buf, the results can be unpredictable.
+//
+// UnmarshalMerge merges into existing data in pb.
+// Most code should use Unmarshal instead.
+func UnmarshalMerge(buf []byte, pb Message) error {
+	if u, ok := pb.(newUnmarshaler); ok {
+		return u.XXX_Unmarshal(buf)
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		// NOTE: The history of proto have unfortunately been inconsistent
+		// whether Unmarshaler should or should not implicitly clear itself.
+		// Some implementations do, most do not.
+		// Thus, calling this here may or may not do what people want.
+		//
+		// See https://github.com/golang/protobuf/issues/424
+		return u.Unmarshal(buf)
+	}
+	return NewBuffer(buf).Unmarshal(pb)
+}
+
+// DecodeMessage reads a count-delimited message from the Buffer.
+func (p *Buffer) DecodeMessage(pb Message) error {
+	enc, err := p.DecodeRawBytes(false)
+	if err != nil {
+		return err
+	}
+	return NewBuffer(enc).Unmarshal(pb)
+}
+
+// DecodeGroup reads a tag-delimited group from the Buffer.
+// StartGroup tag is already consumed. This function consumes
+// EndGroup tag.
+func (p *Buffer) DecodeGroup(pb Message) error {
+	b := p.buf[p.index:]
+	x, y := findEndGroup(b)
+	if x < 0 {
+		return io.ErrUnexpectedEOF
+	}
+	err := Unmarshal(b[:x], pb)
+	p.index += y
+	return err
+}
+
+// Unmarshal parses the protocol buffer representation in the
+// Buffer and places the decoded result in pb.  If the struct
+// underlying pb does not match the data in the buffer, the results can be
+// unpredictable.
+//
+// Unlike proto.Unmarshal, this does not reset pb before starting to unmarshal.
+func (p *Buffer) Unmarshal(pb Message) error {
+	// If the object can unmarshal itself, let it.
+	if u, ok := pb.(newUnmarshaler); ok {
+		err := u.XXX_Unmarshal(p.buf[p.index:])
+		p.index = len(p.buf)
+		return err
+	}
+	if u, ok := pb.(Unmarshaler); ok {
+		// NOTE: The history of proto have unfortunately been inconsistent
+		// whether Unmarshaler should or should not implicitly clear itself.
+		// Some implementations do, most do not.
+		// Thus, calling this here may or may not do what people want.
+		//
+		// See https://github.com/golang/protobuf/issues/424
+		err := u.Unmarshal(p.buf[p.index:])
+		p.index = len(p.buf)
+		return err
+	}
+
+	// Slow workaround for messages that aren't Unmarshalers.
+	// This includes some hand-coded .pb.go files and
+	// bootstrap protos.
+	// TODO: fix all of those and then add Unmarshal to
+	// the Message interface. Then:
+	// The cast above and code below can be deleted.
+	// The old unmarshaler can be deleted.
+	// Clients can call Unmarshal directly (can already do that, actually).
+	var info InternalMessageInfo
+	err := info.Unmarshal(pb, p.buf[p.index:])
+	p.index = len(p.buf)
+	return err
+}
diff --git a/vendor/github.com/golang/protobuf/proto/deprecated.go b/vendor/github.com/golang/protobuf/proto/deprecated.go
new file mode 100644
index 0000000..35b882c
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/deprecated.go
@@ -0,0 +1,63 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2018 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import "errors"
+
+// Deprecated: do not use.
+type Stats struct{ Emalloc, Dmalloc, Encode, Decode, Chit, Cmiss, Size uint64 }
+
+// Deprecated: do not use.
+func GetStats() Stats { return Stats{} }
+
+// Deprecated: do not use.
+func MarshalMessageSet(interface{}) ([]byte, error) {
+	return nil, errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func UnmarshalMessageSet([]byte, interface{}) error {
+	return errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func MarshalMessageSetJSON(interface{}) ([]byte, error) {
+	return nil, errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func UnmarshalMessageSetJSON([]byte, interface{}) error {
+	return errors.New("proto: not implemented")
+}
+
+// Deprecated: do not use.
+func RegisterMessageSetType(Message, int32, string) {}
diff --git a/vendor/github.com/golang/protobuf/proto/discard.go b/vendor/github.com/golang/protobuf/proto/discard.go
new file mode 100644
index 0000000..dea2617
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/discard.go
@@ -0,0 +1,350 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2017 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+	"sync"
+	"sync/atomic"
+)
+
+type generatedDiscarder interface {
+	XXX_DiscardUnknown()
+}
+
+// DiscardUnknown recursively discards all unknown fields from this message
+// and all embedded messages.
+//
+// When unmarshaling a message with unrecognized fields, the tags and values
+// of such fields are preserved in the Message. This allows a later call to
+// marshal to be able to produce a message that continues to have those
+// unrecognized fields. To avoid this, DiscardUnknown is used to
+// explicitly clear the unknown fields after unmarshaling.
+//
+// For proto2 messages, the unknown fields of message extensions are only
+// discarded from messages that have been accessed via GetExtension.
+func DiscardUnknown(m Message) {
+	if m, ok := m.(generatedDiscarder); ok {
+		m.XXX_DiscardUnknown()
+		return
+	}
+	// TODO: Dynamically populate a InternalMessageInfo for legacy messages,
+	// but the master branch has no implementation for InternalMessageInfo,
+	// so it would be more work to replicate that approach.
+	discardLegacy(m)
+}
+
+// DiscardUnknown recursively discards all unknown fields.
+func (a *InternalMessageInfo) DiscardUnknown(m Message) {
+	di := atomicLoadDiscardInfo(&a.discard)
+	if di == nil {
+		di = getDiscardInfo(reflect.TypeOf(m).Elem())
+		atomicStoreDiscardInfo(&a.discard, di)
+	}
+	di.discard(toPointer(&m))
+}
+
+type discardInfo struct {
+	typ reflect.Type
+
+	initialized int32 // 0: only typ is valid, 1: everything is valid
+	lock        sync.Mutex
+
+	fields       []discardFieldInfo
+	unrecognized field
+}
+
+type discardFieldInfo struct {
+	field   field // Offset of field, guaranteed to be valid
+	discard func(src pointer)
+}
+
+var (
+	discardInfoMap  = map[reflect.Type]*discardInfo{}
+	discardInfoLock sync.Mutex
+)
+
+func getDiscardInfo(t reflect.Type) *discardInfo {
+	discardInfoLock.Lock()
+	defer discardInfoLock.Unlock()
+	di := discardInfoMap[t]
+	if di == nil {
+		di = &discardInfo{typ: t}
+		discardInfoMap[t] = di
+	}
+	return di
+}
+
+func (di *discardInfo) discard(src pointer) {
+	if src.isNil() {
+		return // Nothing to do.
+	}
+
+	if atomic.LoadInt32(&di.initialized) == 0 {
+		di.computeDiscardInfo()
+	}
+
+	for _, fi := range di.fields {
+		sfp := src.offset(fi.field)
+		fi.discard(sfp)
+	}
+
+	// For proto2 messages, only discard unknown fields in message extensions
+	// that have been accessed via GetExtension.
+	if em, err := extendable(src.asPointerTo(di.typ).Interface()); err == nil {
+		// Ignore lock since DiscardUnknown is not concurrency safe.
+		emm, _ := em.extensionsRead()
+		for _, mx := range emm {
+			if m, ok := mx.value.(Message); ok {
+				DiscardUnknown(m)
+			}
+		}
+	}
+
+	if di.unrecognized.IsValid() {
+		*src.offset(di.unrecognized).toBytes() = nil
+	}
+}
+
+func (di *discardInfo) computeDiscardInfo() {
+	di.lock.Lock()
+	defer di.lock.Unlock()
+	if di.initialized != 0 {
+		return
+	}
+	t := di.typ
+	n := t.NumField()
+
+	for i := 0; i < n; i++ {
+		f := t.Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+
+		dfi := discardFieldInfo{field: toField(&f)}
+		tf := f.Type
+
+		// Unwrap tf to get its most basic type.
+		var isPointer, isSlice bool
+		if tf.Kind() == reflect.Slice && tf.Elem().Kind() != reflect.Uint8 {
+			isSlice = true
+			tf = tf.Elem()
+		}
+		if tf.Kind() == reflect.Ptr {
+			isPointer = true
+			tf = tf.Elem()
+		}
+		if isPointer && isSlice && tf.Kind() != reflect.Struct {
+			panic(fmt.Sprintf("%v.%s cannot be a slice of pointers to primitive types", t, f.Name))
+		}
+
+		switch tf.Kind() {
+		case reflect.Struct:
+			switch {
+			case !isPointer:
+				panic(fmt.Sprintf("%v.%s cannot be a direct struct value", t, f.Name))
+			case isSlice: // E.g., []*pb.T
+				di := getDiscardInfo(tf)
+				dfi.discard = func(src pointer) {
+					sps := src.getPointerSlice()
+					for _, sp := range sps {
+						if !sp.isNil() {
+							di.discard(sp)
+						}
+					}
+				}
+			default: // E.g., *pb.T
+				di := getDiscardInfo(tf)
+				dfi.discard = func(src pointer) {
+					sp := src.getPointer()
+					if !sp.isNil() {
+						di.discard(sp)
+					}
+				}
+			}
+		case reflect.Map:
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%v.%s cannot be a pointer to a map or a slice of map values", t, f.Name))
+			default: // E.g., map[K]V
+				if tf.Elem().Kind() == reflect.Ptr { // Proto struct (e.g., *T)
+					dfi.discard = func(src pointer) {
+						sm := src.asPointerTo(tf).Elem()
+						if sm.Len() == 0 {
+							return
+						}
+						for _, key := range sm.MapKeys() {
+							val := sm.MapIndex(key)
+							DiscardUnknown(val.Interface().(Message))
+						}
+					}
+				} else {
+					dfi.discard = func(pointer) {} // Noop
+				}
+			}
+		case reflect.Interface:
+			// Must be oneof field.
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%v.%s cannot be a pointer to a interface or a slice of interface values", t, f.Name))
+			default: // E.g., interface{}
+				// TODO: Make this faster?
+				dfi.discard = func(src pointer) {
+					su := src.asPointerTo(tf).Elem()
+					if !su.IsNil() {
+						sv := su.Elem().Elem().Field(0)
+						if sv.Kind() == reflect.Ptr && sv.IsNil() {
+							return
+						}
+						switch sv.Type().Kind() {
+						case reflect.Ptr: // Proto struct (e.g., *T)
+							DiscardUnknown(sv.Interface().(Message))
+						}
+					}
+				}
+			}
+		default:
+			continue
+		}
+		di.fields = append(di.fields, dfi)
+	}
+
+	di.unrecognized = invalidField
+	if f, ok := t.FieldByName("XXX_unrecognized"); ok {
+		if f.Type != reflect.TypeOf([]byte{}) {
+			panic("expected XXX_unrecognized to be of type []byte")
+		}
+		di.unrecognized = toField(&f)
+	}
+
+	atomic.StoreInt32(&di.initialized, 1)
+}
+
+func discardLegacy(m Message) {
+	v := reflect.ValueOf(m)
+	if v.Kind() != reflect.Ptr || v.IsNil() {
+		return
+	}
+	v = v.Elem()
+	if v.Kind() != reflect.Struct {
+		return
+	}
+	t := v.Type()
+
+	for i := 0; i < v.NumField(); i++ {
+		f := t.Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		vf := v.Field(i)
+		tf := f.Type
+
+		// Unwrap tf to get its most basic type.
+		var isPointer, isSlice bool
+		if tf.Kind() == reflect.Slice && tf.Elem().Kind() != reflect.Uint8 {
+			isSlice = true
+			tf = tf.Elem()
+		}
+		if tf.Kind() == reflect.Ptr {
+			isPointer = true
+			tf = tf.Elem()
+		}
+		if isPointer && isSlice && tf.Kind() != reflect.Struct {
+			panic(fmt.Sprintf("%T.%s cannot be a slice of pointers to primitive types", m, f.Name))
+		}
+
+		switch tf.Kind() {
+		case reflect.Struct:
+			switch {
+			case !isPointer:
+				panic(fmt.Sprintf("%T.%s cannot be a direct struct value", m, f.Name))
+			case isSlice: // E.g., []*pb.T
+				for j := 0; j < vf.Len(); j++ {
+					discardLegacy(vf.Index(j).Interface().(Message))
+				}
+			default: // E.g., *pb.T
+				discardLegacy(vf.Interface().(Message))
+			}
+		case reflect.Map:
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%T.%s cannot be a pointer to a map or a slice of map values", m, f.Name))
+			default: // E.g., map[K]V
+				tv := vf.Type().Elem()
+				if tv.Kind() == reflect.Ptr && tv.Implements(protoMessageType) { // Proto struct (e.g., *T)
+					for _, key := range vf.MapKeys() {
+						val := vf.MapIndex(key)
+						discardLegacy(val.Interface().(Message))
+					}
+				}
+			}
+		case reflect.Interface:
+			// Must be oneof field.
+			switch {
+			case isPointer || isSlice:
+				panic(fmt.Sprintf("%T.%s cannot be a pointer to a interface or a slice of interface values", m, f.Name))
+			default: // E.g., test_proto.isCommunique_Union interface
+				if !vf.IsNil() && f.Tag.Get("protobuf_oneof") != "" {
+					vf = vf.Elem() // E.g., *test_proto.Communique_Msg
+					if !vf.IsNil() {
+						vf = vf.Elem()   // E.g., test_proto.Communique_Msg
+						vf = vf.Field(0) // E.g., Proto struct (e.g., *T) or primitive value
+						if vf.Kind() == reflect.Ptr {
+							discardLegacy(vf.Interface().(Message))
+						}
+					}
+				}
+			}
+		}
+	}
+
+	if vf := v.FieldByName("XXX_unrecognized"); vf.IsValid() {
+		if vf.Type() != reflect.TypeOf([]byte{}) {
+			panic("expected XXX_unrecognized to be of type []byte")
+		}
+		vf.Set(reflect.ValueOf([]byte(nil)))
+	}
+
+	// For proto2 messages, only discard unknown fields in message extensions
+	// that have been accessed via GetExtension.
+	if em, err := extendable(m); err == nil {
+		// Ignore lock since discardLegacy is not concurrency safe.
+		emm, _ := em.extensionsRead()
+		for _, mx := range emm {
+			if m, ok := mx.value.(Message); ok {
+				discardLegacy(m)
+			}
+		}
+	}
+}
diff --git a/vendor/github.com/golang/protobuf/proto/encode.go b/vendor/github.com/golang/protobuf/proto/encode.go
new file mode 100644
index 0000000..3abfed2
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/encode.go
@@ -0,0 +1,203 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for encoding data into the wire format for protocol buffers.
+ */
+
+import (
+	"errors"
+	"reflect"
+)
+
+var (
+	// errRepeatedHasNil is the error returned if Marshal is called with
+	// a struct with a repeated field containing a nil element.
+	errRepeatedHasNil = errors.New("proto: repeated field has nil element")
+
+	// errOneofHasNil is the error returned if Marshal is called with
+	// a struct with a oneof field containing a nil element.
+	errOneofHasNil = errors.New("proto: oneof field has nil value")
+
+	// ErrNil is the error returned if Marshal is called with nil.
+	ErrNil = errors.New("proto: Marshal called with nil")
+
+	// ErrTooLarge is the error returned if Marshal is called with a
+	// message that encodes to >2GB.
+	ErrTooLarge = errors.New("proto: message encodes to over 2 GB")
+)
+
+// The fundamental encoders that put bytes on the wire.
+// Those that take integer types all accept uint64 and are
+// therefore of type valueEncoder.
+
+const maxVarintBytes = 10 // maximum length of a varint
+
+// EncodeVarint returns the varint encoding of x.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+// Not used by the package itself, but helpful to clients
+// wishing to use the same encoding.
+func EncodeVarint(x uint64) []byte {
+	var buf [maxVarintBytes]byte
+	var n int
+	for n = 0; x > 127; n++ {
+		buf[n] = 0x80 | uint8(x&0x7F)
+		x >>= 7
+	}
+	buf[n] = uint8(x)
+	n++
+	return buf[0:n]
+}
+
+// EncodeVarint writes a varint-encoded integer to the Buffer.
+// This is the format for the
+// int32, int64, uint32, uint64, bool, and enum
+// protocol buffer types.
+func (p *Buffer) EncodeVarint(x uint64) error {
+	for x >= 1<<7 {
+		p.buf = append(p.buf, uint8(x&0x7f|0x80))
+		x >>= 7
+	}
+	p.buf = append(p.buf, uint8(x))
+	return nil
+}
+
+// SizeVarint returns the varint encoding size of an integer.
+func SizeVarint(x uint64) int {
+	switch {
+	case x < 1<<7:
+		return 1
+	case x < 1<<14:
+		return 2
+	case x < 1<<21:
+		return 3
+	case x < 1<<28:
+		return 4
+	case x < 1<<35:
+		return 5
+	case x < 1<<42:
+		return 6
+	case x < 1<<49:
+		return 7
+	case x < 1<<56:
+		return 8
+	case x < 1<<63:
+		return 9
+	}
+	return 10
+}
+
+// EncodeFixed64 writes a 64-bit integer to the Buffer.
+// This is the format for the
+// fixed64, sfixed64, and double protocol buffer types.
+func (p *Buffer) EncodeFixed64(x uint64) error {
+	p.buf = append(p.buf,
+		uint8(x),
+		uint8(x>>8),
+		uint8(x>>16),
+		uint8(x>>24),
+		uint8(x>>32),
+		uint8(x>>40),
+		uint8(x>>48),
+		uint8(x>>56))
+	return nil
+}
+
+// EncodeFixed32 writes a 32-bit integer to the Buffer.
+// This is the format for the
+// fixed32, sfixed32, and float protocol buffer types.
+func (p *Buffer) EncodeFixed32(x uint64) error {
+	p.buf = append(p.buf,
+		uint8(x),
+		uint8(x>>8),
+		uint8(x>>16),
+		uint8(x>>24))
+	return nil
+}
+
+// EncodeZigzag64 writes a zigzag-encoded 64-bit integer
+// to the Buffer.
+// This is the format used for the sint64 protocol buffer type.
+func (p *Buffer) EncodeZigzag64(x uint64) error {
+	// use signed number to get arithmetic right shift.
+	return p.EncodeVarint(uint64((x << 1) ^ uint64((int64(x) >> 63))))
+}
+
+// EncodeZigzag32 writes a zigzag-encoded 32-bit integer
+// to the Buffer.
+// This is the format used for the sint32 protocol buffer type.
+func (p *Buffer) EncodeZigzag32(x uint64) error {
+	// use signed number to get arithmetic right shift.
+	return p.EncodeVarint(uint64((uint32(x) << 1) ^ uint32((int32(x) >> 31))))
+}
+
+// EncodeRawBytes writes a count-delimited byte buffer to the Buffer.
+// This is the format used for the bytes protocol buffer
+// type and for embedded messages.
+func (p *Buffer) EncodeRawBytes(b []byte) error {
+	p.EncodeVarint(uint64(len(b)))
+	p.buf = append(p.buf, b...)
+	return nil
+}
+
+// EncodeStringBytes writes an encoded string to the Buffer.
+// This is the format used for the proto2 string type.
+func (p *Buffer) EncodeStringBytes(s string) error {
+	p.EncodeVarint(uint64(len(s)))
+	p.buf = append(p.buf, s...)
+	return nil
+}
+
+// Marshaler is the interface representing objects that can marshal themselves.
+type Marshaler interface {
+	Marshal() ([]byte, error)
+}
+
+// EncodeMessage writes the protocol buffer to the Buffer,
+// prefixed by a varint-encoded length.
+func (p *Buffer) EncodeMessage(pb Message) error {
+	siz := Size(pb)
+	p.EncodeVarint(uint64(siz))
+	return p.Marshal(pb)
+}
+
+// All protocol buffer fields are nillable, but be careful.
+func isNil(v reflect.Value) bool {
+	switch v.Kind() {
+	case reflect.Interface, reflect.Map, reflect.Ptr, reflect.Slice:
+		return v.IsNil()
+	}
+	return false
+}
diff --git a/vendor/github.com/golang/protobuf/proto/equal.go b/vendor/github.com/golang/protobuf/proto/equal.go
new file mode 100644
index 0000000..f9b6e41
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/equal.go
@@ -0,0 +1,301 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2011 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Protocol buffer comparison.
+
+package proto
+
+import (
+	"bytes"
+	"log"
+	"reflect"
+	"strings"
+)
+
+/*
+Equal returns true iff protocol buffers a and b are equal.
+The arguments must both be pointers to protocol buffer structs.
+
+Equality is defined in this way:
+  - Two messages are equal iff they are the same type,
+    corresponding fields are equal, unknown field sets
+    are equal, and extensions sets are equal.
+  - Two set scalar fields are equal iff their values are equal.
+    If the fields are of a floating-point type, remember that
+    NaN != x for all x, including NaN. If the message is defined
+    in a proto3 .proto file, fields are not "set"; specifically,
+    zero length proto3 "bytes" fields are equal (nil == {}).
+  - Two repeated fields are equal iff their lengths are the same,
+    and their corresponding elements are equal. Note a "bytes" field,
+    although represented by []byte, is not a repeated field and the
+    rule for the scalar fields described above applies.
+  - Two unset fields are equal.
+  - Two unknown field sets are equal if their current
+    encoded state is equal.
+  - Two extension sets are equal iff they have corresponding
+    elements that are pairwise equal.
+  - Two map fields are equal iff their lengths are the same,
+    and they contain the same set of elements. Zero-length map
+    fields are equal.
+  - Every other combination of things are not equal.
+
+The return value is undefined if a and b are not protocol buffers.
+*/
+func Equal(a, b Message) bool {
+	if a == nil || b == nil {
+		return a == b
+	}
+	v1, v2 := reflect.ValueOf(a), reflect.ValueOf(b)
+	if v1.Type() != v2.Type() {
+		return false
+	}
+	if v1.Kind() == reflect.Ptr {
+		if v1.IsNil() {
+			return v2.IsNil()
+		}
+		if v2.IsNil() {
+			return false
+		}
+		v1, v2 = v1.Elem(), v2.Elem()
+	}
+	if v1.Kind() != reflect.Struct {
+		return false
+	}
+	return equalStruct(v1, v2)
+}
+
+// v1 and v2 are known to have the same type.
+func equalStruct(v1, v2 reflect.Value) bool {
+	sprop := GetProperties(v1.Type())
+	for i := 0; i < v1.NumField(); i++ {
+		f := v1.Type().Field(i)
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		f1, f2 := v1.Field(i), v2.Field(i)
+		if f.Type.Kind() == reflect.Ptr {
+			if n1, n2 := f1.IsNil(), f2.IsNil(); n1 && n2 {
+				// both unset
+				continue
+			} else if n1 != n2 {
+				// set/unset mismatch
+				return false
+			}
+			f1, f2 = f1.Elem(), f2.Elem()
+		}
+		if !equalAny(f1, f2, sprop.Prop[i]) {
+			return false
+		}
+	}
+
+	if em1 := v1.FieldByName("XXX_InternalExtensions"); em1.IsValid() {
+		em2 := v2.FieldByName("XXX_InternalExtensions")
+		if !equalExtensions(v1.Type(), em1.Interface().(XXX_InternalExtensions), em2.Interface().(XXX_InternalExtensions)) {
+			return false
+		}
+	}
+
+	if em1 := v1.FieldByName("XXX_extensions"); em1.IsValid() {
+		em2 := v2.FieldByName("XXX_extensions")
+		if !equalExtMap(v1.Type(), em1.Interface().(map[int32]Extension), em2.Interface().(map[int32]Extension)) {
+			return false
+		}
+	}
+
+	uf := v1.FieldByName("XXX_unrecognized")
+	if !uf.IsValid() {
+		return true
+	}
+
+	u1 := uf.Bytes()
+	u2 := v2.FieldByName("XXX_unrecognized").Bytes()
+	return bytes.Equal(u1, u2)
+}
+
+// v1 and v2 are known to have the same type.
+// prop may be nil.
+func equalAny(v1, v2 reflect.Value, prop *Properties) bool {
+	if v1.Type() == protoMessageType {
+		m1, _ := v1.Interface().(Message)
+		m2, _ := v2.Interface().(Message)
+		return Equal(m1, m2)
+	}
+	switch v1.Kind() {
+	case reflect.Bool:
+		return v1.Bool() == v2.Bool()
+	case reflect.Float32, reflect.Float64:
+		return v1.Float() == v2.Float()
+	case reflect.Int32, reflect.Int64:
+		return v1.Int() == v2.Int()
+	case reflect.Interface:
+		// Probably a oneof field; compare the inner values.
+		n1, n2 := v1.IsNil(), v2.IsNil()
+		if n1 || n2 {
+			return n1 == n2
+		}
+		e1, e2 := v1.Elem(), v2.Elem()
+		if e1.Type() != e2.Type() {
+			return false
+		}
+		return equalAny(e1, e2, nil)
+	case reflect.Map:
+		if v1.Len() != v2.Len() {
+			return false
+		}
+		for _, key := range v1.MapKeys() {
+			val2 := v2.MapIndex(key)
+			if !val2.IsValid() {
+				// This key was not found in the second map.
+				return false
+			}
+			if !equalAny(v1.MapIndex(key), val2, nil) {
+				return false
+			}
+		}
+		return true
+	case reflect.Ptr:
+		// Maps may have nil values in them, so check for nil.
+		if v1.IsNil() && v2.IsNil() {
+			return true
+		}
+		if v1.IsNil() != v2.IsNil() {
+			return false
+		}
+		return equalAny(v1.Elem(), v2.Elem(), prop)
+	case reflect.Slice:
+		if v1.Type().Elem().Kind() == reflect.Uint8 {
+			// short circuit: []byte
+
+			// Edge case: if this is in a proto3 message, a zero length
+			// bytes field is considered the zero value.
+			if prop != nil && prop.proto3 && v1.Len() == 0 && v2.Len() == 0 {
+				return true
+			}
+			if v1.IsNil() != v2.IsNil() {
+				return false
+			}
+			return bytes.Equal(v1.Interface().([]byte), v2.Interface().([]byte))
+		}
+
+		if v1.Len() != v2.Len() {
+			return false
+		}
+		for i := 0; i < v1.Len(); i++ {
+			if !equalAny(v1.Index(i), v2.Index(i), prop) {
+				return false
+			}
+		}
+		return true
+	case reflect.String:
+		return v1.Interface().(string) == v2.Interface().(string)
+	case reflect.Struct:
+		return equalStruct(v1, v2)
+	case reflect.Uint32, reflect.Uint64:
+		return v1.Uint() == v2.Uint()
+	}
+
+	// unknown type, so not a protocol buffer
+	log.Printf("proto: don't know how to compare %v", v1)
+	return false
+}
+
+// base is the struct type that the extensions are based on.
+// x1 and x2 are InternalExtensions.
+func equalExtensions(base reflect.Type, x1, x2 XXX_InternalExtensions) bool {
+	em1, _ := x1.extensionsRead()
+	em2, _ := x2.extensionsRead()
+	return equalExtMap(base, em1, em2)
+}
+
+func equalExtMap(base reflect.Type, em1, em2 map[int32]Extension) bool {
+	if len(em1) != len(em2) {
+		return false
+	}
+
+	for extNum, e1 := range em1 {
+		e2, ok := em2[extNum]
+		if !ok {
+			return false
+		}
+
+		m1 := extensionAsLegacyType(e1.value)
+		m2 := extensionAsLegacyType(e2.value)
+
+		if m1 == nil && m2 == nil {
+			// Both have only encoded form.
+			if bytes.Equal(e1.enc, e2.enc) {
+				continue
+			}
+			// The bytes are different, but the extensions might still be
+			// equal. We need to decode them to compare.
+		}
+
+		if m1 != nil && m2 != nil {
+			// Both are unencoded.
+			if !equalAny(reflect.ValueOf(m1), reflect.ValueOf(m2), nil) {
+				return false
+			}
+			continue
+		}
+
+		// At least one is encoded. To do a semantically correct comparison
+		// we need to unmarshal them first.
+		var desc *ExtensionDesc
+		if m := extensionMaps[base]; m != nil {
+			desc = m[extNum]
+		}
+		if desc == nil {
+			// If both have only encoded form and the bytes are the same,
+			// it is handled above. We get here when the bytes are different.
+			// We don't know how to decode it, so just compare them as byte
+			// slices.
+			log.Printf("proto: don't know how to compare extension %d of %v", extNum, base)
+			return false
+		}
+		var err error
+		if m1 == nil {
+			m1, err = decodeExtension(e1.enc, desc)
+		}
+		if m2 == nil && err == nil {
+			m2, err = decodeExtension(e2.enc, desc)
+		}
+		if err != nil {
+			// The encoded form is invalid.
+			log.Printf("proto: badly encoded extension %d of %v: %v", extNum, base, err)
+			return false
+		}
+		if !equalAny(reflect.ValueOf(m1), reflect.ValueOf(m2), nil) {
+			return false
+		}
+	}
+
+	return true
+}
diff --git a/vendor/github.com/golang/protobuf/proto/extensions.go b/vendor/github.com/golang/protobuf/proto/extensions.go
new file mode 100644
index 0000000..fa88add
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/extensions.go
@@ -0,0 +1,607 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Types and routines for supporting protocol buffer extensions.
+ */
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"reflect"
+	"strconv"
+	"sync"
+)
+
+// ErrMissingExtension is the error returned by GetExtension if the named extension is not in the message.
+var ErrMissingExtension = errors.New("proto: missing extension")
+
+// ExtensionRange represents a range of message extensions for a protocol buffer.
+// Used in code generated by the protocol compiler.
+type ExtensionRange struct {
+	Start, End int32 // both inclusive
+}
+
+// extendableProto is an interface implemented by any protocol buffer generated by the current
+// proto compiler that may be extended.
+type extendableProto interface {
+	Message
+	ExtensionRangeArray() []ExtensionRange
+	extensionsWrite() map[int32]Extension
+	extensionsRead() (map[int32]Extension, sync.Locker)
+}
+
+// extendableProtoV1 is an interface implemented by a protocol buffer generated by the previous
+// version of the proto compiler that may be extended.
+type extendableProtoV1 interface {
+	Message
+	ExtensionRangeArray() []ExtensionRange
+	ExtensionMap() map[int32]Extension
+}
+
+// extensionAdapter is a wrapper around extendableProtoV1 that implements extendableProto.
+type extensionAdapter struct {
+	extendableProtoV1
+}
+
+func (e extensionAdapter) extensionsWrite() map[int32]Extension {
+	return e.ExtensionMap()
+}
+
+func (e extensionAdapter) extensionsRead() (map[int32]Extension, sync.Locker) {
+	return e.ExtensionMap(), notLocker{}
+}
+
+// notLocker is a sync.Locker whose Lock and Unlock methods are nops.
+type notLocker struct{}
+
+func (n notLocker) Lock()   {}
+func (n notLocker) Unlock() {}
+
+// extendable returns the extendableProto interface for the given generated proto message.
+// If the proto message has the old extension format, it returns a wrapper that implements
+// the extendableProto interface.
+func extendable(p interface{}) (extendableProto, error) {
+	switch p := p.(type) {
+	case extendableProto:
+		if isNilPtr(p) {
+			return nil, fmt.Errorf("proto: nil %T is not extendable", p)
+		}
+		return p, nil
+	case extendableProtoV1:
+		if isNilPtr(p) {
+			return nil, fmt.Errorf("proto: nil %T is not extendable", p)
+		}
+		return extensionAdapter{p}, nil
+	}
+	// Don't allocate a specific error containing %T:
+	// this is the hot path for Clone and MarshalText.
+	return nil, errNotExtendable
+}
+
+var errNotExtendable = errors.New("proto: not an extendable proto.Message")
+
+func isNilPtr(x interface{}) bool {
+	v := reflect.ValueOf(x)
+	return v.Kind() == reflect.Ptr && v.IsNil()
+}
+
+// XXX_InternalExtensions is an internal representation of proto extensions.
+//
+// Each generated message struct type embeds an anonymous XXX_InternalExtensions field,
+// thus gaining the unexported 'extensions' method, which can be called only from the proto package.
+//
+// The methods of XXX_InternalExtensions are not concurrency safe in general,
+// but calls to logically read-only methods such as has and get may be executed concurrently.
+type XXX_InternalExtensions struct {
+	// The struct must be indirect so that if a user inadvertently copies a
+	// generated message and its embedded XXX_InternalExtensions, they
+	// avoid the mayhem of a copied mutex.
+	//
+	// The mutex serializes all logically read-only operations to p.extensionMap.
+	// It is up to the client to ensure that write operations to p.extensionMap are
+	// mutually exclusive with other accesses.
+	p *struct {
+		mu           sync.Mutex
+		extensionMap map[int32]Extension
+	}
+}
+
+// extensionsWrite returns the extension map, creating it on first use.
+func (e *XXX_InternalExtensions) extensionsWrite() map[int32]Extension {
+	if e.p == nil {
+		e.p = new(struct {
+			mu           sync.Mutex
+			extensionMap map[int32]Extension
+		})
+		e.p.extensionMap = make(map[int32]Extension)
+	}
+	return e.p.extensionMap
+}
+
+// extensionsRead returns the extensions map for read-only use.  It may be nil.
+// The caller must hold the returned mutex's lock when accessing Elements within the map.
+func (e *XXX_InternalExtensions) extensionsRead() (map[int32]Extension, sync.Locker) {
+	if e.p == nil {
+		return nil, nil
+	}
+	return e.p.extensionMap, &e.p.mu
+}
+
+// ExtensionDesc represents an extension specification.
+// Used in generated code from the protocol compiler.
+type ExtensionDesc struct {
+	ExtendedType  Message     // nil pointer to the type that is being extended
+	ExtensionType interface{} // nil pointer to the extension type
+	Field         int32       // field number
+	Name          string      // fully-qualified name of extension, for text formatting
+	Tag           string      // protobuf tag style
+	Filename      string      // name of the file in which the extension is defined
+}
+
+func (ed *ExtensionDesc) repeated() bool {
+	t := reflect.TypeOf(ed.ExtensionType)
+	return t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8
+}
+
+// Extension represents an extension in a message.
+type Extension struct {
+	// When an extension is stored in a message using SetExtension
+	// only desc and value are set. When the message is marshaled
+	// enc will be set to the encoded form of the message.
+	//
+	// When a message is unmarshaled and contains extensions, each
+	// extension will have only enc set. When such an extension is
+	// accessed using GetExtension (or GetExtensions) desc and value
+	// will be set.
+	desc *ExtensionDesc
+
+	// value is a concrete value for the extension field. Let the type of
+	// desc.ExtensionType be the "API type" and the type of Extension.value
+	// be the "storage type". The API type and storage type are the same except:
+	//	* For scalars (except []byte), the API type uses *T,
+	//	while the storage type uses T.
+	//	* For repeated fields, the API type uses []T, while the storage type
+	//	uses *[]T.
+	//
+	// The reason for the divergence is so that the storage type more naturally
+	// matches what is expected of when retrieving the values through the
+	// protobuf reflection APIs.
+	//
+	// The value may only be populated if desc is also populated.
+	value interface{}
+
+	// enc is the raw bytes for the extension field.
+	enc []byte
+}
+
+// SetRawExtension is for testing only.
+func SetRawExtension(base Message, id int32, b []byte) {
+	epb, err := extendable(base)
+	if err != nil {
+		return
+	}
+	extmap := epb.extensionsWrite()
+	extmap[id] = Extension{enc: b}
+}
+
+// isExtensionField returns true iff the given field number is in an extension range.
+func isExtensionField(pb extendableProto, field int32) bool {
+	for _, er := range pb.ExtensionRangeArray() {
+		if er.Start <= field && field <= er.End {
+			return true
+		}
+	}
+	return false
+}
+
+// checkExtensionTypes checks that the given extension is valid for pb.
+func checkExtensionTypes(pb extendableProto, extension *ExtensionDesc) error {
+	var pbi interface{} = pb
+	// Check the extended type.
+	if ea, ok := pbi.(extensionAdapter); ok {
+		pbi = ea.extendableProtoV1
+	}
+	if a, b := reflect.TypeOf(pbi), reflect.TypeOf(extension.ExtendedType); a != b {
+		return fmt.Errorf("proto: bad extended type; %v does not extend %v", b, a)
+	}
+	// Check the range.
+	if !isExtensionField(pb, extension.Field) {
+		return errors.New("proto: bad extension number; not in declared ranges")
+	}
+	return nil
+}
+
+// extPropKey is sufficient to uniquely identify an extension.
+type extPropKey struct {
+	base  reflect.Type
+	field int32
+}
+
+var extProp = struct {
+	sync.RWMutex
+	m map[extPropKey]*Properties
+}{
+	m: make(map[extPropKey]*Properties),
+}
+
+func extensionProperties(ed *ExtensionDesc) *Properties {
+	key := extPropKey{base: reflect.TypeOf(ed.ExtendedType), field: ed.Field}
+
+	extProp.RLock()
+	if prop, ok := extProp.m[key]; ok {
+		extProp.RUnlock()
+		return prop
+	}
+	extProp.RUnlock()
+
+	extProp.Lock()
+	defer extProp.Unlock()
+	// Check again.
+	if prop, ok := extProp.m[key]; ok {
+		return prop
+	}
+
+	prop := new(Properties)
+	prop.Init(reflect.TypeOf(ed.ExtensionType), "unknown_name", ed.Tag, nil)
+	extProp.m[key] = prop
+	return prop
+}
+
+// HasExtension returns whether the given extension is present in pb.
+func HasExtension(pb Message, extension *ExtensionDesc) bool {
+	// TODO: Check types, field numbers, etc.?
+	epb, err := extendable(pb)
+	if err != nil {
+		return false
+	}
+	extmap, mu := epb.extensionsRead()
+	if extmap == nil {
+		return false
+	}
+	mu.Lock()
+	_, ok := extmap[extension.Field]
+	mu.Unlock()
+	return ok
+}
+
+// ClearExtension removes the given extension from pb.
+func ClearExtension(pb Message, extension *ExtensionDesc) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return
+	}
+	// TODO: Check types, field numbers, etc.?
+	extmap := epb.extensionsWrite()
+	delete(extmap, extension.Field)
+}
+
+// GetExtension retrieves a proto2 extended field from pb.
+//
+// If the descriptor is type complete (i.e., ExtensionDesc.ExtensionType is non-nil),
+// then GetExtension parses the encoded field and returns a Go value of the specified type.
+// If the field is not present, then the default value is returned (if one is specified),
+// otherwise ErrMissingExtension is reported.
+//
+// If the descriptor is not type complete (i.e., ExtensionDesc.ExtensionType is nil),
+// then GetExtension returns the raw encoded bytes of the field extension.
+func GetExtension(pb Message, extension *ExtensionDesc) (interface{}, error) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+
+	if extension.ExtendedType != nil {
+		// can only check type if this is a complete descriptor
+		if err := checkExtensionTypes(epb, extension); err != nil {
+			return nil, err
+		}
+	}
+
+	emap, mu := epb.extensionsRead()
+	if emap == nil {
+		return defaultExtensionValue(extension)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	e, ok := emap[extension.Field]
+	if !ok {
+		// defaultExtensionValue returns the default value or
+		// ErrMissingExtension if there is no default.
+		return defaultExtensionValue(extension)
+	}
+
+	if e.value != nil {
+		// Already decoded. Check the descriptor, though.
+		if e.desc != extension {
+			// This shouldn't happen. If it does, it means that
+			// GetExtension was called twice with two different
+			// descriptors with the same field number.
+			return nil, errors.New("proto: descriptor conflict")
+		}
+		return extensionAsLegacyType(e.value), nil
+	}
+
+	if extension.ExtensionType == nil {
+		// incomplete descriptor
+		return e.enc, nil
+	}
+
+	v, err := decodeExtension(e.enc, extension)
+	if err != nil {
+		return nil, err
+	}
+
+	// Remember the decoded version and drop the encoded version.
+	// That way it is safe to mutate what we return.
+	e.value = extensionAsStorageType(v)
+	e.desc = extension
+	e.enc = nil
+	emap[extension.Field] = e
+	return extensionAsLegacyType(e.value), nil
+}
+
+// defaultExtensionValue returns the default value for extension.
+// If no default for an extension is defined ErrMissingExtension is returned.
+func defaultExtensionValue(extension *ExtensionDesc) (interface{}, error) {
+	if extension.ExtensionType == nil {
+		// incomplete descriptor, so no default
+		return nil, ErrMissingExtension
+	}
+
+	t := reflect.TypeOf(extension.ExtensionType)
+	props := extensionProperties(extension)
+
+	sf, _, err := fieldDefault(t, props)
+	if err != nil {
+		return nil, err
+	}
+
+	if sf == nil || sf.value == nil {
+		// There is no default value.
+		return nil, ErrMissingExtension
+	}
+
+	if t.Kind() != reflect.Ptr {
+		// We do not need to return a Ptr, we can directly return sf.value.
+		return sf.value, nil
+	}
+
+	// We need to return an interface{} that is a pointer to sf.value.
+	value := reflect.New(t).Elem()
+	value.Set(reflect.New(value.Type().Elem()))
+	if sf.kind == reflect.Int32 {
+		// We may have an int32 or an enum, but the underlying data is int32.
+		// Since we can't set an int32 into a non int32 reflect.value directly
+		// set it as a int32.
+		value.Elem().SetInt(int64(sf.value.(int32)))
+	} else {
+		value.Elem().Set(reflect.ValueOf(sf.value))
+	}
+	return value.Interface(), nil
+}
+
+// decodeExtension decodes an extension encoded in b.
+func decodeExtension(b []byte, extension *ExtensionDesc) (interface{}, error) {
+	t := reflect.TypeOf(extension.ExtensionType)
+	unmarshal := typeUnmarshaler(t, extension.Tag)
+
+	// t is a pointer to a struct, pointer to basic type or a slice.
+	// Allocate space to store the pointer/slice.
+	value := reflect.New(t).Elem()
+
+	var err error
+	for {
+		x, n := decodeVarint(b)
+		if n == 0 {
+			return nil, io.ErrUnexpectedEOF
+		}
+		b = b[n:]
+		wire := int(x) & 7
+
+		b, err = unmarshal(b, valToPointer(value.Addr()), wire)
+		if err != nil {
+			return nil, err
+		}
+
+		if len(b) == 0 {
+			break
+		}
+	}
+	return value.Interface(), nil
+}
+
+// GetExtensions returns a slice of the extensions present in pb that are also listed in es.
+// The returned slice has the same length as es; missing extensions will appear as nil elements.
+func GetExtensions(pb Message, es []*ExtensionDesc) (extensions []interface{}, err error) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+	extensions = make([]interface{}, len(es))
+	for i, e := range es {
+		extensions[i], err = GetExtension(epb, e)
+		if err == ErrMissingExtension {
+			err = nil
+		}
+		if err != nil {
+			return
+		}
+	}
+	return
+}
+
+// ExtensionDescs returns a new slice containing pb's extension descriptors, in undefined order.
+// For non-registered extensions, ExtensionDescs returns an incomplete descriptor containing
+// just the Field field, which defines the extension's field number.
+func ExtensionDescs(pb Message) ([]*ExtensionDesc, error) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return nil, err
+	}
+	registeredExtensions := RegisteredExtensions(pb)
+
+	emap, mu := epb.extensionsRead()
+	if emap == nil {
+		return nil, nil
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	extensions := make([]*ExtensionDesc, 0, len(emap))
+	for extid, e := range emap {
+		desc := e.desc
+		if desc == nil {
+			desc = registeredExtensions[extid]
+			if desc == nil {
+				desc = &ExtensionDesc{Field: extid}
+			}
+		}
+
+		extensions = append(extensions, desc)
+	}
+	return extensions, nil
+}
+
+// SetExtension sets the specified extension of pb to the specified value.
+func SetExtension(pb Message, extension *ExtensionDesc, value interface{}) error {
+	epb, err := extendable(pb)
+	if err != nil {
+		return err
+	}
+	if err := checkExtensionTypes(epb, extension); err != nil {
+		return err
+	}
+	typ := reflect.TypeOf(extension.ExtensionType)
+	if typ != reflect.TypeOf(value) {
+		return fmt.Errorf("proto: bad extension value type. got: %T, want: %T", value, extension.ExtensionType)
+	}
+	// nil extension values need to be caught early, because the
+	// encoder can't distinguish an ErrNil due to a nil extension
+	// from an ErrNil due to a missing field. Extensions are
+	// always optional, so the encoder would just swallow the error
+	// and drop all the extensions from the encoded message.
+	if reflect.ValueOf(value).IsNil() {
+		return fmt.Errorf("proto: SetExtension called with nil value of type %T", value)
+	}
+
+	extmap := epb.extensionsWrite()
+	extmap[extension.Field] = Extension{desc: extension, value: extensionAsStorageType(value)}
+	return nil
+}
+
+// ClearAllExtensions clears all extensions from pb.
+func ClearAllExtensions(pb Message) {
+	epb, err := extendable(pb)
+	if err != nil {
+		return
+	}
+	m := epb.extensionsWrite()
+	for k := range m {
+		delete(m, k)
+	}
+}
+
+// A global registry of extensions.
+// The generated code will register the generated descriptors by calling RegisterExtension.
+
+var extensionMaps = make(map[reflect.Type]map[int32]*ExtensionDesc)
+
+// RegisterExtension is called from the generated code.
+func RegisterExtension(desc *ExtensionDesc) {
+	st := reflect.TypeOf(desc.ExtendedType).Elem()
+	m := extensionMaps[st]
+	if m == nil {
+		m = make(map[int32]*ExtensionDesc)
+		extensionMaps[st] = m
+	}
+	if _, ok := m[desc.Field]; ok {
+		panic("proto: duplicate extension registered: " + st.String() + " " + strconv.Itoa(int(desc.Field)))
+	}
+	m[desc.Field] = desc
+}
+
+// RegisteredExtensions returns a map of the registered extensions of a
+// protocol buffer struct, indexed by the extension number.
+// The argument pb should be a nil pointer to the struct type.
+func RegisteredExtensions(pb Message) map[int32]*ExtensionDesc {
+	return extensionMaps[reflect.TypeOf(pb).Elem()]
+}
+
+// extensionAsLegacyType converts an value in the storage type as the API type.
+// See Extension.value.
+func extensionAsLegacyType(v interface{}) interface{} {
+	switch rv := reflect.ValueOf(v); rv.Kind() {
+	case reflect.Bool, reflect.Int32, reflect.Int64, reflect.Uint32, reflect.Uint64, reflect.Float32, reflect.Float64, reflect.String:
+		// Represent primitive types as a pointer to the value.
+		rv2 := reflect.New(rv.Type())
+		rv2.Elem().Set(rv)
+		v = rv2.Interface()
+	case reflect.Ptr:
+		// Represent slice types as the value itself.
+		switch rv.Type().Elem().Kind() {
+		case reflect.Slice:
+			if rv.IsNil() {
+				v = reflect.Zero(rv.Type().Elem()).Interface()
+			} else {
+				v = rv.Elem().Interface()
+			}
+		}
+	}
+	return v
+}
+
+// extensionAsStorageType converts an value in the API type as the storage type.
+// See Extension.value.
+func extensionAsStorageType(v interface{}) interface{} {
+	switch rv := reflect.ValueOf(v); rv.Kind() {
+	case reflect.Ptr:
+		// Represent slice types as the value itself.
+		switch rv.Type().Elem().Kind() {
+		case reflect.Bool, reflect.Int32, reflect.Int64, reflect.Uint32, reflect.Uint64, reflect.Float32, reflect.Float64, reflect.String:
+			if rv.IsNil() {
+				v = reflect.Zero(rv.Type().Elem()).Interface()
+			} else {
+				v = rv.Elem().Interface()
+			}
+		}
+	case reflect.Slice:
+		// Represent slice types as a pointer to the value.
+		if rv.Type().Elem().Kind() != reflect.Uint8 {
+			rv2 := reflect.New(rv.Type())
+			rv2.Elem().Set(rv)
+			v = rv2.Interface()
+		}
+	}
+	return v
+}
diff --git a/vendor/github.com/golang/protobuf/proto/lib.go b/vendor/github.com/golang/protobuf/proto/lib.go
new file mode 100644
index 0000000..fdd328b
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/lib.go
@@ -0,0 +1,965 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+/*
+Package proto converts data structures to and from the wire format of
+protocol buffers.  It works in concert with the Go source code generated
+for .proto files by the protocol compiler.
+
+A summary of the properties of the protocol buffer interface
+for a protocol buffer variable v:
+
+  - Names are turned from camel_case to CamelCase for export.
+  - There are no methods on v to set fields; just treat
+	them as structure fields.
+  - There are getters that return a field's value if set,
+	and return the field's default value if unset.
+	The getters work even if the receiver is a nil message.
+  - The zero value for a struct is its correct initialization state.
+	All desired fields must be set before marshaling.
+  - A Reset() method will restore a protobuf struct to its zero state.
+  - Non-repeated fields are pointers to the values; nil means unset.
+	That is, optional or required field int32 f becomes F *int32.
+  - Repeated fields are slices.
+  - Helper functions are available to aid the setting of fields.
+	msg.Foo = proto.String("hello") // set field
+  - Constants are defined to hold the default values of all fields that
+	have them.  They have the form Default_StructName_FieldName.
+	Because the getter methods handle defaulted values,
+	direct use of these constants should be rare.
+  - Enums are given type names and maps from names to values.
+	Enum values are prefixed by the enclosing message's name, or by the
+	enum's type name if it is a top-level enum. Enum types have a String
+	method, and a Enum method to assist in message construction.
+  - Nested messages, groups and enums have type names prefixed with the name of
+	the surrounding message type.
+  - Extensions are given descriptor names that start with E_,
+	followed by an underscore-delimited list of the nested messages
+	that contain it (if any) followed by the CamelCased name of the
+	extension field itself.  HasExtension, ClearExtension, GetExtension
+	and SetExtension are functions for manipulating extensions.
+  - Oneof field sets are given a single field in their message,
+	with distinguished wrapper types for each possible field value.
+  - Marshal and Unmarshal are functions to encode and decode the wire format.
+
+When the .proto file specifies `syntax="proto3"`, there are some differences:
+
+  - Non-repeated fields of non-message type are values instead of pointers.
+  - Enum types do not get an Enum method.
+
+The simplest way to describe this is to see an example.
+Given file test.proto, containing
+
+	package example;
+
+	enum FOO { X = 17; }
+
+	message Test {
+	  required string label = 1;
+	  optional int32 type = 2 [default=77];
+	  repeated int64 reps = 3;
+	  optional group OptionalGroup = 4 {
+	    required string RequiredField = 5;
+	  }
+	  oneof union {
+	    int32 number = 6;
+	    string name = 7;
+	  }
+	}
+
+The resulting file, test.pb.go, is:
+
+	package example
+
+	import proto "github.com/golang/protobuf/proto"
+	import math "math"
+
+	type FOO int32
+	const (
+		FOO_X FOO = 17
+	)
+	var FOO_name = map[int32]string{
+		17: "X",
+	}
+	var FOO_value = map[string]int32{
+		"X": 17,
+	}
+
+	func (x FOO) Enum() *FOO {
+		p := new(FOO)
+		*p = x
+		return p
+	}
+	func (x FOO) String() string {
+		return proto.EnumName(FOO_name, int32(x))
+	}
+	func (x *FOO) UnmarshalJSON(data []byte) error {
+		value, err := proto.UnmarshalJSONEnum(FOO_value, data)
+		if err != nil {
+			return err
+		}
+		*x = FOO(value)
+		return nil
+	}
+
+	type Test struct {
+		Label         *string             `protobuf:"bytes,1,req,name=label" json:"label,omitempty"`
+		Type          *int32              `protobuf:"varint,2,opt,name=type,def=77" json:"type,omitempty"`
+		Reps          []int64             `protobuf:"varint,3,rep,name=reps" json:"reps,omitempty"`
+		Optionalgroup *Test_OptionalGroup `protobuf:"group,4,opt,name=OptionalGroup" json:"optionalgroup,omitempty"`
+		// Types that are valid to be assigned to Union:
+		//	*Test_Number
+		//	*Test_Name
+		Union            isTest_Union `protobuf_oneof:"union"`
+		XXX_unrecognized []byte       `json:"-"`
+	}
+	func (m *Test) Reset()         { *m = Test{} }
+	func (m *Test) String() string { return proto.CompactTextString(m) }
+	func (*Test) ProtoMessage() {}
+
+	type isTest_Union interface {
+		isTest_Union()
+	}
+
+	type Test_Number struct {
+		Number int32 `protobuf:"varint,6,opt,name=number"`
+	}
+	type Test_Name struct {
+		Name string `protobuf:"bytes,7,opt,name=name"`
+	}
+
+	func (*Test_Number) isTest_Union() {}
+	func (*Test_Name) isTest_Union()   {}
+
+	func (m *Test) GetUnion() isTest_Union {
+		if m != nil {
+			return m.Union
+		}
+		return nil
+	}
+	const Default_Test_Type int32 = 77
+
+	func (m *Test) GetLabel() string {
+		if m != nil && m.Label != nil {
+			return *m.Label
+		}
+		return ""
+	}
+
+	func (m *Test) GetType() int32 {
+		if m != nil && m.Type != nil {
+			return *m.Type
+		}
+		return Default_Test_Type
+	}
+
+	func (m *Test) GetOptionalgroup() *Test_OptionalGroup {
+		if m != nil {
+			return m.Optionalgroup
+		}
+		return nil
+	}
+
+	type Test_OptionalGroup struct {
+		RequiredField *string `protobuf:"bytes,5,req" json:"RequiredField,omitempty"`
+	}
+	func (m *Test_OptionalGroup) Reset()         { *m = Test_OptionalGroup{} }
+	func (m *Test_OptionalGroup) String() string { return proto.CompactTextString(m) }
+
+	func (m *Test_OptionalGroup) GetRequiredField() string {
+		if m != nil && m.RequiredField != nil {
+			return *m.RequiredField
+		}
+		return ""
+	}
+
+	func (m *Test) GetNumber() int32 {
+		if x, ok := m.GetUnion().(*Test_Number); ok {
+			return x.Number
+		}
+		return 0
+	}
+
+	func (m *Test) GetName() string {
+		if x, ok := m.GetUnion().(*Test_Name); ok {
+			return x.Name
+		}
+		return ""
+	}
+
+	func init() {
+		proto.RegisterEnum("example.FOO", FOO_name, FOO_value)
+	}
+
+To create and play with a Test object:
+
+	package main
+
+	import (
+		"log"
+
+		"github.com/golang/protobuf/proto"
+		pb "./example.pb"
+	)
+
+	func main() {
+		test := &pb.Test{
+			Label: proto.String("hello"),
+			Type:  proto.Int32(17),
+			Reps:  []int64{1, 2, 3},
+			Optionalgroup: &pb.Test_OptionalGroup{
+				RequiredField: proto.String("good bye"),
+			},
+			Union: &pb.Test_Name{"fred"},
+		}
+		data, err := proto.Marshal(test)
+		if err != nil {
+			log.Fatal("marshaling error: ", err)
+		}
+		newTest := &pb.Test{}
+		err = proto.Unmarshal(data, newTest)
+		if err != nil {
+			log.Fatal("unmarshaling error: ", err)
+		}
+		// Now test and newTest contain the same data.
+		if test.GetLabel() != newTest.GetLabel() {
+			log.Fatalf("data mismatch %q != %q", test.GetLabel(), newTest.GetLabel())
+		}
+		// Use a type switch to determine which oneof was set.
+		switch u := test.Union.(type) {
+		case *pb.Test_Number: // u.Number contains the number.
+		case *pb.Test_Name: // u.Name contains the string.
+		}
+		// etc.
+	}
+*/
+package proto
+
+import (
+	"encoding/json"
+	"fmt"
+	"log"
+	"reflect"
+	"sort"
+	"strconv"
+	"sync"
+)
+
+// RequiredNotSetError is an error type returned by either Marshal or Unmarshal.
+// Marshal reports this when a required field is not initialized.
+// Unmarshal reports this when a required field is missing from the wire data.
+type RequiredNotSetError struct{ field string }
+
+func (e *RequiredNotSetError) Error() string {
+	if e.field == "" {
+		return fmt.Sprintf("proto: required field not set")
+	}
+	return fmt.Sprintf("proto: required field %q not set", e.field)
+}
+func (e *RequiredNotSetError) RequiredNotSet() bool {
+	return true
+}
+
+type invalidUTF8Error struct{ field string }
+
+func (e *invalidUTF8Error) Error() string {
+	if e.field == "" {
+		return "proto: invalid UTF-8 detected"
+	}
+	return fmt.Sprintf("proto: field %q contains invalid UTF-8", e.field)
+}
+func (e *invalidUTF8Error) InvalidUTF8() bool {
+	return true
+}
+
+// errInvalidUTF8 is a sentinel error to identify fields with invalid UTF-8.
+// This error should not be exposed to the external API as such errors should
+// be recreated with the field information.
+var errInvalidUTF8 = &invalidUTF8Error{}
+
+// isNonFatal reports whether the error is either a RequiredNotSet error
+// or a InvalidUTF8 error.
+func isNonFatal(err error) bool {
+	if re, ok := err.(interface{ RequiredNotSet() bool }); ok && re.RequiredNotSet() {
+		return true
+	}
+	if re, ok := err.(interface{ InvalidUTF8() bool }); ok && re.InvalidUTF8() {
+		return true
+	}
+	return false
+}
+
+type nonFatal struct{ E error }
+
+// Merge merges err into nf and reports whether it was successful.
+// Otherwise it returns false for any fatal non-nil errors.
+func (nf *nonFatal) Merge(err error) (ok bool) {
+	if err == nil {
+		return true // not an error
+	}
+	if !isNonFatal(err) {
+		return false // fatal error
+	}
+	if nf.E == nil {
+		nf.E = err // store first instance of non-fatal error
+	}
+	return true
+}
+
+// Message is implemented by generated protocol buffer messages.
+type Message interface {
+	Reset()
+	String() string
+	ProtoMessage()
+}
+
+// A Buffer is a buffer manager for marshaling and unmarshaling
+// protocol buffers.  It may be reused between invocations to
+// reduce memory usage.  It is not necessary to use a Buffer;
+// the global functions Marshal and Unmarshal create a
+// temporary Buffer and are fine for most applications.
+type Buffer struct {
+	buf   []byte // encode/decode byte stream
+	index int    // read point
+
+	deterministic bool
+}
+
+// NewBuffer allocates a new Buffer and initializes its internal data to
+// the contents of the argument slice.
+func NewBuffer(e []byte) *Buffer {
+	return &Buffer{buf: e}
+}
+
+// Reset resets the Buffer, ready for marshaling a new protocol buffer.
+func (p *Buffer) Reset() {
+	p.buf = p.buf[0:0] // for reading/writing
+	p.index = 0        // for reading
+}
+
+// SetBuf replaces the internal buffer with the slice,
+// ready for unmarshaling the contents of the slice.
+func (p *Buffer) SetBuf(s []byte) {
+	p.buf = s
+	p.index = 0
+}
+
+// Bytes returns the contents of the Buffer.
+func (p *Buffer) Bytes() []byte { return p.buf }
+
+// SetDeterministic sets whether to use deterministic serialization.
+//
+// Deterministic serialization guarantees that for a given binary, equal
+// messages will always be serialized to the same bytes. This implies:
+//
+//   - Repeated serialization of a message will return the same bytes.
+//   - Different processes of the same binary (which may be executing on
+//     different machines) will serialize equal messages to the same bytes.
+//
+// Note that the deterministic serialization is NOT canonical across
+// languages. It is not guaranteed to remain stable over time. It is unstable
+// across different builds with schema changes due to unknown fields.
+// Users who need canonical serialization (e.g., persistent storage in a
+// canonical form, fingerprinting, etc.) should define their own
+// canonicalization specification and implement their own serializer rather
+// than relying on this API.
+//
+// If deterministic serialization is requested, map entries will be sorted
+// by keys in lexographical order. This is an implementation detail and
+// subject to change.
+func (p *Buffer) SetDeterministic(deterministic bool) {
+	p.deterministic = deterministic
+}
+
+/*
+ * Helper routines for simplifying the creation of optional fields of basic type.
+ */
+
+// Bool is a helper routine that allocates a new bool value
+// to store v and returns a pointer to it.
+func Bool(v bool) *bool {
+	return &v
+}
+
+// Int32 is a helper routine that allocates a new int32 value
+// to store v and returns a pointer to it.
+func Int32(v int32) *int32 {
+	return &v
+}
+
+// Int is a helper routine that allocates a new int32 value
+// to store v and returns a pointer to it, but unlike Int32
+// its argument value is an int.
+func Int(v int) *int32 {
+	p := new(int32)
+	*p = int32(v)
+	return p
+}
+
+// Int64 is a helper routine that allocates a new int64 value
+// to store v and returns a pointer to it.
+func Int64(v int64) *int64 {
+	return &v
+}
+
+// Float32 is a helper routine that allocates a new float32 value
+// to store v and returns a pointer to it.
+func Float32(v float32) *float32 {
+	return &v
+}
+
+// Float64 is a helper routine that allocates a new float64 value
+// to store v and returns a pointer to it.
+func Float64(v float64) *float64 {
+	return &v
+}
+
+// Uint32 is a helper routine that allocates a new uint32 value
+// to store v and returns a pointer to it.
+func Uint32(v uint32) *uint32 {
+	return &v
+}
+
+// Uint64 is a helper routine that allocates a new uint64 value
+// to store v and returns a pointer to it.
+func Uint64(v uint64) *uint64 {
+	return &v
+}
+
+// String is a helper routine that allocates a new string value
+// to store v and returns a pointer to it.
+func String(v string) *string {
+	return &v
+}
+
+// EnumName is a helper function to simplify printing protocol buffer enums
+// by name.  Given an enum map and a value, it returns a useful string.
+func EnumName(m map[int32]string, v int32) string {
+	s, ok := m[v]
+	if ok {
+		return s
+	}
+	return strconv.Itoa(int(v))
+}
+
+// UnmarshalJSONEnum is a helper function to simplify recovering enum int values
+// from their JSON-encoded representation. Given a map from the enum's symbolic
+// names to its int values, and a byte buffer containing the JSON-encoded
+// value, it returns an int32 that can be cast to the enum type by the caller.
+//
+// The function can deal with both JSON representations, numeric and symbolic.
+func UnmarshalJSONEnum(m map[string]int32, data []byte, enumName string) (int32, error) {
+	if data[0] == '"' {
+		// New style: enums are strings.
+		var repr string
+		if err := json.Unmarshal(data, &repr); err != nil {
+			return -1, err
+		}
+		val, ok := m[repr]
+		if !ok {
+			return 0, fmt.Errorf("unrecognized enum %s value %q", enumName, repr)
+		}
+		return val, nil
+	}
+	// Old style: enums are ints.
+	var val int32
+	if err := json.Unmarshal(data, &val); err != nil {
+		return 0, fmt.Errorf("cannot unmarshal %#q into enum %s", data, enumName)
+	}
+	return val, nil
+}
+
+// DebugPrint dumps the encoded data in b in a debugging format with a header
+// including the string s. Used in testing but made available for general debugging.
+func (p *Buffer) DebugPrint(s string, b []byte) {
+	var u uint64
+
+	obuf := p.buf
+	index := p.index
+	p.buf = b
+	p.index = 0
+	depth := 0
+
+	fmt.Printf("\n--- %s ---\n", s)
+
+out:
+	for {
+		for i := 0; i < depth; i++ {
+			fmt.Print("  ")
+		}
+
+		index := p.index
+		if index == len(p.buf) {
+			break
+		}
+
+		op, err := p.DecodeVarint()
+		if err != nil {
+			fmt.Printf("%3d: fetching op err %v\n", index, err)
+			break out
+		}
+		tag := op >> 3
+		wire := op & 7
+
+		switch wire {
+		default:
+			fmt.Printf("%3d: t=%3d unknown wire=%d\n",
+				index, tag, wire)
+			break out
+
+		case WireBytes:
+			var r []byte
+
+			r, err = p.DecodeRawBytes(false)
+			if err != nil {
+				break out
+			}
+			fmt.Printf("%3d: t=%3d bytes [%d]", index, tag, len(r))
+			if len(r) <= 6 {
+				for i := 0; i < len(r); i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+			} else {
+				for i := 0; i < 3; i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+				fmt.Printf(" ..")
+				for i := len(r) - 3; i < len(r); i++ {
+					fmt.Printf(" %.2x", r[i])
+				}
+			}
+			fmt.Printf("\n")
+
+		case WireFixed32:
+			u, err = p.DecodeFixed32()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d fix32 err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d fix32 %d\n", index, tag, u)
+
+		case WireFixed64:
+			u, err = p.DecodeFixed64()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d fix64 err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d fix64 %d\n", index, tag, u)
+
+		case WireVarint:
+			u, err = p.DecodeVarint()
+			if err != nil {
+				fmt.Printf("%3d: t=%3d varint err %v\n", index, tag, err)
+				break out
+			}
+			fmt.Printf("%3d: t=%3d varint %d\n", index, tag, u)
+
+		case WireStartGroup:
+			fmt.Printf("%3d: t=%3d start\n", index, tag)
+			depth++
+
+		case WireEndGroup:
+			depth--
+			fmt.Printf("%3d: t=%3d end\n", index, tag)
+		}
+	}
+
+	if depth != 0 {
+		fmt.Printf("%3d: start-end not balanced %d\n", p.index, depth)
+	}
+	fmt.Printf("\n")
+
+	p.buf = obuf
+	p.index = index
+}
+
+// SetDefaults sets unset protocol buffer fields to their default values.
+// It only modifies fields that are both unset and have defined defaults.
+// It recursively sets default values in any non-nil sub-messages.
+func SetDefaults(pb Message) {
+	setDefaults(reflect.ValueOf(pb), true, false)
+}
+
+// v is a pointer to a struct.
+func setDefaults(v reflect.Value, recur, zeros bool) {
+	v = v.Elem()
+
+	defaultMu.RLock()
+	dm, ok := defaults[v.Type()]
+	defaultMu.RUnlock()
+	if !ok {
+		dm = buildDefaultMessage(v.Type())
+		defaultMu.Lock()
+		defaults[v.Type()] = dm
+		defaultMu.Unlock()
+	}
+
+	for _, sf := range dm.scalars {
+		f := v.Field(sf.index)
+		if !f.IsNil() {
+			// field already set
+			continue
+		}
+		dv := sf.value
+		if dv == nil && !zeros {
+			// no explicit default, and don't want to set zeros
+			continue
+		}
+		fptr := f.Addr().Interface() // **T
+		// TODO: Consider batching the allocations we do here.
+		switch sf.kind {
+		case reflect.Bool:
+			b := new(bool)
+			if dv != nil {
+				*b = dv.(bool)
+			}
+			*(fptr.(**bool)) = b
+		case reflect.Float32:
+			f := new(float32)
+			if dv != nil {
+				*f = dv.(float32)
+			}
+			*(fptr.(**float32)) = f
+		case reflect.Float64:
+			f := new(float64)
+			if dv != nil {
+				*f = dv.(float64)
+			}
+			*(fptr.(**float64)) = f
+		case reflect.Int32:
+			// might be an enum
+			if ft := f.Type(); ft != int32PtrType {
+				// enum
+				f.Set(reflect.New(ft.Elem()))
+				if dv != nil {
+					f.Elem().SetInt(int64(dv.(int32)))
+				}
+			} else {
+				// int32 field
+				i := new(int32)
+				if dv != nil {
+					*i = dv.(int32)
+				}
+				*(fptr.(**int32)) = i
+			}
+		case reflect.Int64:
+			i := new(int64)
+			if dv != nil {
+				*i = dv.(int64)
+			}
+			*(fptr.(**int64)) = i
+		case reflect.String:
+			s := new(string)
+			if dv != nil {
+				*s = dv.(string)
+			}
+			*(fptr.(**string)) = s
+		case reflect.Uint8:
+			// exceptional case: []byte
+			var b []byte
+			if dv != nil {
+				db := dv.([]byte)
+				b = make([]byte, len(db))
+				copy(b, db)
+			} else {
+				b = []byte{}
+			}
+			*(fptr.(*[]byte)) = b
+		case reflect.Uint32:
+			u := new(uint32)
+			if dv != nil {
+				*u = dv.(uint32)
+			}
+			*(fptr.(**uint32)) = u
+		case reflect.Uint64:
+			u := new(uint64)
+			if dv != nil {
+				*u = dv.(uint64)
+			}
+			*(fptr.(**uint64)) = u
+		default:
+			log.Printf("proto: can't set default for field %v (sf.kind=%v)", f, sf.kind)
+		}
+	}
+
+	for _, ni := range dm.nested {
+		f := v.Field(ni)
+		// f is *T or []*T or map[T]*T
+		switch f.Kind() {
+		case reflect.Ptr:
+			if f.IsNil() {
+				continue
+			}
+			setDefaults(f, recur, zeros)
+
+		case reflect.Slice:
+			for i := 0; i < f.Len(); i++ {
+				e := f.Index(i)
+				if e.IsNil() {
+					continue
+				}
+				setDefaults(e, recur, zeros)
+			}
+
+		case reflect.Map:
+			for _, k := range f.MapKeys() {
+				e := f.MapIndex(k)
+				if e.IsNil() {
+					continue
+				}
+				setDefaults(e, recur, zeros)
+			}
+		}
+	}
+}
+
+var (
+	// defaults maps a protocol buffer struct type to a slice of the fields,
+	// with its scalar fields set to their proto-declared non-zero default values.
+	defaultMu sync.RWMutex
+	defaults  = make(map[reflect.Type]defaultMessage)
+
+	int32PtrType = reflect.TypeOf((*int32)(nil))
+)
+
+// defaultMessage represents information about the default values of a message.
+type defaultMessage struct {
+	scalars []scalarField
+	nested  []int // struct field index of nested messages
+}
+
+type scalarField struct {
+	index int          // struct field index
+	kind  reflect.Kind // element type (the T in *T or []T)
+	value interface{}  // the proto-declared default value, or nil
+}
+
+// t is a struct type.
+func buildDefaultMessage(t reflect.Type) (dm defaultMessage) {
+	sprop := GetProperties(t)
+	for _, prop := range sprop.Prop {
+		fi, ok := sprop.decoderTags.get(prop.Tag)
+		if !ok {
+			// XXX_unrecognized
+			continue
+		}
+		ft := t.Field(fi).Type
+
+		sf, nested, err := fieldDefault(ft, prop)
+		switch {
+		case err != nil:
+			log.Print(err)
+		case nested:
+			dm.nested = append(dm.nested, fi)
+		case sf != nil:
+			sf.index = fi
+			dm.scalars = append(dm.scalars, *sf)
+		}
+	}
+
+	return dm
+}
+
+// fieldDefault returns the scalarField for field type ft.
+// sf will be nil if the field can not have a default.
+// nestedMessage will be true if this is a nested message.
+// Note that sf.index is not set on return.
+func fieldDefault(ft reflect.Type, prop *Properties) (sf *scalarField, nestedMessage bool, err error) {
+	var canHaveDefault bool
+	switch ft.Kind() {
+	case reflect.Ptr:
+		if ft.Elem().Kind() == reflect.Struct {
+			nestedMessage = true
+		} else {
+			canHaveDefault = true // proto2 scalar field
+		}
+
+	case reflect.Slice:
+		switch ft.Elem().Kind() {
+		case reflect.Ptr:
+			nestedMessage = true // repeated message
+		case reflect.Uint8:
+			canHaveDefault = true // bytes field
+		}
+
+	case reflect.Map:
+		if ft.Elem().Kind() == reflect.Ptr {
+			nestedMessage = true // map with message values
+		}
+	}
+
+	if !canHaveDefault {
+		if nestedMessage {
+			return nil, true, nil
+		}
+		return nil, false, nil
+	}
+
+	// We now know that ft is a pointer or slice.
+	sf = &scalarField{kind: ft.Elem().Kind()}
+
+	// scalar fields without defaults
+	if !prop.HasDefault {
+		return sf, false, nil
+	}
+
+	// a scalar field: either *T or []byte
+	switch ft.Elem().Kind() {
+	case reflect.Bool:
+		x, err := strconv.ParseBool(prop.Default)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default bool %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.Float32:
+		x, err := strconv.ParseFloat(prop.Default, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default float32 %q: %v", prop.Default, err)
+		}
+		sf.value = float32(x)
+	case reflect.Float64:
+		x, err := strconv.ParseFloat(prop.Default, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default float64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.Int32:
+		x, err := strconv.ParseInt(prop.Default, 10, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default int32 %q: %v", prop.Default, err)
+		}
+		sf.value = int32(x)
+	case reflect.Int64:
+		x, err := strconv.ParseInt(prop.Default, 10, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default int64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	case reflect.String:
+		sf.value = prop.Default
+	case reflect.Uint8:
+		// []byte (not *uint8)
+		sf.value = []byte(prop.Default)
+	case reflect.Uint32:
+		x, err := strconv.ParseUint(prop.Default, 10, 32)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default uint32 %q: %v", prop.Default, err)
+		}
+		sf.value = uint32(x)
+	case reflect.Uint64:
+		x, err := strconv.ParseUint(prop.Default, 10, 64)
+		if err != nil {
+			return nil, false, fmt.Errorf("proto: bad default uint64 %q: %v", prop.Default, err)
+		}
+		sf.value = x
+	default:
+		return nil, false, fmt.Errorf("proto: unhandled def kind %v", ft.Elem().Kind())
+	}
+
+	return sf, false, nil
+}
+
+// mapKeys returns a sort.Interface to be used for sorting the map keys.
+// Map fields may have key types of non-float scalars, strings and enums.
+func mapKeys(vs []reflect.Value) sort.Interface {
+	s := mapKeySorter{vs: vs}
+
+	// Type specialization per https://developers.google.com/protocol-buffers/docs/proto#maps.
+	if len(vs) == 0 {
+		return s
+	}
+	switch vs[0].Kind() {
+	case reflect.Int32, reflect.Int64:
+		s.less = func(a, b reflect.Value) bool { return a.Int() < b.Int() }
+	case reflect.Uint32, reflect.Uint64:
+		s.less = func(a, b reflect.Value) bool { return a.Uint() < b.Uint() }
+	case reflect.Bool:
+		s.less = func(a, b reflect.Value) bool { return !a.Bool() && b.Bool() } // false < true
+	case reflect.String:
+		s.less = func(a, b reflect.Value) bool { return a.String() < b.String() }
+	default:
+		panic(fmt.Sprintf("unsupported map key type: %v", vs[0].Kind()))
+	}
+
+	return s
+}
+
+type mapKeySorter struct {
+	vs   []reflect.Value
+	less func(a, b reflect.Value) bool
+}
+
+func (s mapKeySorter) Len() int      { return len(s.vs) }
+func (s mapKeySorter) Swap(i, j int) { s.vs[i], s.vs[j] = s.vs[j], s.vs[i] }
+func (s mapKeySorter) Less(i, j int) bool {
+	return s.less(s.vs[i], s.vs[j])
+}
+
+// isProto3Zero reports whether v is a zero proto3 value.
+func isProto3Zero(v reflect.Value) bool {
+	switch v.Kind() {
+	case reflect.Bool:
+		return !v.Bool()
+	case reflect.Int32, reflect.Int64:
+		return v.Int() == 0
+	case reflect.Uint32, reflect.Uint64:
+		return v.Uint() == 0
+	case reflect.Float32, reflect.Float64:
+		return v.Float() == 0
+	case reflect.String:
+		return v.String() == ""
+	}
+	return false
+}
+
+const (
+	// ProtoPackageIsVersion3 is referenced from generated protocol buffer files
+	// to assert that that code is compatible with this version of the proto package.
+	ProtoPackageIsVersion3 = true
+
+	// ProtoPackageIsVersion2 is referenced from generated protocol buffer files
+	// to assert that that code is compatible with this version of the proto package.
+	ProtoPackageIsVersion2 = true
+
+	// ProtoPackageIsVersion1 is referenced from generated protocol buffer files
+	// to assert that that code is compatible with this version of the proto package.
+	ProtoPackageIsVersion1 = true
+)
+
+// InternalMessageInfo is a type used internally by generated .pb.go files.
+// This type is not intended to be used by non-generated code.
+// This type is not subject to any compatibility guarantee.
+type InternalMessageInfo struct {
+	marshal   *marshalInfo
+	unmarshal *unmarshalInfo
+	merge     *mergeInfo
+	discard   *discardInfo
+}
diff --git a/vendor/github.com/golang/protobuf/proto/message_set.go b/vendor/github.com/golang/protobuf/proto/message_set.go
new file mode 100644
index 0000000..f48a756
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/message_set.go
@@ -0,0 +1,181 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Support for message sets.
+ */
+
+import (
+	"errors"
+)
+
+// errNoMessageTypeID occurs when a protocol buffer does not have a message type ID.
+// A message type ID is required for storing a protocol buffer in a message set.
+var errNoMessageTypeID = errors.New("proto does not have a message type ID")
+
+// The first two types (_MessageSet_Item and messageSet)
+// model what the protocol compiler produces for the following protocol message:
+//   message MessageSet {
+//     repeated group Item = 1 {
+//       required int32 type_id = 2;
+//       required string message = 3;
+//     };
+//   }
+// That is the MessageSet wire format. We can't use a proto to generate these
+// because that would introduce a circular dependency between it and this package.
+
+type _MessageSet_Item struct {
+	TypeId  *int32 `protobuf:"varint,2,req,name=type_id"`
+	Message []byte `protobuf:"bytes,3,req,name=message"`
+}
+
+type messageSet struct {
+	Item             []*_MessageSet_Item `protobuf:"group,1,rep"`
+	XXX_unrecognized []byte
+	// TODO: caching?
+}
+
+// Make sure messageSet is a Message.
+var _ Message = (*messageSet)(nil)
+
+// messageTypeIder is an interface satisfied by a protocol buffer type
+// that may be stored in a MessageSet.
+type messageTypeIder interface {
+	MessageTypeId() int32
+}
+
+func (ms *messageSet) find(pb Message) *_MessageSet_Item {
+	mti, ok := pb.(messageTypeIder)
+	if !ok {
+		return nil
+	}
+	id := mti.MessageTypeId()
+	for _, item := range ms.Item {
+		if *item.TypeId == id {
+			return item
+		}
+	}
+	return nil
+}
+
+func (ms *messageSet) Has(pb Message) bool {
+	return ms.find(pb) != nil
+}
+
+func (ms *messageSet) Unmarshal(pb Message) error {
+	if item := ms.find(pb); item != nil {
+		return Unmarshal(item.Message, pb)
+	}
+	if _, ok := pb.(messageTypeIder); !ok {
+		return errNoMessageTypeID
+	}
+	return nil // TODO: return error instead?
+}
+
+func (ms *messageSet) Marshal(pb Message) error {
+	msg, err := Marshal(pb)
+	if err != nil {
+		return err
+	}
+	if item := ms.find(pb); item != nil {
+		// reuse existing item
+		item.Message = msg
+		return nil
+	}
+
+	mti, ok := pb.(messageTypeIder)
+	if !ok {
+		return errNoMessageTypeID
+	}
+
+	mtid := mti.MessageTypeId()
+	ms.Item = append(ms.Item, &_MessageSet_Item{
+		TypeId:  &mtid,
+		Message: msg,
+	})
+	return nil
+}
+
+func (ms *messageSet) Reset()         { *ms = messageSet{} }
+func (ms *messageSet) String() string { return CompactTextString(ms) }
+func (*messageSet) ProtoMessage()     {}
+
+// Support for the message_set_wire_format message option.
+
+func skipVarint(buf []byte) []byte {
+	i := 0
+	for ; buf[i]&0x80 != 0; i++ {
+	}
+	return buf[i+1:]
+}
+
+// unmarshalMessageSet decodes the extension map encoded in buf in the message set wire format.
+// It is called by Unmarshal methods on protocol buffer messages with the message_set_wire_format option.
+func unmarshalMessageSet(buf []byte, exts interface{}) error {
+	var m map[int32]Extension
+	switch exts := exts.(type) {
+	case *XXX_InternalExtensions:
+		m = exts.extensionsWrite()
+	case map[int32]Extension:
+		m = exts
+	default:
+		return errors.New("proto: not an extension map")
+	}
+
+	ms := new(messageSet)
+	if err := Unmarshal(buf, ms); err != nil {
+		return err
+	}
+	for _, item := range ms.Item {
+		id := *item.TypeId
+		msg := item.Message
+
+		// Restore wire type and field number varint, plus length varint.
+		// Be careful to preserve duplicate items.
+		b := EncodeVarint(uint64(id)<<3 | WireBytes)
+		if ext, ok := m[id]; ok {
+			// Existing data; rip off the tag and length varint
+			// so we join the new data correctly.
+			// We can assume that ext.enc is set because we are unmarshaling.
+			o := ext.enc[len(b):]   // skip wire type and field number
+			_, n := DecodeVarint(o) // calculate length of length varint
+			o = o[n:]               // skip length varint
+			msg = append(o, msg...) // join old data and new data
+		}
+		b = append(b, EncodeVarint(uint64(len(msg)))...)
+		b = append(b, msg...)
+
+		m[id] = Extension{enc: b}
+	}
+	return nil
+}
diff --git a/vendor/github.com/golang/protobuf/proto/pointer_reflect.go b/vendor/github.com/golang/protobuf/proto/pointer_reflect.go
new file mode 100644
index 0000000..94fa919
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/pointer_reflect.go
@@ -0,0 +1,360 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2012 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build purego appengine js
+
+// This file contains an implementation of proto field accesses using package reflect.
+// It is slower than the code in pointer_unsafe.go but it avoids package unsafe and can
+// be used on App Engine.
+
+package proto
+
+import (
+	"reflect"
+	"sync"
+)
+
+const unsafeAllowed = false
+
+// A field identifies a field in a struct, accessible from a pointer.
+// In this implementation, a field is identified by the sequence of field indices
+// passed to reflect's FieldByIndex.
+type field []int
+
+// toField returns a field equivalent to the given reflect field.
+func toField(f *reflect.StructField) field {
+	return f.Index
+}
+
+// invalidField is an invalid field identifier.
+var invalidField = field(nil)
+
+// zeroField is a noop when calling pointer.offset.
+var zeroField = field([]int{})
+
+// IsValid reports whether the field identifier is valid.
+func (f field) IsValid() bool { return f != nil }
+
+// The pointer type is for the table-driven decoder.
+// The implementation here uses a reflect.Value of pointer type to
+// create a generic pointer. In pointer_unsafe.go we use unsafe
+// instead of reflect to implement the same (but faster) interface.
+type pointer struct {
+	v reflect.Value
+}
+
+// toPointer converts an interface of pointer type to a pointer
+// that points to the same target.
+func toPointer(i *Message) pointer {
+	return pointer{v: reflect.ValueOf(*i)}
+}
+
+// toAddrPointer converts an interface to a pointer that points to
+// the interface data.
+func toAddrPointer(i *interface{}, isptr, deref bool) pointer {
+	v := reflect.ValueOf(*i)
+	u := reflect.New(v.Type())
+	u.Elem().Set(v)
+	if deref {
+		u = u.Elem()
+	}
+	return pointer{v: u}
+}
+
+// valToPointer converts v to a pointer.  v must be of pointer type.
+func valToPointer(v reflect.Value) pointer {
+	return pointer{v: v}
+}
+
+// offset converts from a pointer to a structure to a pointer to
+// one of its fields.
+func (p pointer) offset(f field) pointer {
+	return pointer{v: p.v.Elem().FieldByIndex(f).Addr()}
+}
+
+func (p pointer) isNil() bool {
+	return p.v.IsNil()
+}
+
+// grow updates the slice s in place to make it one element longer.
+// s must be addressable.
+// Returns the (addressable) new element.
+func grow(s reflect.Value) reflect.Value {
+	n, m := s.Len(), s.Cap()
+	if n < m {
+		s.SetLen(n + 1)
+	} else {
+		s.Set(reflect.Append(s, reflect.Zero(s.Type().Elem())))
+	}
+	return s.Index(n)
+}
+
+func (p pointer) toInt64() *int64 {
+	return p.v.Interface().(*int64)
+}
+func (p pointer) toInt64Ptr() **int64 {
+	return p.v.Interface().(**int64)
+}
+func (p pointer) toInt64Slice() *[]int64 {
+	return p.v.Interface().(*[]int64)
+}
+
+var int32ptr = reflect.TypeOf((*int32)(nil))
+
+func (p pointer) toInt32() *int32 {
+	return p.v.Convert(int32ptr).Interface().(*int32)
+}
+
+// The toInt32Ptr/Slice methods don't work because of enums.
+// Instead, we must use set/get methods for the int32ptr/slice case.
+/*
+	func (p pointer) toInt32Ptr() **int32 {
+		return p.v.Interface().(**int32)
+}
+	func (p pointer) toInt32Slice() *[]int32 {
+		return p.v.Interface().(*[]int32)
+}
+*/
+func (p pointer) getInt32Ptr() *int32 {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		return p.v.Elem().Interface().(*int32)
+	}
+	// an enum
+	return p.v.Elem().Convert(int32PtrType).Interface().(*int32)
+}
+func (p pointer) setInt32Ptr(v int32) {
+	// Allocate value in a *int32. Possibly convert that to a *enum.
+	// Then assign it to a **int32 or **enum.
+	// Note: we can convert *int32 to *enum, but we can't convert
+	// **int32 to **enum!
+	p.v.Elem().Set(reflect.ValueOf(&v).Convert(p.v.Type().Elem()))
+}
+
+// getInt32Slice copies []int32 from p as a new slice.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) getInt32Slice() []int32 {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		return p.v.Elem().Interface().([]int32)
+	}
+	// an enum
+	// Allocate a []int32, then assign []enum's values into it.
+	// Note: we can't convert []enum to []int32.
+	slice := p.v.Elem()
+	s := make([]int32, slice.Len())
+	for i := 0; i < slice.Len(); i++ {
+		s[i] = int32(slice.Index(i).Int())
+	}
+	return s
+}
+
+// setInt32Slice copies []int32 into p as a new slice.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) setInt32Slice(v []int32) {
+	if p.v.Type().Elem().Elem() == reflect.TypeOf(int32(0)) {
+		// raw int32 type
+		p.v.Elem().Set(reflect.ValueOf(v))
+		return
+	}
+	// an enum
+	// Allocate a []enum, then assign []int32's values into it.
+	// Note: we can't convert []enum to []int32.
+	slice := reflect.MakeSlice(p.v.Type().Elem(), len(v), cap(v))
+	for i, x := range v {
+		slice.Index(i).SetInt(int64(x))
+	}
+	p.v.Elem().Set(slice)
+}
+func (p pointer) appendInt32Slice(v int32) {
+	grow(p.v.Elem()).SetInt(int64(v))
+}
+
+func (p pointer) toUint64() *uint64 {
+	return p.v.Interface().(*uint64)
+}
+func (p pointer) toUint64Ptr() **uint64 {
+	return p.v.Interface().(**uint64)
+}
+func (p pointer) toUint64Slice() *[]uint64 {
+	return p.v.Interface().(*[]uint64)
+}
+func (p pointer) toUint32() *uint32 {
+	return p.v.Interface().(*uint32)
+}
+func (p pointer) toUint32Ptr() **uint32 {
+	return p.v.Interface().(**uint32)
+}
+func (p pointer) toUint32Slice() *[]uint32 {
+	return p.v.Interface().(*[]uint32)
+}
+func (p pointer) toBool() *bool {
+	return p.v.Interface().(*bool)
+}
+func (p pointer) toBoolPtr() **bool {
+	return p.v.Interface().(**bool)
+}
+func (p pointer) toBoolSlice() *[]bool {
+	return p.v.Interface().(*[]bool)
+}
+func (p pointer) toFloat64() *float64 {
+	return p.v.Interface().(*float64)
+}
+func (p pointer) toFloat64Ptr() **float64 {
+	return p.v.Interface().(**float64)
+}
+func (p pointer) toFloat64Slice() *[]float64 {
+	return p.v.Interface().(*[]float64)
+}
+func (p pointer) toFloat32() *float32 {
+	return p.v.Interface().(*float32)
+}
+func (p pointer) toFloat32Ptr() **float32 {
+	return p.v.Interface().(**float32)
+}
+func (p pointer) toFloat32Slice() *[]float32 {
+	return p.v.Interface().(*[]float32)
+}
+func (p pointer) toString() *string {
+	return p.v.Interface().(*string)
+}
+func (p pointer) toStringPtr() **string {
+	return p.v.Interface().(**string)
+}
+func (p pointer) toStringSlice() *[]string {
+	return p.v.Interface().(*[]string)
+}
+func (p pointer) toBytes() *[]byte {
+	return p.v.Interface().(*[]byte)
+}
+func (p pointer) toBytesSlice() *[][]byte {
+	return p.v.Interface().(*[][]byte)
+}
+func (p pointer) toExtensions() *XXX_InternalExtensions {
+	return p.v.Interface().(*XXX_InternalExtensions)
+}
+func (p pointer) toOldExtensions() *map[int32]Extension {
+	return p.v.Interface().(*map[int32]Extension)
+}
+func (p pointer) getPointer() pointer {
+	return pointer{v: p.v.Elem()}
+}
+func (p pointer) setPointer(q pointer) {
+	p.v.Elem().Set(q.v)
+}
+func (p pointer) appendPointer(q pointer) {
+	grow(p.v.Elem()).Set(q.v)
+}
+
+// getPointerSlice copies []*T from p as a new []pointer.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) getPointerSlice() []pointer {
+	if p.v.IsNil() {
+		return nil
+	}
+	n := p.v.Elem().Len()
+	s := make([]pointer, n)
+	for i := 0; i < n; i++ {
+		s[i] = pointer{v: p.v.Elem().Index(i)}
+	}
+	return s
+}
+
+// setPointerSlice copies []pointer into p as a new []*T.
+// This behavior differs from the implementation in pointer_unsafe.go.
+func (p pointer) setPointerSlice(v []pointer) {
+	if v == nil {
+		p.v.Elem().Set(reflect.New(p.v.Elem().Type()).Elem())
+		return
+	}
+	s := reflect.MakeSlice(p.v.Elem().Type(), 0, len(v))
+	for _, p := range v {
+		s = reflect.Append(s, p.v)
+	}
+	p.v.Elem().Set(s)
+}
+
+// getInterfacePointer returns a pointer that points to the
+// interface data of the interface pointed by p.
+func (p pointer) getInterfacePointer() pointer {
+	if p.v.Elem().IsNil() {
+		return pointer{v: p.v.Elem()}
+	}
+	return pointer{v: p.v.Elem().Elem().Elem().Field(0).Addr()} // *interface -> interface -> *struct -> struct
+}
+
+func (p pointer) asPointerTo(t reflect.Type) reflect.Value {
+	// TODO: check that p.v.Type().Elem() == t?
+	return p.v
+}
+
+func atomicLoadUnmarshalInfo(p **unmarshalInfo) *unmarshalInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreUnmarshalInfo(p **unmarshalInfo, v *unmarshalInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadMarshalInfo(p **marshalInfo) *marshalInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreMarshalInfo(p **marshalInfo, v *marshalInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadMergeInfo(p **mergeInfo) *mergeInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreMergeInfo(p **mergeInfo, v *mergeInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+func atomicLoadDiscardInfo(p **discardInfo) *discardInfo {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	return *p
+}
+func atomicStoreDiscardInfo(p **discardInfo, v *discardInfo) {
+	atomicLock.Lock()
+	defer atomicLock.Unlock()
+	*p = v
+}
+
+var atomicLock sync.Mutex
diff --git a/vendor/github.com/golang/protobuf/proto/pointer_unsafe.go b/vendor/github.com/golang/protobuf/proto/pointer_unsafe.go
new file mode 100644
index 0000000..dbfffe0
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/pointer_unsafe.go
@@ -0,0 +1,313 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2012 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// +build !purego,!appengine,!js
+
+// This file contains the implementation of the proto field accesses using package unsafe.
+
+package proto
+
+import (
+	"reflect"
+	"sync/atomic"
+	"unsafe"
+)
+
+const unsafeAllowed = true
+
+// A field identifies a field in a struct, accessible from a pointer.
+// In this implementation, a field is identified by its byte offset from the start of the struct.
+type field uintptr
+
+// toField returns a field equivalent to the given reflect field.
+func toField(f *reflect.StructField) field {
+	return field(f.Offset)
+}
+
+// invalidField is an invalid field identifier.
+const invalidField = ^field(0)
+
+// zeroField is a noop when calling pointer.offset.
+const zeroField = field(0)
+
+// IsValid reports whether the field identifier is valid.
+func (f field) IsValid() bool {
+	return f != invalidField
+}
+
+// The pointer type below is for the new table-driven encoder/decoder.
+// The implementation here uses unsafe.Pointer to create a generic pointer.
+// In pointer_reflect.go we use reflect instead of unsafe to implement
+// the same (but slower) interface.
+type pointer struct {
+	p unsafe.Pointer
+}
+
+// size of pointer
+var ptrSize = unsafe.Sizeof(uintptr(0))
+
+// toPointer converts an interface of pointer type to a pointer
+// that points to the same target.
+func toPointer(i *Message) pointer {
+	// Super-tricky - read pointer out of data word of interface value.
+	// Saves ~25ns over the equivalent:
+	// return valToPointer(reflect.ValueOf(*i))
+	return pointer{p: (*[2]unsafe.Pointer)(unsafe.Pointer(i))[1]}
+}
+
+// toAddrPointer converts an interface to a pointer that points to
+// the interface data.
+func toAddrPointer(i *interface{}, isptr, deref bool) (p pointer) {
+	// Super-tricky - read or get the address of data word of interface value.
+	if isptr {
+		// The interface is of pointer type, thus it is a direct interface.
+		// The data word is the pointer data itself. We take its address.
+		p = pointer{p: unsafe.Pointer(uintptr(unsafe.Pointer(i)) + ptrSize)}
+	} else {
+		// The interface is not of pointer type. The data word is the pointer
+		// to the data.
+		p = pointer{p: (*[2]unsafe.Pointer)(unsafe.Pointer(i))[1]}
+	}
+	if deref {
+		p.p = *(*unsafe.Pointer)(p.p)
+	}
+	return p
+}
+
+// valToPointer converts v to a pointer. v must be of pointer type.
+func valToPointer(v reflect.Value) pointer {
+	return pointer{p: unsafe.Pointer(v.Pointer())}
+}
+
+// offset converts from a pointer to a structure to a pointer to
+// one of its fields.
+func (p pointer) offset(f field) pointer {
+	// For safety, we should panic if !f.IsValid, however calling panic causes
+	// this to no longer be inlineable, which is a serious performance cost.
+	/*
+		if !f.IsValid() {
+			panic("invalid field")
+		}
+	*/
+	return pointer{p: unsafe.Pointer(uintptr(p.p) + uintptr(f))}
+}
+
+func (p pointer) isNil() bool {
+	return p.p == nil
+}
+
+func (p pointer) toInt64() *int64 {
+	return (*int64)(p.p)
+}
+func (p pointer) toInt64Ptr() **int64 {
+	return (**int64)(p.p)
+}
+func (p pointer) toInt64Slice() *[]int64 {
+	return (*[]int64)(p.p)
+}
+func (p pointer) toInt32() *int32 {
+	return (*int32)(p.p)
+}
+
+// See pointer_reflect.go for why toInt32Ptr/Slice doesn't exist.
+/*
+	func (p pointer) toInt32Ptr() **int32 {
+		return (**int32)(p.p)
+	}
+	func (p pointer) toInt32Slice() *[]int32 {
+		return (*[]int32)(p.p)
+	}
+*/
+func (p pointer) getInt32Ptr() *int32 {
+	return *(**int32)(p.p)
+}
+func (p pointer) setInt32Ptr(v int32) {
+	*(**int32)(p.p) = &v
+}
+
+// getInt32Slice loads a []int32 from p.
+// The value returned is aliased with the original slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) getInt32Slice() []int32 {
+	return *(*[]int32)(p.p)
+}
+
+// setInt32Slice stores a []int32 to p.
+// The value set is aliased with the input slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) setInt32Slice(v []int32) {
+	*(*[]int32)(p.p) = v
+}
+
+// TODO: Can we get rid of appendInt32Slice and use setInt32Slice instead?
+func (p pointer) appendInt32Slice(v int32) {
+	s := (*[]int32)(p.p)
+	*s = append(*s, v)
+}
+
+func (p pointer) toUint64() *uint64 {
+	return (*uint64)(p.p)
+}
+func (p pointer) toUint64Ptr() **uint64 {
+	return (**uint64)(p.p)
+}
+func (p pointer) toUint64Slice() *[]uint64 {
+	return (*[]uint64)(p.p)
+}
+func (p pointer) toUint32() *uint32 {
+	return (*uint32)(p.p)
+}
+func (p pointer) toUint32Ptr() **uint32 {
+	return (**uint32)(p.p)
+}
+func (p pointer) toUint32Slice() *[]uint32 {
+	return (*[]uint32)(p.p)
+}
+func (p pointer) toBool() *bool {
+	return (*bool)(p.p)
+}
+func (p pointer) toBoolPtr() **bool {
+	return (**bool)(p.p)
+}
+func (p pointer) toBoolSlice() *[]bool {
+	return (*[]bool)(p.p)
+}
+func (p pointer) toFloat64() *float64 {
+	return (*float64)(p.p)
+}
+func (p pointer) toFloat64Ptr() **float64 {
+	return (**float64)(p.p)
+}
+func (p pointer) toFloat64Slice() *[]float64 {
+	return (*[]float64)(p.p)
+}
+func (p pointer) toFloat32() *float32 {
+	return (*float32)(p.p)
+}
+func (p pointer) toFloat32Ptr() **float32 {
+	return (**float32)(p.p)
+}
+func (p pointer) toFloat32Slice() *[]float32 {
+	return (*[]float32)(p.p)
+}
+func (p pointer) toString() *string {
+	return (*string)(p.p)
+}
+func (p pointer) toStringPtr() **string {
+	return (**string)(p.p)
+}
+func (p pointer) toStringSlice() *[]string {
+	return (*[]string)(p.p)
+}
+func (p pointer) toBytes() *[]byte {
+	return (*[]byte)(p.p)
+}
+func (p pointer) toBytesSlice() *[][]byte {
+	return (*[][]byte)(p.p)
+}
+func (p pointer) toExtensions() *XXX_InternalExtensions {
+	return (*XXX_InternalExtensions)(p.p)
+}
+func (p pointer) toOldExtensions() *map[int32]Extension {
+	return (*map[int32]Extension)(p.p)
+}
+
+// getPointerSlice loads []*T from p as a []pointer.
+// The value returned is aliased with the original slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) getPointerSlice() []pointer {
+	// Super-tricky - p should point to a []*T where T is a
+	// message type. We load it as []pointer.
+	return *(*[]pointer)(p.p)
+}
+
+// setPointerSlice stores []pointer into p as a []*T.
+// The value set is aliased with the input slice.
+// This behavior differs from the implementation in pointer_reflect.go.
+func (p pointer) setPointerSlice(v []pointer) {
+	// Super-tricky - p should point to a []*T where T is a
+	// message type. We store it as []pointer.
+	*(*[]pointer)(p.p) = v
+}
+
+// getPointer loads the pointer at p and returns it.
+func (p pointer) getPointer() pointer {
+	return pointer{p: *(*unsafe.Pointer)(p.p)}
+}
+
+// setPointer stores the pointer q at p.
+func (p pointer) setPointer(q pointer) {
+	*(*unsafe.Pointer)(p.p) = q.p
+}
+
+// append q to the slice pointed to by p.
+func (p pointer) appendPointer(q pointer) {
+	s := (*[]unsafe.Pointer)(p.p)
+	*s = append(*s, q.p)
+}
+
+// getInterfacePointer returns a pointer that points to the
+// interface data of the interface pointed by p.
+func (p pointer) getInterfacePointer() pointer {
+	// Super-tricky - read pointer out of data word of interface value.
+	return pointer{p: (*(*[2]unsafe.Pointer)(p.p))[1]}
+}
+
+// asPointerTo returns a reflect.Value that is a pointer to an
+// object of type t stored at p.
+func (p pointer) asPointerTo(t reflect.Type) reflect.Value {
+	return reflect.NewAt(t, p.p)
+}
+
+func atomicLoadUnmarshalInfo(p **unmarshalInfo) *unmarshalInfo {
+	return (*unmarshalInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreUnmarshalInfo(p **unmarshalInfo, v *unmarshalInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadMarshalInfo(p **marshalInfo) *marshalInfo {
+	return (*marshalInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreMarshalInfo(p **marshalInfo, v *marshalInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadMergeInfo(p **mergeInfo) *mergeInfo {
+	return (*mergeInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreMergeInfo(p **mergeInfo, v *mergeInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
+func atomicLoadDiscardInfo(p **discardInfo) *discardInfo {
+	return (*discardInfo)(atomic.LoadPointer((*unsafe.Pointer)(unsafe.Pointer(p))))
+}
+func atomicStoreDiscardInfo(p **discardInfo, v *discardInfo) {
+	atomic.StorePointer((*unsafe.Pointer)(unsafe.Pointer(p)), unsafe.Pointer(v))
+}
diff --git a/vendor/github.com/golang/protobuf/proto/properties.go b/vendor/github.com/golang/protobuf/proto/properties.go
new file mode 100644
index 0000000..79668ff
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/properties.go
@@ -0,0 +1,545 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2010 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+/*
+ * Routines for encoding data into the wire format for protocol buffers.
+ */
+
+import (
+	"fmt"
+	"log"
+	"os"
+	"reflect"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+)
+
+const debug bool = false
+
+// Constants that identify the encoding of a value on the wire.
+const (
+	WireVarint     = 0
+	WireFixed64    = 1
+	WireBytes      = 2
+	WireStartGroup = 3
+	WireEndGroup   = 4
+	WireFixed32    = 5
+)
+
+// tagMap is an optimization over map[int]int for typical protocol buffer
+// use-cases. Encoded protocol buffers are often in tag order with small tag
+// numbers.
+type tagMap struct {
+	fastTags []int
+	slowTags map[int]int
+}
+
+// tagMapFastLimit is the upper bound on the tag number that will be stored in
+// the tagMap slice rather than its map.
+const tagMapFastLimit = 1024
+
+func (p *tagMap) get(t int) (int, bool) {
+	if t > 0 && t < tagMapFastLimit {
+		if t >= len(p.fastTags) {
+			return 0, false
+		}
+		fi := p.fastTags[t]
+		return fi, fi >= 0
+	}
+	fi, ok := p.slowTags[t]
+	return fi, ok
+}
+
+func (p *tagMap) put(t int, fi int) {
+	if t > 0 && t < tagMapFastLimit {
+		for len(p.fastTags) < t+1 {
+			p.fastTags = append(p.fastTags, -1)
+		}
+		p.fastTags[t] = fi
+		return
+	}
+	if p.slowTags == nil {
+		p.slowTags = make(map[int]int)
+	}
+	p.slowTags[t] = fi
+}
+
+// StructProperties represents properties for all the fields of a struct.
+// decoderTags and decoderOrigNames should only be used by the decoder.
+type StructProperties struct {
+	Prop             []*Properties  // properties for each field
+	reqCount         int            // required count
+	decoderTags      tagMap         // map from proto tag to struct field number
+	decoderOrigNames map[string]int // map from original name to struct field number
+	order            []int          // list of struct field numbers in tag order
+
+	// OneofTypes contains information about the oneof fields in this message.
+	// It is keyed by the original name of a field.
+	OneofTypes map[string]*OneofProperties
+}
+
+// OneofProperties represents information about a specific field in a oneof.
+type OneofProperties struct {
+	Type  reflect.Type // pointer to generated struct type for this oneof field
+	Field int          // struct field number of the containing oneof in the message
+	Prop  *Properties
+}
+
+// Implement the sorting interface so we can sort the fields in tag order, as recommended by the spec.
+// See encode.go, (*Buffer).enc_struct.
+
+func (sp *StructProperties) Len() int { return len(sp.order) }
+func (sp *StructProperties) Less(i, j int) bool {
+	return sp.Prop[sp.order[i]].Tag < sp.Prop[sp.order[j]].Tag
+}
+func (sp *StructProperties) Swap(i, j int) { sp.order[i], sp.order[j] = sp.order[j], sp.order[i] }
+
+// Properties represents the protocol-specific behavior of a single struct field.
+type Properties struct {
+	Name     string // name of the field, for error messages
+	OrigName string // original name before protocol compiler (always set)
+	JSONName string // name to use for JSON; determined by protoc
+	Wire     string
+	WireType int
+	Tag      int
+	Required bool
+	Optional bool
+	Repeated bool
+	Packed   bool   // relevant for repeated primitives only
+	Enum     string // set for enum types only
+	proto3   bool   // whether this is known to be a proto3 field
+	oneof    bool   // whether this is a oneof field
+
+	Default    string // default value
+	HasDefault bool   // whether an explicit default was provided
+
+	stype reflect.Type      // set for struct types only
+	sprop *StructProperties // set for struct types only
+
+	mtype      reflect.Type // set for map types only
+	MapKeyProp *Properties  // set for map types only
+	MapValProp *Properties  // set for map types only
+}
+
+// String formats the properties in the protobuf struct field tag style.
+func (p *Properties) String() string {
+	s := p.Wire
+	s += ","
+	s += strconv.Itoa(p.Tag)
+	if p.Required {
+		s += ",req"
+	}
+	if p.Optional {
+		s += ",opt"
+	}
+	if p.Repeated {
+		s += ",rep"
+	}
+	if p.Packed {
+		s += ",packed"
+	}
+	s += ",name=" + p.OrigName
+	if p.JSONName != p.OrigName {
+		s += ",json=" + p.JSONName
+	}
+	if p.proto3 {
+		s += ",proto3"
+	}
+	if p.oneof {
+		s += ",oneof"
+	}
+	if len(p.Enum) > 0 {
+		s += ",enum=" + p.Enum
+	}
+	if p.HasDefault {
+		s += ",def=" + p.Default
+	}
+	return s
+}
+
+// Parse populates p by parsing a string in the protobuf struct field tag style.
+func (p *Properties) Parse(s string) {
+	// "bytes,49,opt,name=foo,def=hello!"
+	fields := strings.Split(s, ",") // breaks def=, but handled below.
+	if len(fields) < 2 {
+		fmt.Fprintf(os.Stderr, "proto: tag has too few fields: %q\n", s)
+		return
+	}
+
+	p.Wire = fields[0]
+	switch p.Wire {
+	case "varint":
+		p.WireType = WireVarint
+	case "fixed32":
+		p.WireType = WireFixed32
+	case "fixed64":
+		p.WireType = WireFixed64
+	case "zigzag32":
+		p.WireType = WireVarint
+	case "zigzag64":
+		p.WireType = WireVarint
+	case "bytes", "group":
+		p.WireType = WireBytes
+		// no numeric converter for non-numeric types
+	default:
+		fmt.Fprintf(os.Stderr, "proto: tag has unknown wire type: %q\n", s)
+		return
+	}
+
+	var err error
+	p.Tag, err = strconv.Atoi(fields[1])
+	if err != nil {
+		return
+	}
+
+outer:
+	for i := 2; i < len(fields); i++ {
+		f := fields[i]
+		switch {
+		case f == "req":
+			p.Required = true
+		case f == "opt":
+			p.Optional = true
+		case f == "rep":
+			p.Repeated = true
+		case f == "packed":
+			p.Packed = true
+		case strings.HasPrefix(f, "name="):
+			p.OrigName = f[5:]
+		case strings.HasPrefix(f, "json="):
+			p.JSONName = f[5:]
+		case strings.HasPrefix(f, "enum="):
+			p.Enum = f[5:]
+		case f == "proto3":
+			p.proto3 = true
+		case f == "oneof":
+			p.oneof = true
+		case strings.HasPrefix(f, "def="):
+			p.HasDefault = true
+			p.Default = f[4:] // rest of string
+			if i+1 < len(fields) {
+				// Commas aren't escaped, and def is always last.
+				p.Default += "," + strings.Join(fields[i+1:], ",")
+				break outer
+			}
+		}
+	}
+}
+
+var protoMessageType = reflect.TypeOf((*Message)(nil)).Elem()
+
+// setFieldProps initializes the field properties for submessages and maps.
+func (p *Properties) setFieldProps(typ reflect.Type, f *reflect.StructField, lockGetProp bool) {
+	switch t1 := typ; t1.Kind() {
+	case reflect.Ptr:
+		if t1.Elem().Kind() == reflect.Struct {
+			p.stype = t1.Elem()
+		}
+
+	case reflect.Slice:
+		if t2 := t1.Elem(); t2.Kind() == reflect.Ptr && t2.Elem().Kind() == reflect.Struct {
+			p.stype = t2.Elem()
+		}
+
+	case reflect.Map:
+		p.mtype = t1
+		p.MapKeyProp = &Properties{}
+		p.MapKeyProp.init(reflect.PtrTo(p.mtype.Key()), "Key", f.Tag.Get("protobuf_key"), nil, lockGetProp)
+		p.MapValProp = &Properties{}
+		vtype := p.mtype.Elem()
+		if vtype.Kind() != reflect.Ptr && vtype.Kind() != reflect.Slice {
+			// The value type is not a message (*T) or bytes ([]byte),
+			// so we need encoders for the pointer to this type.
+			vtype = reflect.PtrTo(vtype)
+		}
+		p.MapValProp.init(vtype, "Value", f.Tag.Get("protobuf_val"), nil, lockGetProp)
+	}
+
+	if p.stype != nil {
+		if lockGetProp {
+			p.sprop = GetProperties(p.stype)
+		} else {
+			p.sprop = getPropertiesLocked(p.stype)
+		}
+	}
+}
+
+var (
+	marshalerType = reflect.TypeOf((*Marshaler)(nil)).Elem()
+)
+
+// Init populates the properties from a protocol buffer struct tag.
+func (p *Properties) Init(typ reflect.Type, name, tag string, f *reflect.StructField) {
+	p.init(typ, name, tag, f, true)
+}
+
+func (p *Properties) init(typ reflect.Type, name, tag string, f *reflect.StructField, lockGetProp bool) {
+	// "bytes,49,opt,def=hello!"
+	p.Name = name
+	p.OrigName = name
+	if tag == "" {
+		return
+	}
+	p.Parse(tag)
+	p.setFieldProps(typ, f, lockGetProp)
+}
+
+var (
+	propertiesMu  sync.RWMutex
+	propertiesMap = make(map[reflect.Type]*StructProperties)
+)
+
+// GetProperties returns the list of properties for the type represented by t.
+// t must represent a generated struct type of a protocol message.
+func GetProperties(t reflect.Type) *StructProperties {
+	if t.Kind() != reflect.Struct {
+		panic("proto: type must have kind struct")
+	}
+
+	// Most calls to GetProperties in a long-running program will be
+	// retrieving details for types we have seen before.
+	propertiesMu.RLock()
+	sprop, ok := propertiesMap[t]
+	propertiesMu.RUnlock()
+	if ok {
+		return sprop
+	}
+
+	propertiesMu.Lock()
+	sprop = getPropertiesLocked(t)
+	propertiesMu.Unlock()
+	return sprop
+}
+
+type (
+	oneofFuncsIface interface {
+		XXX_OneofFuncs() (func(Message, *Buffer) error, func(Message, int, int, *Buffer) (bool, error), func(Message) int, []interface{})
+	}
+	oneofWrappersIface interface {
+		XXX_OneofWrappers() []interface{}
+	}
+)
+
+// getPropertiesLocked requires that propertiesMu is held.
+func getPropertiesLocked(t reflect.Type) *StructProperties {
+	if prop, ok := propertiesMap[t]; ok {
+		return prop
+	}
+
+	prop := new(StructProperties)
+	// in case of recursive protos, fill this in now.
+	propertiesMap[t] = prop
+
+	// build properties
+	prop.Prop = make([]*Properties, t.NumField())
+	prop.order = make([]int, t.NumField())
+
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		p := new(Properties)
+		name := f.Name
+		p.init(f.Type, name, f.Tag.Get("protobuf"), &f, false)
+
+		oneof := f.Tag.Get("protobuf_oneof") // special case
+		if oneof != "" {
+			// Oneof fields don't use the traditional protobuf tag.
+			p.OrigName = oneof
+		}
+		prop.Prop[i] = p
+		prop.order[i] = i
+		if debug {
+			print(i, " ", f.Name, " ", t.String(), " ")
+			if p.Tag > 0 {
+				print(p.String())
+			}
+			print("\n")
+		}
+	}
+
+	// Re-order prop.order.
+	sort.Sort(prop)
+
+	var oots []interface{}
+	switch m := reflect.Zero(reflect.PtrTo(t)).Interface().(type) {
+	case oneofFuncsIface:
+		_, _, _, oots = m.XXX_OneofFuncs()
+	case oneofWrappersIface:
+		oots = m.XXX_OneofWrappers()
+	}
+	if len(oots) > 0 {
+		// Interpret oneof metadata.
+		prop.OneofTypes = make(map[string]*OneofProperties)
+		for _, oot := range oots {
+			oop := &OneofProperties{
+				Type: reflect.ValueOf(oot).Type(), // *T
+				Prop: new(Properties),
+			}
+			sft := oop.Type.Elem().Field(0)
+			oop.Prop.Name = sft.Name
+			oop.Prop.Parse(sft.Tag.Get("protobuf"))
+			// There will be exactly one interface field that
+			// this new value is assignable to.
+			for i := 0; i < t.NumField(); i++ {
+				f := t.Field(i)
+				if f.Type.Kind() != reflect.Interface {
+					continue
+				}
+				if !oop.Type.AssignableTo(f.Type) {
+					continue
+				}
+				oop.Field = i
+				break
+			}
+			prop.OneofTypes[oop.Prop.OrigName] = oop
+		}
+	}
+
+	// build required counts
+	// build tags
+	reqCount := 0
+	prop.decoderOrigNames = make(map[string]int)
+	for i, p := range prop.Prop {
+		if strings.HasPrefix(p.Name, "XXX_") {
+			// Internal fields should not appear in tags/origNames maps.
+			// They are handled specially when encoding and decoding.
+			continue
+		}
+		if p.Required {
+			reqCount++
+		}
+		prop.decoderTags.put(p.Tag, i)
+		prop.decoderOrigNames[p.OrigName] = i
+	}
+	prop.reqCount = reqCount
+
+	return prop
+}
+
+// A global registry of enum types.
+// The generated code will register the generated maps by calling RegisterEnum.
+
+var enumValueMaps = make(map[string]map[string]int32)
+
+// RegisterEnum is called from the generated code to install the enum descriptor
+// maps into the global table to aid parsing text format protocol buffers.
+func RegisterEnum(typeName string, unusedNameMap map[int32]string, valueMap map[string]int32) {
+	if _, ok := enumValueMaps[typeName]; ok {
+		panic("proto: duplicate enum registered: " + typeName)
+	}
+	enumValueMaps[typeName] = valueMap
+}
+
+// EnumValueMap returns the mapping from names to integers of the
+// enum type enumType, or a nil if not found.
+func EnumValueMap(enumType string) map[string]int32 {
+	return enumValueMaps[enumType]
+}
+
+// A registry of all linked message types.
+// The string is a fully-qualified proto name ("pkg.Message").
+var (
+	protoTypedNils = make(map[string]Message)      // a map from proto names to typed nil pointers
+	protoMapTypes  = make(map[string]reflect.Type) // a map from proto names to map types
+	revProtoTypes  = make(map[reflect.Type]string)
+)
+
+// RegisterType is called from generated code and maps from the fully qualified
+// proto name to the type (pointer to struct) of the protocol buffer.
+func RegisterType(x Message, name string) {
+	if _, ok := protoTypedNils[name]; ok {
+		// TODO: Some day, make this a panic.
+		log.Printf("proto: duplicate proto type registered: %s", name)
+		return
+	}
+	t := reflect.TypeOf(x)
+	if v := reflect.ValueOf(x); v.Kind() == reflect.Ptr && v.Pointer() == 0 {
+		// Generated code always calls RegisterType with nil x.
+		// This check is just for extra safety.
+		protoTypedNils[name] = x
+	} else {
+		protoTypedNils[name] = reflect.Zero(t).Interface().(Message)
+	}
+	revProtoTypes[t] = name
+}
+
+// RegisterMapType is called from generated code and maps from the fully qualified
+// proto name to the native map type of the proto map definition.
+func RegisterMapType(x interface{}, name string) {
+	if reflect.TypeOf(x).Kind() != reflect.Map {
+		panic(fmt.Sprintf("RegisterMapType(%T, %q); want map", x, name))
+	}
+	if _, ok := protoMapTypes[name]; ok {
+		log.Printf("proto: duplicate proto type registered: %s", name)
+		return
+	}
+	t := reflect.TypeOf(x)
+	protoMapTypes[name] = t
+	revProtoTypes[t] = name
+}
+
+// MessageName returns the fully-qualified proto name for the given message type.
+func MessageName(x Message) string {
+	type xname interface {
+		XXX_MessageName() string
+	}
+	if m, ok := x.(xname); ok {
+		return m.XXX_MessageName()
+	}
+	return revProtoTypes[reflect.TypeOf(x)]
+}
+
+// MessageType returns the message type (pointer to struct) for a named message.
+// The type is not guaranteed to implement proto.Message if the name refers to a
+// map entry.
+func MessageType(name string) reflect.Type {
+	if t, ok := protoTypedNils[name]; ok {
+		return reflect.TypeOf(t)
+	}
+	return protoMapTypes[name]
+}
+
+// A registry of all linked proto files.
+var (
+	protoFiles = make(map[string][]byte) // file name => fileDescriptor
+)
+
+// RegisterFile is called from generated code and maps from the
+// full file name of a .proto file to its compressed FileDescriptorProto.
+func RegisterFile(filename string, fileDescriptor []byte) {
+	protoFiles[filename] = fileDescriptor
+}
+
+// FileDescriptor returns the compressed FileDescriptorProto for a .proto file.
+func FileDescriptor(filename string) []byte { return protoFiles[filename] }
diff --git a/vendor/github.com/golang/protobuf/proto/table_marshal.go b/vendor/github.com/golang/protobuf/proto/table_marshal.go
new file mode 100644
index 0000000..5cb11fa
--- /dev/null
+++ b/vendor/github.com/golang/protobuf/proto/table_marshal.go
@@ -0,0 +1,2776 @@
+// Go support for Protocol Buffers - Google's data interchange format
+//
+// Copyright 2016 The Go Authors.  All rights reserved.
+// https://github.com/golang/protobuf
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package proto
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"reflect"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"unicode/utf8"
+)
+
+// a sizer takes a pointer to a field and the size of its tag, computes the size of
+// the encoded data.
+type sizer func(pointer, int) int
+
+// a marshaler takes a byte slice, a pointer to a field, and its tag (in wire format),
+// marshals the field to the end of the slice, returns the slice and error (if any).
+type marshaler func(b []byte, ptr pointer, wiretag uint64, deterministic bool) ([]byte, error)
+
+// marshalInfo is the information used for marshaling a message.
+type marshalInfo struct {
+	typ          reflect.Type
+	fields       []*marshalFieldInfo
+	unrecognized field                      // offset of XXX_unrecognized
+	extensions   field                      // offset of XXX_InternalExtensions
+	v1extensions field                      // offset of XXX_extensions
+	sizecache    field                      // offset of XXX_sizecache
+	initialized  int32                      // 0 -- only typ is set, 1 -- fully initialized
+	messageset   bool                       // uses message set wire format
+	hasmarshaler bool                       // has custom marshaler
+	sync.RWMutex                            // protect extElems map, also for initialization
+	extElems     map[int32]*marshalElemInfo // info of extension elements
+}
+
+// marshalFieldInfo is the information used for marshaling a field of a message.
+type marshalFieldInfo struct {
+	field      field
+	wiretag    uint64 // tag in wire format
+	tagsize    int    // size of tag in wire format
+	sizer      sizer
+	marshaler  marshaler
+	isPointer  bool
+	required   bool                              // field is required
+	name       string                            // name of the field, for error reporting
+	oneofElems map[reflect.Type]*marshalElemInfo // info of oneof elements
+}
+
+// marshalElemInfo is the information used for marshaling an extension or oneof element.
+type marshalElemInfo struct {
+	wiretag   uint64 // tag in wire format
+	tagsize   int    // size of tag in wire format
+	sizer     sizer
+	marshaler marshaler
+	isptr     bool // elem is pointer typed, thus interface of this type is a direct interface (extension only)
+	deref     bool // dereference the pointer before operating on it; implies isptr
+}
+
+var (
+	marshalInfoMap  = map[reflect.Type]*marshalInfo{}
+	marshalInfoLock sync.Mutex
+)
+
+// getMarshalInfo returns the information to marshal a given type of message.
+// The info it returns may not necessarily initialized.
+// t is the type of the message (NOT the pointer to it).
+func getMarshalInfo(t reflect.Type) *marshalInfo {
+	marshalInfoLock.Lock()
+	u, ok := marshalInfoMap[t]
+	if !ok {
+		u = &marshalInfo{typ: t}
+		marshalInfoMap[t] = u
+	}
+	marshalInfoLock.Unlock()
+	return u
+}
+
+// Size is the entry point from generated code,
+// and should be ONLY called by generated code.
+// It computes the size of encoded data of msg.
+// a is a pointer to a place to store cached marshal info.
+func (a *InternalMessageInfo) Size(msg Message) int {
+	u := getMessageMarshalInfo(msg, a)
+	ptr := toPointer(&msg)
+	if ptr.isNil() {
+		// We get here if msg is a typed nil ((*SomeMessage)(nil)),
+		// so it satisfies the interface, and msg == nil wouldn't
+		// catch it. We don't want crash in this case.
+		return 0
+	}
+	return u.size(ptr)
+}
+
+// Marshal is the entry point from generated code,
+// and should be ONLY called by generated code.
+// It marshals msg to the end of b.
+// a is a pointer to a place to store cached marshal info.
+func (a *InternalMessageInfo) Marshal(b []byte, msg Message, deterministic bool) ([]byte, error) {
+	u := getMessageMarshalInfo(msg, a)
+	ptr := toPointer(&msg)
+	if ptr.isNil() {
+		// We get here if msg is a typed nil ((*SomeMessage)(nil)),
+		// so it satisfies the interface, and msg == nil wouldn't
+		// catch it. We don't want crash in this case.
+		return b, ErrNil
+	}
+	return u.marshal(b, ptr, deterministic)
+}
+
+func getMessageMarshalInfo(msg interface{}, a *InternalMessageInfo) *marshalInfo {
+	// u := a.marshal, but atomically.
+	// We use an atomic here to ensure memory consistency.
+	u := atomicLoadMarshalInfo(&a.marshal)
+	if u == nil {
+		// Get marshal information from type of message.
+		t := reflect.ValueOf(msg).Type()
+		if t.Kind() != reflect.Ptr {
+			panic(fmt.Sprintf("cannot handle non-pointer message type %v", t))
+		}
+		u = getMarshalInfo(t.Elem())
+		// Store it in the cache for later users.
+		// a.marshal = u, but atomically.
+		atomicStoreMarshalInfo(&a.marshal, u)
+	}
+	return u
+}
+
+// size is the main function to compute the size of the encoded data of a message.
+// ptr is the pointer to the message.
+func (u *marshalInfo) size(ptr pointer) int {
+	if atomic.LoadInt32(&u.initialized) == 0 {
+		u.computeMarshalInfo()
+	}
+
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if u.hasmarshaler {
+		m := ptr.asPointerTo(u.typ).Interface().(Marshaler)
+		b, _ := m.Marshal()
+		return len(b)
+	}
+
+	n := 0
+	for _, f := range u.fields {
+		if f.isPointer && ptr.offset(f.field).getPointer().isNil() {
+			// nil pointer always marshals to nothing
+			continue
+		}
+		n += f.sizer(ptr.offset(f.field), f.tagsize)
+	}
+	if u.extensions.IsValid() {
+		e := ptr.offset(u.extensions).toExtensions()
+		if u.messageset {
+			n += u.sizeMessageSet(e)
+		} else {
+			n += u.sizeExtensions(e)
+		}
+	}
+	if u.v1extensions.IsValid() {
+		m := *ptr.offset(u.v1extensions).toOldExtensions()
+		n += u.sizeV1Extensions(m)
+	}
+	if u.unrecognized.IsValid() {
+		s := *ptr.offset(u.unrecognized).toBytes()
+		n += len(s)
+	}
+	// cache the result for use in marshal
+	if u.sizecache.IsValid() {
+		atomic.StoreInt32(ptr.offset(u.sizecache).toInt32(), int32(n))
+	}
+	return n
+}
+
+// cachedsize gets the size from cache. If there is no cache (i.e. message is not generated),
+// fall back to compute the size.
+func (u *marshalInfo) cachedsize(ptr pointer) int {
+	if u.sizecache.IsValid() {
+		return int(atomic.LoadInt32(ptr.offset(u.sizecache).toInt32()))
+	}
+	return u.size(ptr)
+}
+
+// marshal is the main function to marshal a message. It takes a byte slice and appends
+// the encoded data to the end of the slice, returns the slice and error (if any).
+// ptr is the pointer to the message.
+// If deterministic is true, map is marshaled in deterministic order.
+func (u *marshalInfo) marshal(b []byte, ptr pointer, deterministic bool) ([]byte, error) {
+	if atomic.LoadInt32(&u.initialized) == 0 {
+		u.computeMarshalInfo()
+	}
+
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if u.hasmarshaler {
+		m := ptr.asPointerTo(u.typ).Interface().(Marshaler)
+		b1, err := m.Marshal()
+		b = append(b, b1...)
+		return b, err
+	}
+
+	var err, errLater error
+	// The old marshaler encodes extensions at beginning.
+	if u.extensions.IsValid() {
+		e := ptr.offset(u.extensions).toExtensions()
+		if u.messageset {
+			b, err = u.appendMessageSet(b, e, deterministic)
+		} else {
+			b, err = u.appendExtensions(b, e, deterministic)
+		}
+		if err != nil {
+			return b, err
+		}
+	}
+	if u.v1extensions.IsValid() {
+		m := *ptr.offset(u.v1extensions).toOldExtensions()
+		b, err = u.appendV1Extensions(b, m, deterministic)
+		if err != nil {
+			return b, err
+		}
+	}
+	for _, f := range u.fields {
+		if f.required {
+			if ptr.offset(f.field).getPointer().isNil() {
+				// Required field is not set.
+				// We record the error but keep going, to give a complete marshaling.
+				if errLater == nil {
+					errLater = &RequiredNotSetError{f.name}
+				}
+				continue
+			}
+		}
+		if f.isPointer && ptr.offset(f.field).getPointer().isNil() {
+			// nil pointer always marshals to nothing
+			continue
+		}
+		b, err = f.marshaler(b, ptr.offset(f.field), f.wiretag, deterministic)
+		if err != nil {
+			if err1, ok := err.(*RequiredNotSetError); ok {
+				// Required field in submessage is not set.
+				// We record the error but keep going, to give a complete marshaling.
+				if errLater == nil {
+					errLater = &RequiredNotSetError{f.name + "." + err1.field}
+				}
+				continue
+			}
+			if err == errRepeatedHasNil {
+				err = errors.New("proto: repeated field " + f.name + " has nil element")
+			}
+			if err == errInvalidUTF8 {
+				if errLater == nil {
+					fullName := revProtoTypes[reflect.PtrTo(u.typ)] + "." + f.name
+					errLater = &invalidUTF8Error{fullName}
+				}
+				continue
+			}
+			return b, err
+		}
+	}
+	if u.unrecognized.IsValid() {
+		s := *ptr.offset(u.unrecognized).toBytes()
+		b = append(b, s...)
+	}
+	return b, errLater
+}
+
+// computeMarshalInfo initializes the marshal info.
+func (u *marshalInfo) computeMarshalInfo() {
+	u.Lock()
+	defer u.Unlock()
+	if u.initialized != 0 { // non-atomic read is ok as it is protected by the lock
+		return
+	}
+
+	t := u.typ
+	u.unrecognized = invalidField
+	u.extensions = invalidField
+	u.v1extensions = invalidField
+	u.sizecache = invalidField
+
+	// If the message can marshal itself, let it do it, for compatibility.
+	// NOTE: This is not efficient.
+	if reflect.PtrTo(t).Implements(marshalerType) {
+		u.hasmarshaler = true
+		atomic.StoreInt32(&u.initialized, 1)
+		return
+	}
+
+	// get oneof implementers
+	var oneofImplementers []interface{}
+	switch m := reflect.Zero(reflect.PtrTo(t)).Interface().(type) {
+	case oneofFuncsIface:
+		_, _, _, oneofImplementers = m.XXX_OneofFuncs()
+	case oneofWrappersIface:
+		oneofImplementers = m.XXX_OneofWrappers()
+	}
+
+	n := t.NumField()
+
+	// deal with XXX fields first
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		if !strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		switch f.Name {
+		case "XXX_sizecache":
+			u.sizecache = toField(&f)
+		case "XXX_unrecognized":
+			u.unrecognized = toField(&f)
+		case "XXX_InternalExtensions":
+			u.extensions = toField(&f)
+			u.messageset = f.Tag.Get("protobuf_messageset") == "1"
+		case "XXX_extensions":
+			u.v1extensions = toField(&f)
+		case "XXX_NoUnkeyedLiteral":
+			// nothing to do
+		default:
+			panic("unknown XXX field: " + f.Name)
+		}
+		n--
+	}
+
+	// normal fields
+	fields := make([]marshalFieldInfo, n) // batch allocation
+	u.fields = make([]*marshalFieldInfo, 0, n)
+	for i, j := 0, 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+
+		if strings.HasPrefix(f.Name, "XXX_") {
+			continue
+		}
+		field := &fields[j]
+		j++
+		field.name = f.Name
+		u.fields = append(u.fields, field)
+		if f.Tag.Get("protobuf_oneof") != "" {
+			field.computeOneofFieldInfo(&f, oneofImplementers)
+			continue
+		}
+		if f.Tag.Get("protobuf") == "" {
+			// field has no tag (not in generated message), ignore it
+			u.fields = u.fields[:len(u.fields)-1]
+			j--
+			continue
+		}
+		field.computeMarshalFieldInfo(&f)
+	}
+
+	// fields are marshaled in tag order on the wire.
+	sort.Sort(byTag(u.fields))
+
+	atomic.StoreInt32(&u.initialized, 1)
+}
+
+// helper for sorting fields by tag
+type byTag []*marshalFieldInfo
+
+func (a byTag) Len() int           { return len(a) }
+func (a byTag) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
+func (a byTag) Less(i, j int) bool { return a[i].wiretag < a[j].wiretag }
+
+// getExtElemInfo returns the information to marshal an extension element.
+// The info it returns is initialized.
+func (u *marshalInfo) getExtElemInfo(desc *ExtensionDesc) *marshalElemInfo {
+	// get from cache first
+	u.RLock()
+	e, ok := u.extElems[desc.Field]
+	u.RUnlock()
+	if ok {
+		return e
+	}
+
+	t := reflect.TypeOf(desc.ExtensionType) // pointer or slice to basic type or struct
+	tags := strings.Split(desc.Tag, ",")
+	tag, err := strconv.Atoi(tags[1])
+	if err != nil {
+		panic("tag is not an integer")
+	}
+	wt := wiretype(tags[0])
+	if t.Kind() == reflect.Ptr && t.Elem().Kind() != reflect.Struct {
+		t = t.Elem()
+	}
+	sizer, marshaler := typeMarshaler(t, tags, false, false)
+	var deref bool
+	if t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8 {
+		t = reflect.PtrTo(t)
+		deref = true
+	}
+	e = &marshalElemInfo{
+		wiretag:   uint64(tag)<<3 | wt,
+		tagsize:   SizeVarint(uint64(tag) << 3),
+		sizer:     sizer,
+		marshaler: marshaler,
+		isptr:     t.Kind() == reflect.Ptr,
+		deref:     deref,
+	}
+
+	// update cache
+	u.Lock()
+	if u.extElems == nil {
+		u.extElems = make(map[int32]*marshalElemInfo)
+	}
+	u.extElems[desc.Field] = e
+	u.Unlock()
+	return e
+}
+
+// computeMarshalFieldInfo fills up the information to marshal a field.
+func (fi *marshalFieldInfo) computeMarshalFieldInfo(f *reflect.StructField) {
+	// parse protobuf tag of the field.
+	// tag has format of "bytes,49,opt,name=foo,def=hello!"
+	tags := strings.Split(f.Tag.Get("protobuf"), ",")
+	if tags[0] == "" {
+		return
+	}
+	tag, err := strconv.Atoi(tags[1])
+	if err != nil {
+		panic("tag is not an integer")
+	}
+	wt := wiretype(tags[0])
+	if tags[2] == "req" {
+		fi.required = true
+	}
+	fi.setTag(f, tag, wt)
+	fi.setMarshaler(f, tags)
+}
+
+func (fi *marshalFieldInfo) computeOneofFieldInfo(f *reflect.StructField, oneofImplementers []interface{}) {
+	fi.field = toField(f)
+	fi.wiretag = math.MaxInt32 // Use a large tag number, make oneofs sorted at the end. This tag will not appear on the wire.
+	fi.isPointer = true
+	fi.sizer, fi.marshaler = makeOneOfMarshaler(fi, f)
+	fi.oneofElems = make(map[reflect.Type]*marshalElemInfo)
+
+	ityp := f.Type // interface type
+	for _, o := range oneofImplementers {
+		t := reflect.TypeOf(o)
+		if !t.Implements(ityp) {
+			continue
+		}
+		sf := t.Elem().Field(0) // oneof implementer is a struct with a single field
+		tags := strings.Split(sf.Tag.Get("protobuf"), ",")
+		tag, err := strconv.Atoi(tags[1])
+		if err != nil {
+			panic("tag is not an integer")
+		}
+		wt := wiretype(tags[0])
+		sizer, marshaler := typeMarshaler(sf.Type, tags, false, true) // oneof should not omit any zero value
+		fi.oneofElems[t.Elem()] = &marshalElemInfo{
+			wiretag:   uint64(tag)<<3 | wt,
+			tagsize:   SizeVarint(uint64(tag) << 3),
+			sizer:     sizer,
+			marshaler: marshaler,
+		}
+	}
+}
+
+// wiretype returns the wire encoding of the type.
+func wiretype(encoding string) uint64 {
+	switch encoding {
+	case "fixed32":
+		return WireFixed32
+	case "fixed64":
+		return WireFixed64
+	case "varint", "zigzag32", "zigzag64":
+		return WireVarint
+	case "bytes":
+		return WireBytes
+	case "group":
+		return WireStartGroup
+	}
+	panic("unknown wire type " + encoding)
+}
+
+// setTag fills up the tag (in wire format) and its size in the info of a field.
+func (fi *marshalFieldInfo) setTag(f *reflect.StructField, tag int, wt uint64) {
+	fi.field = toField(f)
+	fi.wiretag = uint64(tag)<<3 | wt
+	fi.tagsize = SizeVarint(uint64(tag) << 3)
+}
+
+// setMarshaler fills up the sizer and marshaler in the info of a field.
+func (fi *marshalFieldInfo) setMarshaler(f *reflect.StructField, tags []string) {
+	switch f.Type.Kind() {
+	case reflect.Map:
+		// map field
+		fi.isPointer = true
+		fi.sizer, fi.marshaler = makeMapMarshaler(f)
+		return
+	case reflect.Ptr, reflect.Slice:
+		fi.isPointer = true
+	}
+	fi.sizer, fi.marshaler = typeMarshaler(f.Type, tags, true, false)
+}
+
+// typeMarshaler returns the sizer and marshaler of a given field.
+// t is the type of the field.
+// tags is the generated "protobuf" tag of the field.
+// If nozero is true, zero value is not marshaled to the wire.
+// If oneof is true, it is a oneof field.
+func typeMarshaler(t reflect.Type, tags []string, nozero, oneof bool) (sizer, marshaler) {
+	encoding := tags[0]
+
+	pointer := false
+	slice := false
+	if t.Kind() == reflect.Slice && t.Elem().Kind() != reflect.Uint8 {
+		slice = true
+		t = t.Elem()
+	}
+	if t.Kind() == reflect.Ptr {
+		pointer = true
+		t = t.Elem()
+	}
+
+	packed := false
+	proto3 := false
+	validateUTF8 := true
+	for i := 2; i < len(tags); i++ {
+		if tags[i] == "packed" {
+			packed = true
+		}
+		if tags[i] == "proto3" {
+			proto3 = true
+		}
+	}
+	validateUTF8 = validateUTF8 && proto3
+
+	switch t.Kind() {
+	case reflect.Bool:
+		if pointer {
+			return sizeBoolPtr, appendBoolPtr
+		}
+		if slice {
+			if packed {
+				return sizeBoolPackedSlice, appendBoolPackedSlice
+			}
+			return sizeBoolSlice, appendBoolSlice
+		}
+		if nozero {
+			return sizeBoolValueNoZero, appendBoolValueNoZero
+		}
+		return sizeBoolValue, appendBoolValue
+	case reflect.Uint32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return sizeFixed32Ptr, appendFixed32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixed32PackedSlice, appendFixed32PackedSlice
+				}
+				return sizeFixed32Slice, appendFixed32Slice
+			}
+			if nozero {
+				return sizeFixed32ValueNoZero, appendFixed32ValueNoZero
+			}
+			return sizeFixed32Value, appendFixed32Value
+		case "varint":
+			if pointer {
+				return sizeVarint32Ptr, appendVarint32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarint32PackedSlice, appendVarint32PackedSlice
+				}
+				return sizeVarint32Slice, appendVarint32Slice
+			}
+			if nozero {
+				return sizeVarint32ValueNoZero, appendVarint32ValueNoZero
+			}
+			return sizeVarint32Value, appendVarint32Value
+		}
+	case reflect.Int32:
+		switch encoding {
+		case "fixed32":
+			if pointer {
+				return sizeFixedS32Ptr, appendFixedS32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixedS32PackedSlice, appendFixedS32PackedSlice
+				}
+				return sizeFixedS32Slice, appendFixedS32Slice
+			}
+			if nozero {
+				return sizeFixedS32ValueNoZero, appendFixedS32ValueNoZero
+			}
+			return sizeFixedS32Value, appendFixedS32Value
+		case "varint":
+			if pointer {
+				return sizeVarintS32Ptr, appendVarintS32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarintS32PackedSlice, appendVarintS32PackedSlice
+				}
+				return sizeVarintS32Slice, appendVarintS32Slice
+			}
+			if nozero {
+				return sizeVarintS32ValueNoZero, appendVarintS32ValueNoZero
+			}
+			return sizeVarintS32Value, appendVarintS32Value
+		case "zigzag32":
+			if pointer {
+				return sizeZigzag32Ptr, appendZigzag32Ptr
+			}
+			if slice {
+				if packed {
+					return sizeZigzag32PackedSlice, appendZigzag32PackedSlice
+				}
+				return sizeZigzag32Slice, appendZigzag32Slice
+			}
+			if nozero {
+				return sizeZigzag32ValueNoZero, appendZigzag32ValueNoZero
+			}
+			return sizeZigzag32Value, appendZigzag32Value
+		}
+	case reflect.Uint64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return sizeFixed64Ptr, appendFixed64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixed64PackedSlice, appendFixed64PackedSlice
+				}
+				return sizeFixed64Slice, appendFixed64Slice
+			}
+			if nozero {
+				return sizeFixed64ValueNoZero, appendFixed64ValueNoZero
+			}
+			return sizeFixed64Value, appendFixed64Value
+		case "varint":
+			if pointer {
+				return sizeVarint64Ptr, appendVarint64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarint64PackedSlice, appendVarint64PackedSlice
+				}
+				return sizeVarint64Slice, appendVarint64Slice
+			}
+			if nozero {
+				return sizeVarint64ValueNoZero, appendVarint64ValueNoZero
+			}
+			return sizeVarint64Value, appendVarint64Value
+		}
+	case reflect.Int64:
+		switch encoding {
+		case "fixed64":
+			if pointer {
+				return sizeFixedS64Ptr, appendFixedS64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeFixedS64PackedSlice, appendFixedS64PackedSlice
+				}
+				return sizeFixedS64Slice, appendFixedS64Slice
+			}
+			if nozero {
+				return sizeFixedS64ValueNoZero, appendFixedS64ValueNoZero
+			}
+			return sizeFixedS64Value, appendFixedS64Value
+		case "varint":
+			if pointer {
+				return sizeVarintS64Ptr, appendVarintS64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeVarintS64PackedSlice, appendVarintS64PackedSlice
+				}
+				return sizeVarintS64Slice, appendVarintS64Slice
+			}
+			if nozero {
+				return sizeVarintS64ValueNoZero, appendVarintS64ValueNoZero
+			}
+			return sizeVarintS64Value, appendVarintS64Value
+		case "zigzag64":
+			if pointer {
+				return sizeZigzag64Ptr, appendZigzag64Ptr
+			}
+			if slice {
+				if packed {
+					return sizeZigzag64PackedSlice, appendZigzag64PackedSlice
+				}
+				return sizeZigzag64Slice, appendZigzag64Slice
+			}
+			if nozero {
+				return sizeZigzag64ValueNoZero, appendZigzag64ValueNoZero
+			}
+			return sizeZigzag64Value, appendZigzag64Value
+		}
+	case reflect.Float32:
+		if pointer {
+			return sizeFloat32Ptr, appendFloat32Ptr
+		}
+		if slice {
+			if packed {
+				return sizeFloat32PackedSlice, appendFloat32PackedSlice
+			}
+			return sizeFloat32Slice, appendFloat32Slice
+		}
+		if nozero {
+			return sizeFloat32ValueNoZero, appendFloat32ValueNoZero
+		}
+		return sizeFloat32Value, appendFloat32Value
+	case reflect.Float64:
+		if pointer {
+			return sizeFloat64Ptr, appendFloat64Ptr
+		}
+		if slice {
+			if packed {
+				return sizeFloat64PackedSlice, appendFloat64PackedSlice
+			}
+			return sizeFloat64Slice, appendFloat64Slice
+		}
+		if nozero {
+			return sizeFloat64ValueNoZero, appendFloat64ValueNoZero
+		}