diff --git a/vendor/github.com/cespare/xxhash/LICENSE.txt b/vendor/github.com/cespare/xxhash/LICENSE.txt
new file mode 100644
index 0000000..24b5306
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/LICENSE.txt
@@ -0,0 +1,22 @@
+Copyright (c) 2016 Caleb Spare
+
+MIT License
+
+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/cespare/xxhash/README.md b/vendor/github.com/cespare/xxhash/README.md
new file mode 100644
index 0000000..0982fd2
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/README.md
@@ -0,0 +1,50 @@
+# xxhash
+
+[![GoDoc](https://godoc.org/github.com/cespare/xxhash?status.svg)](https://godoc.org/github.com/cespare/xxhash)
+
+xxhash is a Go implementation of the 64-bit
+[xxHash](http://cyan4973.github.io/xxHash/) algorithm, XXH64. This is a
+high-quality hashing algorithm that is much faster than anything in the Go
+standard library.
+
+The API is very small, taking its cue from the other hashing packages in the
+standard library:
+
+    $ go doc github.com/cespare/xxhash                                                                                                                                                                                              !
+    package xxhash // import "github.com/cespare/xxhash"
+
+    Package xxhash implements the 64-bit variant of xxHash (XXH64) as described
+    at http://cyan4973.github.io/xxHash/.
+
+    func New() hash.Hash64
+    func Sum64(b []byte) uint64
+    func Sum64String(s string) uint64
+
+This implementation provides a fast pure-Go implementation and an even faster
+assembly implementation for amd64.
+
+## Benchmarks
+
+Here are some quick benchmarks comparing the pure-Go and assembly
+implementations of Sum64 against another popular Go XXH64 implementation,
+[github.com/OneOfOne/xxhash](https://github.com/OneOfOne/xxhash):
+
+| input size | OneOfOne | cespare (purego) | cespare |
+| --- | --- | --- | --- |
+| 5 B   |  416 MB/s | 720 MB/s |  872 MB/s  |
+| 100 B | 3980 MB/s | 5013 MB/s | 5252 MB/s  |
+| 4 KB  | 12727 MB/s | 12999 MB/s | 13026 MB/s |
+| 10 MB | 9879 MB/s | 10775 MB/s | 10913 MB/s  |
+
+These numbers were generated with:
+
+```
+$ go test -benchtime 10s -bench '/OneOfOne,'
+$ go test -tags purego -benchtime 10s -bench '/xxhash,'
+$ go test -benchtime 10s -bench '/xxhash,'
+```
+
+## Projects using this package
+
+- [InfluxDB](https://github.com/influxdata/influxdb)
+- [Prometheus](https://github.com/prometheus/prometheus)
diff --git a/vendor/github.com/cespare/xxhash/go.mod b/vendor/github.com/cespare/xxhash/go.mod
new file mode 100644
index 0000000..10605a6
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/go.mod
@@ -0,0 +1,6 @@
+module github.com/cespare/xxhash
+
+require (
+	github.com/OneOfOne/xxhash v1.2.2
+	github.com/spaolacci/murmur3 v0.0.0-20180118202830-f09979ecbc72
+)
diff --git a/vendor/github.com/cespare/xxhash/go.sum b/vendor/github.com/cespare/xxhash/go.sum
new file mode 100644
index 0000000..f6b5542
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/go.sum
@@ -0,0 +1,4 @@
+github.com/OneOfOne/xxhash v1.2.2 h1:KMrpdQIwFcEqXDklaen+P1axHaj9BSKzvpUUfnHldSE=
+github.com/OneOfOne/xxhash v1.2.2/go.mod h1:HSdplMjZKSmBqAxg5vPj2TmRDmfkzw+cTzAElWljhcU=
+github.com/spaolacci/murmur3 v0.0.0-20180118202830-f09979ecbc72 h1:qLC7fQah7D6K1B0ujays3HV9gkFtllcxhzImRR7ArPQ=
+github.com/spaolacci/murmur3 v0.0.0-20180118202830-f09979ecbc72/go.mod h1:JwIasOWyU6f++ZhiEuf87xNszmSA2myDM2Kzu9HwQUA=
diff --git a/vendor/github.com/cespare/xxhash/rotate.go b/vendor/github.com/cespare/xxhash/rotate.go
new file mode 100644
index 0000000..f3eac5e
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/rotate.go
@@ -0,0 +1,14 @@
+// +build !go1.9
+
+package xxhash
+
+// TODO(caleb): After Go 1.10 comes out, remove this fallback code.
+
+func rol1(x uint64) uint64  { return (x << 1) | (x >> (64 - 1)) }
+func rol7(x uint64) uint64  { return (x << 7) | (x >> (64 - 7)) }
+func rol11(x uint64) uint64 { return (x << 11) | (x >> (64 - 11)) }
+func rol12(x uint64) uint64 { return (x << 12) | (x >> (64 - 12)) }
+func rol18(x uint64) uint64 { return (x << 18) | (x >> (64 - 18)) }
+func rol23(x uint64) uint64 { return (x << 23) | (x >> (64 - 23)) }
+func rol27(x uint64) uint64 { return (x << 27) | (x >> (64 - 27)) }
+func rol31(x uint64) uint64 { return (x << 31) | (x >> (64 - 31)) }
diff --git a/vendor/github.com/cespare/xxhash/rotate19.go b/vendor/github.com/cespare/xxhash/rotate19.go
new file mode 100644
index 0000000..b99612b
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/rotate19.go
@@ -0,0 +1,14 @@
+// +build go1.9
+
+package xxhash
+
+import "math/bits"
+
+func rol1(x uint64) uint64  { return bits.RotateLeft64(x, 1) }
+func rol7(x uint64) uint64  { return bits.RotateLeft64(x, 7) }
+func rol11(x uint64) uint64 { return bits.RotateLeft64(x, 11) }
+func rol12(x uint64) uint64 { return bits.RotateLeft64(x, 12) }
+func rol18(x uint64) uint64 { return bits.RotateLeft64(x, 18) }
+func rol23(x uint64) uint64 { return bits.RotateLeft64(x, 23) }
+func rol27(x uint64) uint64 { return bits.RotateLeft64(x, 27) }
+func rol31(x uint64) uint64 { return bits.RotateLeft64(x, 31) }
diff --git a/vendor/github.com/cespare/xxhash/xxhash.go b/vendor/github.com/cespare/xxhash/xxhash.go
new file mode 100644
index 0000000..f896bd2
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash.go
@@ -0,0 +1,168 @@
+// Package xxhash implements the 64-bit variant of xxHash (XXH64) as described
+// at http://cyan4973.github.io/xxHash/.
+package xxhash
+
+import (
+	"encoding/binary"
+	"hash"
+)
+
+const (
+	prime1 uint64 = 11400714785074694791
+	prime2 uint64 = 14029467366897019727
+	prime3 uint64 = 1609587929392839161
+	prime4 uint64 = 9650029242287828579
+	prime5 uint64 = 2870177450012600261
+)
+
+// NOTE(caleb): I'm using both consts and vars of the primes. Using consts where
+// possible in the Go code is worth a small (but measurable) performance boost
+// by avoiding some MOVQs. Vars are needed for the asm and also are useful for
+// convenience in the Go code in a few places where we need to intentionally
+// avoid constant arithmetic (e.g., v1 := prime1 + prime2 fails because the
+// result overflows a uint64).
+var (
+	prime1v = prime1
+	prime2v = prime2
+	prime3v = prime3
+	prime4v = prime4
+	prime5v = prime5
+)
+
+type xxh struct {
+	v1    uint64
+	v2    uint64
+	v3    uint64
+	v4    uint64
+	total int
+	mem   [32]byte
+	n     int // how much of mem is used
+}
+
+// New creates a new hash.Hash64 that implements the 64-bit xxHash algorithm.
+func New() hash.Hash64 {
+	var x xxh
+	x.Reset()
+	return &x
+}
+
+func (x *xxh) Reset() {
+	x.n = 0
+	x.total = 0
+	x.v1 = prime1v + prime2
+	x.v2 = prime2
+	x.v3 = 0
+	x.v4 = -prime1v
+}
+
+func (x *xxh) Size() int      { return 8 }
+func (x *xxh) BlockSize() int { return 32 }
+
+// Write adds more data to x. It always returns len(b), nil.
+func (x *xxh) Write(b []byte) (n int, err error) {
+	n = len(b)
+	x.total += len(b)
+
+	if x.n+len(b) < 32 {
+		// This new data doesn't even fill the current block.
+		copy(x.mem[x.n:], b)
+		x.n += len(b)
+		return
+	}
+
+	if x.n > 0 {
+		// Finish off the partial block.
+		copy(x.mem[x.n:], b)
+		x.v1 = round(x.v1, u64(x.mem[0:8]))
+		x.v2 = round(x.v2, u64(x.mem[8:16]))
+		x.v3 = round(x.v3, u64(x.mem[16:24]))
+		x.v4 = round(x.v4, u64(x.mem[24:32]))
+		b = b[32-x.n:]
+		x.n = 0
+	}
+
+	if len(b) >= 32 {
+		// One or more full blocks left.
+		b = writeBlocks(x, b)
+	}
+
+	// Store any remaining partial block.
+	copy(x.mem[:], b)
+	x.n = len(b)
+
+	return
+}
+
+func (x *xxh) Sum(b []byte) []byte {
+	s := x.Sum64()
+	return append(
+		b,
+		byte(s>>56),
+		byte(s>>48),
+		byte(s>>40),
+		byte(s>>32),
+		byte(s>>24),
+		byte(s>>16),
+		byte(s>>8),
+		byte(s),
+	)
+}
+
+func (x *xxh) Sum64() uint64 {
+	var h uint64
+
+	if x.total >= 32 {
+		v1, v2, v3, v4 := x.v1, x.v2, x.v3, x.v4
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = x.v3 + prime5
+	}
+
+	h += uint64(x.total)
+
+	i, end := 0, x.n
+	for ; i+8 <= end; i += 8 {
+		k1 := round(0, u64(x.mem[i:i+8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if i+4 <= end {
+		h ^= uint64(u32(x.mem[i:i+4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		i += 4
+	}
+	for i < end {
+		h ^= uint64(x.mem[i]) * prime5
+		h = rol11(h) * prime1
+		i++
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+func u64(b []byte) uint64 { return binary.LittleEndian.Uint64(b) }
+func u32(b []byte) uint32 { return binary.LittleEndian.Uint32(b) }
+
+func round(acc, input uint64) uint64 {
+	acc += input * prime2
+	acc = rol31(acc)
+	acc *= prime1
+	return acc
+}
+
+func mergeRound(acc, val uint64) uint64 {
+	val = round(0, val)
+	acc ^= val
+	acc = acc*prime1 + prime4
+	return acc
+}
diff --git a/vendor/github.com/cespare/xxhash/xxhash_amd64.go b/vendor/github.com/cespare/xxhash/xxhash_amd64.go
new file mode 100644
index 0000000..d617652
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash_amd64.go
@@ -0,0 +1,12 @@
+// +build !appengine
+// +build gc
+// +build !purego
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b.
+//
+//go:noescape
+func Sum64(b []byte) uint64
+
+func writeBlocks(x *xxh, b []byte) []byte
diff --git a/vendor/github.com/cespare/xxhash/xxhash_amd64.s b/vendor/github.com/cespare/xxhash/xxhash_amd64.s
new file mode 100644
index 0000000..757f201
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash_amd64.s
@@ -0,0 +1,233 @@
+// +build !appengine
+// +build gc
+// +build !purego
+
+#include "textflag.h"
+
+// Register allocation:
+// AX	h
+// CX	pointer to advance through b
+// DX	n
+// BX	loop end
+// R8	v1, k1
+// R9	v2
+// R10	v3
+// R11	v4
+// R12	tmp
+// R13	prime1v
+// R14	prime2v
+// R15	prime4v
+
+// round reads from and advances the buffer pointer in CX.
+// It assumes that R13 has prime1v and R14 has prime2v.
+#define round(r) \
+	MOVQ  (CX), R12 \
+	ADDQ  $8, CX    \
+	IMULQ R14, R12  \
+	ADDQ  R12, r    \
+	ROLQ  $31, r    \
+	IMULQ R13, r
+
+// mergeRound applies a merge round on the two registers acc and val.
+// It assumes that R13 has prime1v, R14 has prime2v, and R15 has prime4v.
+#define mergeRound(acc, val) \
+	IMULQ R14, val \
+	ROLQ  $31, val \
+	IMULQ R13, val \
+	XORQ  val, acc \
+	IMULQ R13, acc \
+	ADDQ  R15, acc
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT, $0-32
+	// Load fixed primes.
+	MOVQ ·prime1v(SB), R13
+	MOVQ ·prime2v(SB), R14
+	MOVQ ·prime4v(SB), R15
+
+	// Load slice.
+	MOVQ b_base+0(FP), CX
+	MOVQ b_len+8(FP), DX
+	LEAQ (CX)(DX*1), BX
+
+	// The first loop limit will be len(b)-32.
+	SUBQ $32, BX
+
+	// Check whether we have at least one block.
+	CMPQ DX, $32
+	JLT  noBlocks
+
+	// Set up initial state (v1, v2, v3, v4).
+	MOVQ R13, R8
+	ADDQ R14, R8
+	MOVQ R14, R9
+	XORQ R10, R10
+	XORQ R11, R11
+	SUBQ R13, R11
+
+	// Loop until CX > BX.
+blockLoop:
+	round(R8)
+	round(R9)
+	round(R10)
+	round(R11)
+
+	CMPQ CX, BX
+	JLE  blockLoop
+
+	MOVQ R8, AX
+	ROLQ $1, AX
+	MOVQ R9, R12
+	ROLQ $7, R12
+	ADDQ R12, AX
+	MOVQ R10, R12
+	ROLQ $12, R12
+	ADDQ R12, AX
+	MOVQ R11, R12
+	ROLQ $18, R12
+	ADDQ R12, AX
+
+	mergeRound(AX, R8)
+	mergeRound(AX, R9)
+	mergeRound(AX, R10)
+	mergeRound(AX, R11)
+
+	JMP afterBlocks
+
+noBlocks:
+	MOVQ ·prime5v(SB), AX
+
+afterBlocks:
+	ADDQ DX, AX
+
+	// Right now BX has len(b)-32, and we want to loop until CX > len(b)-8.
+	ADDQ $24, BX
+
+	CMPQ CX, BX
+	JG   fourByte
+
+wordLoop:
+	// Calculate k1.
+	MOVQ  (CX), R8
+	ADDQ  $8, CX
+	IMULQ R14, R8
+	ROLQ  $31, R8
+	IMULQ R13, R8
+
+	XORQ  R8, AX
+	ROLQ  $27, AX
+	IMULQ R13, AX
+	ADDQ  R15, AX
+
+	CMPQ CX, BX
+	JLE  wordLoop
+
+fourByte:
+	ADDQ $4, BX
+	CMPQ CX, BX
+	JG   singles
+
+	MOVL  (CX), R8
+	ADDQ  $4, CX
+	IMULQ R13, R8
+	XORQ  R8, AX
+
+	ROLQ  $23, AX
+	IMULQ R14, AX
+	ADDQ  ·prime3v(SB), AX
+
+singles:
+	ADDQ $4, BX
+	CMPQ CX, BX
+	JGE  finalize
+
+singlesLoop:
+	MOVBQZX (CX), R12
+	ADDQ    $1, CX
+	IMULQ   ·prime5v(SB), R12
+	XORQ    R12, AX
+
+	ROLQ  $11, AX
+	IMULQ R13, AX
+
+	CMPQ CX, BX
+	JL   singlesLoop
+
+finalize:
+	MOVQ  AX, R12
+	SHRQ  $33, R12
+	XORQ  R12, AX
+	IMULQ R14, AX
+	MOVQ  AX, R12
+	SHRQ  $29, R12
+	XORQ  R12, AX
+	IMULQ ·prime3v(SB), AX
+	MOVQ  AX, R12
+	SHRQ  $32, R12
+	XORQ  R12, AX
+
+	MOVQ AX, ret+24(FP)
+	RET
+
+// writeBlocks uses the same registers as above except that it uses AX to store
+// the x pointer.
+
+// func writeBlocks(x *xxh, b []byte) []byte
+TEXT ·writeBlocks(SB), NOSPLIT, $0-56
+	// Load fixed primes needed for round.
+	MOVQ ·prime1v(SB), R13
+	MOVQ ·prime2v(SB), R14
+
+	// Load slice.
+	MOVQ b_base+8(FP), CX
+	MOVQ CX, ret_base+32(FP) // initialize return base pointer; see NOTE below
+	MOVQ b_len+16(FP), DX
+	LEAQ (CX)(DX*1), BX
+	SUBQ $32, BX
+
+	// Load vN from x.
+	MOVQ x+0(FP), AX
+	MOVQ 0(AX), R8   // v1
+	MOVQ 8(AX), R9   // v2
+	MOVQ 16(AX), R10 // v3
+	MOVQ 24(AX), R11 // v4
+
+	// We don't need to check the loop condition here; this function is
+	// always called with at least one block of data to process.
+blockLoop:
+	round(R8)
+	round(R9)
+	round(R10)
+	round(R11)
+
+	CMPQ CX, BX
+	JLE  blockLoop
+
+	// Copy vN back to x.
+	MOVQ R8, 0(AX)
+	MOVQ R9, 8(AX)
+	MOVQ R10, 16(AX)
+	MOVQ R11, 24(AX)
+
+	// Construct return slice.
+	// NOTE: It's important that we don't construct a slice that has a base
+	// pointer off the end of the original slice, as in Go 1.7+ this will
+	// cause runtime crashes. (See discussion in, for example,
+	// https://github.com/golang/go/issues/16772.)
+	// Therefore, we calculate the length/cap first, and if they're zero, we
+	// keep the old base. This is what the compiler does as well if you
+	// write code like
+	//   b = b[len(b):]
+
+	// New length is 32 - (CX - BX) -> BX+32 - CX.
+	ADDQ $32, BX
+	SUBQ CX, BX
+	JZ   afterSetBase
+
+	MOVQ CX, ret_base+32(FP)
+
+afterSetBase:
+	MOVQ BX, ret_len+40(FP)
+	MOVQ BX, ret_cap+48(FP) // set cap == len
+
+	RET
diff --git a/vendor/github.com/cespare/xxhash/xxhash_other.go b/vendor/github.com/cespare/xxhash/xxhash_other.go
new file mode 100644
index 0000000..c68d13f
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash_other.go
@@ -0,0 +1,75 @@
+// +build !amd64 appengine !gc purego
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b.
+func Sum64(b []byte) uint64 {
+	// A simpler version would be
+	//   x := New()
+	//   x.Write(b)
+	//   return x.Sum64()
+	// but this is faster, particularly for small inputs.
+
+	n := len(b)
+	var h uint64
+
+	if n >= 32 {
+		v1 := prime1v + prime2
+		v2 := prime2
+		v3 := uint64(0)
+		v4 := -prime1v
+		for len(b) >= 32 {
+			v1 = round(v1, u64(b[0:8:len(b)]))
+			v2 = round(v2, u64(b[8:16:len(b)]))
+			v3 = round(v3, u64(b[16:24:len(b)]))
+			v4 = round(v4, u64(b[24:32:len(b)]))
+			b = b[32:len(b):len(b)]
+		}
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = prime5
+	}
+
+	h += uint64(n)
+
+	i, end := 0, len(b)
+	for ; i+8 <= end; i += 8 {
+		k1 := round(0, u64(b[i:i+8:len(b)]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if i+4 <= end {
+		h ^= uint64(u32(b[i:i+4:len(b)])) * prime1
+		h = rol23(h)*prime2 + prime3
+		i += 4
+	}
+	for ; i < end; i++ {
+		h ^= uint64(b[i]) * prime5
+		h = rol11(h) * prime1
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+func writeBlocks(x *xxh, b []byte) []byte {
+	v1, v2, v3, v4 := x.v1, x.v2, x.v3, x.v4
+	for len(b) >= 32 {
+		v1 = round(v1, u64(b[0:8:len(b)]))
+		v2 = round(v2, u64(b[8:16:len(b)]))
+		v3 = round(v3, u64(b[16:24:len(b)]))
+		v4 = round(v4, u64(b[24:32:len(b)]))
+		b = b[32:len(b):len(b)]
+	}
+	x.v1, x.v2, x.v3, x.v4 = v1, v2, v3, v4
+	return b
+}
diff --git a/vendor/github.com/cespare/xxhash/xxhash_safe.go b/vendor/github.com/cespare/xxhash/xxhash_safe.go
new file mode 100644
index 0000000..dfa15ab
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash_safe.go
@@ -0,0 +1,10 @@
+// +build appengine
+
+// This file contains the safe implementations of otherwise unsafe-using code.
+
+package xxhash
+
+// Sum64String computes the 64-bit xxHash digest of s.
+func Sum64String(s string) uint64 {
+	return Sum64([]byte(s))
+}
diff --git a/vendor/github.com/cespare/xxhash/xxhash_unsafe.go b/vendor/github.com/cespare/xxhash/xxhash_unsafe.go
new file mode 100644
index 0000000..d2b64e8
--- /dev/null
+++ b/vendor/github.com/cespare/xxhash/xxhash_unsafe.go
@@ -0,0 +1,30 @@
+// +build !appengine
+
+// This file encapsulates usage of unsafe.
+// xxhash_safe.go contains the safe implementations.
+
+package xxhash
+
+import (
+	"reflect"
+	"unsafe"
+)
+
+// Sum64String computes the 64-bit xxHash digest of s.
+// It may be faster than Sum64([]byte(s)) by avoiding a copy.
+//
+// TODO(caleb): Consider removing this if an optimization is ever added to make
+// it unnecessary: https://golang.org/issue/2205.
+//
+// TODO(caleb): We still have a function call; we could instead write Go/asm
+// copies of Sum64 for strings to squeeze out a bit more speed.
+func Sum64String(s string) uint64 {
+	// See https://groups.google.com/d/msg/golang-nuts/dcjzJy-bSpw/tcZYBzQqAQAJ
+	// for some discussion about this unsafe conversion.
+	var b []byte
+	bh := (*reflect.SliceHeader)(unsafe.Pointer(&b))
+	bh.Data = (*reflect.StringHeader)(unsafe.Pointer(&s)).Data
+	bh.Len = len(s)
+	bh.Cap = len(s)
+	return Sum64(b)
+}