diff --git a/vendor/github.com/DataDog/zstd/zstd_v01.c b/vendor/github.com/DataDog/zstd/zstd_v01.c
index cad2b99..ae8cba2 100644
--- a/vendor/github.com/DataDog/zstd/zstd_v01.c
+++ b/vendor/github.com/DataDog/zstd/zstd_v01.c
@@ -1073,99 +1073,102 @@
     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))
+    if (cSrcSize < 6) return (size_t)-FSE_ERROR_srcSize_wrong;
     {
-#define HUF_DECODE_SYMBOL_0(n, Dstream) \
-        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog);
+        BYTE* const ostart = (BYTE*) dst;
+        BYTE* op = ostart;
+        BYTE* const omax = op + maxDstSize;
+        BYTE* const olimit = omax-15;
 
-#define HUF_DECODE_SYMBOL_1(n, Dstream) \
-        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog); \
-        if (FSE_32bits() && (HUF_MAX_TABLELOG>12)) FSE_reloadDStream(&Dstream)
+        const void* ptr = DTable;
+        const HUF_DElt* const dt = (const HUF_DElt*)(ptr)+1;
+        const U32 dtLog = DTable[0];
+        size_t errorCode;
+        U32 reloadStatus;
 
-#define HUF_DECODE_SYMBOL_2(n, Dstream) \
-        op[n] = HUF_decodeSymbol(&Dstream, dt, dtLog); \
-        if (FSE_32bits()) FSE_reloadDStream(&Dstream)
+        /* Init */
 
-        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);
-    }
+        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 (reloadStatus!=FSE_DStream_completed)   /* not complete : some bitStream might be FSE_DStream_unfinished */
-        return (size_t)-FSE_ERROR_corruptionDetected;
+        if (length1+length2+length3+6 >= cSrcSize) return (size_t)-FSE_ERROR_srcSize_wrong;
 
-    /* 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++)
+        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))
         {
-            HUF_DECODE_SYMBOL_0(0, bitTail);
+    #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 (FSE_endOfDStream(&bitTail))
-            return op-ostart;
+        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;
     }
-
-    if (op==omax) return (size_t)-FSE_ERROR_dstSize_tooSmall;   /* dst buffer is full, but cSrc unfinished */
-
-    return (size_t)-FSE_ERROR_corruptionDetected;
 }
 
 
@@ -1355,8 +1358,6 @@
 
 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); }
@@ -1381,16 +1382,9 @@
     }
 }
 
-
-static U32 ZSTD_readLE32(const void* memPtr)
+static U32 ZSTD_readLE24(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));
-    }
+    return ZSTD_readLE16(memPtr) + (((const BYTE*)memPtr)[2] << 16);
 }
 
 static U32 ZSTD_readBE32(const void* memPtr)
@@ -1704,13 +1698,13 @@
     seqState->prevOffset = seq->offset;
     if (litLength == MaxLL)
     {
-        U32 add = dumps<de ? *dumps++ : 0;
+        const 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 */
+                litLength = ZSTD_readLE24(dumps);
                 dumps += 3;
             }
         }
@@ -1732,13 +1726,13 @@
     matchLength = FSE_decodeSymbol(&(seqState->stateML), &(seqState->DStream));
     if (matchLength == MaxML)
     {
-        U32 add = dumps<de ? *dumps++ : 0;
+        const 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 */
+                matchLength = ZSTD_readLE24(dumps);
                 dumps += 3;
             }
         }