zstdlib 0.7.0-x64-mingw32 → 0.8.0-x64-mingw32

data/ext/zstdlib/{zstd-1.4.5 → zstd-1.5.0}/lib/decompress/zstd_decompress_block.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,7 +14,7 @@
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
-#include <string.h>      /* memcpy, memmove, memset */
+#include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "../common/compiler.h"    /* prefetch */
 #include "../common/cpu.h"         /* bmi2 */
 #include "../common/mem.h"         /* low level memory routines */
@@ -44,7 +44,7 @@
 /*_*******************************************************
 *  Memory operations
 **********************************************************/
-static void ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+static void ZSTD_copy4(void* dst, const void* src) { ZSTD_memcpy(dst, src, 4); }
 
 
 /*-*************************************************************
@@ -166,7 +166,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                 dctx->litSize = litSize;
                 dctx->litEntropy = 1;
                 if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
-                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                ZSTD_memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
                 return litCSize + lhSize;
             }
 
@@ -191,10 +191,10 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
                 if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
                     RETURN_ERROR_IF(litSize+lhSize > srcSize, corruption_detected, "");
-                    memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                    ZSTD_memcpy(dctx->litBuffer, istart+lhSize, litSize);
                     dctx->litPtr = dctx->litBuffer;
                     dctx->litSize = litSize;
-                    memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
+                    ZSTD_memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
                     return lhSize+litSize;
                 }
                 /* direct reference into compressed stream */
@@ -223,7 +223,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     break;
                 }
                 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
-                memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+                ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
                 dctx->litPtr = dctx->litBuffer;
                 dctx->litSize = litSize;
                 return lhSize+1;
@@ -236,7 +236,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
 /* 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
+ * https://github.com/facebook/zstd/blob/release/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()
@@ -364,23 +364,26 @@ static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U32 nbAddB
  * generate FSE decoding table for one symbol (ll, ml or off)
  * cannot fail if input is valid =>
  * all inputs are presumed validated at this stage */
-void
-ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
+FORCE_INLINE_TEMPLATE
+void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
             const short* normalizedCounter, unsigned maxSymbolValue,
             const U32* baseValue, const U32* nbAdditionalBits,
-            unsigned tableLog)
+            unsigned tableLog, void* wksp, size_t wkspSize)
 {
     ZSTD_seqSymbol* const tableDecode = dt+1;
-    U16 symbolNext[MaxSeq+1];
-
     U32 const maxSV1 = maxSymbolValue + 1;
     U32 const tableSize = 1 << tableLog;
-    U32 highThreshold = tableSize-1;
+
+    U16* symbolNext = (U16*)wksp;
+    BYTE* spread = (BYTE*)(symbolNext + MaxSeq + 1);
+    U32 highThreshold = tableSize - 1;
+
 
     /* Sanity Checks */
     assert(maxSymbolValue <= MaxSeq);
     assert(tableLog <= MaxFSELog);
-
+    assert(wkspSize >= ZSTD_BUILD_FSE_TABLE_WKSP_SIZE);
+    (void)wkspSize;
     /* Init, lay down lowprob symbols */
     {   ZSTD_seqSymbol_header DTableH;
         DTableH.tableLog = tableLog;
@@ -396,16 +399,69 @@ ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
                     assert(normalizedCounter[s]>=0);
                     symbolNext[s] = (U16)normalizedCounter[s];
         }   }   }
-        memcpy(dt, &DTableH, sizeof(DTableH));
+        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
 
     /* Spread symbols */
-    {   U32 const tableMask = tableSize-1;
+    assert(tableSize <= 512);
+    /* Specialized symbol spreading for the case when there are
+     * no low probability (-1 count) symbols. When compressing
+     * small blocks we avoid low probability symbols to hit this
+     * case, since header decoding speed matters more.
+     */
+    if (highThreshold == tableSize - 1) {
+        size_t const tableMask = tableSize-1;
+        size_t const step = FSE_TABLESTEP(tableSize);
+        /* First lay down the symbols in order.
+         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+         * misses since small blocks generally have small table logs, so nearly
+         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
+         * our buffer to handle the over-write.
+         */
+        {
+            U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                pos += n;
+            }
+        }
+        /* Now we spread those positions across the table.
+         * The benefit of doing it in two stages is that we avoid the the
+         * variable size inner loop, which caused lots of branch misses.
+         * Now we can run through all the positions without any branch misses.
+         * We unroll the loop twice, since that is what emperically worked best.
+         */
+        {
+            size_t position = 0;
+            size_t s;
+            size_t const unroll = 2;
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableDecode[uPosition].baseValue = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);
+        }
+    } else {
+        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++) {
+            int const n = normalizedCounter[s];
+            for (i=0; i<n; i++) {
                 tableDecode[position].baseValue = s;
                 position = (position + step) & tableMask;
                 while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
@@ -414,7 +470,8 @@ ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
     }
 
     /* Build Decoding table */
-    {   U32 u;
+    {
+        U32 u;
         for (u=0; u<tableSize; u++) {
             U32 const symbol = tableDecode[u].baseValue;
             U32 const nextState = symbolNext[symbol]++;
@@ -423,7 +480,46 @@ ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
             assert(nbAdditionalBits[symbol] < 255);
             tableDecode[u].nbAdditionalBits = (BYTE)nbAdditionalBits[symbol];
             tableDecode[u].baseValue = baseValue[symbol];
-    }   }
+        }
+    }
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static void ZSTD_buildFSETable_body_default(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U32* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize)
+{
+    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+}
+
+#if DYNAMIC_BMI2
+TARGET_ATTRIBUTE("bmi2") static void ZSTD_buildFSETable_body_bmi2(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U32* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize)
+{
+    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+}
+#endif
+
+void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U32* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        ZSTD_buildFSETable_body_bmi2(dt, normalizedCounter, maxSymbolValue,
+                baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+        return;
+    }
+#endif
+    (void)bmi2;
+    ZSTD_buildFSETable_body_default(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
 }
 
 
@@ -435,7 +531,8 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
                                  const void* src, size_t srcSize,
                                  const U32* baseValue, const U32* nbAdditionalBits,
                                  const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
-                                 int ddictIsCold, int nbSeq)
+                                 int ddictIsCold, int nbSeq, U32* wksp, size_t wkspSize,
+                                 int bmi2)
 {
     switch(type)
     {
@@ -467,7 +564,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
             size_t const headerSize = FSE_readNCount(norm, &max, &tableLog, src, srcSize);
             RETURN_ERROR_IF(FSE_isError(headerSize), corruption_detected, "");
             RETURN_ERROR_IF(tableLog > maxLog, corruption_detected, "");
-            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog);
+            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog, wksp, wkspSize, bmi2);
             *DTablePtr = DTableSpace;
             return headerSize;
         }
@@ -480,7 +577,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
 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 istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
     int nbSeq;
@@ -499,7 +596,8 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
     if (nbSeq > 0x7F) {
         if (nbSeq == 0xFF) {
             RETURN_ERROR_IF(ip+2 > iend, srcSize_wrong, "");
-            nbSeq = MEM_readLE16(ip) + LONGNBSEQ, ip+=2;
+            nbSeq = MEM_readLE16(ip) + LONGNBSEQ;
+            ip+=2;
         } else {
             RETURN_ERROR_IF(ip >= iend, srcSize_wrong, "");
             nbSeq = ((nbSeq-0x80)<<8) + *ip++;
@@ -520,7 +618,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       LL_base, LL_bits,
                                                       LL_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      dctx->bmi2);
             RETURN_ERROR_IF(ZSTD_isError(llhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += llhSize;
         }
@@ -530,7 +630,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       OF_base, OF_bits,
                                                       OF_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      dctx->bmi2);
             RETURN_ERROR_IF(ZSTD_isError(ofhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += ofhSize;
         }
@@ -540,7 +642,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       ML_base, ML_bits,
                                                       ML_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      dctx->bmi2);
             RETURN_ERROR_IF(ZSTD_isError(mlhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += mlhSize;
         }
@@ -554,7 +658,6 @@ typedef struct {
     size_t litLength;
     size_t matchLength;
     size_t offset;
-    const BYTE* match;
 } seq_t;
 
 typedef struct {
@@ -568,9 +671,6 @@ typedef struct {
     ZSTD_fseState stateOffb;
     ZSTD_fseState stateML;
     size_t prevOffset[ZSTD_REP_NUM];
-    const BYTE* prefixStart;
-    const BYTE* dictEnd;
-    size_t pos;
 } seqState_t;
 
 /*! ZSTD_overlapCopy8() :
@@ -686,12 +786,12 @@ size_t ZSTD_execSequenceEnd(BYTE* op,
         RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
         match = dictEnd - (prefixStart-match);
         if (match + sequence.matchLength <= dictEnd) {
-            memmove(oLitEnd, match, sequence.matchLength);
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
             return sequenceLength;
         }
         /* span extDict & currentPrefixSegment */
         {   size_t const length1 = dictEnd - match;
-            memmove(oLitEnd, match, length1);
+            ZSTD_memmove(oLitEnd, match, length1);
             op = oLitEnd + length1;
             sequence.matchLength -= length1;
             match = prefixStart;
@@ -752,12 +852,12 @@ size_t ZSTD_execSequence(BYTE* op,
         RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
         match = dictEnd + (match - prefixStart);
         if (match + sequence.matchLength <= dictEnd) {
-            memmove(oLitEnd, match, sequence.matchLength);
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
             return sequenceLength;
         }
         /* span extDict & currentPrefixSegment */
         {   size_t const length1 = dictEnd - match;
-            memmove(oLitEnd, match, length1);
+            ZSTD_memmove(oLitEnd, match, length1);
             op = oLitEnd + length1;
             sequence.matchLength -= length1;
             match = prefixStart;
@@ -832,10 +932,9 @@ ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD
         : 0)
 
 typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
-typedef enum { ZSTD_p_noPrefetch=0, ZSTD_p_prefetch=1 } ZSTD_prefetch_e;
 
 FORCE_INLINE_TEMPLATE seq_t
-ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, const ZSTD_prefetch_e prefetch)
+ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
 {
     seq_t seq;
     ZSTD_seqSymbol const llDInfo = seqState->stateLL.table[seqState->stateLL.state];
@@ -910,14 +1009,6 @@ ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, c
     DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
                 (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
 
-    if (prefetch == ZSTD_p_prefetch) {
-        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, offset is only used for prefetching */
-        seqState->pos = pos + seq.matchLength;
-    }
-
     /* ANS state update
      * gcc-9.0.0 does 2.5% worse with ZSTD_updateFseStateWithDInfo().
      * clang-9.2.0 does 7% worse with ZSTD_updateFseState().
@@ -948,7 +1039,7 @@ ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, c
 }
 
 #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-static int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
+MEM_STATIC int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
 {
     size_t const windowSize = dctx->fParams.windowSize;
     /* No dictionary used. */
@@ -969,6 +1060,7 @@ MEM_STATIC void ZSTD_assertValidSequence(
         seq_t const seq,
         BYTE const* prefixStart, BYTE const* virtualStart)
 {
+#if DEBUGLEVEL >= 1
     size_t const windowSize = dctx->fParams.windowSize;
     size_t const sequenceSize = seq.litLength + seq.matchLength;
     BYTE const* const oLitEnd = op + seq.litLength;
@@ -986,6 +1078,9 @@ MEM_STATIC void ZSTD_assertValidSequence(
         /* Offset must be within our window. */
         assert(seq.offset <= windowSize);
     }
+#else
+    (void)dctx, (void)op, (void)oend, (void)seq, (void)prefixStart, (void)virtualStart;
+#endif
 }
 #endif
 
@@ -1000,7 +1095,7 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -1014,7 +1109,6 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
     /* Regen sequences */
     if (nbSeq) {
         seqState_t seqState;
-        size_t error = 0;
         dctx->fseEntropy = 1;
         { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
         RETURN_ERROR_IF(
@@ -1048,13 +1142,14 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
          * If you see most cycles served out of the DSB you've hit the good case.
          * If it is pretty even then you may be in an okay case.
          *
-         * I've been able to reproduce this issue on the following CPUs:
+         * This issue has been reproduced on the following CPUs:
          *   - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
          *               Use Instruments->Counters to get DSB/MITE cycles.
          *               I never got performance swings, but I was able to
          *               go from the good case of mostly DSB to half of the
          *               cycles served from MITE.
          *   - Coffeelake: Intel i9-9900k
+         *   - Coffeelake: Intel i7-9700k
          *
          * I haven't been able to reproduce the instability or DSB misses on any
          * of the following CPUS:
@@ -1067,33 +1162,35 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
          *
          *   https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
          */
+        __asm__(".p2align 6");
+        __asm__("nop");
         __asm__(".p2align 5");
         __asm__("nop");
+#  if __GNUC__ >= 9
+        /* better for gcc-9 and gcc-10, worse for clang and gcc-8 */
+        __asm__(".p2align 3");
+#  else
         __asm__(".p2align 4");
+#  endif
 #endif
         for ( ; ; ) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_noPrefetch);
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
             if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
+            if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                return oneSeqSize;
             DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+            op += oneSeqSize;
+            if (UNLIKELY(!--nbSeq))
+                break;
             BIT_reloadDStream(&(seqState.DStream));
-            /* gcc and clang both don't like early returns in this loop.
-             * gcc doesn't like early breaks either.
-             * Instead save an error and report it at the end.
-             * When there is an error, don't increment op, so we don't
-             * overwrite.
-             */
-            if (UNLIKELY(ZSTD_isError(oneSeqSize))) error = oneSeqSize;
-            else op += oneSeqSize;
-            if (UNLIKELY(!--nbSeq)) break;
         }
 
         /* check if reached exact end */
         DEBUGLOG(5, "ZSTD_decompressSequences_body: after decode loop, remaining nbSeq : %i", nbSeq);
-        if (ZSTD_isError(error)) return error;
         RETURN_ERROR_IF(nbSeq, corruption_detected, "");
         RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
         /* save reps for next block */
@@ -1104,7 +1201,7 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
     {   size_t const lastLLSize = litEnd - litPtr;
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
-            memcpy(op, litPtr, lastLLSize);
+            ZSTD_memcpy(op, litPtr, lastLLSize);
             op += lastLLSize;
         }
     }
@@ -1124,6 +1221,24 @@ ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
+                   const BYTE* const prefixStart, const BYTE* const dictEnd)
+{
+    prefetchPos += sequence.litLength;
+    {   const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
+        const BYTE* const match = matchBase + prefetchPos - sequence.offset; /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
+                                                                              * No consequence though : memory address is only used for prefetching, not for dereferencing */
+        PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE);   /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+    }
+    return prefetchPos + sequence.matchLength;
+}
+
+/* This decoding function employs prefetching
+ * to reduce latency impact of cache misses.
+ * It's generally employed when block contains a significant portion of long-distance matches
+ * or when coupled with a "cold" dictionary */
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_decompressSequencesLong_body(
                                ZSTD_DCtx* dctx,
@@ -1134,7 +1249,7 @@ ZSTD_decompressSequencesLong_body(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -1146,18 +1261,17 @@ ZSTD_decompressSequencesLong_body(
 
     /* Regen sequences */
     if (nbSeq) {
-#define STORED_SEQS 4
+#define STORED_SEQS 8
 #define STORED_SEQS_MASK (STORED_SEQS-1)
-#define ADVANCED_SEQS 4
+#define ADVANCED_SEQS STORED_SEQS
         seq_t sequences[STORED_SEQS];
         int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
         seqState_t seqState;
         int seqNb;
+        size_t prefetchPos = (size_t)(op-prefixStart); /* track position relative to prefixStart */
+
         dctx->fseEntropy = 1;
         { int 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;
         assert(dst != NULL);
         assert(iend >= ip);
         RETURN_ERROR_IF(
@@ -1169,21 +1283,23 @@ ZSTD_decompressSequencesLong_body(
 
         /* prepare in advance */
         for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
-            sequences[seqNb] = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
-            PREFETCH_L1(sequences[seqNb].match); PREFETCH_L1(sequences[seqNb].match + sequences[seqNb].matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+            sequences[seqNb] = sequence;
         }
         RETURN_ERROR_IF(seqNb<seqAdvance, corruption_detected, "");
 
         /* decode and decompress */
         for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb<nbSeq) ; seqNb++) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
             if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
             if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
-            PREFETCH_L1(sequence.match); PREFETCH_L1(sequence.match + sequence.matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+
+            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
             sequences[seqNb & STORED_SEQS_MASK] = sequence;
             op += oneSeqSize;
         }
@@ -1209,7 +1325,7 @@ ZSTD_decompressSequencesLong_body(
     {   size_t const lastLLSize = litEnd - litPtr;
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
-            memcpy(op, litPtr, lastLLSize);
+            ZSTD_memcpy(op, litPtr, lastLLSize);
             op += lastLLSize;
         }
     }
@@ -1409,9 +1525,9 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 }
 
 
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst)
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
 {
-    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+    if (dst != dctx->previousDstEnd && dstSize > 0) {   /* not contiguous */
         dctx->dictEnd = dctx->previousDstEnd;
         dctx->virtualStart = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
         dctx->prefixStart = dst;
@@ -1425,7 +1541,7 @@ size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
                       const void* src, size_t srcSize)
 {
     size_t dSize;
-    ZSTD_checkContinuity(dctx, dst);
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
     dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0);
     dctx->previousDstEnd = (char*)dst + dSize;
     return dSize;