ext/zstdruby/libzstd/decompress/zstd_decompress_block.c in zstd-ruby-1.5.0.0 vs ext/zstdruby/libzstd/decompress/zstd_decompress_block.c in zstd-ruby-1.5.1.0
- old
+ new
@@ -67,19 +67,60 @@
RETURN_ERROR_IF(bpPtr->blockType == bt_reserved, corruption_detected, "");
return cSize;
}
}
+/* Allocate buffer for literals, either overlapping current dst, or split between dst and litExtraBuffer, or stored entirely within litExtraBuffer */
+static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const size_t dstCapacity, const size_t litSize,
+ const streaming_operation streaming, const size_t expectedWriteSize, const unsigned splitImmediately)
+{
+ if (streaming == not_streaming && dstCapacity > ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH)
+ {
+ /* room for litbuffer to fit without read faulting */
+ dctx->litBuffer = (BYTE*)dst + ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH;
+ dctx->litBufferEnd = dctx->litBuffer + litSize;
+ dctx->litBufferLocation = ZSTD_in_dst;
+ }
+ else if (litSize > ZSTD_LITBUFFEREXTRASIZE)
+ {
+ /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
+ if (splitImmediately) {
+ /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
+ dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
+ dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
+ }
+ else {
+ /* initially this will be stored entirely in dst during huffman decoding, it will partially shifted to litExtraBuffer after */
+ dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
+ dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
+ }
+ dctx->litBufferLocation = ZSTD_split;
+ }
+ else
+ {
+ /* fits entirely within litExtraBuffer, so no split is necessary */
+ dctx->litBuffer = dctx->litExtraBuffer;
+ dctx->litBufferEnd = dctx->litBuffer + litSize;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
+ }
+}
/* Hidden declaration for fullbench */
size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
- const void* src, size_t srcSize);
+ const void* src, size_t srcSize,
+ void* dst, size_t dstCapacity, const streaming_operation streaming);
/*! ZSTD_decodeLiteralsBlock() :
+ * Where it is possible to do so without being stomped by the output during decompression, the literals block will be stored
+ * in the dstBuffer. If there is room to do so, it will be stored in full in the excess dst space after where the current
+ * block will be output. Otherwise it will be stored at the end of the current dst blockspace, with a small portion being
+ * stored in dctx->litExtraBuffer to help keep it "ahead" of the current output write.
+ *
* @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 */
+ const void* src, size_t srcSize, /* note : srcSize < BLOCKSIZE */
+ void* dst, size_t dstCapacity, const streaming_operation streaming)
{
DEBUGLOG(5, "ZSTD_decodeLiteralsBlock");
RETURN_ERROR_IF(srcSize < MIN_CBLOCK_SIZE, corruption_detected, "");
{ const BYTE* const istart = (const BYTE*) src;
@@ -88,19 +129,20 @@
switch(litEncType)
{
case set_repeat:
DEBUGLOG(5, "set_repeat flag : re-using stats from previous compressed literals block");
RETURN_ERROR_IF(dctx->litEntropy==0, dictionary_corrupted, "");
- /* fall-through */
+ ZSTD_FALLTHROUGH;
case set_compressed:
RETURN_ERROR_IF(srcSize < 5, 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);
size_t hufSuccess;
+ size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
switch(lhlCode)
{
case 0: case 1: default: /* note : default is impossible, since lhlCode into [0..3] */
/* 2 - 2 - 10 - 10 */
singleStream = !lhlCode;
@@ -119,27 +161,30 @@
lhSize = 5;
litSize = (lhc >> 4) & 0x3FFFF;
litCSize = (lhc >> 22) + ((size_t)istart[4] << 10);
break;
}
+ RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
+ RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
+ ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);
/* prefetch huffman table if cold */
if (dctx->ddictIsCold && (litSize > 768 /* heuristic */)) {
PREFETCH_AREA(dctx->HUFptr, sizeof(dctx->entropy.hufTable));
}
if (litEncType==set_repeat) {
if (singleStream) {
hufSuccess = HUF_decompress1X_usingDTable_bmi2(
dctx->litBuffer, litSize, istart+lhSize, litCSize,
- dctx->HUFptr, dctx->bmi2);
+ dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
} else {
hufSuccess = HUF_decompress4X_usingDTable_bmi2(
dctx->litBuffer, litSize, istart+lhSize, litCSize,
- dctx->HUFptr, dctx->bmi2);
+ dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
}
} else {
if (singleStream) {
#if defined(HUF_FORCE_DECOMPRESS_X2)
hufSuccess = HUF_decompress1X_DCtx_wksp(
@@ -148,33 +193,40 @@
sizeof(dctx->workspace));
#else
hufSuccess = HUF_decompress1X1_DCtx_wksp_bmi2(
dctx->entropy.hufTable, dctx->litBuffer, litSize,
istart+lhSize, litCSize, dctx->workspace,
- sizeof(dctx->workspace), dctx->bmi2);
+ sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
#endif
} else {
hufSuccess = HUF_decompress4X_hufOnly_wksp_bmi2(
dctx->entropy.hufTable, dctx->litBuffer, litSize,
istart+lhSize, litCSize, dctx->workspace,
- sizeof(dctx->workspace), dctx->bmi2);
+ sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
}
}
+ if (dctx->litBufferLocation == ZSTD_split)
+ {
+ ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
+ ZSTD_memmove(dctx->litBuffer + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH, dctx->litBuffer, litSize - ZSTD_LITBUFFEREXTRASIZE);
+ dctx->litBuffer += ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
+ dctx->litBufferEnd -= WILDCOPY_OVERLENGTH;
+ }
RETURN_ERROR_IF(HUF_isError(hufSuccess), corruption_detected, "");
dctx->litPtr = dctx->litBuffer;
dctx->litSize = litSize;
dctx->litEntropy = 1;
if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
- ZSTD_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;
+ size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
switch(lhlCode)
{
case 0: case 2: default: /* note : default is impossible, since lhlCode into [0..3] */
lhSize = 1;
litSize = istart[0] >> 3;
@@ -187,27 +239,40 @@
lhSize = 3;
litSize = MEM_readLE24(istart) >> 4;
break;
}
+ RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
+ RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
+ ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) { /* risk reading beyond src buffer with wildcopy */
RETURN_ERROR_IF(litSize+lhSize > srcSize, corruption_detected, "");
- ZSTD_memcpy(dctx->litBuffer, istart+lhSize, litSize);
+ if (dctx->litBufferLocation == ZSTD_split)
+ {
+ ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize - ZSTD_LITBUFFEREXTRASIZE);
+ ZSTD_memcpy(dctx->litExtraBuffer, istart + lhSize + litSize - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
+ }
+ else
+ {
+ ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize);
+ }
dctx->litPtr = dctx->litBuffer;
dctx->litSize = litSize;
- ZSTD_memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
return lhSize+litSize;
}
/* direct reference into compressed stream */
dctx->litPtr = istart+lhSize;
dctx->litSize = litSize;
+ dctx->litBufferEnd = dctx->litPtr + litSize;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
return lhSize+litSize;
}
case set_rle:
{ U32 const lhlCode = ((istart[0]) >> 2) & 3;
size_t litSize, lhSize;
+ size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
switch(lhlCode)
{
case 0: case 2: default: /* note : default is impossible, since lhlCode into [0..3] */
lhSize = 1;
litSize = istart[0] >> 3;
@@ -220,12 +285,23 @@
lhSize = 3;
litSize = MEM_readLE24(istart) >> 4;
RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4");
break;
}
+ RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
- ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+ RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
+ ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
+ if (dctx->litBufferLocation == ZSTD_split)
+ {
+ ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize - ZSTD_LITBUFFEREXTRASIZE);
+ ZSTD_memset(dctx->litExtraBuffer, istart[lhSize], ZSTD_LITBUFFEREXTRASIZE);
+ }
+ else
+ {
+ ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize);
+ }
dctx->litPtr = dctx->litBuffer;
dctx->litSize = litSize;
return lhSize+1;
}
default:
@@ -341,11 +417,11 @@
{ 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)
+static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U8 nbAddBits)
{
void* ptr = dt;
ZSTD_seqSymbol_header* const DTableH = (ZSTD_seqSymbol_header*)ptr;
ZSTD_seqSymbol* const cell = dt + 1;
@@ -353,11 +429,11 @@
DTableH->fastMode = 0;
cell->nbBits = 0;
cell->nextState = 0;
assert(nbAddBits < 255);
- cell->nbAdditionalBits = (BYTE)nbAddBits;
+ cell->nbAdditionalBits = nbAddBits;
cell->baseValue = baseValue;
}
/* ZSTD_buildFSETable() :
@@ -365,11 +441,11 @@
* cannot fail if input is valid =>
* all inputs are presumed validated at this stage */
FORCE_INLINE_TEMPLATE
void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
const short* normalizedCounter, unsigned maxSymbolValue,
- const U32* baseValue, const U32* nbAdditionalBits,
+ const U32* baseValue, const U8* nbAdditionalBits,
unsigned tableLog, void* wksp, size_t wkspSize)
{
ZSTD_seqSymbol* const tableDecode = dt+1;
U32 const maxSV1 = maxSymbolValue + 1;
U32 const tableSize = 1 << tableLog;
@@ -476,40 +552,40 @@
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].nbAdditionalBits = 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,
+ const U32* baseValue, const U8* 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,
+BMI2_TARGET_ATTRIBUTE static void ZSTD_buildFSETable_body_bmi2(ZSTD_seqSymbol* dt,
const short* normalizedCounter, unsigned maxSymbolValue,
- const U32* baseValue, const U32* nbAdditionalBits,
+ const U32* baseValue, const U8* 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,
+ const U32* baseValue, const U8* nbAdditionalBits,
unsigned tableLog, void* wksp, size_t wkspSize, int bmi2)
{
#if DYNAMIC_BMI2
if (bmi2) {
ZSTD_buildFSETable_body_bmi2(dt, normalizedCounter, maxSymbolValue,
@@ -527,11 +603,11 @@
* @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, unsigned max, U32 maxLog,
const void* src, size_t srcSize,
- const U32* baseValue, const U32* nbAdditionalBits,
+ const U32* baseValue, const U8* nbAdditionalBits,
const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
int ddictIsCold, int nbSeq, U32* wksp, size_t wkspSize,
int bmi2)
{
switch(type)
@@ -539,11 +615,11 @@
case set_rle :
RETURN_ERROR_IF(!srcSize, srcSize_wrong, "");
RETURN_ERROR_IF((*(const BYTE*)src) > max, corruption_detected, "");
{ U32 const symbol = *(const BYTE*)src;
U32 const baseline = baseValue[symbol];
- U32 const nbBits = nbAdditionalBits[symbol];
+ U8 const nbBits = nbAdditionalBits[symbol];
ZSTD_buildSeqTable_rle(DTableSpace, baseline, nbBits);
}
*DTablePtr = DTableSpace;
return 1;
case set_basic :
@@ -618,11 +694,11 @@
ip, iend-ip,
LL_base, LL_bits,
LL_defaultDTable, dctx->fseEntropy,
dctx->ddictIsCold, nbSeq,
dctx->workspace, sizeof(dctx->workspace),
- dctx->bmi2);
+ ZSTD_DCtx_get_bmi2(dctx));
RETURN_ERROR_IF(ZSTD_isError(llhSize), corruption_detected, "ZSTD_buildSeqTable failed");
ip += llhSize;
}
{ size_t const ofhSize = ZSTD_buildSeqTable(dctx->entropy.OFTable, &dctx->OFTptr,
@@ -630,11 +706,11 @@
ip, iend-ip,
OF_base, OF_bits,
OF_defaultDTable, dctx->fseEntropy,
dctx->ddictIsCold, nbSeq,
dctx->workspace, sizeof(dctx->workspace),
- dctx->bmi2);
+ ZSTD_DCtx_get_bmi2(dctx));
RETURN_ERROR_IF(ZSTD_isError(ofhSize), corruption_detected, "ZSTD_buildSeqTable failed");
ip += ofhSize;
}
{ size_t const mlhSize = ZSTD_buildSeqTable(dctx->entropy.MLTable, &dctx->MLTptr,
@@ -642,11 +718,11 @@
ip, iend-ip,
ML_base, ML_bits,
ML_defaultDTable, dctx->fseEntropy,
dctx->ddictIsCold, nbSeq,
dctx->workspace, sizeof(dctx->workspace),
- dctx->bmi2);
+ ZSTD_DCtx_get_bmi2(dctx));
RETURN_ERROR_IF(ZSTD_isError(mlhSize), corruption_detected, "ZSTD_buildSeqTable failed");
ip += mlhSize;
}
}
@@ -711,11 +787,11 @@
* @param ovtype controls the overlap detection
* - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
* - ZSTD_overlap_src_before_dst: The src and dst may overlap and may be any distance apart.
* The src buffer must be before the dst buffer.
*/
-static void ZSTD_safecopy(BYTE* op, BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
+static void ZSTD_safecopy(BYTE* op, const BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
ptrdiff_t const diff = op - ip;
BYTE* const oend = op + length;
assert((ovtype == ZSTD_no_overlap && (diff <= -8 || diff >= 8 || op >= oend_w)) ||
(ovtype == ZSTD_overlap_src_before_dst && diff >= 0));
@@ -727,10 +803,11 @@
}
if (ovtype == ZSTD_overlap_src_before_dst) {
/* Copy 8 bytes and ensure the offset >= 8 when there can be overlap. */
assert(length >= 8);
ZSTD_overlapCopy8(&op, &ip, diff);
+ length -= 8;
assert(op - ip >= 8);
assert(op <= oend);
}
if (oend <= oend_w) {
@@ -741,29 +818,52 @@
if (op <= oend_w) {
/* Wildcopy until we get close to the end. */
assert(oend > oend_w);
ZSTD_wildcopy(op, ip, oend_w - op, ovtype);
ip += oend_w - op;
- op = oend_w;
+ op += oend_w - op;
}
/* Handle the leftovers. */
while (op < oend) *op++ = *ip++;
}
+/* ZSTD_safecopyDstBeforeSrc():
+ * This version allows overlap with dst before src, or handles the non-overlap case with dst after src
+ * Kept separate from more common ZSTD_safecopy case to avoid performance impact to the safecopy common case */
+static void ZSTD_safecopyDstBeforeSrc(BYTE* op, BYTE const* ip, ptrdiff_t length) {
+ ptrdiff_t const diff = op - ip;
+ BYTE* const oend = op + length;
+
+ if (length < 8 || diff > -8) {
+ /* Handle short lengths, close overlaps, and dst not before src. */
+ while (op < oend) *op++ = *ip++;
+ return;
+ }
+
+ if (op <= oend - WILDCOPY_OVERLENGTH && diff < -WILDCOPY_VECLEN) {
+ ZSTD_wildcopy(op, ip, oend - WILDCOPY_OVERLENGTH - op, ZSTD_no_overlap);
+ ip += oend - WILDCOPY_OVERLENGTH - op;
+ op += oend - WILDCOPY_OVERLENGTH - op;
+ }
+
+ /* Handle the leftovers. */
+ while (op < oend) *op++ = *ip++;
+}
+
/* ZSTD_execSequenceEnd():
* This version handles cases that are near the end of the output buffer. It requires
* more careful checks to make sure there is no overflow. By separating out these hard
* and unlikely cases, we can speed up the common cases.
*
* NOTE: This function needs to be fast for a single long sequence, but doesn't need
* to be optimized for many small sequences, since those fall into ZSTD_execSequence().
*/
FORCE_NOINLINE
size_t ZSTD_execSequenceEnd(BYTE* op,
- BYTE* const oend, seq_t sequence,
- const BYTE** litPtr, const BYTE* const litLimit,
- const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+ BYTE* const oend, seq_t sequence,
+ const BYTE** litPtr, const BYTE* const litLimit,
+ const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
{
BYTE* const oLitEnd = op + sequence.litLength;
size_t const sequenceLength = sequence.litLength + sequence.matchLength;
const BYTE* const iLitEnd = *litPtr + sequence.litLength;
const BYTE* match = oLitEnd - sequence.offset;
@@ -782,31 +882,80 @@
/* copy Match */
if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
/* offset beyond prefix */
RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
- match = dictEnd - (prefixStart-match);
+ match = dictEnd - (prefixStart - match);
if (match + sequence.matchLength <= dictEnd) {
ZSTD_memmove(oLitEnd, match, sequence.matchLength);
return sequenceLength;
}
/* span extDict & currentPrefixSegment */
{ size_t const length1 = dictEnd - match;
- ZSTD_memmove(oLitEnd, match, length1);
- op = oLitEnd + length1;
- sequence.matchLength -= length1;
- match = prefixStart;
- } }
+ ZSTD_memmove(oLitEnd, match, length1);
+ op = oLitEnd + length1;
+ sequence.matchLength -= length1;
+ match = prefixStart;
+ }
+ }
ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
return sequenceLength;
}
+/* ZSTD_execSequenceEndSplitLitBuffer():
+ * This version is intended to be used during instances where the litBuffer is still split. It is kept separate to avoid performance impact for the good case.
+ */
+FORCE_NOINLINE
+size_t ZSTD_execSequenceEndSplitLitBuffer(BYTE* op,
+ BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
+ const BYTE** litPtr, const BYTE* const litLimit,
+ const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+{
+ BYTE* const oLitEnd = op + sequence.litLength;
+ size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+ const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+ const BYTE* match = oLitEnd - sequence.offset;
+
+
+ /* bounds checks : careful of address space overflow in 32-bit mode */
+ RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
+ RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
+ assert(op < op + sequenceLength);
+ assert(oLitEnd < op + sequenceLength);
+
+ /* copy literals */
+ RETURN_ERROR_IF(op > *litPtr && op < *litPtr + sequence.litLength, dstSize_tooSmall, "output should not catch up to and overwrite literal buffer");
+ ZSTD_safecopyDstBeforeSrc(op, *litPtr, sequence.litLength);
+ op = oLitEnd;
+ *litPtr = iLitEnd;
+
+ /* copy Match */
+ if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
+ /* offset beyond prefix */
+ RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
+ match = dictEnd - (prefixStart - match);
+ if (match + sequence.matchLength <= dictEnd) {
+ ZSTD_memmove(oLitEnd, match, sequence.matchLength);
+ return sequenceLength;
+ }
+ /* span extDict & currentPrefixSegment */
+ { size_t const length1 = dictEnd - match;
+ ZSTD_memmove(oLitEnd, match, length1);
+ op = oLitEnd + length1;
+ sequence.matchLength -= length1;
+ match = prefixStart;
+ }
+ }
+ ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
+ 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 prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+ BYTE* const oend, seq_t sequence,
+ const BYTE** litPtr, const BYTE* const litLimit,
+ const BYTE* const prefixStart, const BYTE* const virtualStart, 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; /* risk : address space underflow on oend=NULL */
@@ -819,14 +968,106 @@
* - Read beyond end of literals
* - Match end is within WILDCOPY_OVERLIMIT of oend
* - 32-bit mode and the match length overflows
*/
if (UNLIKELY(
+ iLitEnd > litLimit ||
+ oMatchEnd > oend_w ||
+ (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
+ return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
+
+ /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
+ assert(op <= oLitEnd /* No overflow */);
+ assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
+ assert(oMatchEnd <= oend /* No underflow */);
+ assert(iLitEnd <= litLimit /* Literal length is in bounds */);
+ assert(oLitEnd <= oend_w /* Can wildcopy literals */);
+ assert(oMatchEnd <= oend_w /* Can wildcopy matches */);
+
+ /* Copy Literals:
+ * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
+ * We likely don't need the full 32-byte wildcopy.
+ */
+ assert(WILDCOPY_OVERLENGTH >= 16);
+ ZSTD_copy16(op, (*litPtr));
+ if (UNLIKELY(sequence.litLength > 16)) {
+ ZSTD_wildcopy(op + 16, (*litPtr) + 16, sequence.litLength - 16, ZSTD_no_overlap);
+ }
+ op = oLitEnd;
+ *litPtr = iLitEnd; /* update for next sequence */
+
+ /* Copy Match */
+ if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
+ /* offset beyond prefix -> go into extDict */
+ RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
+ match = dictEnd + (match - prefixStart);
+ if (match + sequence.matchLength <= dictEnd) {
+ ZSTD_memmove(oLitEnd, match, sequence.matchLength);
+ return sequenceLength;
+ }
+ /* span extDict & currentPrefixSegment */
+ { size_t const length1 = dictEnd - match;
+ ZSTD_memmove(oLitEnd, match, length1);
+ op = oLitEnd + length1;
+ sequence.matchLength -= length1;
+ match = prefixStart;
+ }
+ }
+ /* Match within prefix of 1 or more bytes */
+ assert(op <= oMatchEnd);
+ assert(oMatchEnd <= oend_w);
+ assert(match >= prefixStart);
+ assert(sequence.matchLength >= 1);
+
+ /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
+ * without overlap checking.
+ */
+ if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
+ /* We bet on a full wildcopy for matches, since we expect matches to be
+ * longer than literals (in general). In silesia, ~10% of matches are longer
+ * than 16 bytes.
+ */
+ ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
+ return sequenceLength;
+ }
+ assert(sequence.offset < WILDCOPY_VECLEN);
+
+ /* Copy 8 bytes and spread the offset to be >= 8. */
+ ZSTD_overlapCopy8(&op, &match, sequence.offset);
+
+ /* If the match length is > 8 bytes, then continue with the wildcopy. */
+ if (sequence.matchLength > 8) {
+ assert(op < oMatchEnd);
+ ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength - 8, ZSTD_overlap_src_before_dst);
+ }
+ return sequenceLength;
+}
+
+HINT_INLINE
+size_t ZSTD_execSequenceSplitLitBuffer(BYTE* op,
+ BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
+ const BYTE** litPtr, const BYTE* const litLimit,
+ const BYTE* const prefixStart, const BYTE* const virtualStart, 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) */
+ const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+ const BYTE* match = oLitEnd - sequence.offset;
+
+ assert(op != NULL /* Precondition */);
+ assert(oend_w < oend /* No underflow */);
+ /* Handle edge cases in a slow path:
+ * - Read beyond end of literals
+ * - Match end is within WILDCOPY_OVERLIMIT of oend
+ * - 32-bit mode and the match length overflows
+ */
+ if (UNLIKELY(
iLitEnd > litLimit ||
oMatchEnd > oend_w ||
(MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
- return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
+ return ZSTD_execSequenceEndSplitLitBuffer(op, oend, oend_w, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
/* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
assert(op <= oLitEnd /* No overflow */);
assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
assert(oMatchEnd <= oend /* No underflow */);
@@ -890,10 +1131,11 @@
ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8, ZSTD_overlap_src_before_dst);
}
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;
@@ -903,26 +1145,16 @@
BIT_reloadDStream(bitD);
DStatePtr->table = dt + 1;
}
FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD)
+ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16 nextState, U32 nbBits)
{
- 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;
+ DStatePtr->state = nextState + lowBits;
}
-FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD_seqSymbol const DInfo)
-{
- 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 offsets.
*/
@@ -935,106 +1167,104 @@
FORCE_INLINE_TEMPLATE seq_t
ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
{
seq_t seq;
- ZSTD_seqSymbol const llDInfo = seqState->stateLL.table[seqState->stateLL.state];
- ZSTD_seqSymbol const mlDInfo = seqState->stateML.table[seqState->stateML.state];
- ZSTD_seqSymbol const ofDInfo = seqState->stateOffb.table[seqState->stateOffb.state];
- U32 const llBase = llDInfo.baseValue;
- U32 const mlBase = mlDInfo.baseValue;
- U32 const ofBase = ofDInfo.baseValue;
- BYTE const llBits = llDInfo.nbAdditionalBits;
- BYTE const mlBits = mlDInfo.nbAdditionalBits;
- BYTE const ofBits = ofDInfo.nbAdditionalBits;
- BYTE const totalBits = llBits+mlBits+ofBits;
+ const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
+ const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
+ const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
+ seq.matchLength = mlDInfo->baseValue;
+ seq.litLength = llDInfo->baseValue;
+ { U32 const ofBase = ofDInfo->baseValue;
+ BYTE const llBits = llDInfo->nbAdditionalBits;
+ BYTE const mlBits = mlDInfo->nbAdditionalBits;
+ BYTE const ofBits = ofDInfo->nbAdditionalBits;
+ BYTE const totalBits = llBits+mlBits+ofBits;
- /* sequence */
- { size_t offset;
- if (ofBits > 1) {
- 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);
- }
- seqState->prevOffset[2] = seqState->prevOffset[1];
- seqState->prevOffset[1] = seqState->prevOffset[0];
- seqState->prevOffset[0] = offset;
- } else {
- U32 const ll0 = (llBase == 0);
- if (LIKELY((ofBits == 0))) {
- if (LIKELY(!ll0))
- offset = seqState->prevOffset[0];
- else {
- offset = seqState->prevOffset[1];
- seqState->prevOffset[1] = seqState->prevOffset[0];
- seqState->prevOffset[0] = offset;
+ U16 const llNext = llDInfo->nextState;
+ U16 const mlNext = mlDInfo->nextState;
+ U16 const ofNext = ofDInfo->nextState;
+ U32 const llnbBits = llDInfo->nbBits;
+ U32 const mlnbBits = mlDInfo->nbBits;
+ U32 const ofnbBits = ofDInfo->nbBits;
+ /*
+ * As gcc has better branch and block analyzers, sometimes it is only
+ * valuable to mark likelyness for clang, it gives around 3-4% of
+ * performance.
+ */
+
+ /* sequence */
+ { size_t offset;
+ #if defined(__clang__)
+ if (LIKELY(ofBits > 1)) {
+ #else
+ if (ofBits > 1) {
+ #endif
+ 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);
}
+ seqState->prevOffset[2] = seqState->prevOffset[1];
+ seqState->prevOffset[1] = seqState->prevOffset[0];
+ seqState->prevOffset[0] = offset;
} else {
- offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
- { 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;
- } } }
- seq.offset = offset;
- }
+ U32 const ll0 = (llDInfo->baseValue == 0);
+ if (LIKELY((ofBits == 0))) {
+ offset = seqState->prevOffset[ll0];
+ seqState->prevOffset[1] = seqState->prevOffset[!ll0];
+ seqState->prevOffset[0] = offset;
+ } else {
+ offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
+ { 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;
+ } } }
+ seq.offset = offset;
+ }
- seq.matchLength = mlBase;
- if (mlBits > 0)
- seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
+ #if defined(__clang__)
+ if (UNLIKELY(mlBits > 0))
+ #else
+ if (mlBits > 0)
+ #endif
+ seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
- if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
- BIT_reloadDStream(&seqState->DStream);
- if (MEM_64bits() && UNLIKELY(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);
+ if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
+ BIT_reloadDStream(&seqState->DStream);
+ if (MEM_64bits() && UNLIKELY(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;
- if (llBits > 0)
- seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
+ #if defined(__clang__)
+ if (UNLIKELY(llBits > 0))
+ #else
+ if (llBits > 0)
+ #endif
+ seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
- if (MEM_32bits())
- BIT_reloadDStream(&seqState->DStream);
+ 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);
+ DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
+ (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
- /* ANS state update
- * gcc-9.0.0 does 2.5% worse with ZSTD_updateFseStateWithDInfo().
- * clang-9.2.0 does 7% worse with ZSTD_updateFseState().
- * Naturally it seems like ZSTD_updateFseStateWithDInfo() should be the
- * better option, so it is the default for other compilers. But, if you
- * measure that it is worse, please put up a pull request.
- */
- {
-#if defined(__GNUC__) && !defined(__clang__)
- const int kUseUpdateFseState = 1;
-#else
- const int kUseUpdateFseState = 0;
-#endif
- if (kUseUpdateFseState) {
- 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 */
- } else {
- ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llDInfo); /* <= 9 bits */
- ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlDInfo); /* <= 9 bits */
- if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream); /* <= 18 bits */
- ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofDInfo); /* <= 8 bits */
- }
+ ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits); /* <= 9 bits */
+ ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits); /* <= 9 bits */
+ if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream); /* <= 18 bits */
+ ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits); /* <= 8 bits */
}
return seq;
}
@@ -1083,13 +1313,15 @@
#endif
}
#endif
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
+
+
FORCE_INLINE_TEMPLATE size_t
DONT_VECTORIZE
-ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
+ZSTD_decompressSequences_bodySplitLitBuffer( ZSTD_DCtx* dctx,
void* dst, size_t maxDstSize,
const void* seqStart, size_t seqSize, int nbSeq,
const ZSTD_longOffset_e isLongOffset,
const int frame)
{
@@ -1097,15 +1329,15 @@
const BYTE* const iend = ip + seqSize;
BYTE* const ostart = (BYTE*)dst;
BYTE* const oend = ostart + maxDstSize;
BYTE* op = ostart;
const BYTE* litPtr = dctx->litPtr;
- const BYTE* const litEnd = litPtr + dctx->litSize;
+ const BYTE* litBufferEnd = dctx->litBufferEnd;
const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
const BYTE* const vBase = (const BYTE*) (dctx->virtualStart);
const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
- DEBUGLOG(5, "ZSTD_decompressSequences_body");
+ DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer");
(void)frame;
/* Regen sequences */
if (nbSeq) {
seqState_t seqState;
@@ -1122,59 +1354,241 @@
ZSTD_STATIC_ASSERT(
BIT_DStream_unfinished < BIT_DStream_completed &&
BIT_DStream_endOfBuffer < BIT_DStream_completed &&
BIT_DStream_completed < BIT_DStream_overflow);
+ /* decompress without overrunning litPtr begins */
+ {
+ seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+ /* Align the decompression loop to 32 + 16 bytes.
+ *
+ * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
+ * speed swings based on the alignment of the decompression loop. This
+ * performance swing is caused by parts of the decompression loop falling
+ * out of the DSB. The entire decompression loop should fit in the DSB,
+ * when it can't we get much worse performance. You can measure if you've
+ * hit the good case or the bad case with this perf command for some
+ * compressed file test.zst:
+ *
+ * perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
+ * -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
+ *
+ * If you see most cycles served out of the MITE you've hit the bad case.
+ * 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.
+ *
+ * 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:
+ * - Haswell
+ * - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
+ * - Skylake
+ *
+ * Alignment is done for each of the three major decompression loops:
+ * - ZSTD_decompressSequences_bodySplitLitBuffer - presplit section of the literal buffer
+ * - ZSTD_decompressSequences_bodySplitLitBuffer - postsplit section of the literal buffer
+ * - ZSTD_decompressSequences_body
+ * Alignment choices are made to minimize large swings on bad cases and influence on performance
+ * from changes external to this code, rather than to overoptimize on the current commit.
+ *
+ * If you are seeing performance stability this script can help test.
+ * It tests on 4 commits in zstd where I saw performance change.
+ *
+ * https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
+ */
#if defined(__GNUC__) && defined(__x86_64__)
- /* Align the decompression loop to 32 + 16 bytes.
- *
- * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
- * speed swings based on the alignment of the decompression loop. This
- * performance swing is caused by parts of the decompression loop falling
- * out of the DSB. The entire decompression loop should fit in the DSB,
- * when it can't we get much worse performance. You can measure if you've
- * hit the good case or the bad case with this perf command for some
- * compressed file test.zst:
- *
- * perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
- * -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
- *
- * If you see most cycles served out of the MITE you've hit the bad case.
- * 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.
- *
- * 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:
- * - Haswell
- * - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
- * - Skylake
- *
- * If you are seeing performance stability this script can help test.
- * It tests on 4 commits in zstd where I saw performance change.
- *
- * 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");
+ __asm__(".p2align 6");
+# if __GNUC__ >= 7
+ /* good for gcc-7, gcc-9, and gcc-11 */
+ __asm__("nop");
+ __asm__(".p2align 5");
+ __asm__("nop");
+ __asm__(".p2align 4");
+# if __GNUC__ == 8 || __GNUC__ == 10
+ /* good for gcc-8 and gcc-10 */
+ __asm__("nop");
+ __asm__(".p2align 3");
+# endif
+# endif
+#endif
+
+ /* Handle the initial state where litBuffer is currently split between dst and litExtraBuffer */
+ for (; litPtr + sequence.litLength <= dctx->litBufferEnd; ) {
+ size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, 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));
+ sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+ }
+
+ /* If there are more sequences, they will need to read literals from litExtraBuffer; copy over the remainder from dst and update litPtr and litEnd */
+ if (nbSeq > 0) {
+ const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+ if (leftoverLit)
+ {
+ RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+ ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+ sequence.litLength -= leftoverLit;
+ op += leftoverLit;
+ }
+ litPtr = dctx->litExtraBuffer;
+ litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
+ {
+ size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, 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 (--nbSeq)
+ BIT_reloadDStream(&(seqState.DStream));
+ }
+ }
+ }
+
+ if (nbSeq > 0) /* there is remaining lit from extra buffer */
+ {
+
+#if defined(__GNUC__) && defined(__x86_64__)
+ __asm__(".p2align 6");
+ __asm__("nop");
+# if __GNUC__ != 7
+ /* worse for gcc-7 better for gcc-8, gcc-9, and gcc-10 and clang */
+ __asm__(".p2align 4");
+ __asm__("nop");
+ __asm__(".p2align 3");
+# elif __GNUC__ >= 11
+ __asm__(".p2align 3");
# else
- __asm__(".p2align 4");
+ __asm__(".p2align 5");
+ __asm__("nop");
+ __asm__(".p2align 3");
# endif
#endif
+
+ for (; ; ) {
+ seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+ size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, 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));
+ }
+ }
+
+ /* check if reached exact end */
+ DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer: after decode loop, remaining nbSeq : %i", nbSeq);
+ RETURN_ERROR_IF(nbSeq, corruption_detected, "");
+ RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, 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 */
+ if (dctx->litBufferLocation == ZSTD_split) /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
+ {
+ size_t const lastLLSize = litBufferEnd - litPtr;
+ RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
+ if (op != NULL) {
+ ZSTD_memmove(op, litPtr, lastLLSize);
+ op += lastLLSize;
+ }
+ litPtr = dctx->litExtraBuffer;
+ litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
+ }
+ { size_t const lastLLSize = litBufferEnd - litPtr;
+ RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
+ if (op != NULL) {
+ ZSTD_memcpy(op, litPtr, lastLLSize);
+ op += lastLLSize;
+ }
+ }
+
+ return op-ostart;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+DONT_VECTORIZE
+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 int frame)
+{
+ const BYTE* ip = (const BYTE*)seqStart;
+ const BYTE* const iend = ip + seqSize;
+ BYTE* const ostart = (BYTE*)dst;
+ BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ostart + maxDstSize : dctx->litBuffer;
+ BYTE* op = ostart;
+ const BYTE* litPtr = dctx->litPtr;
+ const BYTE* const litEnd = litPtr + dctx->litSize;
+ const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
+ const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
+ const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
+ DEBUGLOG(5, "ZSTD_decompressSequences_body");
+ (void)frame;
+
+ /* 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]; }
+ RETURN_ERROR_IF(
+ ERR_isError(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);
+ assert(dst != NULL);
+
+ ZSTD_STATIC_ASSERT(
+ BIT_DStream_unfinished < BIT_DStream_completed &&
+ BIT_DStream_endOfBuffer < BIT_DStream_completed &&
+ BIT_DStream_completed < BIT_DStream_overflow);
+
+#if defined(__GNUC__) && defined(__x86_64__)
+ __asm__(".p2align 6");
+ __asm__("nop");
+# if __GNUC__ >= 7
+ __asm__(".p2align 5");
+ __asm__("nop");
+ __asm__(".p2align 3");
+# else
+ __asm__(".p2align 4");
+ __asm__("nop");
+ __asm__(".p2align 3");
+# endif
+#endif
+
for ( ; ; ) {
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));
@@ -1216,10 +1630,20 @@
const ZSTD_longOffset_e isLongOffset,
const int frame)
{
return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
+
+static size_t
+ZSTD_decompressSequencesSplitLitBuffer_default(ZSTD_DCtx* dctx,
+ void* dst, size_t maxDstSize,
+ const void* seqStart, size_t seqSize, int nbSeq,
+ const ZSTD_longOffset_e isLongOffset,
+ const int frame)
+{
+ return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
FORCE_INLINE_TEMPLATE size_t
@@ -1248,14 +1672,14 @@
const int frame)
{
const BYTE* ip = (const BYTE*)seqStart;
const BYTE* const iend = ip + seqSize;
BYTE* const ostart = (BYTE*)dst;
- BYTE* const oend = ostart + maxDstSize;
+ BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ostart + maxDstSize;
BYTE* op = ostart;
const BYTE* litPtr = dctx->litPtr;
- const BYTE* const litEnd = litPtr + dctx->litSize;
+ const BYTE* litBufferEnd = dctx->litBufferEnd;
const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
const BYTE* const dictStart = (const BYTE*) (dctx->virtualStart);
const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
(void)frame;
@@ -1287,47 +1711,120 @@
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);
- size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+ /* decompress without stomping litBuffer */
+ for (; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb < nbSeq); seqNb++) {
+ seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+ size_t oneSeqSize;
+
+ if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd)
+ {
+ /* lit buffer is reaching split point, empty out the first buffer and transition to litExtraBuffer */
+ const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+ if (leftoverLit)
+ {
+ RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+ ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+ sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength -= leftoverLit;
+ op += leftoverLit;
+ }
+ litPtr = dctx->litExtraBuffer;
+ litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
+ oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, 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);
+ 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;
+ if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
- prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
- sequences[seqNb & STORED_SEQS_MASK] = sequence;
- op += oneSeqSize;
+ prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+ sequences[seqNb & STORED_SEQS_MASK] = sequence;
+ op += oneSeqSize;
+ }
+ else
+ {
+ /* lit buffer is either wholly contained in first or second split, or not split at all*/
+ oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
+ ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength - WILDCOPY_OVERLENGTH, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
+ ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, 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;
+
+ prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+ sequences[seqNb & STORED_SEQS_MASK] = sequence;
+ op += oneSeqSize;
+ }
}
RETURN_ERROR_IF(seqNb<nbSeq, corruption_detected, "");
/* finish queue */
seqNb -= seqAdvance;
for ( ; seqNb<nbSeq ; seqNb++) {
- size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[seqNb&STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+ seq_t *sequence = &(sequences[seqNb&STORED_SEQS_MASK]);
+ if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd)
+ {
+ const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+ if (leftoverLit)
+ {
+ RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+ ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+ sequence->litLength -= leftoverLit;
+ op += leftoverLit;
+ }
+ litPtr = dctx->litExtraBuffer;
+ litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+ dctx->litBufferLocation = ZSTD_not_in_dst;
+ {
+ size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, 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&STORED_SEQS_MASK], prefixStart, dictStart);
+ assert(!ZSTD_isError(oneSeqSize));
+ if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
#endif
- if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
- op += oneSeqSize;
+ if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+ op += oneSeqSize;
+ }
+ }
+ else
+ {
+ size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
+ ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence->litLength - WILDCOPY_OVERLENGTH, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
+ ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, 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&STORED_SEQS_MASK], prefixStart, dictStart);
+#endif
+ 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]); }
}
/* last literal segment */
- { size_t const lastLLSize = litEnd - litPtr;
+ if (dctx->litBufferLocation == ZSTD_split) /* first deplete literal buffer in dst, then copy litExtraBuffer */
+ {
+ size_t const lastLLSize = litBufferEnd - litPtr;
+ RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
+ if (op != NULL) {
+ ZSTD_memmove(op, litPtr, lastLLSize);
+ op += lastLLSize;
+ }
+ litPtr = dctx->litExtraBuffer;
+ litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+ }
+ { size_t const lastLLSize = litBufferEnd - litPtr;
RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
if (op != NULL) {
- ZSTD_memcpy(op, litPtr, lastLLSize);
+ ZSTD_memmove(op, litPtr, lastLLSize);
op += lastLLSize;
}
}
return op-ostart;
@@ -1347,24 +1844,34 @@
#if DYNAMIC_BMI2
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
DONT_VECTORIZE
ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
void* dst, size_t maxDstSize,
const void* seqStart, size_t seqSize, int nbSeq,
const ZSTD_longOffset_e isLongOffset,
const int frame)
{
return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
+static BMI2_TARGET_ATTRIBUTE size_t
+DONT_VECTORIZE
+ZSTD_decompressSequencesSplitLitBuffer_bmi2(ZSTD_DCtx* dctx,
+ void* dst, size_t maxDstSize,
+ const void* seqStart, size_t seqSize, int nbSeq,
+ const ZSTD_longOffset_e isLongOffset,
+ const int frame)
+{
+ return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE 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,
const int frame)
@@ -1389,16 +1896,30 @@
const ZSTD_longOffset_e isLongOffset,
const int frame)
{
DEBUGLOG(5, "ZSTD_decompressSequences");
#if DYNAMIC_BMI2
- if (dctx->bmi2) {
+ if (ZSTD_DCtx_get_bmi2(dctx)) {
return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
#endif
- return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+ return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
+static size_t
+ZSTD_decompressSequencesSplitLitBuffer(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
+ const void* seqStart, size_t seqSize, int nbSeq,
+ const ZSTD_longOffset_e isLongOffset,
+ const int frame)
+{
+ DEBUGLOG(5, "ZSTD_decompressSequencesSplitLitBuffer");
+#if DYNAMIC_BMI2
+ if (ZSTD_DCtx_get_bmi2(dctx)) {
+ return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+ }
+#endif
+ return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
/* ZSTD_decompressSequencesLong() :
@@ -1413,11 +1934,11 @@
const ZSTD_longOffset_e isLongOffset,
const int frame)
{
DEBUGLOG(5, "ZSTD_decompressSequencesLong");
#if DYNAMIC_BMI2
- if (dctx->bmi2) {
+ if (ZSTD_DCtx_get_bmi2(dctx)) {
return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
#endif
return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
}
@@ -1454,11 +1975,11 @@
#endif
size_t
ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
void* dst, size_t dstCapacity,
- const void* src, size_t srcSize, const int frame)
+ const void* src, size_t srcSize, const int frame, const streaming_operation streaming)
{ /* 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.
@@ -1469,11 +1990,11 @@
DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
RETURN_ERROR_IF(srcSize >= ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
/* Decode literals section */
- { size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize);
+ { size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize);
if (ZSTD_isError(litCSize)) return litCSize;
ip += litCSize;
srcSize -= litCSize;
}
@@ -1517,11 +2038,14 @@
return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
#endif
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
/* else */
- return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+ if (dctx->litBufferLocation == ZSTD_split)
+ return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+ else
+ return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
#endif
}
}
@@ -1540,9 +2064,9 @@
void* dst, size_t dstCapacity,
const void* src, size_t srcSize)
{
size_t dSize;
ZSTD_checkContinuity(dctx, dst, dstCapacity);
- dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0);
+ dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0, not_streaming);
dctx->previousDstEnd = (char*)dst + dSize;
return dSize;
}