ext/zstdruby/libzstd/decompress/zstd_decompress_block.c in zstd-ruby-1.5.2.3 vs ext/zstdruby/libzstd/decompress/zstd_decompress_block.c in zstd-ruby-1.5.4.0

- old
+ new

@@ -1,7 +1,7 @@ /* - * Copyright (c) Yann Collet, Facebook, Inc. + * Copyright (c) Meta Platforms, Inc. and affiliates. * All rights reserved. * * This source code is licensed under both the BSD-style license (found in the * LICENSE file in the root directory of this source tree) and the GPLv2 (found * in the COPYING file in the root directory of this source tree). @@ -18,16 +18,16 @@ #include "../common/compiler.h" /* prefetch */ #include "../common/cpu.h" /* bmi2 */ #include "../common/mem.h" /* low level memory routines */ #define FSE_STATIC_LINKING_ONLY #include "../common/fse.h" -#define HUF_STATIC_LINKING_ONLY #include "../common/huf.h" #include "../common/zstd_internal.h" #include "zstd_decompress_internal.h" /* ZSTD_DCtx */ #include "zstd_ddict.h" /* ZSTD_DDictDictContent */ #include "zstd_decompress_block.h" +#include "../common/bits.h" /* ZSTD_highbit32 */ /*_******************************************************* * Macros **********************************************************/ @@ -87,11 +87,11 @@ /* 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 */ + /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */ dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize; dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize; } dctx->litBufferLocation = ZSTD_split; } @@ -132,17 +132,20 @@ DEBUGLOG(5, "set_repeat flag : re-using stats from previous compressed literals block"); RETURN_ERROR_IF(dctx->litEntropy==0, dictionary_corrupted, ""); 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"); + RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; 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); + int const flags = 0 + | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0) + | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0); switch(lhlCode) { case 0: case 1: default: /* note : default is impossible, since lhlCode into [0..3] */ /* 2 - 2 - 10 - 10 */ singleStream = !lhlCode; @@ -163,10 +166,14 @@ 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, ""); + if (!singleStream) + RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong, + "Not enough literals (%zu) for the 4-streams mode (min %u)", + litSize, MIN_LITERALS_FOR_4_STREAMS); 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 */ @@ -174,36 +181,37 @@ PREFETCH_AREA(dctx->HUFptr, sizeof(dctx->entropy.hufTable)); } if (litEncType==set_repeat) { if (singleStream) { - hufSuccess = HUF_decompress1X_usingDTable_bmi2( + hufSuccess = HUF_decompress1X_usingDTable( dctx->litBuffer, litSize, istart+lhSize, litCSize, - dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx)); + dctx->HUFptr, flags); } else { - hufSuccess = HUF_decompress4X_usingDTable_bmi2( + assert(litSize >= MIN_LITERALS_FOR_4_STREAMS); + hufSuccess = HUF_decompress4X_usingDTable( dctx->litBuffer, litSize, istart+lhSize, litCSize, - dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx)); + dctx->HUFptr, flags); } } else { if (singleStream) { #if defined(HUF_FORCE_DECOMPRESS_X2) hufSuccess = HUF_decompress1X_DCtx_wksp( dctx->entropy.hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->workspace, - sizeof(dctx->workspace)); + sizeof(dctx->workspace), flags); #else - hufSuccess = HUF_decompress1X1_DCtx_wksp_bmi2( + hufSuccess = HUF_decompress1X1_DCtx_wksp( dctx->entropy.hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->workspace, - sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx)); + sizeof(dctx->workspace), flags); #endif } else { - hufSuccess = HUF_decompress4X_hufOnly_wksp_bmi2( + hufSuccess = HUF_decompress4X_hufOnly_wksp( dctx->entropy.hufTable, dctx->litBuffer, litSize, istart+lhSize, litCSize, dctx->workspace, - sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx)); + sizeof(dctx->workspace), flags); } } if (dctx->litBufferLocation == ZSTD_split) { ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE); @@ -235,10 +243,11 @@ lhSize = 2; litSize = MEM_readLE16(istart) >> 4; break; case 3: lhSize = 3; + RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3"); litSize = MEM_readLE24(istart) >> 4; break; } RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled"); @@ -277,16 +286,17 @@ lhSize = 1; litSize = istart[0] >> 3; break; case 1: lhSize = 2; + RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3"); litSize = MEM_readLE16(istart) >> 4; break; case 3: lhSize = 3; + RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4"); 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, ""); RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, ""); @@ -504,18 +514,19 @@ 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; + assert(n>=0); + pos += (size_t)n; } } /* Now we spread those positions across the table. - * The benefit of doing it in two stages is that we avoid the the + * The benefit of doing it in two stages is that we avoid 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. + * We unroll the loop twice, since that is what empirically worked best. */ { size_t position = 0; size_t s; size_t const unroll = 2; @@ -538,22 +549,22 @@ int 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 */ + while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask; /* lowprob area */ } } assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */ } /* Build Decoding table */ { U32 u; for (u=0; u<tableSize; u++) { U32 const symbol = tableDecode[u].baseValue; U32 const nextState = symbolNext[symbol]++; - tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) ); + tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) ); tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize); assert(nbAdditionalBits[symbol] < 255); tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol]; tableDecode[u].baseValue = baseValue[symbol]; } @@ -962,10 +973,15 @@ const BYTE* const iLitEnd = *litPtr + sequence.litLength; const BYTE* match = oLitEnd - sequence.offset; assert(op != NULL /* Precondition */); assert(oend_w < oend /* No underflow */); + +#if defined(__aarch64__) + /* prefetch sequence starting from match that will be used for copy later */ + PREFETCH_L1(match); +#endif /* 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 */ @@ -1152,11 +1168,11 @@ size_t const lowBits = BIT_readBits(bitD, nbBits); DStatePtr->state = 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) + * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32 * bits before reloading. This value is the maximum number of bytes we read * after reloading when we are decoding long offsets. */ #define LONG_OFFSETS_MAX_EXTRA_BITS_32 \ (ZSTD_WINDOWLOG_MAX_32 > STREAM_ACCUMULATOR_MIN_32 \ @@ -1167,13 +1183,31 @@ FORCE_INLINE_TEMPLATE seq_t ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets) { seq_t seq; + /* + * ZSTD_seqSymbol is a structure with a total of 64 bits wide. So it can be + * loaded in one operation and extracted its fields by simply shifting or + * bit-extracting on aarch64. + * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh + * operations that cause performance drop. This can be avoided by using this + * ZSTD_memcpy hack. + */ +#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__)) + ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS; + ZSTD_seqSymbol* const llDInfo = &llDInfoS; + ZSTD_seqSymbol* const mlDInfo = &mlDInfoS; + ZSTD_seqSymbol* const ofDInfo = &ofDInfoS; + ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol)); + ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol)); + ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol)); +#else 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; +#endif seq.matchLength = mlDInfo->baseValue; seq.litLength = llDInfo->baseValue; { U32 const ofBase = ofDInfo->baseValue; BYTE const llBits = llDInfo->nbAdditionalBits; BYTE const mlBits = mlDInfo->nbAdditionalBits; @@ -1184,13 +1218,17 @@ 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; + + assert(llBits <= MaxLLBits); + assert(mlBits <= MaxMLBits); + assert(ofBits <= MaxOff); /* * As gcc has better branch and block analyzers, sometimes it is only - * valuable to mark likelyness for clang, it gives around 3-4% of + * valuable to mark likeliness for clang, it gives around 3-4% of * performance. */ /* sequence */ { size_t offset; @@ -1199,17 +1237,20 @@ #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); + ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32); + ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits); if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) { - U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed); + /* Always read extra bits, this keeps the logic simple, + * avoids branches, and avoids accidentally reading 0 bits. + */ + U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32; 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 */ + offset += BIT_readBitsFast(&seqState->DStream, extraBits); } 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]; @@ -1550,11 +1591,11 @@ 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"); + DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq); (void)frame; /* Regen sequences */ if (nbSeq) { seqState_t seqState; @@ -1943,101 +1984,181 @@ return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame); } #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */ +/** + * @returns The total size of the history referencable by zstd, including + * both the prefix and the extDict. At @p op any offset larger than this + * is invalid. + */ +static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart) +{ + return (size_t)(op - virtualStart); +} -#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \ - !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG) -/* ZSTD_getLongOffsetsShare() : +typedef struct { + unsigned longOffsetShare; + unsigned maxNbAdditionalBits; +} ZSTD_OffsetInfo; + +/* ZSTD_getOffsetInfo() : * condition : offTable must be valid * @return : "share" of long offsets (arbitrarily defined as > (1<<23)) - * compared to maximum possible of (1<<OffFSELog) */ -static unsigned -ZSTD_getLongOffsetsShare(const ZSTD_seqSymbol* offTable) + * compared to maximum possible of (1<<OffFSELog), + * as well as the maximum number additional bits required. + */ +static ZSTD_OffsetInfo +ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq) { - const void* ptr = offTable; - U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog; - const ZSTD_seqSymbol* table = offTable + 1; - U32 const max = 1 << tableLog; - U32 u, total = 0; - DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog); + ZSTD_OffsetInfo info = {0, 0}; + /* If nbSeq == 0, then the offTable is uninitialized, but we have + * no sequences, so both values should be 0. + */ + if (nbSeq != 0) { + const void* ptr = offTable; + U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog; + const ZSTD_seqSymbol* table = offTable + 1; + U32 const max = 1 << tableLog; + U32 u; + DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog); - assert(max <= (1 << OffFSELog)); /* max not too large */ - for (u=0; u<max; u++) { - if (table[u].nbAdditionalBits > 22) total += 1; + assert(max <= (1 << OffFSELog)); /* max not too large */ + for (u=0; u<max; u++) { + info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits); + if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1; + } + + assert(tableLog <= OffFSELog); + info.longOffsetShare <<= (OffFSELog - tableLog); /* scale to OffFSELog */ } - assert(tableLog <= OffFSELog); - total <<= (OffFSELog - tableLog); /* scale to OffFSELog */ + return info; +} - return total; +/** + * @returns The maximum offset we can decode in one read of our bitstream, without + * reloading more bits in the middle of the offset bits read. Any offsets larger + * than this must use the long offset decoder. + */ +static size_t ZSTD_maxShortOffset(void) +{ + if (MEM_64bits()) { + /* We can decode any offset without reloading bits. + * This might change if the max window size grows. + */ + ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31); + return (size_t)-1; + } else { + /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1. + * This offBase would require STREAM_ACCUMULATOR_MIN extra bits. + * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset. + */ + size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1; + size_t const maxOffset = maxOffbase - ZSTD_REP_NUM; + assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN); + return maxOffset; + } } -#endif size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, 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. - * In block mode, window size is not known, so we have to be conservative. - * (note: but it could be evaluated from current-lowLimit) - */ - ZSTD_longOffset_e const isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (!frame || (dctx->fParams.windowSize > (1ULL << STREAM_ACCUMULATOR_MIN)))); DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize); - RETURN_ERROR_IF(srcSize >= ZSTD_BLOCKSIZE_MAX, srcSize_wrong, ""); + /* Note : the wording of the specification + * allows compressed block to be sized exactly ZSTD_BLOCKSIZE_MAX. + * This generally does not happen, as it makes little sense, + * since an uncompressed block would feature same size and have no decompression cost. + * Also, note that decoder from reference libzstd before < v1.5.4 + * would consider this edge case as an error. + * As a consequence, avoid generating compressed blocks of size ZSTD_BLOCKSIZE_MAX + * for broader compatibility with the deployed ecosystem of zstd decoders */ + RETURN_ERROR_IF(srcSize > ZSTD_BLOCKSIZE_MAX, srcSize_wrong, ""); /* Decode literals section */ { size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming); - DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize); + DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize); if (ZSTD_isError(litCSize)) return litCSize; ip += litCSize; srcSize -= litCSize; } /* Build Decoding Tables */ { + /* Compute the maximum block size, which must also work when !frame and fParams are unset. + * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t. + */ + size_t const blockSizeMax = MIN(dstCapacity, (frame ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX)); + size_t const totalHistorySize = ZSTD_totalHistorySize((BYTE*)dst + blockSizeMax, (BYTE const*)dctx->virtualStart); + /* isLongOffset must be true if there are long offsets. + * Offsets are long if they are larger than ZSTD_maxShortOffset(). + * We don't expect that to be the case in 64-bit mode. + * + * We check here to see if our history is large enough to allow long offsets. + * If it isn't, then we can't possible have (valid) long offsets. If the offset + * is invalid, then it is okay to read it incorrectly. + * + * If isLongOffsets is true, then we will later check our decoding table to see + * if it is even possible to generate long offsets. + */ + ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset())); /* These macros control at build-time which decompressor implementation * we use. If neither is defined, we do some inspection and dispatch at * runtime. */ #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \ !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG) int usePrefetchDecoder = dctx->ddictIsCold; +#else + /* Set to 1 to avoid computing offset info if we don't need to. + * Otherwise this value is ignored. + */ + int usePrefetchDecoder = 1; #endif int nbSeq; size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize); if (ZSTD_isError(seqHSize)) return seqHSize; ip += seqHSize; srcSize -= seqHSize; RETURN_ERROR_IF(dst == NULL && nbSeq > 0, dstSize_tooSmall, "NULL not handled"); -#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \ - !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG) - if ( !usePrefetchDecoder - && (!frame || (dctx->fParams.windowSize > (1<<24))) - && (nbSeq>ADVANCED_SEQS) ) { /* could probably use a larger nbSeq limit */ - U32 const shareLongOffsets = ZSTD_getLongOffsetsShare(dctx->OFTptr); - U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */ - usePrefetchDecoder = (shareLongOffsets >= minShare); + /* If we could potentially have long offsets, or we might want to use the prefetch decoder, + * compute information about the share of long offsets, and the maximum nbAdditionalBits. + * NOTE: could probably use a larger nbSeq limit + */ + if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) { + ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq); + if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) { + /* If isLongOffset, but the maximum number of additional bits that we see in our table is small + * enough, then we know it is impossible to have too long an offset in this block, so we can + * use the regular offset decoder. + */ + isLongOffset = ZSTD_lo_isRegularOffset; + } + if (!usePrefetchDecoder) { + U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */ + usePrefetchDecoder = (info.longOffsetShare >= minShare); + } } -#endif dctx->ddictIsCold = 0; #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \ !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG) - if (usePrefetchDecoder) + if (usePrefetchDecoder) { +#else + (void)usePrefetchDecoder; + { #endif #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame); #endif + } #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG /* else */ if (dctx->litBufferLocation == ZSTD_split) return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);