diff options
Diffstat (limited to 'thirdparty/zstd/compress')
27 files changed, 7637 insertions, 3124 deletions
diff --git a/thirdparty/zstd/compress/clevels.h b/thirdparty/zstd/compress/clevels.h new file mode 100644 index 0000000000..7ed2e00490 --- /dev/null +++ b/thirdparty/zstd/compress/clevels.h @@ -0,0 +1,134 @@ +/* + * Copyright (c) Yann Collet, Facebook, Inc. + * 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). + * You may select, at your option, one of the above-listed licenses. + */ + +#ifndef ZSTD_CLEVELS_H +#define ZSTD_CLEVELS_H + +#define ZSTD_STATIC_LINKING_ONLY /* ZSTD_compressionParameters */ +#include "../zstd.h" + +/*-===== Pre-defined compression levels =====-*/ + +#define ZSTD_MAX_CLEVEL 22 + +#ifdef __GNUC__ +__attribute__((__unused__)) +#endif + +static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = { +{ /* "default" - for any srcSize > 256 KB */ + /* W, C, H, S, L, TL, strat */ + { 19, 12, 13, 1, 6, 1, ZSTD_fast }, /* base for negative levels */ + { 19, 13, 14, 1, 7, 0, ZSTD_fast }, /* level 1 */ + { 20, 15, 16, 1, 6, 0, ZSTD_fast }, /* level 2 */ + { 21, 16, 17, 1, 5, 0, ZSTD_dfast }, /* level 3 */ + { 21, 18, 18, 1, 5, 0, ZSTD_dfast }, /* level 4 */ + { 21, 18, 19, 3, 5, 2, ZSTD_greedy }, /* level 5 */ + { 21, 18, 19, 3, 5, 4, ZSTD_lazy }, /* level 6 */ + { 21, 19, 20, 4, 5, 8, ZSTD_lazy }, /* level 7 */ + { 21, 19, 20, 4, 5, 16, ZSTD_lazy2 }, /* level 8 */ + { 22, 20, 21, 4, 5, 16, ZSTD_lazy2 }, /* level 9 */ + { 22, 21, 22, 5, 5, 16, ZSTD_lazy2 }, /* level 10 */ + { 22, 21, 22, 6, 5, 16, ZSTD_lazy2 }, /* level 11 */ + { 22, 22, 23, 6, 5, 32, ZSTD_lazy2 }, /* level 12 */ + { 22, 22, 22, 4, 5, 32, ZSTD_btlazy2 }, /* level 13 */ + { 22, 22, 23, 5, 5, 32, ZSTD_btlazy2 }, /* level 14 */ + { 22, 23, 23, 6, 5, 32, ZSTD_btlazy2 }, /* level 15 */ + { 22, 22, 22, 5, 5, 48, ZSTD_btopt }, /* level 16 */ + { 23, 23, 22, 5, 4, 64, ZSTD_btopt }, /* level 17 */ + { 23, 23, 22, 6, 3, 64, ZSTD_btultra }, /* level 18 */ + { 23, 24, 22, 7, 3,256, ZSTD_btultra2}, /* level 19 */ + { 25, 25, 23, 7, 3,256, ZSTD_btultra2}, /* level 20 */ + { 26, 26, 24, 7, 3,512, ZSTD_btultra2}, /* level 21 */ + { 27, 27, 25, 9, 3,999, ZSTD_btultra2}, /* level 22 */ +}, +{ /* for srcSize <= 256 KB */ + /* W, C, H, S, L, T, strat */ + { 18, 12, 13, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ + { 18, 13, 14, 1, 6, 0, ZSTD_fast }, /* level 1 */ + { 18, 14, 14, 1, 5, 0, ZSTD_dfast }, /* level 2 */ + { 18, 16, 16, 1, 4, 0, ZSTD_dfast }, /* level 3 */ + { 18, 16, 17, 3, 5, 2, ZSTD_greedy }, /* level 4.*/ + { 18, 17, 18, 5, 5, 2, ZSTD_greedy }, /* level 5.*/ + { 18, 18, 19, 3, 5, 4, ZSTD_lazy }, /* level 6.*/ + { 18, 18, 19, 4, 4, 4, ZSTD_lazy }, /* level 7 */ + { 18, 18, 19, 4, 4, 8, ZSTD_lazy2 }, /* level 8 */ + { 18, 18, 19, 5, 4, 8, ZSTD_lazy2 }, /* level 9 */ + { 18, 18, 19, 6, 4, 8, ZSTD_lazy2 }, /* level 10 */ + { 18, 18, 19, 5, 4, 12, ZSTD_btlazy2 }, /* level 11.*/ + { 18, 19, 19, 7, 4, 12, ZSTD_btlazy2 }, /* level 12.*/ + { 18, 18, 19, 4, 4, 16, ZSTD_btopt }, /* level 13 */ + { 18, 18, 19, 4, 3, 32, ZSTD_btopt }, /* level 14.*/ + { 18, 18, 19, 6, 3,128, ZSTD_btopt }, /* level 15.*/ + { 18, 19, 19, 6, 3,128, ZSTD_btultra }, /* level 16.*/ + { 18, 19, 19, 8, 3,256, ZSTD_btultra }, /* level 17.*/ + { 18, 19, 19, 6, 3,128, ZSTD_btultra2}, /* level 18.*/ + { 18, 19, 19, 8, 3,256, ZSTD_btultra2}, /* level 19.*/ + { 18, 19, 19, 10, 3,512, ZSTD_btultra2}, /* level 20.*/ + { 18, 19, 19, 12, 3,512, ZSTD_btultra2}, /* level 21.*/ + { 18, 19, 19, 13, 3,999, ZSTD_btultra2}, /* level 22.*/ +}, +{ /* for srcSize <= 128 KB */ + /* W, C, H, S, L, T, strat */ + { 17, 12, 12, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ + { 17, 12, 13, 1, 6, 0, ZSTD_fast }, /* level 1 */ + { 17, 13, 15, 1, 5, 0, ZSTD_fast }, /* level 2 */ + { 17, 15, 16, 2, 5, 0, ZSTD_dfast }, /* level 3 */ + { 17, 17, 17, 2, 4, 0, ZSTD_dfast }, /* level 4 */ + { 17, 16, 17, 3, 4, 2, ZSTD_greedy }, /* level 5 */ + { 17, 16, 17, 3, 4, 4, ZSTD_lazy }, /* level 6 */ + { 17, 16, 17, 3, 4, 8, ZSTD_lazy2 }, /* level 7 */ + { 17, 16, 17, 4, 4, 8, ZSTD_lazy2 }, /* level 8 */ + { 17, 16, 17, 5, 4, 8, ZSTD_lazy2 }, /* level 9 */ + { 17, 16, 17, 6, 4, 8, ZSTD_lazy2 }, /* level 10 */ + { 17, 17, 17, 5, 4, 8, ZSTD_btlazy2 }, /* level 11 */ + { 17, 18, 17, 7, 4, 12, ZSTD_btlazy2 }, /* level 12 */ + { 17, 18, 17, 3, 4, 12, ZSTD_btopt }, /* level 13.*/ + { 17, 18, 17, 4, 3, 32, ZSTD_btopt }, /* level 14.*/ + { 17, 18, 17, 6, 3,256, ZSTD_btopt }, /* level 15.*/ + { 17, 18, 17, 6, 3,128, ZSTD_btultra }, /* level 16.*/ + { 17, 18, 17, 8, 3,256, ZSTD_btultra }, /* level 17.*/ + { 17, 18, 17, 10, 3,512, ZSTD_btultra }, /* level 18.*/ + { 17, 18, 17, 5, 3,256, ZSTD_btultra2}, /* level 19.*/ + { 17, 18, 17, 7, 3,512, ZSTD_btultra2}, /* level 20.*/ + { 17, 18, 17, 9, 3,512, ZSTD_btultra2}, /* level 21.*/ + { 17, 18, 17, 11, 3,999, ZSTD_btultra2}, /* level 22.*/ +}, +{ /* for srcSize <= 16 KB */ + /* W, C, H, S, L, T, strat */ + { 14, 12, 13, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ + { 14, 14, 15, 1, 5, 0, ZSTD_fast }, /* level 1 */ + { 14, 14, 15, 1, 4, 0, ZSTD_fast }, /* level 2 */ + { 14, 14, 15, 2, 4, 0, ZSTD_dfast }, /* level 3 */ + { 14, 14, 14, 4, 4, 2, ZSTD_greedy }, /* level 4 */ + { 14, 14, 14, 3, 4, 4, ZSTD_lazy }, /* level 5.*/ + { 14, 14, 14, 4, 4, 8, ZSTD_lazy2 }, /* level 6 */ + { 14, 14, 14, 6, 4, 8, ZSTD_lazy2 }, /* level 7 */ + { 14, 14, 14, 8, 4, 8, ZSTD_lazy2 }, /* level 8.*/ + { 14, 15, 14, 5, 4, 8, ZSTD_btlazy2 }, /* level 9.*/ + { 14, 15, 14, 9, 4, 8, ZSTD_btlazy2 }, /* level 10.*/ + { 14, 15, 14, 3, 4, 12, ZSTD_btopt }, /* level 11.*/ + { 14, 15, 14, 4, 3, 24, ZSTD_btopt }, /* level 12.*/ + { 14, 15, 14, 5, 3, 32, ZSTD_btultra }, /* level 13.*/ + { 14, 15, 15, 6, 3, 64, ZSTD_btultra }, /* level 14.*/ + { 14, 15, 15, 7, 3,256, ZSTD_btultra }, /* level 15.*/ + { 14, 15, 15, 5, 3, 48, ZSTD_btultra2}, /* level 16.*/ + { 14, 15, 15, 6, 3,128, ZSTD_btultra2}, /* level 17.*/ + { 14, 15, 15, 7, 3,256, ZSTD_btultra2}, /* level 18.*/ + { 14, 15, 15, 8, 3,256, ZSTD_btultra2}, /* level 19.*/ + { 14, 15, 15, 8, 3,512, ZSTD_btultra2}, /* level 20.*/ + { 14, 15, 15, 9, 3,512, ZSTD_btultra2}, /* level 21.*/ + { 14, 15, 15, 10, 3,999, ZSTD_btultra2}, /* level 22.*/ +}, +}; + + + +#endif /* ZSTD_CLEVELS_H */ diff --git a/thirdparty/zstd/compress/fse_compress.c b/thirdparty/zstd/compress/fse_compress.c index a42759814f..5547b4ac09 100644 --- a/thirdparty/zstd/compress/fse_compress.c +++ b/thirdparty/zstd/compress/fse_compress.c @@ -1,6 +1,6 @@ /* ****************************************************************** * FSE : Finite State Entropy encoder - * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc. + * Copyright (c) Yann Collet, Facebook, Inc. * * You can contact the author at : * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy @@ -15,8 +15,6 @@ /* ************************************************************** * Includes ****************************************************************/ -#include <stdlib.h> /* malloc, free, qsort */ -#include <string.h> /* memcpy, memset */ #include "../common/compiler.h" #include "../common/mem.h" /* U32, U16, etc. */ #include "../common/debug.h" /* assert, DEBUGLOG */ @@ -25,6 +23,9 @@ #define FSE_STATIC_LINKING_ONLY #include "../common/fse.h" #include "../common/error_private.h" +#define ZSTD_DEPS_NEED_MALLOC +#define ZSTD_DEPS_NEED_MATH64 +#include "../common/zstd_deps.h" /* ZSTD_malloc, ZSTD_free, ZSTD_memcpy, ZSTD_memset */ /* ************************************************************** @@ -74,13 +75,16 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ; FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT); U32 const step = FSE_TABLESTEP(tableSize); - U32 cumul[FSE_MAX_SYMBOL_VALUE+2]; + U32 const maxSV1 = maxSymbolValue+1; + + U16* cumul = (U16*)workSpace; /* size = maxSV1 */ + FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1)); /* size = tableSize */ - FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)workSpace; U32 highThreshold = tableSize-1; + assert(((size_t)workSpace & 1) == 0); /* Must be 2 bytes-aligned */ + if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge); /* CTable header */ - if (((size_t)1 << tableLog) * sizeof(FSE_FUNCTION_TYPE) > wkspSize) return ERROR(tableLog_tooLarge); tableU16[-2] = (U16) tableLog; tableU16[-1] = (U16) maxSymbolValue; assert(tableLog < 16); /* required for threshold strategy to work */ @@ -89,26 +93,67 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, * http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */ #ifdef __clang_analyzer__ - memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize); /* useless initialization, just to keep scan-build happy */ + ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize); /* useless initialization, just to keep scan-build happy */ #endif /* symbol start positions */ { U32 u; cumul[0] = 0; - for (u=1; u <= maxSymbolValue+1; u++) { + for (u=1; u <= maxSV1; u++) { if (normalizedCounter[u-1]==-1) { /* Low proba symbol */ cumul[u] = cumul[u-1] + 1; tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1); } else { - cumul[u] = cumul[u-1] + normalizedCounter[u-1]; + assert(normalizedCounter[u-1] >= 0); + cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1]; + assert(cumul[u] >= cumul[u-1]); /* no overflow */ } } - cumul[maxSymbolValue+1] = tableSize+1; + cumul[maxSV1] = (U16)(tableSize+1); } /* Spread symbols */ - { U32 position = 0; + if (highThreshold == tableSize - 1) { + /* Case for no low prob count symbols. Lay down 8 bytes at a time + * to reduce branch misses since we are operating on a small block + */ + BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */ + { 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); + } + assert(n>=0); + pos += (size_t)n; + } + } + /* Spread symbols across the table. Lack of lowprob symbols means that + * we don't need variable sized inner loop, so we can unroll the loop and + * reduce branch misses. + */ + { size_t position = 0; + size_t s; + size_t const unroll = 2; /* Experimentally determined optimal unroll */ + 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; + tableSymbol[uPosition] = spread[s + u]; + } + position = (position + (unroll * step)) & tableMask; + } + assert(position == 0); /* Must have initialized all positions */ + } + } else { + U32 position = 0; U32 symbol; - for (symbol=0; symbol<=maxSymbolValue; symbol++) { + for (symbol=0; symbol<maxSV1; symbol++) { int nbOccurrences; int const freq = normalizedCounter[symbol]; for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) { @@ -117,7 +162,6 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, while (position > highThreshold) position = (position + step) & tableMask; /* Low proba area */ } } - assert(position==0); /* Must have initialized all positions */ } @@ -141,16 +185,17 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, case -1: case 1: symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog); - symbolTT[s].deltaFindState = total - 1; + assert(total <= INT_MAX); + symbolTT[s].deltaFindState = (int)(total - 1); total ++; break; default : - { - U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1); - U32 const minStatePlus = normalizedCounter[s] << maxBitsOut; + assert(normalizedCounter[s] > 1); + { U32 const maxBitsOut = tableLog - BIT_highbit32 ((U32)normalizedCounter[s]-1); + U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut; symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus; - symbolTT[s].deltaFindState = total - normalizedCounter[s]; - total += normalizedCounter[s]; + symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]); + total += (unsigned)normalizedCounter[s]; } } } } #if 0 /* debug : symbol costs */ @@ -161,31 +206,26 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, symbol, normalizedCounter[symbol], FSE_getMaxNbBits(symbolTT, symbol), (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256); - } - } + } } #endif return 0; } -size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog) -{ - FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE]; /* memset() is not necessary, even if static analyzer complain about it */ - return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol)); -} - - #ifndef FSE_COMMONDEFS_ONLY - /*-************************************************************** * FSE NCount encoding ****************************************************************/ size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog) { - size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3; + size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog + + 4 /* bitCount initialized at 4 */ + + 2 /* first two symbols may use one additional bit each */) / 8) + + 1 /* round up to whole nb bytes */ + + 2 /* additional two bytes for bitstream flush */; return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND; /* maxSymbolValue==0 ? use default */ } @@ -307,10 +347,10 @@ FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog) size_t size; if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX; size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32); - return (FSE_CTable*)malloc(size); + return (FSE_CTable*)ZSTD_malloc(size); } -void FSE_freeCTable (FSE_CTable* ct) { free(ct); } +void FSE_freeCTable (FSE_CTable* ct) { ZSTD_free(ct); } /* provides the minimum logSize to safely represent a distribution */ static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue) @@ -341,11 +381,10 @@ unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxS return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2); } - /* Secondary normalization method. To be used when primary method fails. */ -static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue) +static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount) { short const NOT_YET_ASSIGNED = -2; U32 s; @@ -362,7 +401,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, continue; } if (count[s] <= lowThreshold) { - norm[s] = -1; + norm[s] = lowProbCount; distributed++; total -= count[s]; continue; @@ -414,7 +453,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, { U64 const vStepLog = 62 - tableLog; U64 const mid = (1ULL << (vStepLog-1)) - 1; - U64 const rStep = ((((U64)1<<vStepLog) * ToDistribute) + mid) / total; /* scale on remaining */ + U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total); /* scale on remaining */ U64 tmpTotal = mid; for (s=0; s<=maxSymbolValue; s++) { if (norm[s]==NOT_YET_ASSIGNED) { @@ -431,10 +470,9 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, return 0; } - size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog, const unsigned* count, size_t total, - unsigned maxSymbolValue) + unsigned maxSymbolValue, unsigned useLowProbCount) { /* Sanity checks */ if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG; @@ -443,8 +481,9 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog, if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC); /* Too small tableLog, compression potentially impossible */ { static U32 const rtbTable[] = { 0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 }; + short const lowProbCount = useLowProbCount ? -1 : 1; U64 const scale = 62 - tableLog; - U64 const step = ((U64)1<<62) / total; /* <== here, one division ! */ + U64 const step = ZSTD_div64((U64)1<<62, (U32)total); /* <== here, one division ! */ U64 const vStep = 1ULL<<(scale-20); int stillToDistribute = 1<<tableLog; unsigned s; @@ -456,7 +495,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog, if (count[s] == total) return 0; /* rle special case */ if (count[s] == 0) { normalizedCounter[s]=0; continue; } if (count[s] <= lowThreshold) { - normalizedCounter[s] = -1; + normalizedCounter[s] = lowProbCount; stillToDistribute--; } else { short proba = (short)((count[s]*step) >> scale); @@ -470,7 +509,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog, } } if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) { /* corner case, need another normalization method */ - size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue); + size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount); if (FSE_isError(errorCode)) return errorCode; } else normalizedCounter[largest] += (short)stillToDistribute; @@ -625,6 +664,7 @@ size_t FSE_compress_usingCTable (void* dst, size_t dstSize, size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); } +#ifndef ZSTD_NO_UNUSED_FUNCTIONS /* FSE_compress_wksp() : * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`). * `wkspSize` size must be `(1<<tableLog)`. @@ -643,7 +683,7 @@ size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t src size_t const scratchBufferSize = wkspSize - (CTableSize * sizeof(FSE_CTable)); /* init conditions */ - if (wkspSize < FSE_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge); + if (wkspSize < FSE_COMPRESS_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge); if (srcSize <= 1) return 0; /* Not compressible */ if (!maxSymbolValue) maxSymbolValue = FSE_MAX_SYMBOL_VALUE; if (!tableLog) tableLog = FSE_DEFAULT_TABLELOG; @@ -656,7 +696,7 @@ size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t src } tableLog = FSE_optimalTableLog(tableLog, srcSize, maxSymbolValue); - CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue) ); + CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue, /* useLowProbCount */ srcSize >= 2048) ); /* Write table description header */ { CHECK_V_F(nc_err, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) ); @@ -678,13 +718,16 @@ size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t src typedef struct { FSE_CTable CTable_max[FSE_CTABLE_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)]; - BYTE scratchBuffer[1 << FSE_MAX_TABLELOG]; + union { + U32 hist_wksp[HIST_WKSP_SIZE_U32]; + BYTE scratchBuffer[1 << FSE_MAX_TABLELOG]; + } workspace; } fseWkspMax_t; size_t FSE_compress2 (void* dst, size_t dstCapacity, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog) { fseWkspMax_t scratchBuffer; - DEBUG_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)); /* compilation failures here means scratchBuffer is not large enough */ + DEBUG_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_COMPRESS_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)); /* compilation failures here means scratchBuffer is not large enough */ if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge); return FSE_compress_wksp(dst, dstCapacity, src, srcSize, maxSymbolValue, tableLog, &scratchBuffer, sizeof(scratchBuffer)); } @@ -693,6 +736,6 @@ size_t FSE_compress (void* dst, size_t dstCapacity, const void* src, size_t srcS { return FSE_compress2(dst, dstCapacity, src, srcSize, FSE_MAX_SYMBOL_VALUE, FSE_DEFAULT_TABLELOG); } - +#endif #endif /* FSE_COMMONDEFS_ONLY */ diff --git a/thirdparty/zstd/compress/hist.c b/thirdparty/zstd/compress/hist.c index 61e08c7968..073c57e752 100644 --- a/thirdparty/zstd/compress/hist.c +++ b/thirdparty/zstd/compress/hist.c @@ -1,7 +1,7 @@ /* ****************************************************************** * hist : Histogram functions * part of Finite State Entropy project - * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc. + * Copyright (c) Yann Collet, Facebook, Inc. * * You can contact the author at : * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy @@ -34,7 +34,7 @@ unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr, unsigned maxSymbolValue = *maxSymbolValuePtr; unsigned largestCount=0; - memset(count, 0, (maxSymbolValue+1) * sizeof(*count)); + ZSTD_memset(count, 0, (maxSymbolValue+1) * sizeof(*count)); if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; } while (ip<end) { @@ -60,9 +60,9 @@ typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e; * this design makes better use of OoO cpus, * and is noticeably faster when some values are heavily repeated. * But it needs some additional workspace for intermediate tables. - * `workSpace` size must be a table of size >= HIST_WKSP_SIZE_U32. + * `workSpace` must be a U32 table of size >= HIST_WKSP_SIZE_U32. * @return : largest histogram frequency, - * or an error code (notably when histogram would be larger than *maxSymbolValuePtr). */ + * or an error code (notably when histogram's alphabet is larger than *maxSymbolValuePtr) */ static size_t HIST_count_parallel_wksp( unsigned* count, unsigned* maxSymbolValuePtr, const void* source, size_t sourceSize, @@ -71,22 +71,21 @@ static size_t HIST_count_parallel_wksp( { const BYTE* ip = (const BYTE*)source; const BYTE* const iend = ip+sourceSize; - unsigned maxSymbolValue = *maxSymbolValuePtr; + size_t const countSize = (*maxSymbolValuePtr + 1) * sizeof(*count); unsigned max=0; U32* const Counting1 = workSpace; U32* const Counting2 = Counting1 + 256; U32* const Counting3 = Counting2 + 256; U32* const Counting4 = Counting3 + 256; - memset(workSpace, 0, 4*256*sizeof(unsigned)); - /* safety checks */ + assert(*maxSymbolValuePtr <= 255); if (!sourceSize) { - memset(count, 0, maxSymbolValue + 1); + ZSTD_memset(count, 0, countSize); *maxSymbolValuePtr = 0; return 0; } - if (!maxSymbolValue) maxSymbolValue = 255; /* 0 == default */ + ZSTD_memset(workSpace, 0, 4*256*sizeof(unsigned)); /* by stripes of 16 bytes */ { U32 cached = MEM_read32(ip); ip += 4; @@ -118,21 +117,18 @@ static size_t HIST_count_parallel_wksp( /* finish last symbols */ while (ip<iend) Counting1[*ip++]++; - if (check) { /* verify stats will fit into destination table */ - U32 s; for (s=255; s>maxSymbolValue; s--) { - Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s]; - if (Counting1[s]) return ERROR(maxSymbolValue_tooSmall); - } } - { U32 s; - if (maxSymbolValue > 255) maxSymbolValue = 255; - for (s=0; s<=maxSymbolValue; s++) { - count[s] = Counting1[s] + Counting2[s] + Counting3[s] + Counting4[s]; - if (count[s] > max) max = count[s]; + for (s=0; s<256; s++) { + Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s]; + if (Counting1[s] > max) max = Counting1[s]; } } - while (!count[maxSymbolValue]) maxSymbolValue--; - *maxSymbolValuePtr = maxSymbolValue; + { unsigned maxSymbolValue = 255; + while (!Counting1[maxSymbolValue]) maxSymbolValue--; + if (check && maxSymbolValue > *maxSymbolValuePtr) return ERROR(maxSymbolValue_tooSmall); + *maxSymbolValuePtr = maxSymbolValue; + ZSTD_memmove(count, Counting1, countSize); /* in case count & Counting1 are overlapping */ + } return (size_t)max; } @@ -152,14 +148,6 @@ size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr, return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace); } -/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */ -size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr, - const void* source, size_t sourceSize) -{ - unsigned tmpCounters[HIST_WKSP_SIZE_U32]; - return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters)); -} - /* HIST_count_wksp() : * Same as HIST_count(), but using an externally provided scratch buffer. * `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */ @@ -175,9 +163,19 @@ size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr, return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize); } +#ifndef ZSTD_NO_UNUSED_FUNCTIONS +/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */ +size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr, + const void* source, size_t sourceSize) +{ + unsigned tmpCounters[HIST_WKSP_SIZE_U32]; + return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters)); +} + size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize) { unsigned tmpCounters[HIST_WKSP_SIZE_U32]; return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters)); } +#endif diff --git a/thirdparty/zstd/compress/hist.h b/thirdparty/zstd/compress/hist.h index 77e3ec4fb1..228ed48a71 100644 --- a/thirdparty/zstd/compress/hist.h +++ b/thirdparty/zstd/compress/hist.h @@ -1,7 +1,7 @@ /* ****************************************************************** * hist : Histogram functions * part of Finite State Entropy project - * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc. + * Copyright (c) Yann Collet, Facebook, Inc. * * You can contact the author at : * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy @@ -14,7 +14,7 @@ ****************************************************************** */ /* --- dependencies --- */ -#include <stddef.h> /* size_t */ +#include "../common/zstd_deps.h" /* size_t */ /* --- simple histogram functions --- */ diff --git a/thirdparty/zstd/compress/huf_compress.c b/thirdparty/zstd/compress/huf_compress.c index 546879868a..2b3d6adc2a 100644 --- a/thirdparty/zstd/compress/huf_compress.c +++ b/thirdparty/zstd/compress/huf_compress.c @@ -1,6 +1,6 @@ /* ****************************************************************** * Huffman encoder, part of New Generation Entropy library - * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc. + * Copyright (c) Yann Collet, Facebook, Inc. * * You can contact the author at : * - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy @@ -23,8 +23,7 @@ /* ************************************************************** * Includes ****************************************************************/ -#include <string.h> /* memcpy, memset */ -#include <stdio.h> /* printf (debug) */ +#include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset */ #include "../common/compiler.h" #include "../common/bitstream.h" #include "hist.h" @@ -54,13 +53,43 @@ unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxS /* ******************************************************* * HUF : Huffman block compression *********************************************************/ +#define HUF_WORKSPACE_MAX_ALIGNMENT 8 + +static void* HUF_alignUpWorkspace(void* workspace, size_t* workspaceSizePtr, size_t align) +{ + size_t const mask = align - 1; + size_t const rem = (size_t)workspace & mask; + size_t const add = (align - rem) & mask; + BYTE* const aligned = (BYTE*)workspace + add; + assert((align & (align - 1)) == 0); /* pow 2 */ + assert(align <= HUF_WORKSPACE_MAX_ALIGNMENT); + if (*workspaceSizePtr >= add) { + assert(add < align); + assert(((size_t)aligned & mask) == 0); + *workspaceSizePtr -= add; + return aligned; + } else { + *workspaceSizePtr = 0; + return NULL; + } +} + + /* HUF_compressWeights() : * Same as FSE_compress(), but dedicated to huff0's weights compression. * The use case needs much less stack memory. * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX. */ #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6 -static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weightTable, size_t wtSize) + +typedef struct { + FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)]; + U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)]; + unsigned count[HUF_TABLELOG_MAX+1]; + S16 norm[HUF_TABLELOG_MAX+1]; +} HUF_CompressWeightsWksp; + +static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightTable, size_t wtSize, void* workspace, size_t workspaceSize) { BYTE* const ostart = (BYTE*) dst; BYTE* op = ostart; @@ -68,33 +97,30 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight unsigned maxSymbolValue = HUF_TABLELOG_MAX; U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER; + HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32)); - FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)]; - BYTE scratchBuffer[1<<MAX_FSE_TABLELOG_FOR_HUFF_HEADER]; - - unsigned count[HUF_TABLELOG_MAX+1]; - S16 norm[HUF_TABLELOG_MAX+1]; + if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC); /* init conditions */ if (wtSize <= 1) return 0; /* Not compressible */ /* Scan input and build symbol stats */ - { unsigned const maxCount = HIST_count_simple(count, &maxSymbolValue, weightTable, wtSize); /* never fails */ + { unsigned const maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize); /* never fails */ if (maxCount == wtSize) return 1; /* only a single symbol in src : rle */ if (maxCount == 1) return 0; /* each symbol present maximum once => not compressible */ } tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue); - CHECK_F( FSE_normalizeCount(norm, tableLog, count, wtSize, maxSymbolValue) ); + CHECK_F( FSE_normalizeCount(wksp->norm, tableLog, wksp->count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) ); /* Write table description header */ - { CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), norm, maxSymbolValue, tableLog) ); + { CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), wksp->norm, maxSymbolValue, tableLog) ); op += hSize; } /* Compress */ - CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, sizeof(scratchBuffer)) ); - { CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, CTable) ); + CHECK_F( FSE_buildCTable_wksp(wksp->CTable, wksp->norm, maxSymbolValue, tableLog, wksp->scratchBuffer, sizeof(wksp->scratchBuffer)) ); + { CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, wksp->CTable) ); if (cSize == 0) return 0; /* not enough space for compressed data */ op += cSize; } @@ -102,35 +128,70 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight return (size_t)(op-ostart); } +static size_t HUF_getNbBits(HUF_CElt elt) +{ + return elt & 0xFF; +} -struct HUF_CElt_s { - U16 val; - BYTE nbBits; -}; /* typedef'd to HUF_CElt within "huf.h" */ +static size_t HUF_getNbBitsFast(HUF_CElt elt) +{ + return elt; +} -/*! HUF_writeCTable() : - `CTable` : Huffman tree to save, using huf representation. - @return : size of saved CTable */ -size_t HUF_writeCTable (void* dst, size_t maxDstSize, - const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog) +static size_t HUF_getValue(HUF_CElt elt) +{ + return elt & ~0xFF; +} + +static size_t HUF_getValueFast(HUF_CElt elt) +{ + return elt; +} + +static void HUF_setNbBits(HUF_CElt* elt, size_t nbBits) { + assert(nbBits <= HUF_TABLELOG_ABSOLUTEMAX); + *elt = nbBits; +} + +static void HUF_setValue(HUF_CElt* elt, size_t value) +{ + size_t const nbBits = HUF_getNbBits(*elt); + if (nbBits > 0) { + assert((value >> nbBits) == 0); + *elt |= value << (sizeof(HUF_CElt) * 8 - nbBits); + } +} + +typedef struct { + HUF_CompressWeightsWksp wksp; BYTE bitsToWeight[HUF_TABLELOG_MAX + 1]; /* precomputed conversion table */ BYTE huffWeight[HUF_SYMBOLVALUE_MAX]; +} HUF_WriteCTableWksp; + +size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, + const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, + void* workspace, size_t workspaceSize) +{ + HUF_CElt const* const ct = CTable + 1; BYTE* op = (BYTE*)dst; U32 n; + HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32)); - /* check conditions */ + /* check conditions */ + if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC); if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge); /* convert to weight */ - bitsToWeight[0] = 0; + wksp->bitsToWeight[0] = 0; for (n=1; n<huffLog+1; n++) - bitsToWeight[n] = (BYTE)(huffLog + 1 - n); + wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n); for (n=0; n<maxSymbolValue; n++) - huffWeight[n] = bitsToWeight[CTable[n].nbBits]; + wksp->huffWeight[n] = wksp->bitsToWeight[HUF_getNbBits(ct[n])]; /* attempt weights compression by FSE */ - { CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, huffWeight, maxSymbolValue) ); + if (maxDstSize < 1) return ERROR(dstSize_tooSmall); + { CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, wksp->huffWeight, maxSymbolValue, &wksp->wksp, sizeof(wksp->wksp)) ); if ((hSize>1) & (hSize < maxSymbolValue/2)) { /* FSE compressed */ op[0] = (BYTE)hSize; return hSize+1; @@ -140,12 +201,22 @@ size_t HUF_writeCTable (void* dst, size_t maxDstSize, if (maxSymbolValue > (256-128)) return ERROR(GENERIC); /* should not happen : likely means source cannot be compressed */ if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall); /* not enough space within dst buffer */ op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1)); - huffWeight[maxSymbolValue] = 0; /* to be sure it doesn't cause msan issue in final combination */ + wksp->huffWeight[maxSymbolValue] = 0; /* to be sure it doesn't cause msan issue in final combination */ for (n=0; n<maxSymbolValue; n+=2) - op[(n/2)+1] = (BYTE)((huffWeight[n] << 4) + huffWeight[n+1]); + op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]); return ((maxSymbolValue+1)/2) + 1; } +/*! HUF_writeCTable() : + `CTable` : Huffman tree to save, using huf representation. + @return : size of saved CTable */ +size_t HUF_writeCTable (void* dst, size_t maxDstSize, + const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog) +{ + HUF_WriteCTableWksp wksp; + return HUF_writeCTable_wksp(dst, maxDstSize, CTable, maxSymbolValue, huffLog, &wksp, sizeof(wksp)); +} + size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights) { @@ -153,34 +224,36 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1]; /* large enough for values from 0 to 16 */ U32 tableLog = 0; U32 nbSymbols = 0; + HUF_CElt* const ct = CTable + 1; /* get symbol weights */ CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize)); + *hasZeroWeights = (rankVal[0] > 0); /* check result */ if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge); if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall); + CTable[0] = tableLog; + /* Prepare base value per rank */ { U32 n, nextRankStart = 0; for (n=1; n<=tableLog; n++) { - U32 current = nextRankStart; + U32 curr = nextRankStart; nextRankStart += (rankVal[n] << (n-1)); - rankVal[n] = current; + rankVal[n] = curr; } } /* fill nbBits */ - *hasZeroWeights = 0; { U32 n; for (n=0; n<nbSymbols; n++) { const U32 w = huffWeight[n]; - *hasZeroWeights |= (w == 0); - CTable[n].nbBits = (BYTE)(tableLog + 1 - w) & -(w != 0); + HUF_setNbBits(ct + n, (BYTE)(tableLog + 1 - w) & -(w != 0)); } } /* fill val */ { U16 nbPerRank[HUF_TABLELOG_MAX+2] = {0}; /* support w=0=>n=tableLog+1 */ U16 valPerRank[HUF_TABLELOG_MAX+2] = {0}; - { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; } + { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[HUF_getNbBits(ct[n])]++; } /* determine stating value per rank */ valPerRank[tableLog+1] = 0; /* for w==0 */ { U16 min = 0; @@ -190,18 +263,18 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void min >>= 1; } } /* assign value within rank, symbol order */ - { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; } + { U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); } } *maxSymbolValuePtr = nbSymbols - 1; return readSize; } -U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue) +U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue) { - const HUF_CElt* table = (const HUF_CElt*)symbolTable; + const HUF_CElt* ct = CTable + 1; assert(symbolValue <= HUF_SYMBOLVALUE_MAX); - return table[symbolValue].nbBits; + return (U32)HUF_getNbBits(ct[symbolValue]); } @@ -212,32 +285,63 @@ typedef struct nodeElt_s { BYTE nbBits; } nodeElt; +/** + * HUF_setMaxHeight(): + * Enforces maxNbBits on the Huffman tree described in huffNode. + * + * It sets all nodes with nbBits > maxNbBits to be maxNbBits. Then it adjusts + * the tree to so that it is a valid canonical Huffman tree. + * + * @pre The sum of the ranks of each symbol == 2^largestBits, + * where largestBits == huffNode[lastNonNull].nbBits. + * @post The sum of the ranks of each symbol == 2^largestBits, + * where largestBits is the return value <= maxNbBits. + * + * @param huffNode The Huffman tree modified in place to enforce maxNbBits. + * @param lastNonNull The symbol with the lowest count in the Huffman tree. + * @param maxNbBits The maximum allowed number of bits, which the Huffman tree + * may not respect. After this function the Huffman tree will + * respect maxNbBits. + * @return The maximum number of bits of the Huffman tree after adjustment, + * necessarily no more than maxNbBits. + */ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits) { const U32 largestBits = huffNode[lastNonNull].nbBits; - if (largestBits <= maxNbBits) return largestBits; /* early exit : no elt > maxNbBits */ + /* early exit : no elt > maxNbBits, so the tree is already valid. */ + if (largestBits <= maxNbBits) return largestBits; /* there are several too large elements (at least >= 2) */ { int totalCost = 0; const U32 baseCost = 1 << (largestBits - maxNbBits); int n = (int)lastNonNull; + /* Adjust any ranks > maxNbBits to maxNbBits. + * Compute totalCost, which is how far the sum of the ranks is + * we are over 2^largestBits after adjust the offending ranks. + */ while (huffNode[n].nbBits > maxNbBits) { totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits)); huffNode[n].nbBits = (BYTE)maxNbBits; - n --; - } /* n stops at huffNode[n].nbBits <= maxNbBits */ - while (huffNode[n].nbBits == maxNbBits) n--; /* n end at index of smallest symbol using < maxNbBits */ + n--; + } + /* n stops at huffNode[n].nbBits <= maxNbBits */ + assert(huffNode[n].nbBits <= maxNbBits); + /* n end at index of smallest symbol using < maxNbBits */ + while (huffNode[n].nbBits == maxNbBits) --n; - /* renorm totalCost */ - totalCost >>= (largestBits - maxNbBits); /* note : totalCost is necessarily a multiple of baseCost */ + /* renorm totalCost from 2^largestBits to 2^maxNbBits + * note : totalCost is necessarily a multiple of baseCost */ + assert((totalCost & (baseCost - 1)) == 0); + totalCost >>= (largestBits - maxNbBits); + assert(totalCost > 0); /* repay normalized cost */ { U32 const noSymbol = 0xF0F0F0F0; U32 rankLast[HUF_TABLELOG_MAX+2]; - /* Get pos of last (smallest) symbol per rank */ - memset(rankLast, 0xF0, sizeof(rankLast)); + /* Get pos of last (smallest = lowest cum. count) symbol per rank */ + ZSTD_memset(rankLast, 0xF0, sizeof(rankLast)); { U32 currentNbBits = maxNbBits; int pos; for (pos=n ; pos >= 0; pos--) { @@ -247,34 +351,65 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits) } } while (totalCost > 0) { + /* Try to reduce the next power of 2 above totalCost because we + * gain back half the rank. + */ U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1; for ( ; nBitsToDecrease > 1; nBitsToDecrease--) { U32 const highPos = rankLast[nBitsToDecrease]; U32 const lowPos = rankLast[nBitsToDecrease-1]; if (highPos == noSymbol) continue; + /* Decrease highPos if no symbols of lowPos or if it is + * not cheaper to remove 2 lowPos than highPos. + */ if (lowPos == noSymbol) break; { U32 const highTotal = huffNode[highPos].count; U32 const lowTotal = 2 * huffNode[lowPos].count; if (highTotal <= lowTotal) break; } } /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */ + assert(rankLast[nBitsToDecrease] != noSymbol || nBitsToDecrease == 1); /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */ while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol)) - nBitsToDecrease ++; + nBitsToDecrease++; + assert(rankLast[nBitsToDecrease] != noSymbol); + /* Increase the number of bits to gain back half the rank cost. */ totalCost -= 1 << (nBitsToDecrease-1); + huffNode[rankLast[nBitsToDecrease]].nbBits++; + + /* Fix up the new rank. + * If the new rank was empty, this symbol is now its smallest. + * Otherwise, this symbol will be the largest in the new rank so no adjustment. + */ if (rankLast[nBitsToDecrease-1] == noSymbol) - rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]; /* this rank is no longer empty */ - huffNode[rankLast[nBitsToDecrease]].nbBits ++; + rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]; + /* Fix up the old rank. + * If the symbol was at position 0, meaning it was the highest weight symbol in the tree, + * it must be the only symbol in its rank, so the old rank now has no symbols. + * Otherwise, since the Huffman nodes are sorted by count, the previous position is now + * the smallest node in the rank. If the previous position belongs to a different rank, + * then the rank is now empty. + */ if (rankLast[nBitsToDecrease] == 0) /* special case, reached largest symbol */ rankLast[nBitsToDecrease] = noSymbol; else { rankLast[nBitsToDecrease]--; if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease) rankLast[nBitsToDecrease] = noSymbol; /* this rank is now empty */ - } } /* while (totalCost > 0) */ - + } + } /* while (totalCost > 0) */ + + /* If we've removed too much weight, then we have to add it back. + * To avoid overshooting again, we only adjust the smallest rank. + * We take the largest nodes from the lowest rank 0 and move them + * to rank 1. There's guaranteed to be enough rank 0 symbols because + * TODO. + */ while (totalCost < 0) { /* Sometimes, cost correction overshoot */ - if (rankLast[1] == noSymbol) { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */ + /* special case : no rank 1 symbol (using maxNbBits-1); + * let's create one from largest rank 0 (using maxNbBits). + */ + if (rankLast[1] == noSymbol) { while (huffNode[n].nbBits == maxNbBits) n--; huffNode[n+1].nbBits--; assert(n >= 0); @@ -285,49 +420,179 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits) huffNode[ rankLast[1] + 1 ].nbBits--; rankLast[1]++; totalCost ++; - } } } /* there are several too large elements (at least >= 2) */ + } + } /* repay normalized cost */ + } /* there are several too large elements (at least >= 2) */ return maxNbBits; } typedef struct { - U32 base; - U32 current; + U16 base; + U16 curr; } rankPos; typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32]; -#define RANK_POSITION_TABLE_SIZE 32 +/* Number of buckets available for HUF_sort() */ +#define RANK_POSITION_TABLE_SIZE 192 typedef struct { huffNodeTable huffNodeTbl; rankPos rankPosition[RANK_POSITION_TABLE_SIZE]; } HUF_buildCTable_wksp_tables; -static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValue, rankPos* rankPosition) -{ +/* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing. + * Strategy is to use as many buckets as possible for representing distinct + * counts while using the remainder to represent all "large" counts. + * + * To satisfy this requirement for 192 buckets, we can do the following: + * Let buckets 0-166 represent distinct counts of [0, 166] + * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing. + */ +#define RANK_POSITION_MAX_COUNT_LOG 32 +#define RANK_POSITION_LOG_BUCKETS_BEGIN (RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */ +#define RANK_POSITION_DISTINCT_COUNT_CUTOFF RANK_POSITION_LOG_BUCKETS_BEGIN + BIT_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */ + +/* Return the appropriate bucket index for a given count. See definition of + * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy. + */ +static U32 HUF_getIndex(U32 const count) { + return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF) + ? count + : BIT_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN; +} + +/* Helper swap function for HUF_quickSortPartition() */ +static void HUF_swapNodes(nodeElt* a, nodeElt* b) { + nodeElt tmp = *a; + *a = *b; + *b = tmp; +} + +/* Returns 0 if the huffNode array is not sorted by descending count */ +MEM_STATIC int HUF_isSorted(nodeElt huffNode[], U32 const maxSymbolValue1) { + U32 i; + for (i = 1; i < maxSymbolValue1; ++i) { + if (huffNode[i].count > huffNode[i-1].count) { + return 0; + } + } + return 1; +} + +/* Insertion sort by descending order */ +HINT_INLINE void HUF_insertionSort(nodeElt huffNode[], int const low, int const high) { + int i; + int const size = high-low+1; + huffNode += low; + for (i = 1; i < size; ++i) { + nodeElt const key = huffNode[i]; + int j = i - 1; + while (j >= 0 && huffNode[j].count < key.count) { + huffNode[j + 1] = huffNode[j]; + j--; + } + huffNode[j + 1] = key; + } +} + +/* Pivot helper function for quicksort. */ +static int HUF_quickSortPartition(nodeElt arr[], int const low, int const high) { + /* Simply select rightmost element as pivot. "Better" selectors like + * median-of-three don't experimentally appear to have any benefit. + */ + U32 const pivot = arr[high].count; + int i = low - 1; + int j = low; + for ( ; j < high; j++) { + if (arr[j].count > pivot) { + i++; + HUF_swapNodes(&arr[i], &arr[j]); + } + } + HUF_swapNodes(&arr[i + 1], &arr[high]); + return i + 1; +} + +/* Classic quicksort by descending with partially iterative calls + * to reduce worst case callstack size. + */ +static void HUF_simpleQuickSort(nodeElt arr[], int low, int high) { + int const kInsertionSortThreshold = 8; + if (high - low < kInsertionSortThreshold) { + HUF_insertionSort(arr, low, high); + return; + } + while (low < high) { + int const idx = HUF_quickSortPartition(arr, low, high); + if (idx - low < high - idx) { + HUF_simpleQuickSort(arr, low, idx - 1); + low = idx + 1; + } else { + HUF_simpleQuickSort(arr, idx + 1, high); + high = idx - 1; + } + } +} + +/** + * HUF_sort(): + * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order. + * This is a typical bucket sorting strategy that uses either quicksort or insertion sort to sort each bucket. + * + * @param[out] huffNode Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled. + * Must have (maxSymbolValue + 1) entries. + * @param[in] count Histogram of the symbols. + * @param[in] maxSymbolValue Maximum symbol value. + * @param rankPosition This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries. + */ +static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) { U32 n; + U32 const maxSymbolValue1 = maxSymbolValue+1; + + /* Compute base and set curr to base. + * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1. + * See HUF_getIndex to see bucketing strategy. + * We attribute each symbol to lowerRank's base value, because we want to know where + * each rank begins in the output, so for rank R we want to count ranks R+1 and above. + */ + ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE); + for (n = 0; n < maxSymbolValue1; ++n) { + U32 lowerRank = HUF_getIndex(count[n]); + assert(lowerRank < RANK_POSITION_TABLE_SIZE - 1); + rankPosition[lowerRank].base++; + } - memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE); - for (n=0; n<=maxSymbolValue; n++) { - U32 r = BIT_highbit32(count[n] + 1); - rankPosition[r].base ++; + assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0); + /* Set up the rankPosition table */ + for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) { + rankPosition[n-1].base += rankPosition[n].base; + rankPosition[n-1].curr = rankPosition[n-1].base; } - for (n=30; n>0; n--) rankPosition[n-1].base += rankPosition[n].base; - for (n=0; n<32; n++) rankPosition[n].current = rankPosition[n].base; - for (n=0; n<=maxSymbolValue; n++) { + + /* Insert each symbol into their appropriate bucket, setting up rankPosition table. */ + for (n = 0; n < maxSymbolValue1; ++n) { U32 const c = count[n]; - U32 const r = BIT_highbit32(c+1) + 1; - U32 pos = rankPosition[r].current++; - while ((pos > rankPosition[r].base) && (c > huffNode[pos-1].count)) { - huffNode[pos] = huffNode[pos-1]; - pos--; - } + U32 const r = HUF_getIndex(c) + 1; + U32 const pos = rankPosition[r].curr++; + assert(pos < maxSymbolValue1); huffNode[pos].count = c; huffNode[pos].byte = (BYTE)n; } -} + /* Sort each bucket. */ + for (n = RANK_POSITION_DISTINCT_COUNT_CUTOFF; n < RANK_POSITION_TABLE_SIZE - 1; ++n) { + U32 const bucketSize = rankPosition[n].curr-rankPosition[n].base; + U32 const bucketStartIdx = rankPosition[n].base; + if (bucketSize > 1) { + assert(bucketStartIdx < maxSymbolValue1); + HUF_simpleQuickSort(huffNode + bucketStartIdx, 0, bucketSize-1); + } + } + + assert(HUF_isSorted(huffNode, maxSymbolValue1)); +} /** HUF_buildCTable_wksp() : * Same as HUF_buildCTable(), but using externally allocated scratch buffer. @@ -335,28 +600,20 @@ static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValu */ #define STARTNODE (HUF_SYMBOLVALUE_MAX+1) -size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize) +/* HUF_buildTree(): + * Takes the huffNode array sorted by HUF_sort() and builds an unlimited-depth Huffman tree. + * + * @param huffNode The array sorted by HUF_sort(). Builds the Huffman tree in this array. + * @param maxSymbolValue The maximum symbol value. + * @return The smallest node in the Huffman tree (by count). + */ +static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue) { - HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)workSpace; - nodeElt* const huffNode0 = wksp_tables->huffNodeTbl; - nodeElt* const huffNode = huffNode0+1; + nodeElt* const huffNode0 = huffNode - 1; int nonNullRank; int lowS, lowN; int nodeNb = STARTNODE; int n, nodeRoot; - - /* safety checks */ - if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC); /* must be aligned on 4-bytes boundaries */ - if (wkspSize < sizeof(HUF_buildCTable_wksp_tables)) - return ERROR(workSpace_tooSmall); - if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT; - if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) - return ERROR(maxSymbolValue_tooLarge); - memset(huffNode0, 0, sizeof(huffNodeTable)); - - /* sort, decreasing order */ - HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition); - /* init for parents */ nonNullRank = (int)maxSymbolValue; while(huffNode[nonNullRank].count == 0) nonNullRank--; @@ -383,127 +640,396 @@ size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbo for (n=0; n<=nonNullRank; n++) huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1; + return nonNullRank; +} + +/** + * HUF_buildCTableFromTree(): + * Build the CTable given the Huffman tree in huffNode. + * + * @param[out] CTable The output Huffman CTable. + * @param huffNode The Huffman tree. + * @param nonNullRank The last and smallest node in the Huffman tree. + * @param maxSymbolValue The maximum symbol value. + * @param maxNbBits The exact maximum number of bits used in the Huffman tree. + */ +static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits) +{ + HUF_CElt* const ct = CTable + 1; + /* fill result into ctable (val, nbBits) */ + int n; + U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0}; + U16 valPerRank[HUF_TABLELOG_MAX+1] = {0}; + int const alphabetSize = (int)(maxSymbolValue + 1); + for (n=0; n<=nonNullRank; n++) + nbPerRank[huffNode[n].nbBits]++; + /* determine starting value per rank */ + { U16 min = 0; + for (n=(int)maxNbBits; n>0; n--) { + valPerRank[n] = min; /* get starting value within each rank */ + min += nbPerRank[n]; + min >>= 1; + } } + for (n=0; n<alphabetSize; n++) + HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits); /* push nbBits per symbol, symbol order */ + for (n=0; n<alphabetSize; n++) + HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); /* assign value within rank, symbol order */ + CTable[0] = maxNbBits; +} + +size_t HUF_buildCTable_wksp (HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize) +{ + HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32)); + nodeElt* const huffNode0 = wksp_tables->huffNodeTbl; + nodeElt* const huffNode = huffNode0+1; + int nonNullRank; + + /* safety checks */ + if (wkspSize < sizeof(HUF_buildCTable_wksp_tables)) + return ERROR(workSpace_tooSmall); + if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT; + if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) + return ERROR(maxSymbolValue_tooLarge); + ZSTD_memset(huffNode0, 0, sizeof(huffNodeTable)); + + /* sort, decreasing order */ + HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition); + + /* build tree */ + nonNullRank = HUF_buildTree(huffNode, maxSymbolValue); + /* enforce maxTableLog */ maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits); + if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC); /* check fit into table */ - /* fill result into tree (val, nbBits) */ - { U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0}; - U16 valPerRank[HUF_TABLELOG_MAX+1] = {0}; - int const alphabetSize = (int)(maxSymbolValue + 1); - if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC); /* check fit into table */ - for (n=0; n<=nonNullRank; n++) - nbPerRank[huffNode[n].nbBits]++; - /* determine stating value per rank */ - { U16 min = 0; - for (n=(int)maxNbBits; n>0; n--) { - valPerRank[n] = min; /* get starting value within each rank */ - min += nbPerRank[n]; - min >>= 1; - } } - for (n=0; n<alphabetSize; n++) - tree[huffNode[n].byte].nbBits = huffNode[n].nbBits; /* push nbBits per symbol, symbol order */ - for (n=0; n<alphabetSize; n++) - tree[n].val = valPerRank[tree[n].nbBits]++; /* assign value within rank, symbol order */ - } + HUF_buildCTableFromTree(CTable, huffNode, nonNullRank, maxSymbolValue, maxNbBits); return maxNbBits; } -/** HUF_buildCTable() : - * @return : maxNbBits - * Note : count is used before tree is written, so they can safely overlap - */ -size_t HUF_buildCTable (HUF_CElt* tree, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits) -{ - HUF_buildCTable_wksp_tables workspace; - return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, &workspace, sizeof(workspace)); -} - size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) { + HUF_CElt const* ct = CTable + 1; size_t nbBits = 0; int s; for (s = 0; s <= (int)maxSymbolValue; ++s) { - nbBits += CTable[s].nbBits * count[s]; + nbBits += HUF_getNbBits(ct[s]) * count[s]; } return nbBits >> 3; } int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) { + HUF_CElt const* ct = CTable + 1; int bad = 0; int s; for (s = 0; s <= (int)maxSymbolValue; ++s) { - bad |= (count[s] != 0) & (CTable[s].nbBits == 0); + bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0); } return !bad; } size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); } +/** HUF_CStream_t: + * Huffman uses its own BIT_CStream_t implementation. + * There are three major differences from BIT_CStream_t: + * 1. HUF_addBits() takes a HUF_CElt (size_t) which is + * the pair (nbBits, value) in the format: + * format: + * - Bits [0, 4) = nbBits + * - Bits [4, 64 - nbBits) = 0 + * - Bits [64 - nbBits, 64) = value + * 2. The bitContainer is built from the upper bits and + * right shifted. E.g. to add a new value of N bits + * you right shift the bitContainer by N, then or in + * the new value into the N upper bits. + * 3. The bitstream has two bit containers. You can add + * bits to the second container and merge them into + * the first container. + */ + +#define HUF_BITS_IN_CONTAINER (sizeof(size_t) * 8) + +typedef struct { + size_t bitContainer[2]; + size_t bitPos[2]; + + BYTE* startPtr; + BYTE* ptr; + BYTE* endPtr; +} HUF_CStream_t; + +/**! HUF_initCStream(): + * Initializes the bitstream. + * @returns 0 or an error code. + */ +static size_t HUF_initCStream(HUF_CStream_t* bitC, + void* startPtr, size_t dstCapacity) +{ + ZSTD_memset(bitC, 0, sizeof(*bitC)); + bitC->startPtr = (BYTE*)startPtr; + bitC->ptr = bitC->startPtr; + bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer[0]); + if (dstCapacity <= sizeof(bitC->bitContainer[0])) return ERROR(dstSize_tooSmall); + return 0; +} + +/*! HUF_addBits(): + * Adds the symbol stored in HUF_CElt elt to the bitstream. + * + * @param elt The element we're adding. This is a (nbBits, value) pair. + * See the HUF_CStream_t docs for the format. + * @param idx Insert into the bitstream at this idx. + * @param kFast This is a template parameter. If the bitstream is guaranteed + * to have at least 4 unused bits after this call it may be 1, + * otherwise it must be 0. HUF_addBits() is faster when fast is set. + */ +FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int idx, int kFast) +{ + assert(idx <= 1); + assert(HUF_getNbBits(elt) <= HUF_TABLELOG_ABSOLUTEMAX); + /* This is efficient on x86-64 with BMI2 because shrx + * only reads the low 6 bits of the register. The compiler + * knows this and elides the mask. When fast is set, + * every operation can use the same value loaded from elt. + */ + bitC->bitContainer[idx] >>= HUF_getNbBits(elt); + bitC->bitContainer[idx] |= kFast ? HUF_getValueFast(elt) : HUF_getValue(elt); + /* We only read the low 8 bits of bitC->bitPos[idx] so it + * doesn't matter that the high bits have noise from the value. + */ + bitC->bitPos[idx] += HUF_getNbBitsFast(elt); + assert((bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER); + /* The last 4-bits of elt are dirty if fast is set, + * so we must not be overwriting bits that have already been + * inserted into the bit container. + */ +#if DEBUGLEVEL >= 1 + { + size_t const nbBits = HUF_getNbBits(elt); + size_t const dirtyBits = nbBits == 0 ? 0 : BIT_highbit32((U32)nbBits) + 1; + (void)dirtyBits; + /* Middle bits are 0. */ + assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0); + /* We didn't overwrite any bits in the bit container. */ + assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER); + (void)dirtyBits; + } +#endif +} + +FORCE_INLINE_TEMPLATE void HUF_zeroIndex1(HUF_CStream_t* bitC) +{ + bitC->bitContainer[1] = 0; + bitC->bitPos[1] = 0; +} + +/*! HUF_mergeIndex1() : + * Merges the bit container @ index 1 into the bit container @ index 0 + * and zeros the bit container @ index 1. + */ +FORCE_INLINE_TEMPLATE void HUF_mergeIndex1(HUF_CStream_t* bitC) +{ + assert((bitC->bitPos[1] & 0xFF) < HUF_BITS_IN_CONTAINER); + bitC->bitContainer[0] >>= (bitC->bitPos[1] & 0xFF); + bitC->bitContainer[0] |= bitC->bitContainer[1]; + bitC->bitPos[0] += bitC->bitPos[1]; + assert((bitC->bitPos[0] & 0xFF) <= HUF_BITS_IN_CONTAINER); +} + +/*! HUF_flushBits() : +* Flushes the bits in the bit container @ index 0. +* +* @post bitPos will be < 8. +* @param kFast If kFast is set then we must know a-priori that +* the bit container will not overflow. +*/ +FORCE_INLINE_TEMPLATE void HUF_flushBits(HUF_CStream_t* bitC, int kFast) +{ + /* The upper bits of bitPos are noisy, so we must mask by 0xFF. */ + size_t const nbBits = bitC->bitPos[0] & 0xFF; + size_t const nbBytes = nbBits >> 3; + /* The top nbBits bits of bitContainer are the ones we need. */ + size_t const bitContainer = bitC->bitContainer[0] >> (HUF_BITS_IN_CONTAINER - nbBits); + /* Mask bitPos to account for the bytes we consumed. */ + bitC->bitPos[0] &= 7; + assert(nbBits > 0); + assert(nbBits <= sizeof(bitC->bitContainer[0]) * 8); + assert(bitC->ptr <= bitC->endPtr); + MEM_writeLEST(bitC->ptr, bitContainer); + bitC->ptr += nbBytes; + assert(!kFast || bitC->ptr <= bitC->endPtr); + if (!kFast && bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr; + /* bitContainer doesn't need to be modified because the leftover + * bits are already the top bitPos bits. And we don't care about + * noise in the lower values. + */ +} + +/*! HUF_endMark() + * @returns The Huffman stream end mark: A 1-bit value = 1. + */ +static HUF_CElt HUF_endMark(void) +{ + HUF_CElt endMark; + HUF_setNbBits(&endMark, 1); + HUF_setValue(&endMark, 1); + return endMark; +} + +/*! HUF_closeCStream() : + * @return Size of CStream, in bytes, + * or 0 if it could not fit into dstBuffer */ +static size_t HUF_closeCStream(HUF_CStream_t* bitC) +{ + HUF_addBits(bitC, HUF_endMark(), /* idx */ 0, /* kFast */ 0); + HUF_flushBits(bitC, /* kFast */ 0); + { + size_t const nbBits = bitC->bitPos[0] & 0xFF; + if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */ + return (bitC->ptr - bitC->startPtr) + (nbBits > 0); + } +} + FORCE_INLINE_TEMPLATE void -HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable) +HUF_encodeSymbol(HUF_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable, int idx, int fast) { - BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits); + HUF_addBits(bitCPtr, CTable[symbol], idx, fast); } -#define HUF_FLUSHBITS(s) BIT_flushBits(s) +FORCE_INLINE_TEMPLATE void +HUF_compress1X_usingCTable_internal_body_loop(HUF_CStream_t* bitC, + const BYTE* ip, size_t srcSize, + const HUF_CElt* ct, + int kUnroll, int kFastFlush, int kLastFast) +{ + /* Join to kUnroll */ + int n = (int)srcSize; + int rem = n % kUnroll; + if (rem > 0) { + for (; rem > 0; --rem) { + HUF_encodeSymbol(bitC, ip[--n], ct, 0, /* fast */ 0); + } + HUF_flushBits(bitC, kFastFlush); + } + assert(n % kUnroll == 0); + + /* Join to 2 * kUnroll */ + if (n % (2 * kUnroll)) { + int u; + for (u = 1; u < kUnroll; ++u) { + HUF_encodeSymbol(bitC, ip[n - u], ct, 0, 1); + } + HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, 0, kLastFast); + HUF_flushBits(bitC, kFastFlush); + n -= kUnroll; + } + assert(n % (2 * kUnroll) == 0); -#define HUF_FLUSHBITS_1(stream) \ - if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream) + for (; n>0; n-= 2 * kUnroll) { + /* Encode kUnroll symbols into the bitstream @ index 0. */ + int u; + for (u = 1; u < kUnroll; ++u) { + HUF_encodeSymbol(bitC, ip[n - u], ct, /* idx */ 0, /* fast */ 1); + } + HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, /* idx */ 0, /* fast */ kLastFast); + HUF_flushBits(bitC, kFastFlush); + /* Encode kUnroll symbols into the bitstream @ index 1. + * This allows us to start filling the bit container + * without any data dependencies. + */ + HUF_zeroIndex1(bitC); + for (u = 1; u < kUnroll; ++u) { + HUF_encodeSymbol(bitC, ip[n - kUnroll - u], ct, /* idx */ 1, /* fast */ 1); + } + HUF_encodeSymbol(bitC, ip[n - kUnroll - kUnroll], ct, /* idx */ 1, /* fast */ kLastFast); + /* Merge bitstream @ index 1 into the bitstream @ index 0 */ + HUF_mergeIndex1(bitC); + HUF_flushBits(bitC, kFastFlush); + } + assert(n == 0); + +} + +/** + * Returns a tight upper bound on the output space needed by Huffman + * with 8 bytes buffer to handle over-writes. If the output is at least + * this large we don't need to do bounds checks during Huffman encoding. + */ +static size_t HUF_tightCompressBound(size_t srcSize, size_t tableLog) +{ + return ((srcSize * tableLog) >> 3) + 8; +} -#define HUF_FLUSHBITS_2(stream) \ - if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream) FORCE_INLINE_TEMPLATE size_t HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable) { + U32 const tableLog = (U32)CTable[0]; + HUF_CElt const* ct = CTable + 1; const BYTE* ip = (const BYTE*) src; BYTE* const ostart = (BYTE*)dst; BYTE* const oend = ostart + dstSize; BYTE* op = ostart; - size_t n; - BIT_CStream_t bitC; + HUF_CStream_t bitC; /* init */ if (dstSize < 8) return 0; /* not enough space to compress */ - { size_t const initErr = BIT_initCStream(&bitC, op, (size_t)(oend-op)); + { size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op)); if (HUF_isError(initErr)) return 0; } - n = srcSize & ~3; /* join to mod 4 */ - switch (srcSize & 3) - { - case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable); - HUF_FLUSHBITS_2(&bitC); - /* fall-through */ - case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable); - HUF_FLUSHBITS_1(&bitC); - /* fall-through */ - case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable); - HUF_FLUSHBITS(&bitC); - /* fall-through */ - case 0 : /* fall-through */ - default: break; - } - - for (; n>0; n-=4) { /* note : n&3==0 at this stage */ - HUF_encodeSymbol(&bitC, ip[n- 1], CTable); - HUF_FLUSHBITS_1(&bitC); - HUF_encodeSymbol(&bitC, ip[n- 2], CTable); - HUF_FLUSHBITS_2(&bitC); - HUF_encodeSymbol(&bitC, ip[n- 3], CTable); - HUF_FLUSHBITS_1(&bitC); - HUF_encodeSymbol(&bitC, ip[n- 4], CTable); - HUF_FLUSHBITS(&bitC); + if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11) + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ MEM_32bits() ? 2 : 4, /* kFast */ 0, /* kLastFast */ 0); + else { + if (MEM_32bits()) { + switch (tableLog) { + case 11: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 0); + break; + case 10: ZSTD_FALLTHROUGH; + case 9: ZSTD_FALLTHROUGH; + case 8: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 1); + break; + case 7: ZSTD_FALLTHROUGH; + default: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 3, /* kFastFlush */ 1, /* kLastFast */ 1); + break; + } + } else { + switch (tableLog) { + case 11: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 0); + break; + case 10: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 1); + break; + case 9: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 6, /* kFastFlush */ 1, /* kLastFast */ 0); + break; + case 8: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 7, /* kFastFlush */ 1, /* kLastFast */ 0); + break; + case 7: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 8, /* kFastFlush */ 1, /* kLastFast */ 0); + break; + case 6: ZSTD_FALLTHROUGH; + default: + HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 9, /* kFastFlush */ 1, /* kLastFast */ 1); + break; + } + } } + assert(bitC.ptr <= bitC.endPtr); - return BIT_closeCStream(&bitC); + return HUF_closeCStream(&bitC); } #if DYNAMIC_BMI2 -static TARGET_ATTRIBUTE("bmi2") size_t +static BMI2_TARGET_ATTRIBUTE size_t HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable) @@ -545,9 +1071,13 @@ HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize, size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable) { - return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0); + return HUF_compress1X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0); } +size_t HUF_compress1X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2) +{ + return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2); +} static size_t HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, @@ -567,8 +1097,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, assert(op <= oend); { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) ); - if (cSize==0) return 0; - assert(cSize <= 65535); + if (cSize == 0 || cSize > 65535) return 0; MEM_writeLE16(ostart, (U16)cSize); op += cSize; } @@ -576,8 +1105,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, ip += segmentSize; assert(op <= oend); { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) ); - if (cSize==0) return 0; - assert(cSize <= 65535); + if (cSize == 0 || cSize > 65535) return 0; MEM_writeLE16(ostart+2, (U16)cSize); op += cSize; } @@ -585,8 +1113,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, ip += segmentSize; assert(op <= oend); { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) ); - if (cSize==0) return 0; - assert(cSize <= 65535); + if (cSize == 0 || cSize > 65535) return 0; MEM_writeLE16(ostart+4, (U16)cSize); op += cSize; } @@ -595,7 +1122,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, assert(op <= oend); assert(ip <= iend); { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) ); - if (cSize==0) return 0; + if (cSize == 0 || cSize > 65535) return 0; op += cSize; } @@ -604,7 +1131,12 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize, size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable) { - return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0); + return HUF_compress4X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0); +} + +size_t HUF_compress4X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2) +{ + return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2); } typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e; @@ -628,12 +1160,20 @@ static size_t HUF_compressCTable_internal( typedef struct { unsigned count[HUF_SYMBOLVALUE_MAX + 1]; - HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1]; - HUF_buildCTable_wksp_tables buildCTable_wksp; + HUF_CElt CTable[HUF_CTABLE_SIZE_ST(HUF_SYMBOLVALUE_MAX)]; + union { + HUF_buildCTable_wksp_tables buildCTable_wksp; + HUF_WriteCTableWksp writeCTable_wksp; + U32 hist_wksp[HIST_WKSP_SIZE_U32]; + } wksps; } HUF_compress_tables_t; +#define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096 +#define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10 /* Must be >= 2 */ + /* HUF_compress_internal() : - * `workSpace` must a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */ + * `workSpace_align4` must be aligned on 4-bytes boundaries, + * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */ static size_t HUF_compress_internal (void* dst, size_t dstSize, const void* src, size_t srcSize, @@ -641,18 +1181,17 @@ HUF_compress_internal (void* dst, size_t dstSize, HUF_nbStreams_e nbStreams, void* workSpace, size_t wkspSize, HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat, - const int bmi2) + const int bmi2, unsigned suspectUncompressible) { - HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace; + HUF_compress_tables_t* const table = (HUF_compress_tables_t*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(size_t)); BYTE* const ostart = (BYTE*)dst; BYTE* const oend = ostart + dstSize; BYTE* op = ostart; - HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE); + HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE); /* checks & inits */ - if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC); /* must be aligned on 4-bytes boundaries */ - if (wkspSize < HUF_WORKSPACE_SIZE) return ERROR(workSpace_tooSmall); + if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall); if (!srcSize) return 0; /* Uncompressed */ if (!dstSize) return 0; /* cannot fit anything within dst budget */ if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong); /* current block size limit */ @@ -668,8 +1207,23 @@ HUF_compress_internal (void* dst, size_t dstSize, nbStreams, oldHufTable, bmi2); } + /* If uncompressible data is suspected, do a smaller sampling first */ + DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2); + if (suspectUncompressible && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) { + size_t largestTotal = 0; + { unsigned maxSymbolValueBegin = maxSymbolValue; + CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) ); + largestTotal += largestBegin; + } + { unsigned maxSymbolValueEnd = maxSymbolValue; + CHECK_V_F(largestEnd, HIST_count_simple (table->count, &maxSymbolValueEnd, (const BYTE*)src + srcSize - SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) ); + largestTotal += largestEnd; + } + if (largestTotal <= ((2 * SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) >> 7)+4) return 0; /* heuristic : probably not compressible enough */ + } + /* Scan input and build symbol stats */ - { CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, workSpace, wkspSize) ); + { CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, table->wksps.hist_wksp, sizeof(table->wksps.hist_wksp)) ); if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; } /* single symbol, rle */ if (largest <= (srcSize >> 7)+4) return 0; /* heuristic : probably not compressible enough */ } @@ -691,16 +1245,20 @@ HUF_compress_internal (void* dst, size_t dstSize, huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue); { size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count, maxSymbolValue, huffLog, - &table->buildCTable_wksp, sizeof(table->buildCTable_wksp)); + &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp)); CHECK_F(maxBits); huffLog = (U32)maxBits; - /* Zero unused symbols in CTable, so we can check it for validity */ - memset(table->CTable + (maxSymbolValue + 1), 0, - sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt))); + } + /* Zero unused symbols in CTable, so we can check it for validity */ + { + size_t const ctableSize = HUF_CTABLE_SIZE_ST(maxSymbolValue); + size_t const unusedSize = sizeof(table->CTable) - ctableSize * sizeof(HUF_CElt); + ZSTD_memset(table->CTable + ctableSize, 0, unusedSize); } /* Write table description header */ - { CHECK_V_F(hSize, HUF_writeCTable (op, dstSize, table->CTable, maxSymbolValue, huffLog) ); + { CHECK_V_F(hSize, HUF_writeCTable_wksp(op, dstSize, table->CTable, maxSymbolValue, huffLog, + &table->wksps.writeCTable_wksp, sizeof(table->wksps.writeCTable_wksp)) ); /* Check if using previous huffman table is beneficial */ if (repeat && *repeat != HUF_repeat_none) { size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue); @@ -716,7 +1274,7 @@ HUF_compress_internal (void* dst, size_t dstSize, op += hSize; if (repeat) { *repeat = HUF_repeat_none; } if (oldHufTable) - memcpy(oldHufTable, table->CTable, sizeof(table->CTable)); /* Save new table */ + ZSTD_memcpy(oldHufTable, table->CTable, sizeof(table->CTable)); /* Save new table */ } return HUF_compressCTable_internal(ostart, op, oend, src, srcSize, @@ -732,27 +1290,20 @@ size_t HUF_compress1X_wksp (void* dst, size_t dstSize, return HUF_compress_internal(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, HUF_singleStream, workSpace, wkspSize, - NULL, NULL, 0, 0 /*bmi2*/); + NULL, NULL, 0, 0 /*bmi2*/, 0); } size_t HUF_compress1X_repeat (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned huffLog, void* workSpace, size_t wkspSize, - HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2) + HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, + int bmi2, unsigned suspectUncompressible) { return HUF_compress_internal(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, HUF_singleStream, workSpace, wkspSize, hufTable, - repeat, preferRepeat, bmi2); -} - -size_t HUF_compress1X (void* dst, size_t dstSize, - const void* src, size_t srcSize, - unsigned maxSymbolValue, unsigned huffLog) -{ - unsigned workSpace[HUF_WORKSPACE_SIZE_U32]; - return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace)); + repeat, preferRepeat, bmi2, suspectUncompressible); } /* HUF_compress4X_repeat(): @@ -766,29 +1317,49 @@ size_t HUF_compress4X_wksp (void* dst, size_t dstSize, return HUF_compress_internal(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, HUF_fourStreams, workSpace, wkspSize, - NULL, NULL, 0, 0 /*bmi2*/); + NULL, NULL, 0, 0 /*bmi2*/, 0); } /* HUF_compress4X_repeat(): * compress input using 4 streams. + * consider skipping quickly * re-use an existing huffman compression table */ size_t HUF_compress4X_repeat (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned huffLog, void* workSpace, size_t wkspSize, - HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2) + HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible) { return HUF_compress_internal(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, HUF_fourStreams, workSpace, wkspSize, - hufTable, repeat, preferRepeat, bmi2); + hufTable, repeat, preferRepeat, bmi2, suspectUncompressible); +} + +#ifndef ZSTD_NO_UNUSED_FUNCTIONS +/** HUF_buildCTable() : + * @return : maxNbBits + * Note : count is used before tree is written, so they can safely overlap + */ +size_t HUF_buildCTable (HUF_CElt* tree, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits) +{ + HUF_buildCTable_wksp_tables workspace; + return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, &workspace, sizeof(workspace)); +} + +size_t HUF_compress1X (void* dst, size_t dstSize, + const void* src, size_t srcSize, + unsigned maxSymbolValue, unsigned huffLog) +{ + U64 workSpace[HUF_WORKSPACE_SIZE_U64]; + return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace)); } size_t HUF_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned huffLog) { - unsigned workSpace[HUF_WORKSPACE_SIZE_U32]; + U64 workSpace[HUF_WORKSPACE_SIZE_U64]; return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace)); } @@ -796,3 +1367,4 @@ size_t HUF_compress (void* dst, size_t maxDstSize, const void* src, size_t srcSi { return HUF_compress2(dst, maxDstSize, src, srcSize, 255, HUF_TABLELOG_DEFAULT); } +#endif diff --git a/thirdparty/zstd/compress/zstd_compress.c b/thirdparty/zstd/compress/zstd_compress.c index 3f963b1cff..f06456af92 100644 --- a/thirdparty/zstd/compress/zstd_compress.c +++ b/thirdparty/zstd/compress/zstd_compress.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 @@ -11,9 +11,7 @@ /*-************************************* * Dependencies ***************************************/ -#include <limits.h> /* INT_MAX */ -#include <string.h> /* memset */ -#include "../common/cpu.h" +#include "../common/zstd_deps.h" /* INT_MAX, ZSTD_memset, ZSTD_memcpy */ #include "../common/mem.h" #include "hist.h" /* HIST_countFast_wksp */ #define FSE_STATIC_LINKING_ONLY /* FSE_encodeSymbol */ @@ -30,6 +28,31 @@ #include "zstd_ldm.h" #include "zstd_compress_superblock.h" +/* *************************************************************** +* Tuning parameters +*****************************************************************/ +/*! + * COMPRESS_HEAPMODE : + * Select how default decompression function ZSTD_compress() allocates its context, + * on stack (0, default), or into heap (1). + * Note that functions with explicit context such as ZSTD_compressCCtx() are unaffected. + */ +#ifndef ZSTD_COMPRESS_HEAPMODE +# define ZSTD_COMPRESS_HEAPMODE 0 +#endif + +/*! + * ZSTD_HASHLOG3_MAX : + * Maximum size of the hash table dedicated to find 3-bytes matches, + * in log format, aka 17 => 1 << 17 == 128Ki positions. + * This structure is only used in zstd_opt. + * Since allocation is centralized for all strategies, it has to be known here. + * The actual (selected) size of the hash table is then stored in ZSTD_matchState_t.hashLog3, + * so that zstd_opt.c doesn't need to know about this constant. + */ +#ifndef ZSTD_HASHLOG3_MAX +# define ZSTD_HASHLOG3_MAX 17 +#endif /*-************************************* * Helper functions @@ -52,6 +75,7 @@ size_t ZSTD_compressBound(size_t srcSize) { struct ZSTD_CDict_s { const void* dictContent; size_t dictContentSize; + ZSTD_dictContentType_e dictContentType; /* The dictContentType the CDict was created with */ U32* entropyWorkspace; /* entropy workspace of HUF_WORKSPACE_SIZE bytes */ ZSTD_cwksp workspace; ZSTD_matchState_t matchState; @@ -59,6 +83,10 @@ struct ZSTD_CDict_s { ZSTD_customMem customMem; U32 dictID; int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */ + ZSTD_paramSwitch_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use + * row-based matchfinder. Unless the cdict is reloaded, we will use + * the same greedy/lazy matchfinder at compression time. + */ }; /* typedef'd to ZSTD_CDict within "zstd.h" */ ZSTD_CCtx* ZSTD_createCCtx(void) @@ -69,9 +97,9 @@ ZSTD_CCtx* ZSTD_createCCtx(void) static void ZSTD_initCCtx(ZSTD_CCtx* cctx, ZSTD_customMem memManager) { assert(cctx != NULL); - memset(cctx, 0, sizeof(*cctx)); + ZSTD_memset(cctx, 0, sizeof(*cctx)); cctx->customMem = memManager; - cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid()); + cctx->bmi2 = ZSTD_cpuSupportsBmi2(); { size_t const err = ZSTD_CCtx_reset(cctx, ZSTD_reset_parameters); assert(!ZSTD_isError(err)); (void)err; @@ -82,8 +110,8 @@ ZSTD_CCtx* ZSTD_createCCtx_advanced(ZSTD_customMem customMem) { ZSTD_STATIC_ASSERT(zcss_init==0); ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN==(0ULL - 1)); - if (!customMem.customAlloc ^ !customMem.customFree) return NULL; - { ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_malloc(sizeof(ZSTD_CCtx), customMem); + if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL; + { ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_customMalloc(sizeof(ZSTD_CCtx), customMem); if (!cctx) return NULL; ZSTD_initCCtx(cctx, customMem); return cctx; @@ -96,20 +124,20 @@ ZSTD_CCtx* ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize) ZSTD_CCtx* cctx; if (workspaceSize <= sizeof(ZSTD_CCtx)) return NULL; /* minimum size */ if ((size_t)workspace & 7) return NULL; /* must be 8-aligned */ - ZSTD_cwksp_init(&ws, workspace, workspaceSize); + ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc); cctx = (ZSTD_CCtx*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CCtx)); if (cctx == NULL) return NULL; - memset(cctx, 0, sizeof(ZSTD_CCtx)); + ZSTD_memset(cctx, 0, sizeof(ZSTD_CCtx)); ZSTD_cwksp_move(&cctx->workspace, &ws); cctx->staticSize = workspaceSize; /* statically sized space. entropyWorkspace never moves (but prev/next block swap places) */ - if (!ZSTD_cwksp_check_available(&cctx->workspace, HUF_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL; + if (!ZSTD_cwksp_check_available(&cctx->workspace, ENTROPY_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL; cctx->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t)); cctx->blockState.nextCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t)); - cctx->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cctx->workspace, HUF_WORKSPACE_SIZE); + cctx->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cctx->workspace, ENTROPY_WORKSPACE_SIZE); cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid()); return cctx; } @@ -119,10 +147,10 @@ ZSTD_CCtx* ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize) */ static void ZSTD_clearAllDicts(ZSTD_CCtx* cctx) { - ZSTD_free(cctx->localDict.dictBuffer, cctx->customMem); + ZSTD_customFree(cctx->localDict.dictBuffer, cctx->customMem); ZSTD_freeCDict(cctx->localDict.cdict); - memset(&cctx->localDict, 0, sizeof(cctx->localDict)); - memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict)); + ZSTD_memset(&cctx->localDict, 0, sizeof(cctx->localDict)); + ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict)); cctx->cdict = NULL; } @@ -153,7 +181,7 @@ size_t ZSTD_freeCCtx(ZSTD_CCtx* cctx) int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx); ZSTD_freeCCtxContent(cctx); if (!cctxInWorkspace) { - ZSTD_free(cctx, cctx->customMem); + ZSTD_customFree(cctx, cctx->customMem); } } return 0; @@ -189,15 +217,84 @@ size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs) /* private API call, for dictBuilder only */ const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); } +/* Returns true if the strategy supports using a row based matchfinder */ +static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) { + return (strategy >= ZSTD_greedy && strategy <= ZSTD_lazy2); +} + +/* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder + * for this compression. + */ +static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_paramSwitch_e mode) { + assert(mode != ZSTD_ps_auto); + return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_ps_enable); +} + +/* Returns row matchfinder usage given an initial mode and cParams */ +static ZSTD_paramSwitch_e ZSTD_resolveRowMatchFinderMode(ZSTD_paramSwitch_e mode, + const ZSTD_compressionParameters* const cParams) { +#if defined(ZSTD_ARCH_X86_SSE2) || defined(ZSTD_ARCH_ARM_NEON) + int const kHasSIMD128 = 1; +#else + int const kHasSIMD128 = 0; +#endif + if (mode != ZSTD_ps_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */ + mode = ZSTD_ps_disable; + if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode; + if (kHasSIMD128) { + if (cParams->windowLog > 14) mode = ZSTD_ps_enable; + } else { + if (cParams->windowLog > 17) mode = ZSTD_ps_enable; + } + return mode; +} + +/* Returns block splitter usage (generally speaking, when using slower/stronger compression modes) */ +static ZSTD_paramSwitch_e ZSTD_resolveBlockSplitterMode(ZSTD_paramSwitch_e mode, + const ZSTD_compressionParameters* const cParams) { + if (mode != ZSTD_ps_auto) return mode; + return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 17) ? ZSTD_ps_enable : ZSTD_ps_disable; +} + +/* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */ +static int ZSTD_allocateChainTable(const ZSTD_strategy strategy, + const ZSTD_paramSwitch_e useRowMatchFinder, + const U32 forDDSDict) { + assert(useRowMatchFinder != ZSTD_ps_auto); + /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate. + * We do not allocate a chaintable if we are using ZSTD_fast, or are using the row-based matchfinder. + */ + return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder)); +} + +/* Returns 1 if compression parameters are such that we should + * enable long distance matching (wlog >= 27, strategy >= btopt). + * Returns 0 otherwise. + */ +static ZSTD_paramSwitch_e ZSTD_resolveEnableLdm(ZSTD_paramSwitch_e mode, + const ZSTD_compressionParameters* const cParams) { + if (mode != ZSTD_ps_auto) return mode; + return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable; +} + static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams( ZSTD_compressionParameters cParams) { ZSTD_CCtx_params cctxParams; - memset(&cctxParams, 0, sizeof(cctxParams)); + /* should not matter, as all cParams are presumed properly defined */ + ZSTD_CCtxParams_init(&cctxParams, ZSTD_CLEVEL_DEFAULT); cctxParams.cParams = cParams; - cctxParams.compressionLevel = ZSTD_CLEVEL_DEFAULT; /* should not matter, as all cParams are presumed properly defined */ + + /* Adjust advanced params according to cParams */ + cctxParams.ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams.ldmParams.enableLdm, &cParams); + if (cctxParams.ldmParams.enableLdm == ZSTD_ps_enable) { + ZSTD_ldm_adjustParameters(&cctxParams.ldmParams, &cParams); + assert(cctxParams.ldmParams.hashLog >= cctxParams.ldmParams.bucketSizeLog); + assert(cctxParams.ldmParams.hashRateLog < 32); + } + cctxParams.useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.useBlockSplitter, &cParams); + cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams); assert(!ZSTD_checkCParams(cParams)); - cctxParams.fParams.contentSizeFlag = 1; return cctxParams; } @@ -205,13 +302,12 @@ static ZSTD_CCtx_params* ZSTD_createCCtxParams_advanced( ZSTD_customMem customMem) { ZSTD_CCtx_params* params; - if (!customMem.customAlloc ^ !customMem.customFree) return NULL; - params = (ZSTD_CCtx_params*)ZSTD_calloc( + if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL; + params = (ZSTD_CCtx_params*)ZSTD_customCalloc( sizeof(ZSTD_CCtx_params), customMem); if (!params) { return NULL; } + ZSTD_CCtxParams_init(params, ZSTD_CLEVEL_DEFAULT); params->customMem = customMem; - params->compressionLevel = ZSTD_CLEVEL_DEFAULT; - params->fParams.contentSizeFlag = 1; return params; } @@ -223,7 +319,7 @@ ZSTD_CCtx_params* ZSTD_createCCtxParams(void) size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params) { if (params == NULL) { return 0; } - ZSTD_free(params, params->customMem); + ZSTD_customFree(params, params->customMem); return 0; } @@ -234,35 +330,57 @@ size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params) size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel) { RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!"); - memset(cctxParams, 0, sizeof(*cctxParams)); + ZSTD_memset(cctxParams, 0, sizeof(*cctxParams)); cctxParams->compressionLevel = compressionLevel; cctxParams->fParams.contentSizeFlag = 1; return 0; } +#define ZSTD_NO_CLEVEL 0 + +/** + * Initializes the cctxParams from params and compressionLevel. + * @param compressionLevel If params are derived from a compression level then that compression level, otherwise ZSTD_NO_CLEVEL. + */ +static void ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams, ZSTD_parameters const* params, int compressionLevel) +{ + assert(!ZSTD_checkCParams(params->cParams)); + ZSTD_memset(cctxParams, 0, sizeof(*cctxParams)); + cctxParams->cParams = params->cParams; + cctxParams->fParams = params->fParams; + /* Should not matter, as all cParams are presumed properly defined. + * But, set it for tracing anyway. + */ + cctxParams->compressionLevel = compressionLevel; + cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, ¶ms->cParams); + cctxParams->useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->useBlockSplitter, ¶ms->cParams); + cctxParams->ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams->ldmParams.enableLdm, ¶ms->cParams); + DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d, useBlockSplitter=%d ldm=%d", + cctxParams->useRowMatchFinder, cctxParams->useBlockSplitter, cctxParams->ldmParams.enableLdm); +} + size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params) { RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!"); FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , ""); - memset(cctxParams, 0, sizeof(*cctxParams)); - assert(!ZSTD_checkCParams(params.cParams)); - cctxParams->cParams = params.cParams; - cctxParams->fParams = params.fParams; - cctxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT; /* should not matter, as all cParams are presumed properly defined */ + ZSTD_CCtxParams_init_internal(cctxParams, ¶ms, ZSTD_NO_CLEVEL); return 0; } -/* ZSTD_assignParamsToCCtxParams() : - * params is presumed valid at this stage */ -static ZSTD_CCtx_params ZSTD_assignParamsToCCtxParams( - const ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params) +/** + * Sets cctxParams' cParams and fParams from params, but otherwise leaves them alone. + * @param param Validated zstd parameters. + */ +static void ZSTD_CCtxParams_setZstdParams( + ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params) { - ZSTD_CCtx_params ret = *cctxParams; assert(!ZSTD_checkCParams(params->cParams)); - ret.cParams = params->cParams; - ret.fParams = params->fParams; - ret.compressionLevel = ZSTD_CLEVEL_DEFAULT; /* should not matter, as all cParams are presumed properly defined */ - return ret; + cctxParams->cParams = params->cParams; + cctxParams->fParams = params->fParams; + /* Should not matter, as all cParams are presumed properly defined. + * But, set it for tracing anyway. + */ + cctxParams->compressionLevel = ZSTD_NO_CLEVEL; } ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param) @@ -354,6 +472,11 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param) #endif return bounds; + case ZSTD_c_enableDedicatedDictSearch: + bounds.lowerBound = 0; + bounds.upperBound = 1; + return bounds; + case ZSTD_c_enableLongDistanceMatching: bounds.lowerBound = 0; bounds.upperBound = 1; @@ -397,15 +520,15 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param) return bounds; case ZSTD_c_forceAttachDict: - ZSTD_STATIC_ASSERT(ZSTD_dictDefaultAttach < ZSTD_dictForceCopy); + ZSTD_STATIC_ASSERT(ZSTD_dictDefaultAttach < ZSTD_dictForceLoad); bounds.lowerBound = ZSTD_dictDefaultAttach; bounds.upperBound = ZSTD_dictForceLoad; /* note : how to ensure at compile time that this is the highest value enum ? */ return bounds; case ZSTD_c_literalCompressionMode: - ZSTD_STATIC_ASSERT(ZSTD_lcm_auto < ZSTD_lcm_huffman && ZSTD_lcm_huffman < ZSTD_lcm_uncompressed); - bounds.lowerBound = ZSTD_lcm_auto; - bounds.upperBound = ZSTD_lcm_uncompressed; + ZSTD_STATIC_ASSERT(ZSTD_ps_auto < ZSTD_ps_enable && ZSTD_ps_enable < ZSTD_ps_disable); + bounds.lowerBound = (int)ZSTD_ps_auto; + bounds.upperBound = (int)ZSTD_ps_disable; return bounds; case ZSTD_c_targetCBlockSize: @@ -418,6 +541,37 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param) bounds.upperBound = ZSTD_SRCSIZEHINT_MAX; return bounds; + case ZSTD_c_stableInBuffer: + case ZSTD_c_stableOutBuffer: + bounds.lowerBound = (int)ZSTD_bm_buffered; + bounds.upperBound = (int)ZSTD_bm_stable; + return bounds; + + case ZSTD_c_blockDelimiters: + bounds.lowerBound = (int)ZSTD_sf_noBlockDelimiters; + bounds.upperBound = (int)ZSTD_sf_explicitBlockDelimiters; + return bounds; + + case ZSTD_c_validateSequences: + bounds.lowerBound = 0; + bounds.upperBound = 1; + return bounds; + + case ZSTD_c_useBlockSplitter: + bounds.lowerBound = (int)ZSTD_ps_auto; + bounds.upperBound = (int)ZSTD_ps_disable; + return bounds; + + case ZSTD_c_useRowMatchFinder: + bounds.lowerBound = (int)ZSTD_ps_auto; + bounds.upperBound = (int)ZSTD_ps_disable; + return bounds; + + case ZSTD_c_deterministicRefPrefix: + bounds.lowerBound = 0; + bounds.upperBound = 1; + return bounds; + default: bounds.error = ERROR(parameter_unsupported); return bounds; @@ -465,6 +619,7 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param) case ZSTD_c_jobSize: case ZSTD_c_overlapLog: case ZSTD_c_rsyncable: + case ZSTD_c_enableDedicatedDictSearch: case ZSTD_c_enableLongDistanceMatching: case ZSTD_c_ldmHashLog: case ZSTD_c_ldmMinMatch: @@ -474,6 +629,13 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param) case ZSTD_c_literalCompressionMode: case ZSTD_c_targetCBlockSize: case ZSTD_c_srcSizeHint: + case ZSTD_c_stableInBuffer: + case ZSTD_c_stableOutBuffer: + case ZSTD_c_blockDelimiters: + case ZSTD_c_validateSequences: + case ZSTD_c_useBlockSplitter: + case ZSTD_c_useRowMatchFinder: + case ZSTD_c_deterministicRefPrefix: default: return 0; } @@ -515,12 +677,20 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value) case ZSTD_c_jobSize: case ZSTD_c_overlapLog: case ZSTD_c_rsyncable: + case ZSTD_c_enableDedicatedDictSearch: case ZSTD_c_enableLongDistanceMatching: case ZSTD_c_ldmHashLog: case ZSTD_c_ldmMinMatch: case ZSTD_c_ldmBucketSizeLog: case ZSTD_c_targetCBlockSize: case ZSTD_c_srcSizeHint: + case ZSTD_c_stableInBuffer: + case ZSTD_c_stableOutBuffer: + case ZSTD_c_blockDelimiters: + case ZSTD_c_validateSequences: + case ZSTD_c_useBlockSplitter: + case ZSTD_c_useRowMatchFinder: + case ZSTD_c_deterministicRefPrefix: break; default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); @@ -541,9 +711,10 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, case ZSTD_c_compressionLevel : { FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(param, &value), ""); - if (value) { /* 0 : does not change current level */ + if (value == 0) + CCtxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT; /* 0 == default */ + else CCtxParams->compressionLevel = value; - } if (CCtxParams->compressionLevel >= 0) return (size_t)CCtxParams->compressionLevel; return 0; /* return type (size_t) cannot represent negative values */ } @@ -617,7 +788,7 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, } case ZSTD_c_literalCompressionMode : { - const ZSTD_literalCompressionMode_e lcm = (ZSTD_literalCompressionMode_e)value; + const ZSTD_paramSwitch_e lcm = (ZSTD_paramSwitch_e)value; BOUNDCHECK(ZSTD_c_literalCompressionMode, lcm); CCtxParams->literalCompressionMode = lcm; return CCtxParams->literalCompressionMode; @@ -667,8 +838,12 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, return CCtxParams->rsyncable; #endif + case ZSTD_c_enableDedicatedDictSearch : + CCtxParams->enableDedicatedDictSearch = (value!=0); + return CCtxParams->enableDedicatedDictSearch; + case ZSTD_c_enableLongDistanceMatching : - CCtxParams->ldmParams.enableLdm = (value!=0); + CCtxParams->ldmParams.enableLdm = (ZSTD_paramSwitch_e)value; return CCtxParams->ldmParams.enableLdm; case ZSTD_c_ldmHashLog : @@ -690,8 +865,8 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, return CCtxParams->ldmParams.bucketSizeLog; case ZSTD_c_ldmHashRateLog : - RETURN_ERROR_IF(value > ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN, - parameter_outOfBound, "Param out of bounds!"); + if (value!=0) /* 0 ==> default */ + BOUNDCHECK(ZSTD_c_ldmHashRateLog, value); CCtxParams->ldmParams.hashRateLog = value; return CCtxParams->ldmParams.hashRateLog; @@ -707,17 +882,52 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams, CCtxParams->srcSizeHint = value; return CCtxParams->srcSizeHint; + case ZSTD_c_stableInBuffer: + BOUNDCHECK(ZSTD_c_stableInBuffer, value); + CCtxParams->inBufferMode = (ZSTD_bufferMode_e)value; + return CCtxParams->inBufferMode; + + case ZSTD_c_stableOutBuffer: + BOUNDCHECK(ZSTD_c_stableOutBuffer, value); + CCtxParams->outBufferMode = (ZSTD_bufferMode_e)value; + return CCtxParams->outBufferMode; + + case ZSTD_c_blockDelimiters: + BOUNDCHECK(ZSTD_c_blockDelimiters, value); + CCtxParams->blockDelimiters = (ZSTD_sequenceFormat_e)value; + return CCtxParams->blockDelimiters; + + case ZSTD_c_validateSequences: + BOUNDCHECK(ZSTD_c_validateSequences, value); + CCtxParams->validateSequences = value; + return CCtxParams->validateSequences; + + case ZSTD_c_useBlockSplitter: + BOUNDCHECK(ZSTD_c_useBlockSplitter, value); + CCtxParams->useBlockSplitter = (ZSTD_paramSwitch_e)value; + return CCtxParams->useBlockSplitter; + + case ZSTD_c_useRowMatchFinder: + BOUNDCHECK(ZSTD_c_useRowMatchFinder, value); + CCtxParams->useRowMatchFinder = (ZSTD_paramSwitch_e)value; + return CCtxParams->useRowMatchFinder; + + case ZSTD_c_deterministicRefPrefix: + BOUNDCHECK(ZSTD_c_deterministicRefPrefix, value); + CCtxParams->deterministicRefPrefix = !!value; + return CCtxParams->deterministicRefPrefix; + default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); } } -size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value) +size_t ZSTD_CCtx_getParameter(ZSTD_CCtx const* cctx, ZSTD_cParameter param, int* value) { return ZSTD_CCtxParams_getParameter(&cctx->requestedParams, param, value); } size_t ZSTD_CCtxParams_getParameter( - ZSTD_CCtx_params* CCtxParams, ZSTD_cParameter param, int* value) + ZSTD_CCtx_params const* CCtxParams, ZSTD_cParameter param, int* value) { switch(param) { @@ -794,6 +1004,9 @@ size_t ZSTD_CCtxParams_getParameter( *value = CCtxParams->rsyncable; break; #endif + case ZSTD_c_enableDedicatedDictSearch : + *value = CCtxParams->enableDedicatedDictSearch; + break; case ZSTD_c_enableLongDistanceMatching : *value = CCtxParams->ldmParams.enableLdm; break; @@ -815,6 +1028,27 @@ size_t ZSTD_CCtxParams_getParameter( case ZSTD_c_srcSizeHint : *value = (int)CCtxParams->srcSizeHint; break; + case ZSTD_c_stableInBuffer : + *value = (int)CCtxParams->inBufferMode; + break; + case ZSTD_c_stableOutBuffer : + *value = (int)CCtxParams->outBufferMode; + break; + case ZSTD_c_blockDelimiters : + *value = (int)CCtxParams->blockDelimiters; + break; + case ZSTD_c_validateSequences : + *value = (int)CCtxParams->validateSequences; + break; + case ZSTD_c_useBlockSplitter : + *value = (int)CCtxParams->useBlockSplitter; + break; + case ZSTD_c_useRowMatchFinder : + *value = (int)CCtxParams->useRowMatchFinder; + break; + case ZSTD_c_deterministicRefPrefix: + *value = (int)CCtxParams->deterministicRefPrefix; + break; default: RETURN_ERROR(parameter_unsupported, "unknown parameter"); } return 0; @@ -841,7 +1075,7 @@ size_t ZSTD_CCtx_setParametersUsingCCtxParams( return 0; } -ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize) +size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize) { DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %u bytes", (U32)pledgedSrcSize); RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong, @@ -850,6 +1084,14 @@ ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long lo return 0; } +static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams( + int const compressionLevel, + size_t const dictSize); +static int ZSTD_dedicatedDictSearch_isSupported( + const ZSTD_compressionParameters* cParams); +static void ZSTD_dedicatedDictSearch_revertCParams( + ZSTD_compressionParameters* cParams); + /** * Initializes the local dict using the requested parameters. * NOTE: This does not use the pledged src size, because it may be used for more @@ -858,8 +1100,6 @@ ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long lo static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx) { ZSTD_localDict* const dl = &cctx->localDict; - ZSTD_compressionParameters const cParams = ZSTD_getCParamsFromCCtxParams( - &cctx->requestedParams, ZSTD_CONTENTSIZE_UNKNOWN, dl->dictSize); if (dl->dict == NULL) { /* No local dictionary. */ assert(dl->dictBuffer == NULL); @@ -876,12 +1116,12 @@ static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx) assert(cctx->cdict == NULL); assert(cctx->prefixDict.dict == NULL); - dl->cdict = ZSTD_createCDict_advanced( + dl->cdict = ZSTD_createCDict_advanced2( dl->dict, dl->dictSize, ZSTD_dlm_byRef, dl->dictContentType, - cParams, + &cctx->requestedParams, cctx->customMem); RETURN_ERROR_IF(!dl->cdict, memory_allocation, "ZSTD_createCDict_advanced failed"); cctx->cdict = dl->cdict; @@ -894,8 +1134,6 @@ size_t ZSTD_CCtx_loadDictionary_advanced( { RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong, "Can't load a dictionary when ctx is not in init stage."); - RETURN_ERROR_IF(cctx->staticSize, memory_allocation, - "no malloc for static CCtx"); DEBUGLOG(4, "ZSTD_CCtx_loadDictionary_advanced (size: %u)", (U32)dictSize); ZSTD_clearAllDicts(cctx); /* in case one already exists */ if (dict == NULL || dictSize == 0) /* no dictionary mode */ @@ -903,9 +1141,12 @@ size_t ZSTD_CCtx_loadDictionary_advanced( if (dictLoadMethod == ZSTD_dlm_byRef) { cctx->localDict.dict = dict; } else { - void* dictBuffer = ZSTD_malloc(dictSize, cctx->customMem); + void* dictBuffer; + RETURN_ERROR_IF(cctx->staticSize, memory_allocation, + "no malloc for static CCtx"); + dictBuffer = ZSTD_customMalloc(dictSize, cctx->customMem); RETURN_ERROR_IF(!dictBuffer, memory_allocation, "NULL pointer!"); - memcpy(dictBuffer, dict, dictSize); + ZSTD_memcpy(dictBuffer, dict, dictSize); cctx->localDict.dictBuffer = dictBuffer; cctx->localDict.dict = dictBuffer; } @@ -914,14 +1155,14 @@ size_t ZSTD_CCtx_loadDictionary_advanced( return 0; } -ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference( +size_t ZSTD_CCtx_loadDictionary_byReference( ZSTD_CCtx* cctx, const void* dict, size_t dictSize) { return ZSTD_CCtx_loadDictionary_advanced( cctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto); } -ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize) +size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize) { return ZSTD_CCtx_loadDictionary_advanced( cctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto); @@ -938,6 +1179,14 @@ size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict) return 0; } +size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool) +{ + RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong, + "Can't ref a pool when ctx not in init stage."); + cctx->pool = pool; + return 0; +} + size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize) { return ZSTD_CCtx_refPrefix_advanced(cctx, prefix, prefixSize, ZSTD_dct_rawContent); @@ -1022,23 +1271,83 @@ U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat) return hashLog - btScale; } +/** ZSTD_dictAndWindowLog() : + * Returns an adjusted window log that is large enough to fit the source and the dictionary. + * The zstd format says that the entire dictionary is valid if one byte of the dictionary + * is within the window. So the hashLog and chainLog should be large enough to reference both + * the dictionary and the window. So we must use this adjusted dictAndWindowLog when downsizing + * the hashLog and windowLog. + * NOTE: srcSize must not be ZSTD_CONTENTSIZE_UNKNOWN. + */ +static U32 ZSTD_dictAndWindowLog(U32 windowLog, U64 srcSize, U64 dictSize) +{ + const U64 maxWindowSize = 1ULL << ZSTD_WINDOWLOG_MAX; + /* No dictionary ==> No change */ + if (dictSize == 0) { + return windowLog; + } + assert(windowLog <= ZSTD_WINDOWLOG_MAX); + assert(srcSize != ZSTD_CONTENTSIZE_UNKNOWN); /* Handled in ZSTD_adjustCParams_internal() */ + { + U64 const windowSize = 1ULL << windowLog; + U64 const dictAndWindowSize = dictSize + windowSize; + /* If the window size is already large enough to fit both the source and the dictionary + * then just use the window size. Otherwise adjust so that it fits the dictionary and + * the window. + */ + if (windowSize >= dictSize + srcSize) { + return windowLog; /* Window size large enough already */ + } else if (dictAndWindowSize >= maxWindowSize) { + return ZSTD_WINDOWLOG_MAX; /* Larger than max window log */ + } else { + return ZSTD_highbit32((U32)dictAndWindowSize - 1) + 1; + } + } +} + /** ZSTD_adjustCParams_internal() : * optimize `cPar` for a specified input (`srcSize` and `dictSize`). * mostly downsize to reduce memory consumption and initialization latency. * `srcSize` can be ZSTD_CONTENTSIZE_UNKNOWN when not known. + * `mode` is the mode for parameter adjustment. See docs for `ZSTD_cParamMode_e`. * note : `srcSize==0` means 0! * condition : cPar is presumed validated (can be checked using ZSTD_checkCParams()). */ static ZSTD_compressionParameters ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar, unsigned long long srcSize, - size_t dictSize) + size_t dictSize, + ZSTD_cParamMode_e mode) { - static const U64 minSrcSize = 513; /* (1<<9) + 1 */ - static const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1); + const U64 minSrcSize = 513; /* (1<<9) + 1 */ + const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1); assert(ZSTD_checkCParams(cPar)==0); - if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN) - srcSize = minSrcSize; + switch (mode) { + case ZSTD_cpm_unknown: + case ZSTD_cpm_noAttachDict: + /* If we don't know the source size, don't make any + * assumptions about it. We will already have selected + * smaller parameters if a dictionary is in use. + */ + break; + case ZSTD_cpm_createCDict: + /* Assume a small source size when creating a dictionary + * with an unknown source size. + */ + if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN) + srcSize = minSrcSize; + break; + case ZSTD_cpm_attachDict: + /* Dictionary has its own dedicated parameters which have + * already been selected. We are selecting parameters + * for only the source. + */ + dictSize = 0; + break; + default: + assert(0); + break; + } /* resize windowLog if input is small enough, to use less memory */ if ( (srcSize < maxWindowResize) @@ -1049,10 +1358,12 @@ ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar, ZSTD_highbit32(tSize-1) + 1; if (cPar.windowLog > srcLog) cPar.windowLog = srcLog; } - if (cPar.hashLog > cPar.windowLog+1) cPar.hashLog = cPar.windowLog+1; - { U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy); - if (cycleLog > cPar.windowLog) - cPar.chainLog -= (cycleLog - cPar.windowLog); + if (srcSize != ZSTD_CONTENTSIZE_UNKNOWN) { + U32 const dictAndWindowLog = ZSTD_dictAndWindowLog(cPar.windowLog, (U64)srcSize, (U64)dictSize); + U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy); + if (cPar.hashLog > dictAndWindowLog+1) cPar.hashLog = dictAndWindowLog+1; + if (cycleLog > dictAndWindowLog) + cPar.chainLog -= (cycleLog - dictAndWindowLog); } if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN) @@ -1068,38 +1379,50 @@ ZSTD_adjustCParams(ZSTD_compressionParameters cPar, { cPar = ZSTD_clampCParams(cPar); /* resulting cPar is necessarily valid (all parameters within range) */ if (srcSize == 0) srcSize = ZSTD_CONTENTSIZE_UNKNOWN; - return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize); + return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize, ZSTD_cpm_unknown); } -static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize); -static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize); +static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode); +static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode); + +static void ZSTD_overrideCParams( + ZSTD_compressionParameters* cParams, + const ZSTD_compressionParameters* overrides) +{ + if (overrides->windowLog) cParams->windowLog = overrides->windowLog; + if (overrides->hashLog) cParams->hashLog = overrides->hashLog; + if (overrides->chainLog) cParams->chainLog = overrides->chainLog; + if (overrides->searchLog) cParams->searchLog = overrides->searchLog; + if (overrides->minMatch) cParams->minMatch = overrides->minMatch; + if (overrides->targetLength) cParams->targetLength = overrides->targetLength; + if (overrides->strategy) cParams->strategy = overrides->strategy; +} ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams( - const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize) + const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) { ZSTD_compressionParameters cParams; if (srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN && CCtxParams->srcSizeHint > 0) { srcSizeHint = CCtxParams->srcSizeHint; } - cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize); - if (CCtxParams->ldmParams.enableLdm) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG; - if (CCtxParams->cParams.windowLog) cParams.windowLog = CCtxParams->cParams.windowLog; - if (CCtxParams->cParams.hashLog) cParams.hashLog = CCtxParams->cParams.hashLog; - if (CCtxParams->cParams.chainLog) cParams.chainLog = CCtxParams->cParams.chainLog; - if (CCtxParams->cParams.searchLog) cParams.searchLog = CCtxParams->cParams.searchLog; - if (CCtxParams->cParams.minMatch) cParams.minMatch = CCtxParams->cParams.minMatch; - if (CCtxParams->cParams.targetLength) cParams.targetLength = CCtxParams->cParams.targetLength; - if (CCtxParams->cParams.strategy) cParams.strategy = CCtxParams->cParams.strategy; + cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize, mode); + if (CCtxParams->ldmParams.enableLdm == ZSTD_ps_enable) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG; + ZSTD_overrideCParams(&cParams, &CCtxParams->cParams); assert(!ZSTD_checkCParams(cParams)); /* srcSizeHint == 0 means 0 */ - return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize); + return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize, mode); } static size_t ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams, + const ZSTD_paramSwitch_e useRowMatchFinder, + const U32 enableDedicatedDictSearch, const U32 forCCtx) { - size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog); + /* chain table size should be 0 for fast or row-hash strategies */ + size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, enableDedicatedDictSearch && !forCCtx) + ? ((size_t)1 << cParams->chainLog) + : 0; size_t const hSize = ((size_t)1) << cParams->hashLog; U32 const hashLog3 = (forCCtx && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0; size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0; @@ -1109,71 +1432,117 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams, + hSize * sizeof(U32) + h3Size * sizeof(U32); size_t const optPotentialSpace = - ZSTD_cwksp_alloc_size((MaxML+1) * sizeof(U32)) - + ZSTD_cwksp_alloc_size((MaxLL+1) * sizeof(U32)) - + ZSTD_cwksp_alloc_size((MaxOff+1) * sizeof(U32)) - + ZSTD_cwksp_alloc_size((1<<Litbits) * sizeof(U32)) - + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t)) - + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t)); + ZSTD_cwksp_aligned_alloc_size((MaxML+1) * sizeof(U32)) + + ZSTD_cwksp_aligned_alloc_size((MaxLL+1) * sizeof(U32)) + + ZSTD_cwksp_aligned_alloc_size((MaxOff+1) * sizeof(U32)) + + ZSTD_cwksp_aligned_alloc_size((1<<Litbits) * sizeof(U32)) + + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t)) + + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t)); + size_t const lazyAdditionalSpace = ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder) + ? ZSTD_cwksp_aligned_alloc_size(hSize*sizeof(U16)) + : 0; size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt)) ? optPotentialSpace : 0; + size_t const slackSpace = ZSTD_cwksp_slack_space_required(); + + /* tables are guaranteed to be sized in multiples of 64 bytes (or 16 uint32_t) */ + ZSTD_STATIC_ASSERT(ZSTD_HASHLOG_MIN >= 4 && ZSTD_WINDOWLOG_MIN >= 4 && ZSTD_CHAINLOG_MIN >= 4); + assert(useRowMatchFinder != ZSTD_ps_auto); + DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u", (U32)chainSize, (U32)hSize, (U32)h3Size); - return tableSpace + optSpace; + return tableSpace + optSpace + slackSpace + lazyAdditionalSpace; +} + +static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal( + const ZSTD_compressionParameters* cParams, + const ldmParams_t* ldmParams, + const int isStatic, + const ZSTD_paramSwitch_e useRowMatchFinder, + const size_t buffInSize, + const size_t buffOutSize, + const U64 pledgedSrcSize) +{ + size_t const windowSize = (size_t) BOUNDED(1ULL, 1ULL << cParams->windowLog, pledgedSrcSize); + size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize); + U32 const divider = (cParams->minMatch==3) ? 3 : 4; + size_t const maxNbSeq = blockSize / divider; + size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize) + + ZSTD_cwksp_aligned_alloc_size(maxNbSeq * sizeof(seqDef)) + + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE)); + size_t const entropySpace = ZSTD_cwksp_alloc_size(ENTROPY_WORKSPACE_SIZE); + size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t)); + size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1); + + size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams); + size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize); + size_t const ldmSeqSpace = ldmParams->enableLdm == ZSTD_ps_enable ? + ZSTD_cwksp_aligned_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0; + + + size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize) + + ZSTD_cwksp_alloc_size(buffOutSize); + + size_t const cctxSpace = isStatic ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0; + + size_t const neededSpace = + cctxSpace + + entropySpace + + blockStateSpace + + ldmSpace + + ldmSeqSpace + + matchStateSize + + tokenSpace + + bufferSpace; + + DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace); + return neededSpace; } size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params) { - RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only."); - { ZSTD_compressionParameters const cParams = - ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0); - size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog); - U32 const divider = (cParams.minMatch==3) ? 3 : 4; - size_t const maxNbSeq = blockSize / divider; - size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize) - + ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(seqDef)) - + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE)); - size_t const entropySpace = ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE); - size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t)); - size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 1); - - size_t const ldmSpace = ZSTD_ldm_getTableSize(params->ldmParams); - size_t const ldmSeqSpace = ZSTD_cwksp_alloc_size(ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize) * sizeof(rawSeq)); - - /* estimateCCtxSize is for one-shot compression. So no buffers should - * be needed. However, we still allocate two 0-sized buffers, which can - * take space under ASAN. */ - size_t const bufferSpace = ZSTD_cwksp_alloc_size(0) - + ZSTD_cwksp_alloc_size(0); - - size_t const cctxSpace = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)); - - size_t const neededSpace = - cctxSpace + - entropySpace + - blockStateSpace + - ldmSpace + - ldmSeqSpace + - matchStateSize + - tokenSpace + - bufferSpace; + ZSTD_compressionParameters const cParams = + ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict); + ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, + &cParams); - DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace); - return neededSpace; - } + RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only."); + /* estimateCCtxSize is for one-shot compression. So no buffers should + * be needed. However, we still allocate two 0-sized buffers, which can + * take space under ASAN. */ + return ZSTD_estimateCCtxSize_usingCCtxParams_internal( + &cParams, ¶ms->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN); } size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams) { - ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams); - return ZSTD_estimateCCtxSize_usingCCtxParams(¶ms); + ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams); + if (ZSTD_rowMatchFinderSupported(cParams.strategy)) { + /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */ + size_t noRowCCtxSize; + size_t rowCCtxSize; + initialParams.useRowMatchFinder = ZSTD_ps_disable; + noRowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + initialParams.useRowMatchFinder = ZSTD_ps_enable; + rowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + return MAX(noRowCCtxSize, rowCCtxSize); + } else { + return ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams); + } } static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel) { - ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0); - return ZSTD_estimateCCtxSize_usingCParams(cParams); + int tier = 0; + size_t largestSize = 0; + static const unsigned long long srcSizeTiers[4] = {16 KB, 128 KB, 256 KB, ZSTD_CONTENTSIZE_UNKNOWN}; + for (; tier < 4; ++tier) { + /* Choose the set of cParams for a given level across all srcSizes that give the largest cctxSize */ + ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeTiers[tier], 0, ZSTD_cpm_noAttachDict); + largestSize = MAX(ZSTD_estimateCCtxSize_usingCParams(cParams), largestSize); + } + return largestSize; } size_t ZSTD_estimateCCtxSize(int compressionLevel) @@ -1181,6 +1550,7 @@ size_t ZSTD_estimateCCtxSize(int compressionLevel) int level; size_t memBudget = 0; for (level=MIN(compressionLevel, 1); level<=compressionLevel; level++) { + /* Ensure monotonically increasing memory usage as compression level increases */ size_t const newMB = ZSTD_estimateCCtxSize_internal(level); if (newMB > memBudget) memBudget = newMB; } @@ -1191,27 +1561,42 @@ size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params) { RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only."); { ZSTD_compressionParameters const cParams = - ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0); - size_t const CCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(params); + ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict); size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog); - size_t const inBuffSize = ((size_t)1 << cParams.windowLog) + blockSize; - size_t const outBuffSize = ZSTD_compressBound(blockSize) + 1; - size_t const streamingSize = ZSTD_cwksp_alloc_size(inBuffSize) - + ZSTD_cwksp_alloc_size(outBuffSize); - - return CCtxSize + streamingSize; + size_t const inBuffSize = (params->inBufferMode == ZSTD_bm_buffered) + ? ((size_t)1 << cParams.windowLog) + blockSize + : 0; + size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered) + ? ZSTD_compressBound(blockSize) + 1 + : 0; + ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, ¶ms->cParams); + + return ZSTD_estimateCCtxSize_usingCCtxParams_internal( + &cParams, ¶ms->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize, + ZSTD_CONTENTSIZE_UNKNOWN); } } size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams) { - ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams); - return ZSTD_estimateCStreamSize_usingCCtxParams(¶ms); + ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams); + if (ZSTD_rowMatchFinderSupported(cParams.strategy)) { + /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */ + size_t noRowCCtxSize; + size_t rowCCtxSize; + initialParams.useRowMatchFinder = ZSTD_ps_disable; + noRowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + initialParams.useRowMatchFinder = ZSTD_ps_enable; + rowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + return MAX(noRowCCtxSize, rowCCtxSize); + } else { + return ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams); + } } static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel) { - ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0); + ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict); return ZSTD_estimateCStreamSize_usingCParams(cParams); } @@ -1305,16 +1690,6 @@ static void ZSTD_invalidateMatchState(ZSTD_matchState_t* ms) } /** - * Indicates whether this compression proceeds directly from user-provided - * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or - * whether the context needs to buffer the input/output (ZSTDb_buffered). - */ -typedef enum { - ZSTDb_not_buffered, - ZSTDb_buffered -} ZSTD_buffered_policy_e; - -/** * Controls, for this matchState reset, whether the tables need to be cleared / * prepared for the coming compression (ZSTDcrp_makeClean), or whether the * tables can be left unclean (ZSTDcrp_leaveDirty), because we know that a @@ -1341,20 +1716,27 @@ typedef enum { ZSTD_resetTarget_CCtx } ZSTD_resetTarget_e; + static size_t ZSTD_reset_matchState(ZSTD_matchState_t* ms, ZSTD_cwksp* ws, const ZSTD_compressionParameters* cParams, + const ZSTD_paramSwitch_e useRowMatchFinder, const ZSTD_compResetPolicy_e crp, const ZSTD_indexResetPolicy_e forceResetIndex, const ZSTD_resetTarget_e forWho) { - size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog); + /* disable chain table allocation for fast or row-based strategies */ + size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, + ms->dedicatedDictSearch && (forWho == ZSTD_resetTarget_CDict)) + ? ((size_t)1 << cParams->chainLog) + : 0; size_t const hSize = ((size_t)1) << cParams->hashLog; U32 const hashLog3 = ((forWho == ZSTD_resetTarget_CCtx) && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0; size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0; DEBUGLOG(4, "reset indices : %u", forceResetIndex == ZSTDirp_reset); + assert(useRowMatchFinder != ZSTD_ps_auto); if (forceResetIndex == ZSTDirp_reset) { ZSTD_window_init(&ms->window); ZSTD_cwksp_mark_tables_dirty(ws); @@ -1393,11 +1775,23 @@ ZSTD_reset_matchState(ZSTD_matchState_t* ms, ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t)); } + if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) { + { /* Row match finder needs an additional table of hashes ("tags") */ + size_t const tagTableSize = hSize*sizeof(U16); + ms->tagTable = (U16*)ZSTD_cwksp_reserve_aligned(ws, tagTableSize); + if (ms->tagTable) ZSTD_memset(ms->tagTable, 0, tagTableSize); + } + { /* Switch to 32-entry rows if searchLog is 5 (or more) */ + U32 const rowLog = BOUNDED(4, cParams->searchLog, 6); + assert(cParams->hashLog >= rowLog); + ms->rowHashLog = cParams->hashLog - rowLog; + } + } + ms->cParams = *cParams; RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation, "failed a workspace allocation in ZSTD_reset_matchState"); - return 0; } @@ -1414,75 +1808,87 @@ static int ZSTD_indexTooCloseToMax(ZSTD_window_t w) return (size_t)(w.nextSrc - w.base) > (ZSTD_CURRENT_MAX - ZSTD_INDEXOVERFLOW_MARGIN); } +/** ZSTD_dictTooBig(): + * When dictionaries are larger than ZSTD_CHUNKSIZE_MAX they can't be loaded in + * one go generically. So we ensure that in that case we reset the tables to zero, + * so that we can load as much of the dictionary as possible. + */ +static int ZSTD_dictTooBig(size_t const loadedDictSize) +{ + return loadedDictSize > ZSTD_CHUNKSIZE_MAX; +} + /*! ZSTD_resetCCtx_internal() : - note : `params` are assumed fully validated at this stage */ + * @param loadedDictSize The size of the dictionary to be loaded + * into the context, if any. If no dictionary is used, or the + * dictionary is being attached / copied, then pass 0. + * note : `params` are assumed fully validated at this stage. + */ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, - ZSTD_CCtx_params params, + ZSTD_CCtx_params const* params, U64 const pledgedSrcSize, + size_t const loadedDictSize, ZSTD_compResetPolicy_e const crp, ZSTD_buffered_policy_e const zbuff) { ZSTD_cwksp* const ws = &zc->workspace; - DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u", - (U32)pledgedSrcSize, params.cParams.windowLog); - assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams))); + DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d useBlockSplitter=%d", + (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder, (int)params->useBlockSplitter); + assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams))); zc->isFirstBlock = 1; - if (params.ldmParams.enableLdm) { + /* Set applied params early so we can modify them for LDM, + * and point params at the applied params. + */ + zc->appliedParams = *params; + params = &zc->appliedParams; + + assert(params->useRowMatchFinder != ZSTD_ps_auto); + assert(params->useBlockSplitter != ZSTD_ps_auto); + assert(params->ldmParams.enableLdm != ZSTD_ps_auto); + if (params->ldmParams.enableLdm == ZSTD_ps_enable) { /* Adjust long distance matching parameters */ - ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams); - assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog); - assert(params.ldmParams.hashRateLog < 32); - zc->ldmState.hashPower = ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength); + ZSTD_ldm_adjustParameters(&zc->appliedParams.ldmParams, ¶ms->cParams); + assert(params->ldmParams.hashLog >= params->ldmParams.bucketSizeLog); + assert(params->ldmParams.hashRateLog < 32); } - { size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params.cParams.windowLog), pledgedSrcSize)); + { size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params->cParams.windowLog), pledgedSrcSize)); size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize); - U32 const divider = (params.cParams.minMatch==3) ? 3 : 4; + U32 const divider = (params->cParams.minMatch==3) ? 3 : 4; size_t const maxNbSeq = blockSize / divider; - size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize) - + ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(seqDef)) - + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE)); - size_t const buffOutSize = (zbuff==ZSTDb_buffered) ? ZSTD_compressBound(blockSize)+1 : 0; - size_t const buffInSize = (zbuff==ZSTDb_buffered) ? windowSize + blockSize : 0; - size_t const matchStateSize = ZSTD_sizeof_matchState(¶ms.cParams, /* forCCtx */ 1); - size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params.ldmParams, blockSize); - - ZSTD_indexResetPolicy_e needsIndexReset = zc->initialized ? ZSTDirp_continue : ZSTDirp_reset; - - if (ZSTD_indexTooCloseToMax(zc->blockState.matchState.window)) { - needsIndexReset = ZSTDirp_reset; - } + size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered) + ? ZSTD_compressBound(blockSize) + 1 + : 0; + size_t const buffInSize = (zbuff == ZSTDb_buffered && params->inBufferMode == ZSTD_bm_buffered) + ? windowSize + blockSize + : 0; + size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize); + + int const indexTooClose = ZSTD_indexTooCloseToMax(zc->blockState.matchState.window); + int const dictTooBig = ZSTD_dictTooBig(loadedDictSize); + ZSTD_indexResetPolicy_e needsIndexReset = + (indexTooClose || dictTooBig || !zc->initialized) ? ZSTDirp_reset : ZSTDirp_continue; - if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0); + size_t const neededSpace = + ZSTD_estimateCCtxSize_usingCCtxParams_internal( + ¶ms->cParams, ¶ms->ldmParams, zc->staticSize != 0, params->useRowMatchFinder, + buffInSize, buffOutSize, pledgedSrcSize); + int resizeWorkspace; + + FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!"); - /* Check if workspace is large enough, alloc a new one if needed */ - { size_t const cctxSpace = zc->staticSize ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0; - size_t const entropySpace = ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE); - size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t)); - size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize) + ZSTD_cwksp_alloc_size(buffOutSize); - size_t const ldmSpace = ZSTD_ldm_getTableSize(params.ldmParams); - size_t const ldmSeqSpace = ZSTD_cwksp_alloc_size(maxNbLdmSeq * sizeof(rawSeq)); - - size_t const neededSpace = - cctxSpace + - entropySpace + - blockStateSpace + - ldmSpace + - ldmSeqSpace + - matchStateSize + - tokenSpace + - bufferSpace; + if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0); + { /* Check if workspace is large enough, alloc a new one if needed */ int const workspaceTooSmall = ZSTD_cwksp_sizeof(ws) < neededSpace; int const workspaceWasteful = ZSTD_cwksp_check_wasteful(ws, neededSpace); - - DEBUGLOG(4, "Need %zuKB workspace, including %zuKB for match state, and %zuKB for buffers", - neededSpace>>10, matchStateSize>>10, bufferSpace>>10); + resizeWorkspace = workspaceTooSmall || workspaceWasteful; + DEBUGLOG(4, "Need %zu B workspace", neededSpace); DEBUGLOG(4, "windowSize: %zu - blockSize: %zu", windowSize, blockSize); - if (workspaceTooSmall || workspaceWasteful) { + if (resizeWorkspace) { DEBUGLOG(4, "Resize workspaceSize from %zuKB to %zuKB", ZSTD_cwksp_sizeof(ws) >> 10, neededSpace >> 10); @@ -1503,15 +1909,14 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, RETURN_ERROR_IF(zc->blockState.prevCBlock == NULL, memory_allocation, "couldn't allocate prevCBlock"); zc->blockState.nextCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t)); RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate nextCBlock"); - zc->entropyWorkspace = (U32*) ZSTD_cwksp_reserve_object(ws, HUF_WORKSPACE_SIZE); - RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate entropyWorkspace"); + zc->entropyWorkspace = (U32*) ZSTD_cwksp_reserve_object(ws, ENTROPY_WORKSPACE_SIZE); + RETURN_ERROR_IF(zc->entropyWorkspace == NULL, memory_allocation, "couldn't allocate entropyWorkspace"); } } ZSTD_cwksp_clear(ws); /* init params */ - zc->appliedParams = params; - zc->blockState.matchState.cParams = params.cParams; + zc->blockState.matchState.cParams = params->cParams; zc->pledgedSrcSizePlusOne = pledgedSrcSize+1; zc->consumedSrcSize = 0; zc->producedCSize = 0; @@ -1524,6 +1929,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, XXH64_reset(&zc->xxhState, 0); zc->stage = ZSTDcs_init; zc->dictID = 0; + zc->dictContentSize = 0; ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock); @@ -1534,19 +1940,20 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, zc->seqStore.maxNbLit = blockSize; /* buffers */ + zc->bufferedPolicy = zbuff; zc->inBuffSize = buffInSize; zc->inBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffInSize); zc->outBuffSize = buffOutSize; zc->outBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffOutSize); /* ldm bucketOffsets table */ - if (params.ldmParams.enableLdm) { + if (params->ldmParams.enableLdm == ZSTD_ps_enable) { /* TODO: avoid memset? */ - size_t const ldmBucketSize = - ((size_t)1) << (params.ldmParams.hashLog - - params.ldmParams.bucketSizeLog); - zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, ldmBucketSize); - memset(zc->ldmState.bucketOffsets, 0, ldmBucketSize); + size_t const numBuckets = + ((size_t)1) << (params->ldmParams.hashLog - + params->ldmParams.bucketSizeLog); + zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, numBuckets); + ZSTD_memset(zc->ldmState.bucketOffsets, 0, numBuckets); } /* sequences storage */ @@ -1560,26 +1967,28 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc, FORWARD_IF_ERROR(ZSTD_reset_matchState( &zc->blockState.matchState, ws, - ¶ms.cParams, + ¶ms->cParams, + params->useRowMatchFinder, crp, needsIndexReset, ZSTD_resetTarget_CCtx), ""); /* ldm hash table */ - if (params.ldmParams.enableLdm) { + if (params->ldmParams.enableLdm == ZSTD_ps_enable) { /* TODO: avoid memset? */ - size_t const ldmHSize = ((size_t)1) << params.ldmParams.hashLog; + size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog; zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned(ws, ldmHSize * sizeof(ldmEntry_t)); - memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t)); + ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t)); zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned(ws, maxNbLdmSeq * sizeof(rawSeq)); zc->maxNbLdmSequences = maxNbLdmSeq; ZSTD_window_init(&zc->ldmState.window); - ZSTD_window_clear(&zc->ldmState.window); zc->ldmState.loadedDictEnd = 0; } DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws)); + assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace, resizeWorkspace)); + zc->initialized = 1; return 0; @@ -1618,12 +2027,14 @@ static int ZSTD_shouldAttachDict(const ZSTD_CDict* cdict, U64 pledgedSrcSize) { size_t cutoff = attachDictSizeCutoffs[cdict->matchState.cParams.strategy]; - return ( pledgedSrcSize <= cutoff - || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN - || params->attachDictPref == ZSTD_dictForceAttach ) - && params->attachDictPref != ZSTD_dictForceCopy - && !params->forceWindow; /* dictMatchState isn't correctly - * handled in _enforceMaxDist */ + int const dedicatedDictSearch = cdict->matchState.dedicatedDictSearch; + return dedicatedDictSearch + || ( ( pledgedSrcSize <= cutoff + || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN + || params->attachDictPref == ZSTD_dictForceAttach ) + && params->attachDictPref != ZSTD_dictForceCopy + && !params->forceWindow ); /* dictMatchState isn't correctly + * handled in _enforceMaxDist */ } static size_t @@ -1633,17 +2044,28 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx, U64 pledgedSrcSize, ZSTD_buffered_policy_e zbuff) { - { const ZSTD_compressionParameters* const cdict_cParams = &cdict->matchState.cParams; + DEBUGLOG(4, "ZSTD_resetCCtx_byAttachingCDict() pledgedSrcSize=%llu", + (unsigned long long)pledgedSrcSize); + { + ZSTD_compressionParameters adjusted_cdict_cParams = cdict->matchState.cParams; unsigned const windowLog = params.cParams.windowLog; assert(windowLog != 0); /* Resize working context table params for input only, since the dict * has its own tables. */ - /* pledgeSrcSize == 0 means 0! */ - params.cParams = ZSTD_adjustCParams_internal(*cdict_cParams, pledgedSrcSize, 0); + /* pledgedSrcSize == 0 means 0! */ + + if (cdict->matchState.dedicatedDictSearch) { + ZSTD_dedicatedDictSearch_revertCParams(&adjusted_cdict_cParams); + } + + params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize, + cdict->dictContentSize, ZSTD_cpm_attachDict); params.cParams.windowLog = windowLog; - FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize, + params.useRowMatchFinder = cdict->useRowMatchFinder; /* cdict overrides */ + FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, ¶ms, pledgedSrcSize, + /* loadedDictSize */ 0, ZSTDcrp_makeClean, zbuff), ""); - assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy); + assert(cctx->appliedParams.cParams.strategy == adjusted_cdict_cParams.strategy); } { const U32 cdictEnd = (U32)( cdict->matchState.window.nextSrc @@ -1668,9 +2090,10 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx, } } cctx->dictID = cdict->dictID; + cctx->dictContentSize = cdict->dictContentSize; /* copy block state */ - memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState)); + ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState)); return 0; } @@ -1683,14 +2106,18 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, { const ZSTD_compressionParameters *cdict_cParams = &cdict->matchState.cParams; - DEBUGLOG(4, "copying dictionary into context"); + assert(!cdict->matchState.dedicatedDictSearch); + DEBUGLOG(4, "ZSTD_resetCCtx_byCopyingCDict() pledgedSrcSize=%llu", + (unsigned long long)pledgedSrcSize); { unsigned const windowLog = params.cParams.windowLog; assert(windowLog != 0); /* Copy only compression parameters related to tables. */ params.cParams = *cdict_cParams; params.cParams.windowLog = windowLog; - FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize, + params.useRowMatchFinder = cdict->useRowMatchFinder; + FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, ¶ms, pledgedSrcSize, + /* loadedDictSize */ 0, ZSTDcrp_leaveDirty, zbuff), ""); assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy); assert(cctx->appliedParams.cParams.hashLog == cdict_cParams->hashLog); @@ -1698,24 +2125,37 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, } ZSTD_cwksp_mark_tables_dirty(&cctx->workspace); + assert(params.useRowMatchFinder != ZSTD_ps_auto); /* copy tables */ - { size_t const chainSize = (cdict_cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cdict_cParams->chainLog); + { size_t const chainSize = ZSTD_allocateChainTable(cdict_cParams->strategy, cdict->useRowMatchFinder, 0 /* DDS guaranteed disabled */) + ? ((size_t)1 << cdict_cParams->chainLog) + : 0; size_t const hSize = (size_t)1 << cdict_cParams->hashLog; - memcpy(cctx->blockState.matchState.hashTable, + ZSTD_memcpy(cctx->blockState.matchState.hashTable, cdict->matchState.hashTable, hSize * sizeof(U32)); - memcpy(cctx->blockState.matchState.chainTable, + /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */ + if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) { + ZSTD_memcpy(cctx->blockState.matchState.chainTable, cdict->matchState.chainTable, chainSize * sizeof(U32)); + } + /* copy tag table */ + if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) { + size_t const tagTableSize = hSize*sizeof(U16); + ZSTD_memcpy(cctx->blockState.matchState.tagTable, + cdict->matchState.tagTable, + tagTableSize); + } } /* Zero the hashTable3, since the cdict never fills it */ { int const h3log = cctx->blockState.matchState.hashLog3; size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0; assert(cdict->matchState.hashLog3 == 0); - memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32)); + ZSTD_memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32)); } ZSTD_cwksp_mark_tables_clean(&cctx->workspace); @@ -1729,9 +2169,10 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, } cctx->dictID = cdict->dictID; + cctx->dictContentSize = cdict->dictContentSize; /* copy block state */ - memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState)); + ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState)); return 0; } @@ -1771,16 +2212,22 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, U64 pledgedSrcSize, ZSTD_buffered_policy_e zbuff) { - DEBUGLOG(5, "ZSTD_copyCCtx_internal"); RETURN_ERROR_IF(srcCCtx->stage!=ZSTDcs_init, stage_wrong, "Can't copy a ctx that's not in init stage."); - - memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem)); + DEBUGLOG(5, "ZSTD_copyCCtx_internal"); + ZSTD_memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem)); { ZSTD_CCtx_params params = dstCCtx->requestedParams; /* Copy only compression parameters related to tables. */ params.cParams = srcCCtx->appliedParams.cParams; + assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_ps_auto); + assert(srcCCtx->appliedParams.useBlockSplitter != ZSTD_ps_auto); + assert(srcCCtx->appliedParams.ldmParams.enableLdm != ZSTD_ps_auto); + params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder; + params.useBlockSplitter = srcCCtx->appliedParams.useBlockSplitter; + params.ldmParams = srcCCtx->appliedParams.ldmParams; params.fParams = fParams; - ZSTD_resetCCtx_internal(dstCCtx, params, pledgedSrcSize, + ZSTD_resetCCtx_internal(dstCCtx, ¶ms, pledgedSrcSize, + /* loadedDictSize */ 0, ZSTDcrp_leaveDirty, zbuff); assert(dstCCtx->appliedParams.cParams.windowLog == srcCCtx->appliedParams.cParams.windowLog); assert(dstCCtx->appliedParams.cParams.strategy == srcCCtx->appliedParams.cParams.strategy); @@ -1792,18 +2239,22 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, ZSTD_cwksp_mark_tables_dirty(&dstCCtx->workspace); /* copy tables */ - { size_t const chainSize = (srcCCtx->appliedParams.cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog); + { size_t const chainSize = ZSTD_allocateChainTable(srcCCtx->appliedParams.cParams.strategy, + srcCCtx->appliedParams.useRowMatchFinder, + 0 /* forDDSDict */) + ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog) + : 0; size_t const hSize = (size_t)1 << srcCCtx->appliedParams.cParams.hashLog; int const h3log = srcCCtx->blockState.matchState.hashLog3; size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0; - memcpy(dstCCtx->blockState.matchState.hashTable, + ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable, srcCCtx->blockState.matchState.hashTable, hSize * sizeof(U32)); - memcpy(dstCCtx->blockState.matchState.chainTable, + ZSTD_memcpy(dstCCtx->blockState.matchState.chainTable, srcCCtx->blockState.matchState.chainTable, chainSize * sizeof(U32)); - memcpy(dstCCtx->blockState.matchState.hashTable3, + ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable3, srcCCtx->blockState.matchState.hashTable3, h3Size * sizeof(U32)); } @@ -1819,9 +2270,10 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd; } dstCCtx->dictID = srcCCtx->dictID; + dstCCtx->dictContentSize = srcCCtx->dictContentSize; /* copy block state */ - memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock)); + ZSTD_memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock)); return 0; } @@ -1834,7 +2286,7 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx, size_t ZSTD_copyCCtx(ZSTD_CCtx* dstCCtx, const ZSTD_CCtx* srcCCtx, unsigned long long pledgedSrcSize) { ZSTD_frameParameters fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ }; - ZSTD_buffered_policy_e const zbuff = (ZSTD_buffered_policy_e)(srcCCtx->inBuffSize>0); + ZSTD_buffered_policy_e const zbuff = srcCCtx->bufferedPolicy; ZSTD_STATIC_ASSERT((U32)ZSTDb_buffered==1); if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN; fParams.contentSizeFlag = (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN); @@ -1858,10 +2310,12 @@ ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerVa int const nbRows = (int)size / ZSTD_ROWSIZE; int cellNb = 0; int rowNb; + /* Protect special index values < ZSTD_WINDOW_START_INDEX. */ + U32 const reducerThreshold = reducerValue + ZSTD_WINDOW_START_INDEX; assert((size & (ZSTD_ROWSIZE-1)) == 0); /* multiple of ZSTD_ROWSIZE */ assert(size < (1U<<31)); /* can be casted to int */ -#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) +#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) /* To validate that the table re-use logic is sound, and that we don't * access table space that we haven't cleaned, we re-"poison" the table * space every time we mark it dirty. @@ -1877,12 +2331,17 @@ ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerVa for (rowNb=0 ; rowNb < nbRows ; rowNb++) { int column; for (column=0; column<ZSTD_ROWSIZE; column++) { - if (preserveMark) { - U32 const adder = (table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) ? reducerValue : 0; - table[cellNb] += adder; + U32 newVal; + if (preserveMark && table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) { + /* This write is pointless, but is required(?) for the compiler + * to auto-vectorize the loop. */ + newVal = ZSTD_DUBT_UNSORTED_MARK; + } else if (table[cellNb] < reducerThreshold) { + newVal = 0; + } else { + newVal = table[cellNb] - reducerValue; } - if (table[cellNb] < reducerValue) table[cellNb] = 0; - else table[cellNb] -= reducerValue; + table[cellNb] = newVal; cellNb++; } } } @@ -1905,7 +2364,7 @@ static void ZSTD_reduceIndex (ZSTD_matchState_t* ms, ZSTD_CCtx_params const* par ZSTD_reduceTable(ms->hashTable, hSize, reducerValue); } - if (params->cParams.strategy != ZSTD_fast) { + if (ZSTD_allocateChainTable(params->cParams.strategy, params->useRowMatchFinder, (U32)ms->dedicatedDictSearch)) { U32 const chainSize = (U32)1 << params->cParams.chainLog; if (params->cParams.strategy == ZSTD_btlazy2) ZSTD_reduceTable_btlazy2(ms->chainTable, chainSize, reducerValue); @@ -1937,14 +2396,14 @@ void ZSTD_seqToCodes(const seqStore_t* seqStorePtr) assert(nbSeq <= seqStorePtr->maxNbSeq); for (u=0; u<nbSeq; u++) { U32 const llv = sequences[u].litLength; - U32 const mlv = sequences[u].matchLength; + U32 const mlv = sequences[u].mlBase; llCodeTable[u] = (BYTE)ZSTD_LLcode(llv); - ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offset); + ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offBase); mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv); } - if (seqStorePtr->longLengthID==1) + if (seqStorePtr->longLengthType==ZSTD_llt_literalLength) llCodeTable[seqStorePtr->longLengthPos] = MaxLL; - if (seqStorePtr->longLengthID==2) + if (seqStorePtr->longLengthType==ZSTD_llt_matchLength) mlCodeTable[seqStorePtr->longLengthPos] = MaxML; } @@ -1958,10 +2417,161 @@ static int ZSTD_useTargetCBlockSize(const ZSTD_CCtx_params* cctxParams) return (cctxParams->targetCBlockSize != 0); } -/* ZSTD_compressSequences_internal(): - * actually compresses both literals and sequences */ +/* ZSTD_blockSplitterEnabled(): + * Returns if block splitting param is being used + * If used, compression will do best effort to split a block in order to improve compression ratio. + * At the time this function is called, the parameter must be finalized. + * Returns 1 if true, 0 otherwise. */ +static int ZSTD_blockSplitterEnabled(ZSTD_CCtx_params* cctxParams) +{ + DEBUGLOG(5, "ZSTD_blockSplitterEnabled (useBlockSplitter=%d)", cctxParams->useBlockSplitter); + assert(cctxParams->useBlockSplitter != ZSTD_ps_auto); + return (cctxParams->useBlockSplitter == ZSTD_ps_enable); +} + +/* Type returned by ZSTD_buildSequencesStatistics containing finalized symbol encoding types + * and size of the sequences statistics + */ +typedef struct { + U32 LLtype; + U32 Offtype; + U32 MLtype; + size_t size; + size_t lastCountSize; /* Accounts for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */ +} ZSTD_symbolEncodingTypeStats_t; + +/* ZSTD_buildSequencesStatistics(): + * Returns a ZSTD_symbolEncodingTypeStats_t, or a zstd error code in the `size` field. + * Modifies `nextEntropy` to have the appropriate values as a side effect. + * nbSeq must be greater than 0. + * + * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32) + */ +static ZSTD_symbolEncodingTypeStats_t +ZSTD_buildSequencesStatistics(seqStore_t* seqStorePtr, size_t nbSeq, + const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy, + BYTE* dst, const BYTE* const dstEnd, + ZSTD_strategy strategy, unsigned* countWorkspace, + void* entropyWorkspace, size_t entropyWkspSize) { + BYTE* const ostart = dst; + const BYTE* const oend = dstEnd; + BYTE* op = ostart; + FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable; + FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable; + FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable; + const BYTE* const ofCodeTable = seqStorePtr->ofCode; + const BYTE* const llCodeTable = seqStorePtr->llCode; + const BYTE* const mlCodeTable = seqStorePtr->mlCode; + ZSTD_symbolEncodingTypeStats_t stats; + + stats.lastCountSize = 0; + /* convert length/distances into codes */ + ZSTD_seqToCodes(seqStorePtr); + assert(op <= oend); + assert(nbSeq != 0); /* ZSTD_selectEncodingType() divides by nbSeq */ + /* build CTable for Literal Lengths */ + { unsigned max = MaxLL; + size_t const mostFrequent = HIST_countFast_wksp(countWorkspace, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ + DEBUGLOG(5, "Building LL table"); + nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode; + stats.LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode, + countWorkspace, max, mostFrequent, nbSeq, + LLFSELog, prevEntropy->litlengthCTable, + LL_defaultNorm, LL_defaultNormLog, + ZSTD_defaultAllowed, strategy); + assert(set_basic < set_compressed && set_rle < set_compressed); + assert(!(stats.LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ + { size_t const countSize = ZSTD_buildCTable( + op, (size_t)(oend - op), + CTable_LitLength, LLFSELog, (symbolEncodingType_e)stats.LLtype, + countWorkspace, max, llCodeTable, nbSeq, + LL_defaultNorm, LL_defaultNormLog, MaxLL, + prevEntropy->litlengthCTable, + sizeof(prevEntropy->litlengthCTable), + entropyWorkspace, entropyWkspSize); + if (ZSTD_isError(countSize)) { + DEBUGLOG(3, "ZSTD_buildCTable for LitLens failed"); + stats.size = countSize; + return stats; + } + if (stats.LLtype == set_compressed) + stats.lastCountSize = countSize; + op += countSize; + assert(op <= oend); + } } + /* build CTable for Offsets */ + { unsigned max = MaxOff; + size_t const mostFrequent = HIST_countFast_wksp( + countWorkspace, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ + /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */ + ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed; + DEBUGLOG(5, "Building OF table"); + nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode; + stats.Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode, + countWorkspace, max, mostFrequent, nbSeq, + OffFSELog, prevEntropy->offcodeCTable, + OF_defaultNorm, OF_defaultNormLog, + defaultPolicy, strategy); + assert(!(stats.Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */ + { size_t const countSize = ZSTD_buildCTable( + op, (size_t)(oend - op), + CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)stats.Offtype, + countWorkspace, max, ofCodeTable, nbSeq, + OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff, + prevEntropy->offcodeCTable, + sizeof(prevEntropy->offcodeCTable), + entropyWorkspace, entropyWkspSize); + if (ZSTD_isError(countSize)) { + DEBUGLOG(3, "ZSTD_buildCTable for Offsets failed"); + stats.size = countSize; + return stats; + } + if (stats.Offtype == set_compressed) + stats.lastCountSize = countSize; + op += countSize; + assert(op <= oend); + } } + /* build CTable for MatchLengths */ + { unsigned max = MaxML; + size_t const mostFrequent = HIST_countFast_wksp( + countWorkspace, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ + DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op)); + nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode; + stats.MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode, + countWorkspace, max, mostFrequent, nbSeq, + MLFSELog, prevEntropy->matchlengthCTable, + ML_defaultNorm, ML_defaultNormLog, + ZSTD_defaultAllowed, strategy); + assert(!(stats.MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ + { size_t const countSize = ZSTD_buildCTable( + op, (size_t)(oend - op), + CTable_MatchLength, MLFSELog, (symbolEncodingType_e)stats.MLtype, + countWorkspace, max, mlCodeTable, nbSeq, + ML_defaultNorm, ML_defaultNormLog, MaxML, + prevEntropy->matchlengthCTable, + sizeof(prevEntropy->matchlengthCTable), + entropyWorkspace, entropyWkspSize); + if (ZSTD_isError(countSize)) { + DEBUGLOG(3, "ZSTD_buildCTable for MatchLengths failed"); + stats.size = countSize; + return stats; + } + if (stats.MLtype == set_compressed) + stats.lastCountSize = countSize; + op += countSize; + assert(op <= oend); + } } + stats.size = (size_t)(op-ostart); + return stats; +} + +/* ZSTD_entropyCompressSeqStore_internal(): + * compresses both literals and sequences + * Returns compressed size of block, or a zstd error. + */ +#define SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO 20 MEM_STATIC size_t -ZSTD_compressSequences_internal(seqStore_t* seqStorePtr, +ZSTD_entropyCompressSeqStore_internal(seqStore_t* seqStorePtr, const ZSTD_entropyCTables_t* prevEntropy, ZSTD_entropyCTables_t* nextEntropy, const ZSTD_CCtx_params* cctxParams, @@ -1971,36 +2581,42 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr, { const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN; ZSTD_strategy const strategy = cctxParams->cParams.strategy; - unsigned count[MaxSeq+1]; + unsigned* count = (unsigned*)entropyWorkspace; FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable; FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable; FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable; - U32 LLtype, Offtype, MLtype; /* compressed, raw or rle */ const seqDef* const sequences = seqStorePtr->sequencesStart; + const size_t nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart; const BYTE* const ofCodeTable = seqStorePtr->ofCode; const BYTE* const llCodeTable = seqStorePtr->llCode; const BYTE* const mlCodeTable = seqStorePtr->mlCode; BYTE* const ostart = (BYTE*)dst; BYTE* const oend = ostart + dstCapacity; BYTE* op = ostart; - size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart); - BYTE* seqHead; - BYTE* lastNCount = NULL; + size_t lastCountSize; + + entropyWorkspace = count + (MaxSeq + 1); + entropyWkspSize -= (MaxSeq + 1) * sizeof(*count); - DEBUGLOG(5, "ZSTD_compressSequences_internal (nbSeq=%zu)", nbSeq); + DEBUGLOG(4, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu)", nbSeq); ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog))); + assert(entropyWkspSize >= HUF_WORKSPACE_SIZE); /* Compress literals */ { const BYTE* const literals = seqStorePtr->litStart; + size_t const numSequences = seqStorePtr->sequences - seqStorePtr->sequencesStart; + size_t const numLiterals = seqStorePtr->lit - seqStorePtr->litStart; + /* Base suspicion of uncompressibility on ratio of literals to sequences */ + unsigned const suspectUncompressible = (numSequences == 0) || (numLiterals / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO); size_t const litSize = (size_t)(seqStorePtr->lit - literals); size_t const cSize = ZSTD_compressLiterals( &prevEntropy->huf, &nextEntropy->huf, cctxParams->cParams.strategy, - ZSTD_disableLiteralsCompression(cctxParams), + ZSTD_literalsCompressionIsDisabled(cctxParams), op, dstCapacity, literals, litSize, entropyWorkspace, entropyWkspSize, - bmi2); + bmi2, suspectUncompressible); FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed"); assert(cSize <= dstCapacity); op += cSize; @@ -2023,98 +2639,23 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr, assert(op <= oend); if (nbSeq==0) { /* Copy the old tables over as if we repeated them */ - memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse)); + ZSTD_memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse)); return (size_t)(op - ostart); } - - /* seqHead : flags for FSE encoding type */ - seqHead = op++; - assert(op <= oend); - - /* convert length/distances into codes */ - ZSTD_seqToCodes(seqStorePtr); - /* build CTable for Literal Lengths */ - { unsigned max = MaxLL; - size_t const mostFrequent = HIST_countFast_wksp(count, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ - DEBUGLOG(5, "Building LL table"); - nextEntropy->fse.litlength_repeatMode = prevEntropy->fse.litlength_repeatMode; - LLtype = ZSTD_selectEncodingType(&nextEntropy->fse.litlength_repeatMode, - count, max, mostFrequent, nbSeq, - LLFSELog, prevEntropy->fse.litlengthCTable, - LL_defaultNorm, LL_defaultNormLog, - ZSTD_defaultAllowed, strategy); - assert(set_basic < set_compressed && set_rle < set_compressed); - assert(!(LLtype < set_compressed && nextEntropy->fse.litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable( - op, (size_t)(oend - op), - CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype, - count, max, llCodeTable, nbSeq, - LL_defaultNorm, LL_defaultNormLog, MaxLL, - prevEntropy->fse.litlengthCTable, - sizeof(prevEntropy->fse.litlengthCTable), - entropyWorkspace, entropyWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed"); - if (LLtype == set_compressed) - lastNCount = op; - op += countSize; - assert(op <= oend); - } } - /* build CTable for Offsets */ - { unsigned max = MaxOff; - size_t const mostFrequent = HIST_countFast_wksp( - count, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ - /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */ - ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed; - DEBUGLOG(5, "Building OF table"); - nextEntropy->fse.offcode_repeatMode = prevEntropy->fse.offcode_repeatMode; - Offtype = ZSTD_selectEncodingType(&nextEntropy->fse.offcode_repeatMode, - count, max, mostFrequent, nbSeq, - OffFSELog, prevEntropy->fse.offcodeCTable, - OF_defaultNorm, OF_defaultNormLog, - defaultPolicy, strategy); - assert(!(Offtype < set_compressed && nextEntropy->fse.offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable( - op, (size_t)(oend - op), - CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype, - count, max, ofCodeTable, nbSeq, - OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff, - prevEntropy->fse.offcodeCTable, - sizeof(prevEntropy->fse.offcodeCTable), - entropyWorkspace, entropyWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed"); - if (Offtype == set_compressed) - lastNCount = op; - op += countSize; - assert(op <= oend); - } } - /* build CTable for MatchLengths */ - { unsigned max = MaxML; - size_t const mostFrequent = HIST_countFast_wksp( - count, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize); /* can't fail */ - DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op)); - nextEntropy->fse.matchlength_repeatMode = prevEntropy->fse.matchlength_repeatMode; - MLtype = ZSTD_selectEncodingType(&nextEntropy->fse.matchlength_repeatMode, - count, max, mostFrequent, nbSeq, - MLFSELog, prevEntropy->fse.matchlengthCTable, - ML_defaultNorm, ML_defaultNormLog, - ZSTD_defaultAllowed, strategy); - assert(!(MLtype < set_compressed && nextEntropy->fse.matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable( - op, (size_t)(oend - op), - CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype, - count, max, mlCodeTable, nbSeq, - ML_defaultNorm, ML_defaultNormLog, MaxML, - prevEntropy->fse.matchlengthCTable, - sizeof(prevEntropy->fse.matchlengthCTable), - entropyWorkspace, entropyWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed"); - if (MLtype == set_compressed) - lastNCount = op; - op += countSize; - assert(op <= oend); - } } - - *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2)); + { + ZSTD_symbolEncodingTypeStats_t stats; + BYTE* seqHead = op++; + /* build stats for sequences */ + stats = ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq, + &prevEntropy->fse, &nextEntropy->fse, + op, oend, + strategy, count, + entropyWorkspace, entropyWkspSize); + FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!"); + *seqHead = (BYTE)((stats.LLtype<<6) + (stats.Offtype<<4) + (stats.MLtype<<2)); + lastCountSize = stats.lastCountSize; + op += stats.size; + } { size_t const bitstreamSize = ZSTD_encodeSequences( op, (size_t)(oend - op), @@ -2134,9 +2675,9 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr, * In this exceedingly rare case, we will simply emit an uncompressed * block, since it isn't worth optimizing. */ - if (lastNCount && (op - lastNCount) < 4) { - /* NCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */ - assert(op - lastNCount == 3); + if (lastCountSize && (lastCountSize + bitstreamSize) < 4) { + /* lastCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */ + assert(lastCountSize + bitstreamSize == 3); DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.3.4 by " "emitting an uncompressed block."); return 0; @@ -2148,7 +2689,7 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr, } MEM_STATIC size_t -ZSTD_compressSequences(seqStore_t* seqStorePtr, +ZSTD_entropyCompressSeqStore(seqStore_t* seqStorePtr, const ZSTD_entropyCTables_t* prevEntropy, ZSTD_entropyCTables_t* nextEntropy, const ZSTD_CCtx_params* cctxParams, @@ -2157,7 +2698,7 @@ ZSTD_compressSequences(seqStore_t* seqStorePtr, void* entropyWorkspace, size_t entropyWkspSize, int bmi2) { - size_t const cSize = ZSTD_compressSequences_internal( + size_t const cSize = ZSTD_entropyCompressSeqStore_internal( seqStorePtr, prevEntropy, nextEntropy, cctxParams, dst, dstCapacity, entropyWorkspace, entropyWkspSize, bmi2); @@ -2167,22 +2708,22 @@ ZSTD_compressSequences(seqStore_t* seqStorePtr, */ if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity)) return 0; /* block not compressed */ - FORWARD_IF_ERROR(cSize, "ZSTD_compressSequences_internal failed"); + FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSeqStore_internal failed"); /* Check compressibility */ { size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, cctxParams->cParams.strategy); if (cSize >= maxCSize) return 0; /* block not compressed */ } - + DEBUGLOG(4, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize); return cSize; } /* ZSTD_selectBlockCompressor() : * Not static, but internal use only (used by long distance matcher) * assumption : strat is a valid strategy */ -ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode) +ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e useRowMatchFinder, ZSTD_dictMode_e dictMode) { - static const ZSTD_blockCompressor blockCompressor[3][ZSTD_STRATEGY_MAX+1] = { + static const ZSTD_blockCompressor blockCompressor[4][ZSTD_STRATEGY_MAX+1] = { { ZSTD_compressBlock_fast /* default for 0 */, ZSTD_compressBlock_fast, ZSTD_compressBlock_doubleFast, @@ -2212,13 +2753,44 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo ZSTD_compressBlock_btlazy2_dictMatchState, ZSTD_compressBlock_btopt_dictMatchState, ZSTD_compressBlock_btultra_dictMatchState, - ZSTD_compressBlock_btultra_dictMatchState } + ZSTD_compressBlock_btultra_dictMatchState }, + { NULL /* default for 0 */, + NULL, + NULL, + ZSTD_compressBlock_greedy_dedicatedDictSearch, + ZSTD_compressBlock_lazy_dedicatedDictSearch, + ZSTD_compressBlock_lazy2_dedicatedDictSearch, + NULL, + NULL, + NULL, + NULL } }; ZSTD_blockCompressor selectedCompressor; ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1); assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat)); - selectedCompressor = blockCompressor[(int)dictMode][(int)strat]; + DEBUGLOG(4, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder); + if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) { + static const ZSTD_blockCompressor rowBasedBlockCompressors[4][3] = { + { ZSTD_compressBlock_greedy_row, + ZSTD_compressBlock_lazy_row, + ZSTD_compressBlock_lazy2_row }, + { ZSTD_compressBlock_greedy_extDict_row, + ZSTD_compressBlock_lazy_extDict_row, + ZSTD_compressBlock_lazy2_extDict_row }, + { ZSTD_compressBlock_greedy_dictMatchState_row, + ZSTD_compressBlock_lazy_dictMatchState_row, + ZSTD_compressBlock_lazy2_dictMatchState_row }, + { ZSTD_compressBlock_greedy_dedicatedDictSearch_row, + ZSTD_compressBlock_lazy_dedicatedDictSearch_row, + ZSTD_compressBlock_lazy2_dedicatedDictSearch_row } + }; + DEBUGLOG(4, "Selecting a row-based matchfinder"); + assert(useRowMatchFinder != ZSTD_ps_auto); + selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy]; + } else { + selectedCompressor = blockCompressor[(int)dictMode][(int)strat]; + } assert(selectedCompressor != NULL); return selectedCompressor; } @@ -2226,7 +2798,7 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo static void ZSTD_storeLastLiterals(seqStore_t* seqStorePtr, const BYTE* anchor, size_t lastLLSize) { - memcpy(seqStorePtr->lit, anchor, lastLLSize); + ZSTD_memcpy(seqStorePtr->lit, anchor, lastLLSize); seqStorePtr->lit += lastLLSize; } @@ -2234,7 +2806,7 @@ void ZSTD_resetSeqStore(seqStore_t* ssPtr) { ssPtr->lit = ssPtr->litStart; ssPtr->sequences = ssPtr->sequencesStart; - ssPtr->longLengthID = 0; + ssPtr->longLengthType = ZSTD_llt_none; } typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_buildSeqStore_e; @@ -2247,7 +2819,11 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) /* Assert that we have correctly flushed the ctx params into the ms's copy */ ZSTD_assertEqualCParams(zc->appliedParams.cParams, ms->cParams); if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) { - ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.minMatch); + if (zc->appliedParams.cParams.strategy >= ZSTD_btopt) { + ZSTD_ldm_skipRawSeqStoreBytes(&zc->externSeqStore, srcSize); + } else { + ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.minMatch); + } return ZSTDbss_noCompress; /* don't even attempt compression below a certain srcSize */ } ZSTD_resetSeqStore(&(zc->seqStore)); @@ -2263,10 +2839,10 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) /* limited update after a very long match */ { const BYTE* const base = ms->window.base; const BYTE* const istart = (const BYTE*)src; - const U32 current = (U32)(istart-base); + const U32 curr = (U32)(istart-base); if (sizeof(ptrdiff_t)==8) assert(istart - base < (ptrdiff_t)(U32)(-1)); /* ensure no overflow */ - if (current > ms->nextToUpdate + 384) - ms->nextToUpdate = current - MIN(192, (U32)(current - ms->nextToUpdate - 384)); + if (curr > ms->nextToUpdate + 384) + ms->nextToUpdate = curr - MIN(192, (U32)(curr - ms->nextToUpdate - 384)); } /* select and store sequences */ @@ -2277,16 +2853,17 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) zc->blockState.nextCBlock->rep[i] = zc->blockState.prevCBlock->rep[i]; } if (zc->externSeqStore.pos < zc->externSeqStore.size) { - assert(!zc->appliedParams.ldmParams.enableLdm); + assert(zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_disable); /* Updates ldmSeqStore.pos */ lastLLSize = ZSTD_ldm_blockCompress(&zc->externSeqStore, ms, &zc->seqStore, zc->blockState.nextCBlock->rep, + zc->appliedParams.useRowMatchFinder, src, srcSize); assert(zc->externSeqStore.pos <= zc->externSeqStore.size); - } else if (zc->appliedParams.ldmParams.enableLdm) { - rawSeqStore_t ldmSeqStore = {NULL, 0, 0, 0}; + } else if (zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) { + rawSeqStore_t ldmSeqStore = kNullRawSeqStore; ldmSeqStore.seq = zc->ldmSequences; ldmSeqStore.capacity = zc->maxNbLdmSequences; @@ -2299,10 +2876,14 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) ZSTD_ldm_blockCompress(&ldmSeqStore, ms, &zc->seqStore, zc->blockState.nextCBlock->rep, + zc->appliedParams.useRowMatchFinder, src, srcSize); assert(ldmSeqStore.pos == ldmSeqStore.size); } else { /* not long range mode */ - ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, dictMode); + ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, + zc->appliedParams.useRowMatchFinder, + dictMode); + ms->ldmSeqStore = NULL; lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize); } { const BYTE* const lastLiterals = (const BYTE*)src + srcSize - lastLLSize; @@ -2314,59 +2895,72 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize) static void ZSTD_copyBlockSequences(ZSTD_CCtx* zc) { const seqStore_t* seqStore = ZSTD_getSeqStore(zc); - const seqDef* seqs = seqStore->sequencesStart; - size_t seqsSize = seqStore->sequences - seqs; + const seqDef* seqStoreSeqs = seqStore->sequencesStart; + size_t seqStoreSeqSize = seqStore->sequences - seqStoreSeqs; + size_t seqStoreLiteralsSize = (size_t)(seqStore->lit - seqStore->litStart); + size_t literalsRead = 0; + size_t lastLLSize; ZSTD_Sequence* outSeqs = &zc->seqCollector.seqStart[zc->seqCollector.seqIndex]; - size_t i; size_t position; int repIdx; + size_t i; + repcodes_t updatedRepcodes; assert(zc->seqCollector.seqIndex + 1 < zc->seqCollector.maxSequences); - for (i = 0, position = 0; i < seqsSize; ++i) { - outSeqs[i].offset = seqs[i].offset; - outSeqs[i].litLength = seqs[i].litLength; - outSeqs[i].matchLength = seqs[i].matchLength + MINMATCH; + /* Ensure we have enough space for last literals "sequence" */ + assert(zc->seqCollector.maxSequences >= seqStoreSeqSize + 1); + ZSTD_memcpy(updatedRepcodes.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t)); + for (i = 0; i < seqStoreSeqSize; ++i) { + U32 rawOffset = seqStoreSeqs[i].offBase - ZSTD_REP_NUM; + outSeqs[i].litLength = seqStoreSeqs[i].litLength; + outSeqs[i].matchLength = seqStoreSeqs[i].mlBase + MINMATCH; + outSeqs[i].rep = 0; if (i == seqStore->longLengthPos) { - if (seqStore->longLengthID == 1) { + if (seqStore->longLengthType == ZSTD_llt_literalLength) { outSeqs[i].litLength += 0x10000; - } else if (seqStore->longLengthID == 2) { + } else if (seqStore->longLengthType == ZSTD_llt_matchLength) { outSeqs[i].matchLength += 0x10000; } } - if (outSeqs[i].offset <= ZSTD_REP_NUM) { - outSeqs[i].rep = outSeqs[i].offset; - repIdx = (unsigned int)i - outSeqs[i].offset; - - if (outSeqs[i].litLength == 0) { - if (outSeqs[i].offset < 3) { - --repIdx; + if (seqStoreSeqs[i].offBase <= ZSTD_REP_NUM) { + /* Derive the correct offset corresponding to a repcode */ + outSeqs[i].rep = seqStoreSeqs[i].offBase; + if (outSeqs[i].litLength != 0) { + rawOffset = updatedRepcodes.rep[outSeqs[i].rep - 1]; + } else { + if (outSeqs[i].rep == 3) { + rawOffset = updatedRepcodes.rep[0] - 1; } else { - repIdx = (unsigned int)i - 1; + rawOffset = updatedRepcodes.rep[outSeqs[i].rep]; } - ++outSeqs[i].rep; } - assert(repIdx >= -3); - outSeqs[i].offset = repIdx >= 0 ? outSeqs[repIdx].offset : repStartValue[-repIdx - 1]; - if (outSeqs[i].rep == 4) { - --outSeqs[i].offset; - } - } else { - outSeqs[i].offset -= ZSTD_REP_NUM; } - - position += outSeqs[i].litLength; - outSeqs[i].matchPos = (unsigned int)position; - position += outSeqs[i].matchLength; + outSeqs[i].offset = rawOffset; + /* seqStoreSeqs[i].offset == offCode+1, and ZSTD_updateRep() expects offCode + so we provide seqStoreSeqs[i].offset - 1 */ + ZSTD_updateRep(updatedRepcodes.rep, + seqStoreSeqs[i].offBase - 1, + seqStoreSeqs[i].litLength == 0); + literalsRead += outSeqs[i].litLength; } - zc->seqCollector.seqIndex += seqsSize; + /* Insert last literals (if any exist) in the block as a sequence with ml == off == 0. + * If there are no last literals, then we'll emit (of: 0, ml: 0, ll: 0), which is a marker + * for the block boundary, according to the API. + */ + assert(seqStoreLiteralsSize >= literalsRead); + lastLLSize = seqStoreLiteralsSize - literalsRead; + outSeqs[i].litLength = (U32)lastLLSize; + outSeqs[i].matchLength = outSeqs[i].offset = outSeqs[i].rep = 0; + seqStoreSeqSize++; + zc->seqCollector.seqIndex += seqStoreSeqSize; } -size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs, - size_t outSeqsSize, const void* src, size_t srcSize) +size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs, + size_t outSeqsSize, const void* src, size_t srcSize) { const size_t dstCapacity = ZSTD_compressBound(srcSize); - void* dst = ZSTD_malloc(dstCapacity, ZSTD_defaultCMem); + void* dst = ZSTD_customMalloc(dstCapacity, ZSTD_defaultCMem); SeqCollector seqCollector; RETURN_ERROR_IF(dst == NULL, memory_allocation, "NULL pointer!"); @@ -2378,16 +2972,47 @@ size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs, zc->seqCollector = seqCollector; ZSTD_compress2(zc, dst, dstCapacity, src, srcSize); - ZSTD_free(dst, ZSTD_defaultCMem); + ZSTD_customFree(dst, ZSTD_defaultCMem); return zc->seqCollector.seqIndex; } -/* Returns true if the given block is a RLE block */ -static int ZSTD_isRLE(const BYTE *ip, size_t length) { +size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize) { + size_t in = 0; + size_t out = 0; + for (; in < seqsSize; ++in) { + if (sequences[in].offset == 0 && sequences[in].matchLength == 0) { + if (in != seqsSize - 1) { + sequences[in+1].litLength += sequences[in].litLength; + } + } else { + sequences[out] = sequences[in]; + ++out; + } + } + return out; +} + +/* Unrolled loop to read four size_ts of input at a time. Returns 1 if is RLE, 0 if not. */ +static int ZSTD_isRLE(const BYTE* src, size_t length) { + const BYTE* ip = src; + const BYTE value = ip[0]; + const size_t valueST = (size_t)((U64)value * 0x0101010101010101ULL); + const size_t unrollSize = sizeof(size_t) * 4; + const size_t unrollMask = unrollSize - 1; + const size_t prefixLength = length & unrollMask; size_t i; - if (length < 2) return 1; - for (i = 1; i < length; ++i) { - if (ip[0] != ip[i]) return 0; + size_t u; + if (length == 1) return 1; + /* Check if prefix is RLE first before using unrolled loop */ + if (prefixLength && ZSTD_count(ip+1, ip, ip+prefixLength) != prefixLength-1) { + return 0; + } + for (i = prefixLength; i != length; i += unrollSize) { + for (u = 0; u < unrollSize; u += sizeof(size_t)) { + if (MEM_readST(ip + i + u) != valueST) { + return 0; + } + } } return 1; } @@ -2404,16 +3029,740 @@ static int ZSTD_maybeRLE(seqStore_t const* seqStore) return nbSeqs < 4 && nbLits < 10; } -static void ZSTD_confirmRepcodesAndEntropyTables(ZSTD_CCtx* zc) +static void ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs) +{ + ZSTD_compressedBlockState_t* const tmp = bs->prevCBlock; + bs->prevCBlock = bs->nextCBlock; + bs->nextCBlock = tmp; +} + +/* Writes the block header */ +static void writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock) { + U32 const cBlockHeader = cSize == 1 ? + lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) : + lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3); + MEM_writeLE24(op, cBlockHeader); + DEBUGLOG(3, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock); +} + +/** ZSTD_buildBlockEntropyStats_literals() : + * Builds entropy for the literals. + * Stores literals block type (raw, rle, compressed, repeat) and + * huffman description table to hufMetadata. + * Requires ENTROPY_WORKSPACE_SIZE workspace + * @return : size of huffman description table or error code */ +static size_t ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize, + const ZSTD_hufCTables_t* prevHuf, + ZSTD_hufCTables_t* nextHuf, + ZSTD_hufCTablesMetadata_t* hufMetadata, + const int literalsCompressionIsDisabled, + void* workspace, size_t wkspSize) +{ + BYTE* const wkspStart = (BYTE*)workspace; + BYTE* const wkspEnd = wkspStart + wkspSize; + BYTE* const countWkspStart = wkspStart; + unsigned* const countWksp = (unsigned*)workspace; + const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned); + BYTE* const nodeWksp = countWkspStart + countWkspSize; + const size_t nodeWkspSize = wkspEnd-nodeWksp; + unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX; + unsigned huffLog = HUF_TABLELOG_DEFAULT; + HUF_repeat repeat = prevHuf->repeatMode; + DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_literals (srcSize=%zu)", srcSize); + + /* Prepare nextEntropy assuming reusing the existing table */ + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + + if (literalsCompressionIsDisabled) { + DEBUGLOG(5, "set_basic - disabled"); + hufMetadata->hType = set_basic; + return 0; + } + + /* small ? don't even attempt compression (speed opt) */ +#ifndef COMPRESS_LITERALS_SIZE_MIN +#define COMPRESS_LITERALS_SIZE_MIN 63 +#endif + { size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN; + if (srcSize <= minLitSize) { + DEBUGLOG(5, "set_basic - too small"); + hufMetadata->hType = set_basic; + return 0; + } + } + + /* Scan input and build symbol stats */ + { size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize); + FORWARD_IF_ERROR(largest, "HIST_count_wksp failed"); + if (largest == srcSize) { + DEBUGLOG(5, "set_rle"); + hufMetadata->hType = set_rle; + return 0; + } + if (largest <= (srcSize >> 7)+4) { + DEBUGLOG(5, "set_basic - no gain"); + hufMetadata->hType = set_basic; + return 0; + } + } + + /* Validate the previous Huffman table */ + if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) { + repeat = HUF_repeat_none; + } + + /* Build Huffman Tree */ + ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable)); + huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue); + { size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp, + maxSymbolValue, huffLog, + nodeWksp, nodeWkspSize); + FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp"); + huffLog = (U32)maxBits; + { /* Build and write the CTable */ + size_t const newCSize = HUF_estimateCompressedSize( + (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue); + size_t const hSize = HUF_writeCTable_wksp( + hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer), + (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog, + nodeWksp, nodeWkspSize); + /* Check against repeating the previous CTable */ + if (repeat != HUF_repeat_none) { + size_t const oldCSize = HUF_estimateCompressedSize( + (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue); + if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) { + DEBUGLOG(5, "set_repeat - smaller"); + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + hufMetadata->hType = set_repeat; + return 0; + } + } + if (newCSize + hSize >= srcSize) { + DEBUGLOG(5, "set_basic - no gains"); + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + hufMetadata->hType = set_basic; + return 0; + } + DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize); + hufMetadata->hType = set_compressed; + nextHuf->repeatMode = HUF_repeat_check; + return hSize; + } + } +} + + +/* ZSTD_buildDummySequencesStatistics(): + * Returns a ZSTD_symbolEncodingTypeStats_t with all encoding types as set_basic, + * and updates nextEntropy to the appropriate repeatMode. + */ +static ZSTD_symbolEncodingTypeStats_t +ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy) { + ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0}; + nextEntropy->litlength_repeatMode = FSE_repeat_none; + nextEntropy->offcode_repeatMode = FSE_repeat_none; + nextEntropy->matchlength_repeatMode = FSE_repeat_none; + return stats; +} + +/** ZSTD_buildBlockEntropyStats_sequences() : + * Builds entropy for the sequences. + * Stores symbol compression modes and fse table to fseMetadata. + * Requires ENTROPY_WORKSPACE_SIZE wksp. + * @return : size of fse tables or error code */ +static size_t ZSTD_buildBlockEntropyStats_sequences(seqStore_t* seqStorePtr, + const ZSTD_fseCTables_t* prevEntropy, + ZSTD_fseCTables_t* nextEntropy, + const ZSTD_CCtx_params* cctxParams, + ZSTD_fseCTablesMetadata_t* fseMetadata, + void* workspace, size_t wkspSize) +{ + ZSTD_strategy const strategy = cctxParams->cParams.strategy; + size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart; + BYTE* const ostart = fseMetadata->fseTablesBuffer; + BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer); + BYTE* op = ostart; + unsigned* countWorkspace = (unsigned*)workspace; + unsigned* entropyWorkspace = countWorkspace + (MaxSeq + 1); + size_t entropyWorkspaceSize = wkspSize - (MaxSeq + 1) * sizeof(*countWorkspace); + ZSTD_symbolEncodingTypeStats_t stats; + + DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_sequences (nbSeq=%zu)", nbSeq); + stats = nbSeq != 0 ? ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq, + prevEntropy, nextEntropy, op, oend, + strategy, countWorkspace, + entropyWorkspace, entropyWorkspaceSize) + : ZSTD_buildDummySequencesStatistics(nextEntropy); + FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!"); + fseMetadata->llType = (symbolEncodingType_e) stats.LLtype; + fseMetadata->ofType = (symbolEncodingType_e) stats.Offtype; + fseMetadata->mlType = (symbolEncodingType_e) stats.MLtype; + fseMetadata->lastCountSize = stats.lastCountSize; + return stats.size; +} + + +/** ZSTD_buildBlockEntropyStats() : + * Builds entropy for the block. + * Requires workspace size ENTROPY_WORKSPACE_SIZE + * + * @return : 0 on success or error code + */ +size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr, + const ZSTD_entropyCTables_t* prevEntropy, + ZSTD_entropyCTables_t* nextEntropy, + const ZSTD_CCtx_params* cctxParams, + ZSTD_entropyCTablesMetadata_t* entropyMetadata, + void* workspace, size_t wkspSize) +{ + size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart; + entropyMetadata->hufMetadata.hufDesSize = + ZSTD_buildBlockEntropyStats_literals(seqStorePtr->litStart, litSize, + &prevEntropy->huf, &nextEntropy->huf, + &entropyMetadata->hufMetadata, + ZSTD_literalsCompressionIsDisabled(cctxParams), + workspace, wkspSize); + FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildBlockEntropyStats_literals failed"); + entropyMetadata->fseMetadata.fseTablesSize = + ZSTD_buildBlockEntropyStats_sequences(seqStorePtr, + &prevEntropy->fse, &nextEntropy->fse, + cctxParams, + &entropyMetadata->fseMetadata, + workspace, wkspSize); + FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildBlockEntropyStats_sequences failed"); + return 0; +} + +/* Returns the size estimate for the literals section (header + content) of a block */ +static size_t ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize, + const ZSTD_hufCTables_t* huf, + const ZSTD_hufCTablesMetadata_t* hufMetadata, + void* workspace, size_t wkspSize, + int writeEntropy) +{ + unsigned* const countWksp = (unsigned*)workspace; + unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX; + size_t literalSectionHeaderSize = 3 + (litSize >= 1 KB) + (litSize >= 16 KB); + U32 singleStream = litSize < 256; + + if (hufMetadata->hType == set_basic) return litSize; + else if (hufMetadata->hType == set_rle) return 1; + else if (hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat) { + size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)literals, litSize, workspace, wkspSize); + if (ZSTD_isError(largest)) return litSize; + { size_t cLitSizeEstimate = HUF_estimateCompressedSize((const HUF_CElt*)huf->CTable, countWksp, maxSymbolValue); + if (writeEntropy) cLitSizeEstimate += hufMetadata->hufDesSize; + if (!singleStream) cLitSizeEstimate += 6; /* multi-stream huffman uses 6-byte jump table */ + return cLitSizeEstimate + literalSectionHeaderSize; + } } + assert(0); /* impossible */ + return 0; +} + +/* Returns the size estimate for the FSE-compressed symbols (of, ml, ll) of a block */ +static size_t ZSTD_estimateBlockSize_symbolType(symbolEncodingType_e type, + const BYTE* codeTable, size_t nbSeq, unsigned maxCode, + const FSE_CTable* fseCTable, + const U8* additionalBits, + short const* defaultNorm, U32 defaultNormLog, U32 defaultMax, + void* workspace, size_t wkspSize) +{ + unsigned* const countWksp = (unsigned*)workspace; + const BYTE* ctp = codeTable; + const BYTE* const ctStart = ctp; + const BYTE* const ctEnd = ctStart + nbSeq; + size_t cSymbolTypeSizeEstimateInBits = 0; + unsigned max = maxCode; + + HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize); /* can't fail */ + if (type == set_basic) { + /* We selected this encoding type, so it must be valid. */ + assert(max <= defaultMax); + (void)defaultMax; + cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max); + } else if (type == set_rle) { + cSymbolTypeSizeEstimateInBits = 0; + } else if (type == set_compressed || type == set_repeat) { + cSymbolTypeSizeEstimateInBits = ZSTD_fseBitCost(fseCTable, countWksp, max); + } + if (ZSTD_isError(cSymbolTypeSizeEstimateInBits)) { + return nbSeq * 10; + } + while (ctp < ctEnd) { + if (additionalBits) cSymbolTypeSizeEstimateInBits += additionalBits[*ctp]; + else cSymbolTypeSizeEstimateInBits += *ctp; /* for offset, offset code is also the number of additional bits */ + ctp++; + } + return cSymbolTypeSizeEstimateInBits >> 3; +} + +/* Returns the size estimate for the sequences section (header + content) of a block */ +static size_t ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable, + const BYTE* llCodeTable, + const BYTE* mlCodeTable, + size_t nbSeq, + const ZSTD_fseCTables_t* fseTables, + const ZSTD_fseCTablesMetadata_t* fseMetadata, + void* workspace, size_t wkspSize, + int writeEntropy) +{ + size_t sequencesSectionHeaderSize = 1 /* seqHead */ + 1 /* min seqSize size */ + (nbSeq >= 128) + (nbSeq >= LONGNBSEQ); + size_t cSeqSizeEstimate = 0; + cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, nbSeq, MaxOff, + fseTables->offcodeCTable, NULL, + OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff, + workspace, wkspSize); + cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->llType, llCodeTable, nbSeq, MaxLL, + fseTables->litlengthCTable, LL_bits, + LL_defaultNorm, LL_defaultNormLog, MaxLL, + workspace, wkspSize); + cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, nbSeq, MaxML, + fseTables->matchlengthCTable, ML_bits, + ML_defaultNorm, ML_defaultNormLog, MaxML, + workspace, wkspSize); + if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize; + return cSeqSizeEstimate + sequencesSectionHeaderSize; +} + +/* Returns the size estimate for a given stream of literals, of, ll, ml */ +static size_t ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize, + const BYTE* ofCodeTable, + const BYTE* llCodeTable, + const BYTE* mlCodeTable, + size_t nbSeq, + const ZSTD_entropyCTables_t* entropy, + const ZSTD_entropyCTablesMetadata_t* entropyMetadata, + void* workspace, size_t wkspSize, + int writeLitEntropy, int writeSeqEntropy) { + size_t const literalsSize = ZSTD_estimateBlockSize_literal(literals, litSize, + &entropy->huf, &entropyMetadata->hufMetadata, + workspace, wkspSize, writeLitEntropy); + size_t const seqSize = ZSTD_estimateBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable, + nbSeq, &entropy->fse, &entropyMetadata->fseMetadata, + workspace, wkspSize, writeSeqEntropy); + return seqSize + literalsSize + ZSTD_blockHeaderSize; +} + +/* Builds entropy statistics and uses them for blocksize estimation. + * + * Returns the estimated compressed size of the seqStore, or a zstd error. + */ +static size_t ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(seqStore_t* seqStore, ZSTD_CCtx* zc) { + ZSTD_entropyCTablesMetadata_t* entropyMetadata = &zc->blockSplitCtx.entropyMetadata; + DEBUGLOG(6, "ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize()"); + FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(seqStore, + &zc->blockState.prevCBlock->entropy, + &zc->blockState.nextCBlock->entropy, + &zc->appliedParams, + entropyMetadata, + zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), ""); + return ZSTD_estimateBlockSize(seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart), + seqStore->ofCode, seqStore->llCode, seqStore->mlCode, + (size_t)(seqStore->sequences - seqStore->sequencesStart), + &zc->blockState.nextCBlock->entropy, entropyMetadata, zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE, + (int)(entropyMetadata->hufMetadata.hType == set_compressed), 1); +} + +/* Returns literals bytes represented in a seqStore */ +static size_t ZSTD_countSeqStoreLiteralsBytes(const seqStore_t* const seqStore) { + size_t literalsBytes = 0; + size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart; + size_t i; + for (i = 0; i < nbSeqs; ++i) { + seqDef seq = seqStore->sequencesStart[i]; + literalsBytes += seq.litLength; + if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_literalLength) { + literalsBytes += 0x10000; + } + } + return literalsBytes; +} + +/* Returns match bytes represented in a seqStore */ +static size_t ZSTD_countSeqStoreMatchBytes(const seqStore_t* const seqStore) { + size_t matchBytes = 0; + size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart; + size_t i; + for (i = 0; i < nbSeqs; ++i) { + seqDef seq = seqStore->sequencesStart[i]; + matchBytes += seq.mlBase + MINMATCH; + if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_matchLength) { + matchBytes += 0x10000; + } + } + return matchBytes; +} + +/* Derives the seqStore that is a chunk of the originalSeqStore from [startIdx, endIdx). + * Stores the result in resultSeqStore. + */ +static void ZSTD_deriveSeqStoreChunk(seqStore_t* resultSeqStore, + const seqStore_t* originalSeqStore, + size_t startIdx, size_t endIdx) { + BYTE* const litEnd = originalSeqStore->lit; + size_t literalsBytes; + size_t literalsBytesPreceding = 0; + + *resultSeqStore = *originalSeqStore; + if (startIdx > 0) { + resultSeqStore->sequences = originalSeqStore->sequencesStart + startIdx; + literalsBytesPreceding = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore); + } + + /* Move longLengthPos into the correct position if necessary */ + if (originalSeqStore->longLengthType != ZSTD_llt_none) { + if (originalSeqStore->longLengthPos < startIdx || originalSeqStore->longLengthPos > endIdx) { + resultSeqStore->longLengthType = ZSTD_llt_none; + } else { + resultSeqStore->longLengthPos -= (U32)startIdx; + } + } + resultSeqStore->sequencesStart = originalSeqStore->sequencesStart + startIdx; + resultSeqStore->sequences = originalSeqStore->sequencesStart + endIdx; + literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore); + resultSeqStore->litStart += literalsBytesPreceding; + if (endIdx == (size_t)(originalSeqStore->sequences - originalSeqStore->sequencesStart)) { + /* This accounts for possible last literals if the derived chunk reaches the end of the block */ + resultSeqStore->lit = litEnd; + } else { + resultSeqStore->lit = resultSeqStore->litStart+literalsBytes; + } + resultSeqStore->llCode += startIdx; + resultSeqStore->mlCode += startIdx; + resultSeqStore->ofCode += startIdx; +} + +/** + * Returns the raw offset represented by the combination of offCode, ll0, and repcode history. + * offCode must represent a repcode in the numeric representation of ZSTD_storeSeq(). + */ +static U32 +ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offCode, const U32 ll0) +{ + U32 const adjustedOffCode = STORED_REPCODE(offCode) - 1 + ll0; /* [ 0 - 3 ] */ + assert(STORED_IS_REPCODE(offCode)); + if (adjustedOffCode == ZSTD_REP_NUM) { + /* litlength == 0 and offCode == 2 implies selection of first repcode - 1 */ + assert(rep[0] > 0); + return rep[0] - 1; + } + return rep[adjustedOffCode]; +} + +/** + * ZSTD_seqStore_resolveOffCodes() reconciles any possible divergences in offset history that may arise + * due to emission of RLE/raw blocks that disturb the offset history, + * and replaces any repcodes within the seqStore that may be invalid. + * + * dRepcodes are updated as would be on the decompression side. + * cRepcodes are updated exactly in accordance with the seqStore. + * + * Note : this function assumes seq->offBase respects the following numbering scheme : + * 0 : invalid + * 1-3 : repcode 1-3 + * 4+ : real_offset+3 + */ +static void ZSTD_seqStore_resolveOffCodes(repcodes_t* const dRepcodes, repcodes_t* const cRepcodes, + seqStore_t* const seqStore, U32 const nbSeq) { + U32 idx = 0; + for (; idx < nbSeq; ++idx) { + seqDef* const seq = seqStore->sequencesStart + idx; + U32 const ll0 = (seq->litLength == 0); + U32 const offCode = OFFBASE_TO_STORED(seq->offBase); + assert(seq->offBase > 0); + if (STORED_IS_REPCODE(offCode)) { + U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offCode, ll0); + U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offCode, ll0); + /* Adjust simulated decompression repcode history if we come across a mismatch. Replace + * the repcode with the offset it actually references, determined by the compression + * repcode history. + */ + if (dRawOffset != cRawOffset) { + seq->offBase = cRawOffset + ZSTD_REP_NUM; + } + } + /* Compression repcode history is always updated with values directly from the unmodified seqStore. + * Decompression repcode history may use modified seq->offset value taken from compression repcode history. + */ + ZSTD_updateRep(dRepcodes->rep, OFFBASE_TO_STORED(seq->offBase), ll0); + ZSTD_updateRep(cRepcodes->rep, offCode, ll0); + } +} + +/* ZSTD_compressSeqStore_singleBlock(): + * Compresses a seqStore into a block with a block header, into the buffer dst. + * + * Returns the total size of that block (including header) or a ZSTD error code. + */ +static size_t +ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc, seqStore_t* const seqStore, + repcodes_t* const dRep, repcodes_t* const cRep, + void* dst, size_t dstCapacity, + const void* src, size_t srcSize, + U32 lastBlock, U32 isPartition) +{ + const U32 rleMaxLength = 25; + BYTE* op = (BYTE*)dst; + const BYTE* ip = (const BYTE*)src; + size_t cSize; + size_t cSeqsSize; + + /* In case of an RLE or raw block, the simulated decompression repcode history must be reset */ + repcodes_t const dRepOriginal = *dRep; + DEBUGLOG(5, "ZSTD_compressSeqStore_singleBlock"); + if (isPartition) + ZSTD_seqStore_resolveOffCodes(dRep, cRep, seqStore, (U32)(seqStore->sequences - seqStore->sequencesStart)); + + RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "Block header doesn't fit"); + cSeqsSize = ZSTD_entropyCompressSeqStore(seqStore, + &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy, + &zc->appliedParams, + op + ZSTD_blockHeaderSize, dstCapacity - ZSTD_blockHeaderSize, + srcSize, + zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */, + zc->bmi2); + FORWARD_IF_ERROR(cSeqsSize, "ZSTD_entropyCompressSeqStore failed!"); + + if (!zc->isFirstBlock && + cSeqsSize < rleMaxLength && + ZSTD_isRLE((BYTE const*)src, srcSize)) { + /* We don't want to emit our first block as a RLE even if it qualifies because + * doing so will cause the decoder (cli only) to throw a "should consume all input error." + * This is only an issue for zstd <= v1.4.3 + */ + cSeqsSize = 1; + } + + if (zc->seqCollector.collectSequences) { + ZSTD_copyBlockSequences(zc); + ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState); + return 0; + } + + if (cSeqsSize == 0) { + cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock); + FORWARD_IF_ERROR(cSize, "Nocompress block failed"); + DEBUGLOG(4, "Writing out nocompress block, size: %zu", cSize); + *dRep = dRepOriginal; /* reset simulated decompression repcode history */ + } else if (cSeqsSize == 1) { + cSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, srcSize, lastBlock); + FORWARD_IF_ERROR(cSize, "RLE compress block failed"); + DEBUGLOG(4, "Writing out RLE block, size: %zu", cSize); + *dRep = dRepOriginal; /* reset simulated decompression repcode history */ + } else { + ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState); + writeBlockHeader(op, cSeqsSize, srcSize, lastBlock); + cSize = ZSTD_blockHeaderSize + cSeqsSize; + DEBUGLOG(4, "Writing out compressed block, size: %zu", cSize); + } + + if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid) + zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check; + + return cSize; +} + +/* Struct to keep track of where we are in our recursive calls. */ +typedef struct { + U32* splitLocations; /* Array of split indices */ + size_t idx; /* The current index within splitLocations being worked on */ +} seqStoreSplits; + +#define MIN_SEQUENCES_BLOCK_SPLITTING 300 + +/* Helper function to perform the recursive search for block splits. + * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half. + * If advantageous to split, then we recurse down the two sub-blocks. If not, or if an error occurred in estimation, then + * we do not recurse. + * + * Note: The recursion depth is capped by a heuristic minimum number of sequences, defined by MIN_SEQUENCES_BLOCK_SPLITTING. + * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING). + * In practice, recursion depth usually doesn't go beyond 4. + * + * Furthermore, the number of splits is capped by ZSTD_MAX_NB_BLOCK_SPLITS. At ZSTD_MAX_NB_BLOCK_SPLITS == 196 with the current existing blockSize + * maximum of 128 KB, this value is actually impossible to reach. + */ +static void +ZSTD_deriveBlockSplitsHelper(seqStoreSplits* splits, size_t startIdx, size_t endIdx, + ZSTD_CCtx* zc, const seqStore_t* origSeqStore) +{ + seqStore_t* fullSeqStoreChunk = &zc->blockSplitCtx.fullSeqStoreChunk; + seqStore_t* firstHalfSeqStore = &zc->blockSplitCtx.firstHalfSeqStore; + seqStore_t* secondHalfSeqStore = &zc->blockSplitCtx.secondHalfSeqStore; + size_t estimatedOriginalSize; + size_t estimatedFirstHalfSize; + size_t estimatedSecondHalfSize; + size_t midIdx = (startIdx + endIdx)/2; + + if (endIdx - startIdx < MIN_SEQUENCES_BLOCK_SPLITTING || splits->idx >= ZSTD_MAX_NB_BLOCK_SPLITS) { + DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences"); + return; + } + DEBUGLOG(4, "ZSTD_deriveBlockSplitsHelper: startIdx=%zu endIdx=%zu", startIdx, endIdx); + ZSTD_deriveSeqStoreChunk(fullSeqStoreChunk, origSeqStore, startIdx, endIdx); + ZSTD_deriveSeqStoreChunk(firstHalfSeqStore, origSeqStore, startIdx, midIdx); + ZSTD_deriveSeqStoreChunk(secondHalfSeqStore, origSeqStore, midIdx, endIdx); + estimatedOriginalSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(fullSeqStoreChunk, zc); + estimatedFirstHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(firstHalfSeqStore, zc); + estimatedSecondHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(secondHalfSeqStore, zc); + DEBUGLOG(4, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu", + estimatedOriginalSize, estimatedFirstHalfSize, estimatedSecondHalfSize); + if (ZSTD_isError(estimatedOriginalSize) || ZSTD_isError(estimatedFirstHalfSize) || ZSTD_isError(estimatedSecondHalfSize)) { + return; + } + if (estimatedFirstHalfSize + estimatedSecondHalfSize < estimatedOriginalSize) { + ZSTD_deriveBlockSplitsHelper(splits, startIdx, midIdx, zc, origSeqStore); + splits->splitLocations[splits->idx] = (U32)midIdx; + splits->idx++; + ZSTD_deriveBlockSplitsHelper(splits, midIdx, endIdx, zc, origSeqStore); + } +} + +/* Base recursive function. Populates a table with intra-block partition indices that can improve compression ratio. + * + * Returns the number of splits made (which equals the size of the partition table - 1). + */ +static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq) { + seqStoreSplits splits = {partitions, 0}; + if (nbSeq <= 4) { + DEBUGLOG(4, "ZSTD_deriveBlockSplits: Too few sequences to split"); + /* Refuse to try and split anything with less than 4 sequences */ + return 0; + } + ZSTD_deriveBlockSplitsHelper(&splits, 0, nbSeq, zc, &zc->seqStore); + splits.splitLocations[splits.idx] = nbSeq; + DEBUGLOG(5, "ZSTD_deriveBlockSplits: final nb partitions: %zu", splits.idx+1); + return splits.idx; +} + +/* ZSTD_compressBlock_splitBlock(): + * Attempts to split a given block into multiple blocks to improve compression ratio. + * + * Returns combined size of all blocks (which includes headers), or a ZSTD error code. + */ +static size_t +ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc, void* dst, size_t dstCapacity, + const void* src, size_t blockSize, U32 lastBlock, U32 nbSeq) +{ + size_t cSize = 0; + const BYTE* ip = (const BYTE*)src; + BYTE* op = (BYTE*)dst; + size_t i = 0; + size_t srcBytesTotal = 0; + U32* partitions = zc->blockSplitCtx.partitions; /* size == ZSTD_MAX_NB_BLOCK_SPLITS */ + seqStore_t* nextSeqStore = &zc->blockSplitCtx.nextSeqStore; + seqStore_t* currSeqStore = &zc->blockSplitCtx.currSeqStore; + size_t numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq); + + /* If a block is split and some partitions are emitted as RLE/uncompressed, then repcode history + * may become invalid. In order to reconcile potentially invalid repcodes, we keep track of two + * separate repcode histories that simulate repcode history on compression and decompression side, + * and use the histories to determine whether we must replace a particular repcode with its raw offset. + * + * 1) cRep gets updated for each partition, regardless of whether the block was emitted as uncompressed + * or RLE. This allows us to retrieve the offset value that an invalid repcode references within + * a nocompress/RLE block. + * 2) dRep gets updated only for compressed partitions, and when a repcode gets replaced, will use + * the replacement offset value rather than the original repcode to update the repcode history. + * dRep also will be the final repcode history sent to the next block. + * + * See ZSTD_seqStore_resolveOffCodes() for more details. + */ + repcodes_t dRep; + repcodes_t cRep; + ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t)); + ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t)); + ZSTD_memset(nextSeqStore, 0, sizeof(seqStore_t)); + + DEBUGLOG(4, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)", + (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit, + (unsigned)zc->blockState.matchState.nextToUpdate); + + if (numSplits == 0) { + size_t cSizeSingleBlock = ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore, + &dRep, &cRep, + op, dstCapacity, + ip, blockSize, + lastBlock, 0 /* isPartition */); + FORWARD_IF_ERROR(cSizeSingleBlock, "Compressing single block from splitBlock_internal() failed!"); + DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal: No splits"); + assert(cSizeSingleBlock <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize); + return cSizeSingleBlock; + } + + ZSTD_deriveSeqStoreChunk(currSeqStore, &zc->seqStore, 0, partitions[0]); + for (i = 0; i <= numSplits; ++i) { + size_t srcBytes; + size_t cSizeChunk; + U32 const lastPartition = (i == numSplits); + U32 lastBlockEntireSrc = 0; + + srcBytes = ZSTD_countSeqStoreLiteralsBytes(currSeqStore) + ZSTD_countSeqStoreMatchBytes(currSeqStore); + srcBytesTotal += srcBytes; + if (lastPartition) { + /* This is the final partition, need to account for possible last literals */ + srcBytes += blockSize - srcBytesTotal; + lastBlockEntireSrc = lastBlock; + } else { + ZSTD_deriveSeqStoreChunk(nextSeqStore, &zc->seqStore, partitions[i], partitions[i+1]); + } + + cSizeChunk = ZSTD_compressSeqStore_singleBlock(zc, currSeqStore, + &dRep, &cRep, + op, dstCapacity, + ip, srcBytes, + lastBlockEntireSrc, 1 /* isPartition */); + DEBUGLOG(5, "Estimated size: %zu actual size: %zu", ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(currSeqStore, zc), cSizeChunk); + FORWARD_IF_ERROR(cSizeChunk, "Compressing chunk failed!"); + + ip += srcBytes; + op += cSizeChunk; + dstCapacity -= cSizeChunk; + cSize += cSizeChunk; + *currSeqStore = *nextSeqStore; + assert(cSizeChunk <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize); + } + /* cRep and dRep may have diverged during the compression. If so, we use the dRep repcodes + * for the next block. + */ + ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(repcodes_t)); + return cSize; +} + +static size_t +ZSTD_compressBlock_splitBlock(ZSTD_CCtx* zc, + void* dst, size_t dstCapacity, + const void* src, size_t srcSize, U32 lastBlock) { - ZSTD_compressedBlockState_t* const tmp = zc->blockState.prevCBlock; - zc->blockState.prevCBlock = zc->blockState.nextCBlock; - zc->blockState.nextCBlock = tmp; + const BYTE* ip = (const BYTE*)src; + BYTE* op = (BYTE*)dst; + U32 nbSeq; + size_t cSize; + DEBUGLOG(4, "ZSTD_compressBlock_splitBlock"); + assert(zc->appliedParams.useBlockSplitter == ZSTD_ps_enable); + + { const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize); + FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed"); + if (bss == ZSTDbss_noCompress) { + if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid) + zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check; + cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock); + FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed"); + DEBUGLOG(4, "ZSTD_compressBlock_splitBlock: Nocompress block"); + return cSize; + } + nbSeq = (U32)(zc->seqStore.sequences - zc->seqStore.sequencesStart); + } + + cSize = ZSTD_compressBlock_splitBlock_internal(zc, dst, dstCapacity, src, srcSize, lastBlock, nbSeq); + FORWARD_IF_ERROR(cSize, "Splitting blocks failed!"); + return cSize; } -static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, U32 frame) +static size_t +ZSTD_compressBlock_internal(ZSTD_CCtx* zc, + void* dst, size_t dstCapacity, + const void* src, size_t srcSize, U32 frame) { /* This the upper bound for the length of an rle block. * This isn't the actual upper bound. Finding the real threshold @@ -2434,16 +3783,17 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc, if (zc->seqCollector.collectSequences) { ZSTD_copyBlockSequences(zc); + ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState); return 0; } /* encode sequences and literals */ - cSize = ZSTD_compressSequences(&zc->seqStore, + cSize = ZSTD_entropyCompressSeqStore(&zc->seqStore, &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy, &zc->appliedParams, dst, dstCapacity, srcSize, - zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */, + zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */, zc->bmi2); if (frame && @@ -2461,7 +3811,7 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc, out: if (!ZSTD_isError(cSize) && cSize > 1) { - ZSTD_confirmRepcodesAndEntropyTables(zc); + ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState); } /* We check that dictionaries have offset codes available for the first * block. After the first block, the offcode table might not have large @@ -2514,7 +3864,7 @@ static size_t ZSTD_compressBlock_targetCBlockSize_body(ZSTD_CCtx* zc, size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy); FORWARD_IF_ERROR(cSize, "ZSTD_compressSuperBlock failed"); if (cSize != 0 && cSize < maxCSize + ZSTD_blockHeaderSize) { - ZSTD_confirmRepcodesAndEntropyTables(zc); + ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState); return cSize; } } @@ -2554,9 +3904,9 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms, void const* ip, void const* iend) { - if (ZSTD_window_needOverflowCorrection(ms->window, iend)) { - U32 const maxDist = (U32)1 << params->cParams.windowLog; - U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy); + U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy); + U32 const maxDist = (U32)1 << params->cParams.windowLog; + if (ZSTD_window_needOverflowCorrection(ms->window, cycleLog, maxDist, ms->loadedDictEnd, ip, iend)) { U32 const correction = ZSTD_window_correctOverflow(&ms->window, cycleLog, maxDist, ip); ZSTD_STATIC_ASSERT(ZSTD_CHAINLOG_MAX <= 30); ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX_32 <= 30); @@ -2579,7 +3929,7 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms, * Frame is supposed already started (header already produced) * @return : compressed size, or an error code */ -static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx, +static size_t ZSTD_compress_frameChunk(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastFrameChunk) @@ -2593,7 +3943,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx, assert(cctx->appliedParams.cParams.windowLog <= ZSTD_WINDOWLOG_MAX); - DEBUGLOG(5, "ZSTD_compress_frameChunk (blockSize=%u)", (unsigned)blockSize); + DEBUGLOG(4, "ZSTD_compress_frameChunk (blockSize=%u)", (unsigned)blockSize); if (cctx->appliedParams.fParams.checksumFlag && srcSize) XXH64_update(&cctx->xxhState, src, srcSize); @@ -2609,6 +3959,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx, ZSTD_overflowCorrectIfNeeded( ms, &cctx->workspace, &cctx->appliedParams, ip, ip + blockSize); ZSTD_checkDictValidity(&ms->window, ip + blockSize, maxDist, &ms->loadedDictEnd, &ms->dictMatchState); + ZSTD_window_enforceMaxDist(&ms->window, ip, maxDist, &ms->loadedDictEnd, &ms->dictMatchState); /* Ensure hash/chain table insertion resumes no sooner than lowlimit */ if (ms->nextToUpdate < ms->window.lowLimit) ms->nextToUpdate = ms->window.lowLimit; @@ -2619,6 +3970,10 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx, FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_targetCBlockSize failed"); assert(cSize > 0); assert(cSize <= blockSize + ZSTD_blockHeaderSize); + } else if (ZSTD_blockSplitterEnabled(&cctx->appliedParams)) { + cSize = ZSTD_compressBlock_splitBlock(cctx, op, dstCapacity, ip, blockSize, lastBlock); + FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_splitBlock failed"); + assert(cSize > 0 || cctx->seqCollector.collectSequences == 1); } else { cSize = ZSTD_compressBlock_internal(cctx, op+ZSTD_blockHeaderSize, dstCapacity-ZSTD_blockHeaderSize, @@ -2673,7 +4028,6 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity, "dst buf is too small to fit worst-case frame header size."); DEBUGLOG(4, "ZSTD_writeFrameHeader : dictIDFlag : %u ; dictID : %u ; dictIDSizeCode : %u", !params->fParams.noDictIDFlag, (unsigned)dictID, (unsigned)dictIDSizeCode); - if (params->format == ZSTD_f_zstd1) { MEM_writeLE32(dst, ZSTD_MAGICNUMBER); pos = 4; @@ -2682,7 +4036,9 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity, if (!singleSegment) op[pos++] = windowLogByte; switch(dictIDSizeCode) { - default: assert(0); /* impossible */ + default: + assert(0); /* impossible */ + ZSTD_FALLTHROUGH; case 0 : break; case 1 : op[pos] = (BYTE)(dictID); pos++; break; case 2 : MEM_writeLE16(op+pos, (U16)dictID); pos+=2; break; @@ -2690,7 +4046,9 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity, } switch(fcsCode) { - default: assert(0); /* impossible */ + default: + assert(0); /* impossible */ + ZSTD_FALLTHROUGH; case 0 : if (singleSegment) op[pos++] = (BYTE)(pledgedSrcSize); break; case 1 : MEM_writeLE16(op+pos, (U16)(pledgedSrcSize-256)); pos+=2; break; case 2 : MEM_writeLE32(op+pos, (U32)(pledgedSrcSize)); pos+=4; break; @@ -2699,6 +4057,26 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity, return pos; } +/* ZSTD_writeSkippableFrame_advanced() : + * Writes out a skippable frame with the specified magic number variant (16 are supported), + * from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15, and the desired source data. + * + * Returns the total number of bytes written, or a ZSTD error code. + */ +size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity, + const void* src, size_t srcSize, unsigned magicVariant) { + BYTE* op = (BYTE*)dst; + RETURN_ERROR_IF(dstCapacity < srcSize + ZSTD_SKIPPABLEHEADERSIZE /* Skippable frame overhead */, + dstSize_tooSmall, "Not enough room for skippable frame"); + RETURN_ERROR_IF(srcSize > (unsigned)0xFFFFFFFF, srcSize_wrong, "Src size too large for skippable frame"); + RETURN_ERROR_IF(magicVariant > 15, parameter_outOfBound, "Skippable frame magic number variant not supported"); + + MEM_writeLE32(op, (U32)(ZSTD_MAGIC_SKIPPABLE_START + magicVariant)); + MEM_writeLE32(op+4, (U32)srcSize); + ZSTD_memcpy(op+8, src, srcSize); + return srcSize + ZSTD_SKIPPABLEHEADERSIZE; +} + /* ZSTD_writeLastEmptyBlock() : * output an empty Block with end-of-frame mark to complete a frame * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h)) @@ -2718,13 +4096,14 @@ size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSe { RETURN_ERROR_IF(cctx->stage != ZSTDcs_init, stage_wrong, "wrong cctx stage"); - RETURN_ERROR_IF(cctx->appliedParams.ldmParams.enableLdm, + RETURN_ERROR_IF(cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable, parameter_unsupported, "incompatible with ldm"); cctx->externSeqStore.seq = seq; cctx->externSeqStore.size = nbSeq; cctx->externSeqStore.capacity = nbSeq; cctx->externSeqStore.pos = 0; + cctx->externSeqStore.posInSequence = 0; return 0; } @@ -2754,11 +4133,12 @@ static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx, if (!srcSize) return fhSize; /* do not generate an empty block if no input */ - if (!ZSTD_window_update(&ms->window, src, srcSize)) { + if (!ZSTD_window_update(&ms->window, src, srcSize, ms->forceNonContiguous)) { + ms->forceNonContiguous = 0; ms->nextToUpdate = ms->window.dictLimit; } - if (cctx->appliedParams.ldmParams.enableLdm) { - ZSTD_window_update(&cctx->ldmState.window, src, srcSize); + if (cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) { + ZSTD_window_update(&cctx->ldmState.window, src, srcSize, /* forceNonContiguous */ 0); } if (!frame) { @@ -2826,59 +4206,86 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, { const BYTE* ip = (const BYTE*) src; const BYTE* const iend = ip + srcSize; + int const loadLdmDict = params->ldmParams.enableLdm == ZSTD_ps_enable && ls != NULL; + + /* Assert that we the ms params match the params we're being given */ + ZSTD_assertEqualCParams(params->cParams, ms->cParams); - ZSTD_window_update(&ms->window, src, srcSize); + if (srcSize > ZSTD_CHUNKSIZE_MAX) { + /* Allow the dictionary to set indices up to exactly ZSTD_CURRENT_MAX. + * Dictionaries right at the edge will immediately trigger overflow + * correction, but I don't want to insert extra constraints here. + */ + U32 const maxDictSize = ZSTD_CURRENT_MAX - 1; + /* We must have cleared our windows when our source is this large. */ + assert(ZSTD_window_isEmpty(ms->window)); + if (loadLdmDict) + assert(ZSTD_window_isEmpty(ls->window)); + /* If the dictionary is too large, only load the suffix of the dictionary. */ + if (srcSize > maxDictSize) { + ip = iend - maxDictSize; + src = ip; + srcSize = maxDictSize; + } + } + + DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder); + ZSTD_window_update(&ms->window, src, srcSize, /* forceNonContiguous */ 0); ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base); + ms->forceNonContiguous = params->deterministicRefPrefix; - if (params->ldmParams.enableLdm && ls != NULL) { - ZSTD_window_update(&ls->window, src, srcSize); + if (loadLdmDict) { + ZSTD_window_update(&ls->window, src, srcSize, /* forceNonContiguous */ 0); ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base); } - /* Assert that we the ms params match the params we're being given */ - ZSTD_assertEqualCParams(params->cParams, ms->cParams); - if (srcSize <= HASH_READ_SIZE) return 0; - while (iend - ip > HASH_READ_SIZE) { - size_t const remaining = (size_t)(iend - ip); - size_t const chunk = MIN(remaining, ZSTD_CHUNKSIZE_MAX); - const BYTE* const ichunk = ip + chunk; + ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, iend); - ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, ichunk); + if (loadLdmDict) + ZSTD_ldm_fillHashTable(ls, ip, iend, ¶ms->ldmParams); - if (params->ldmParams.enableLdm && ls != NULL) - ZSTD_ldm_fillHashTable(ls, (const BYTE*)src, (const BYTE*)src + srcSize, ¶ms->ldmParams); - - switch(params->cParams.strategy) - { - case ZSTD_fast: - ZSTD_fillHashTable(ms, ichunk, dtlm); - break; - case ZSTD_dfast: - ZSTD_fillDoubleHashTable(ms, ichunk, dtlm); - break; - - case ZSTD_greedy: - case ZSTD_lazy: - case ZSTD_lazy2: - if (chunk >= HASH_READ_SIZE) - ZSTD_insertAndFindFirstIndex(ms, ichunk-HASH_READ_SIZE); - break; - - case ZSTD_btlazy2: /* we want the dictionary table fully sorted */ - case ZSTD_btopt: - case ZSTD_btultra: - case ZSTD_btultra2: - if (chunk >= HASH_READ_SIZE) - ZSTD_updateTree(ms, ichunk-HASH_READ_SIZE, ichunk); - break; + switch(params->cParams.strategy) + { + case ZSTD_fast: + ZSTD_fillHashTable(ms, iend, dtlm); + break; + case ZSTD_dfast: + ZSTD_fillDoubleHashTable(ms, iend, dtlm); + break; - default: - assert(0); /* not possible : not a valid strategy id */ + case ZSTD_greedy: + case ZSTD_lazy: + case ZSTD_lazy2: + assert(srcSize >= HASH_READ_SIZE); + if (ms->dedicatedDictSearch) { + assert(ms->chainTable != NULL); + ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, iend-HASH_READ_SIZE); + } else { + assert(params->useRowMatchFinder != ZSTD_ps_auto); + if (params->useRowMatchFinder == ZSTD_ps_enable) { + size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog) * sizeof(U16); + ZSTD_memset(ms->tagTable, 0, tagTableSize); + ZSTD_row_update(ms, iend-HASH_READ_SIZE); + DEBUGLOG(4, "Using row-based hash table for lazy dict"); + } else { + ZSTD_insertAndFindFirstIndex(ms, iend-HASH_READ_SIZE); + DEBUGLOG(4, "Using chain-based hash table for lazy dict"); + } } + break; - ip = ichunk; + case ZSTD_btlazy2: /* we want the dictionary table fully sorted */ + case ZSTD_btopt: + case ZSTD_btultra: + case ZSTD_btultra2: + assert(srcSize >= HASH_READ_SIZE); + ZSTD_updateTree(ms, iend-HASH_READ_SIZE, iend); + break; + + default: + assert(0); /* not possible : not a valid strategy id */ } ms->nextToUpdate = (U32)(iend - ms->window.base); @@ -2887,22 +4294,28 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, /* Dictionaries that assign zero probability to symbols that show up causes problems - when FSE encoding. Refuse dictionaries that assign zero probability to symbols - that we may encounter during compression. - NOTE: This behavior is not standard and could be improved in the future. */ -static size_t ZSTD_checkDictNCount(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue) { + * when FSE encoding. Mark dictionaries with zero probability symbols as FSE_repeat_check + * and only dictionaries with 100% valid symbols can be assumed valid. + */ +static FSE_repeat ZSTD_dictNCountRepeat(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue) +{ U32 s; - RETURN_ERROR_IF(dictMaxSymbolValue < maxSymbolValue, dictionary_corrupted, "dict fse tables don't have all symbols"); + if (dictMaxSymbolValue < maxSymbolValue) { + return FSE_repeat_check; + } for (s = 0; s <= maxSymbolValue; ++s) { - RETURN_ERROR_IF(normalizedCounter[s] == 0, dictionary_corrupted, "dict fse tables don't have all symbols"); + if (normalizedCounter[s] == 0) { + return FSE_repeat_check; + } } - return 0; + return FSE_repeat_valid; } size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, - short* offcodeNCount, unsigned* offcodeMaxValue, const void* const dict, size_t dictSize) { + short offcodeNCount[MaxOff+1]; + unsigned offcodeMaxValue = MaxOff; const BYTE* dictPtr = (const BYTE*)dict; /* skip magic num and dict ID */ const BYTE* const dictEnd = dictPtr + dictSize; dictPtr += 8; @@ -2924,16 +4337,16 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, } { unsigned offcodeLog; - size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr); + size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr); RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, ""); RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, ""); - /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */ /* fill all offset symbols to avoid garbage at end of table */ RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp( bs->entropy.fse.offcodeCTable, offcodeNCount, MaxOff, offcodeLog, workspace, HUF_WORKSPACE_SIZE)), dictionary_corrupted, ""); + /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */ dictPtr += offcodeHeaderSize; } @@ -2942,13 +4355,12 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr); RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, ""); RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, ""); - /* Every match length code must have non-zero probability */ - FORWARD_IF_ERROR( ZSTD_checkDictNCount(matchlengthNCount, matchlengthMaxValue, MaxML), ""); RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp( bs->entropy.fse.matchlengthCTable, matchlengthNCount, matchlengthMaxValue, matchlengthLog, workspace, HUF_WORKSPACE_SIZE)), dictionary_corrupted, ""); + bs->entropy.fse.matchlength_repeatMode = ZSTD_dictNCountRepeat(matchlengthNCount, matchlengthMaxValue, MaxML); dictPtr += matchlengthHeaderSize; } @@ -2957,13 +4369,12 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr); RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, ""); RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, ""); - /* Every literal length code must have non-zero probability */ - FORWARD_IF_ERROR( ZSTD_checkDictNCount(litlengthNCount, litlengthMaxValue, MaxLL), ""); RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp( bs->entropy.fse.litlengthCTable, litlengthNCount, litlengthMaxValue, litlengthLog, workspace, HUF_WORKSPACE_SIZE)), dictionary_corrupted, ""); + bs->entropy.fse.litlength_repeatMode = ZSTD_dictNCountRepeat(litlengthNCount, litlengthMaxValue, MaxLL); dictPtr += litlengthHeaderSize; } @@ -2973,12 +4384,28 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, bs->rep[2] = MEM_readLE32(dictPtr+8); dictPtr += 12; + { size_t const dictContentSize = (size_t)(dictEnd - dictPtr); + U32 offcodeMax = MaxOff; + if (dictContentSize <= ((U32)-1) - 128 KB) { + U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */ + offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */ + } + /* All offset values <= dictContentSize + 128 KB must be representable for a valid table */ + bs->entropy.fse.offcode_repeatMode = ZSTD_dictNCountRepeat(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff)); + + /* All repCodes must be <= dictContentSize and != 0 */ + { U32 u; + for (u=0; u<3; u++) { + RETURN_ERROR_IF(bs->rep[u] == 0, dictionary_corrupted, ""); + RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, ""); + } } } + return dictPtr - (const BYTE*)dict; } /* Dictionary format : * See : - * https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#dictionary-format + * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#dictionary-format */ /*! ZSTD_loadZstdDictionary() : * @return : dictID, or an error code @@ -2995,42 +4422,23 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs, { const BYTE* dictPtr = (const BYTE*)dict; const BYTE* const dictEnd = dictPtr + dictSize; - short offcodeNCount[MaxOff+1]; - unsigned offcodeMaxValue = MaxOff; size_t dictID; size_t eSize; - ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog))); assert(dictSize >= 8); assert(MEM_readLE32(dictPtr) == ZSTD_MAGIC_DICTIONARY); dictID = params->fParams.noDictIDFlag ? 0 : MEM_readLE32(dictPtr + 4 /* skip magic number */ ); - eSize = ZSTD_loadCEntropy(bs, workspace, offcodeNCount, &offcodeMaxValue, dict, dictSize); + eSize = ZSTD_loadCEntropy(bs, workspace, dict, dictSize); FORWARD_IF_ERROR(eSize, "ZSTD_loadCEntropy failed"); dictPtr += eSize; - { size_t const dictContentSize = (size_t)(dictEnd - dictPtr); - U32 offcodeMax = MaxOff; - if (dictContentSize <= ((U32)-1) - 128 KB) { - U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */ - offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */ - } - /* All offset values <= dictContentSize + 128 KB must be representable */ - FORWARD_IF_ERROR(ZSTD_checkDictNCount(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff)), ""); - /* All repCodes must be <= dictContentSize and != 0*/ - { U32 u; - for (u=0; u<3; u++) { - RETURN_ERROR_IF(bs->rep[u] == 0, dictionary_corrupted, ""); - RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, ""); - } } - - bs->entropy.fse.offcode_repeatMode = FSE_repeat_valid; - bs->entropy.fse.matchlength_repeatMode = FSE_repeat_valid; - bs->entropy.fse.litlength_repeatMode = FSE_repeat_valid; + { + size_t const dictContentSize = (size_t)(dictEnd - dictPtr); FORWARD_IF_ERROR(ZSTD_loadDictionaryContent( ms, NULL, ws, params, dictPtr, dictContentSize, dtlm), ""); - return dictID; } + return dictID; } /** ZSTD_compress_insertDictionary() : @@ -3074,7 +4482,7 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs, } #define ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF (128 KB) -#define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6) +#define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6ULL) /*! ZSTD_compressBegin_internal() : * @return : 0, or an error code */ @@ -3086,6 +4494,10 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params, U64 pledgedSrcSize, ZSTD_buffered_policy_e zbuff) { + size_t const dictContentSize = cdict ? cdict->dictContentSize : dictSize; +#if ZSTD_TRACE + cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0; +#endif DEBUGLOG(4, "ZSTD_compressBegin_internal: wlog=%u", params->cParams.windowLog); /* params are supposed to be fully validated at this point */ assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams))); @@ -3100,13 +4512,14 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx, return ZSTD_resetCCtx_usingCDict(cctx, cdict, params, pledgedSrcSize, zbuff); } - FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, *params, pledgedSrcSize, + FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize, + dictContentSize, ZSTDcrp_makeClean, zbuff) , ""); { size_t const dictID = cdict ? ZSTD_compress_insertDictionary( cctx->blockState.prevCBlock, &cctx->blockState.matchState, &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, cdict->dictContent, - cdict->dictContentSize, dictContentType, dtlm, + cdict->dictContentSize, cdict->dictContentType, dtlm, cctx->entropyWorkspace) : ZSTD_compress_insertDictionary( cctx->blockState.prevCBlock, &cctx->blockState.matchState, @@ -3115,6 +4528,7 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx, FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed"); assert(dictID <= UINT_MAX); cctx->dictID = (U32)dictID; + cctx->dictContentSize = dictContentSize; } return 0; } @@ -3143,8 +4557,8 @@ size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize) { - ZSTD_CCtx_params const cctxParams = - ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, ¶ms); + ZSTD_CCtx_params cctxParams; + ZSTD_CCtxParams_init_internal(&cctxParams, ¶ms, ZSTD_NO_CLEVEL); return ZSTD_compressBegin_advanced_internal(cctx, dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL /*cdict*/, @@ -3153,9 +4567,11 @@ size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel) { - ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize); - ZSTD_CCtx_params const cctxParams = - ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, ¶ms); + ZSTD_CCtx_params cctxParams; + { + ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict); + ZSTD_CCtxParams_init_internal(&cctxParams, ¶ms, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel); + } DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (unsigned)dictSize); return ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL, &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered); @@ -3209,6 +4625,30 @@ static size_t ZSTD_writeEpilogue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity) return op-ostart; } +void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize) +{ +#if ZSTD_TRACE + if (cctx->traceCtx && ZSTD_trace_compress_end != NULL) { + int const streaming = cctx->inBuffSize > 0 || cctx->outBuffSize > 0 || cctx->appliedParams.nbWorkers > 0; + ZSTD_Trace trace; + ZSTD_memset(&trace, 0, sizeof(trace)); + trace.version = ZSTD_VERSION_NUMBER; + trace.streaming = streaming; + trace.dictionaryID = cctx->dictID; + trace.dictionarySize = cctx->dictContentSize; + trace.uncompressedSize = cctx->consumedSrcSize; + trace.compressedSize = cctx->producedCSize + extraCSize; + trace.params = &cctx->appliedParams; + trace.cctx = cctx; + ZSTD_trace_compress_end(cctx->traceCtx, &trace); + } + cctx->traceCtx = 0; +#else + (void)cctx; + (void)extraCSize; +#endif +} + size_t ZSTD_compressEnd (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize) @@ -3231,26 +4671,10 @@ size_t ZSTD_compressEnd (ZSTD_CCtx* cctx, (unsigned)cctx->pledgedSrcSizePlusOne-1, (unsigned)cctx->consumedSrcSize); } + ZSTD_CCtx_trace(cctx, endResult); return cSize + endResult; } - -static size_t ZSTD_compress_internal (ZSTD_CCtx* cctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - const void* dict,size_t dictSize, - const ZSTD_parameters* params) -{ - ZSTD_CCtx_params const cctxParams = - ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, params); - DEBUGLOG(4, "ZSTD_compress_internal"); - return ZSTD_compress_advanced_internal(cctx, - dst, dstCapacity, - src, srcSize, - dict, dictSize, - &cctxParams); -} - size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize, @@ -3259,11 +4683,12 @@ size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx, { DEBUGLOG(4, "ZSTD_compress_advanced"); FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), ""); - return ZSTD_compress_internal(cctx, - dst, dstCapacity, - src, srcSize, - dict, dictSize, - ¶ms); + ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, ¶ms, ZSTD_NO_CLEVEL); + return ZSTD_compress_advanced_internal(cctx, + dst, dstCapacity, + src, srcSize, + dict, dictSize, + &cctx->simpleApiParams); } /* Internal */ @@ -3287,11 +4712,13 @@ size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel) { - ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0); - ZSTD_CCtx_params cctxParams = ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, ¶ms); + { + ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0, ZSTD_cpm_noAttachDict); + assert(params.fParams.contentSizeFlag == 1); + ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, ¶ms, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT: compressionLevel); + } DEBUGLOG(4, "ZSTD_compress_usingDict (srcSize=%u)", (unsigned)srcSize); - assert(params.fParams.contentSizeFlag == 1); - return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctxParams); + return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctx->simpleApiParams); } size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx, @@ -3309,10 +4736,17 @@ size_t ZSTD_compress(void* dst, size_t dstCapacity, int compressionLevel) { size_t result; +#if ZSTD_COMPRESS_HEAPMODE + ZSTD_CCtx* cctx = ZSTD_createCCtx(); + RETURN_ERROR_IF(!cctx, memory_allocation, "ZSTD_createCCtx failed"); + result = ZSTD_compressCCtx(cctx, dst, dstCapacity, src, srcSize, compressionLevel); + ZSTD_freeCCtx(cctx); +#else ZSTD_CCtx ctxBody; ZSTD_initCCtx(&ctxBody, ZSTD_defaultCMem); result = ZSTD_compressCCtx(&ctxBody, dst, dstCapacity, src, srcSize, compressionLevel); ZSTD_freeCCtxContent(&ctxBody); /* can't free ctxBody itself, as it's on stack; free only heap content */ +#endif return result; } @@ -3328,14 +4762,17 @@ size_t ZSTD_estimateCDictSize_advanced( DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (unsigned)sizeof(ZSTD_CDict)); return ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) - + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) + /* enableDedicatedDictSearch == 1 ensures that CDict estimation will not be too small + * in case we are using DDS with row-hash. */ + + ZSTD_sizeof_matchState(&cParams, ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams), + /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void *)))); } size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel) { - ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize); + ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); return ZSTD_estimateCDictSize_advanced(dictSize, cParams, ZSTD_dlm_byCopy); } @@ -3353,20 +4790,22 @@ static size_t ZSTD_initCDict_internal( const void* dictBuffer, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType, - ZSTD_compressionParameters cParams) + ZSTD_CCtx_params params) { DEBUGLOG(3, "ZSTD_initCDict_internal (dictContentType:%u)", (unsigned)dictContentType); - assert(!ZSTD_checkCParams(cParams)); - cdict->matchState.cParams = cParams; + assert(!ZSTD_checkCParams(params.cParams)); + cdict->matchState.cParams = params.cParams; + cdict->matchState.dedicatedDictSearch = params.enableDedicatedDictSearch; if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dictBuffer) || (!dictSize)) { cdict->dictContent = dictBuffer; } else { void *internalBuffer = ZSTD_cwksp_reserve_object(&cdict->workspace, ZSTD_cwksp_align(dictSize, sizeof(void*))); RETURN_ERROR_IF(!internalBuffer, memory_allocation, "NULL pointer!"); cdict->dictContent = internalBuffer; - memcpy(internalBuffer, dictBuffer, dictSize); + ZSTD_memcpy(internalBuffer, dictBuffer, dictSize); } cdict->dictContentSize = dictSize; + cdict->dictContentType = dictContentType; cdict->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cdict->workspace, HUF_WORKSPACE_SIZE); @@ -3376,18 +4815,16 @@ static size_t ZSTD_initCDict_internal( FORWARD_IF_ERROR(ZSTD_reset_matchState( &cdict->matchState, &cdict->workspace, - &cParams, + ¶ms.cParams, + params.useRowMatchFinder, ZSTDcrp_makeClean, ZSTDirp_reset, ZSTD_resetTarget_CDict), ""); /* (Maybe) load the dictionary * Skips loading the dictionary if it is < 8 bytes. */ - { ZSTD_CCtx_params params; - memset(¶ms, 0, sizeof(params)); - params.compressionLevel = ZSTD_CLEVEL_DEFAULT; + { params.compressionLevel = ZSTD_CLEVEL_DEFAULT; params.fParams.contentSizeFlag = 1; - params.cParams = cParams; { size_t const dictID = ZSTD_compress_insertDictionary( &cdict->cBlockState, &cdict->matchState, NULL, &cdict->workspace, ¶ms, cdict->dictContent, cdict->dictContentSize, @@ -3401,66 +4838,129 @@ static size_t ZSTD_initCDict_internal( return 0; } -ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize, +static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, - ZSTD_dictContentType_e dictContentType, - ZSTD_compressionParameters cParams, ZSTD_customMem customMem) + ZSTD_compressionParameters cParams, + ZSTD_paramSwitch_e useRowMatchFinder, + U32 enableDedicatedDictSearch, + ZSTD_customMem customMem) { - DEBUGLOG(3, "ZSTD_createCDict_advanced, mode %u", (unsigned)dictContentType); - if (!customMem.customAlloc ^ !customMem.customFree) return NULL; + if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL; { size_t const workspaceSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) + - ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) + + ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, enableDedicatedDictSearch, /* forCCtx */ 0) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*)))); - void* const workspace = ZSTD_malloc(workspaceSize, customMem); + void* const workspace = ZSTD_customMalloc(workspaceSize, customMem); ZSTD_cwksp ws; ZSTD_CDict* cdict; if (!workspace) { - ZSTD_free(workspace, customMem); + ZSTD_customFree(workspace, customMem); return NULL; } - ZSTD_cwksp_init(&ws, workspace, workspaceSize); + ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_dynamic_alloc); cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict)); assert(cdict != NULL); ZSTD_cwksp_move(&cdict->workspace, &ws); cdict->customMem = customMem; - cdict->compressionLevel = 0; /* signals advanced API usage */ + cdict->compressionLevel = ZSTD_NO_CLEVEL; /* signals advanced API usage */ + cdict->useRowMatchFinder = useRowMatchFinder; + return cdict; + } +} - if (ZSTD_isError( ZSTD_initCDict_internal(cdict, - dictBuffer, dictSize, - dictLoadMethod, dictContentType, - cParams) )) { - ZSTD_freeCDict(cdict); - return NULL; - } +ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize, + ZSTD_dictLoadMethod_e dictLoadMethod, + ZSTD_dictContentType_e dictContentType, + ZSTD_compressionParameters cParams, + ZSTD_customMem customMem) +{ + ZSTD_CCtx_params cctxParams; + ZSTD_memset(&cctxParams, 0, sizeof(cctxParams)); + ZSTD_CCtxParams_init(&cctxParams, 0); + cctxParams.cParams = cParams; + cctxParams.customMem = customMem; + return ZSTD_createCDict_advanced2( + dictBuffer, dictSize, + dictLoadMethod, dictContentType, + &cctxParams, customMem); +} + +ZSTD_CDict* ZSTD_createCDict_advanced2( + const void* dict, size_t dictSize, + ZSTD_dictLoadMethod_e dictLoadMethod, + ZSTD_dictContentType_e dictContentType, + const ZSTD_CCtx_params* originalCctxParams, + ZSTD_customMem customMem) +{ + ZSTD_CCtx_params cctxParams = *originalCctxParams; + ZSTD_compressionParameters cParams; + ZSTD_CDict* cdict; - return cdict; + DEBUGLOG(3, "ZSTD_createCDict_advanced2, mode %u", (unsigned)dictContentType); + if (!customMem.customAlloc ^ !customMem.customFree) return NULL; + + if (cctxParams.enableDedicatedDictSearch) { + cParams = ZSTD_dedicatedDictSearch_getCParams( + cctxParams.compressionLevel, dictSize); + ZSTD_overrideCParams(&cParams, &cctxParams.cParams); + } else { + cParams = ZSTD_getCParamsFromCCtxParams( + &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); + } + + if (!ZSTD_dedicatedDictSearch_isSupported(&cParams)) { + /* Fall back to non-DDSS params */ + cctxParams.enableDedicatedDictSearch = 0; + cParams = ZSTD_getCParamsFromCCtxParams( + &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); + } + + DEBUGLOG(3, "ZSTD_createCDict_advanced2: DDS: %u", cctxParams.enableDedicatedDictSearch); + cctxParams.cParams = cParams; + cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams); + + cdict = ZSTD_createCDict_advanced_internal(dictSize, + dictLoadMethod, cctxParams.cParams, + cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch, + customMem); + + if (ZSTD_isError( ZSTD_initCDict_internal(cdict, + dict, dictSize, + dictLoadMethod, dictContentType, + cctxParams) )) { + ZSTD_freeCDict(cdict); + return NULL; } + + return cdict; } ZSTD_CDict* ZSTD_createCDict(const void* dict, size_t dictSize, int compressionLevel) { - ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize); - ZSTD_CDict* cdict = ZSTD_createCDict_advanced(dict, dictSize, + ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); + ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto, cParams, ZSTD_defaultCMem); if (cdict) - cdict->compressionLevel = compressionLevel == 0 ? ZSTD_CLEVEL_DEFAULT : compressionLevel; + cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel; return cdict; } ZSTD_CDict* ZSTD_createCDict_byReference(const void* dict, size_t dictSize, int compressionLevel) { - ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize); - return ZSTD_createCDict_advanced(dict, dictSize, + ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict); + ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto, cParams, ZSTD_defaultCMem); + if (cdict) + cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel; + return cdict; } size_t ZSTD_freeCDict(ZSTD_CDict* cdict) @@ -3470,7 +4970,7 @@ size_t ZSTD_freeCDict(ZSTD_CDict* cdict) int cdictInWorkspace = ZSTD_cwksp_owns_buffer(&cdict->workspace, cdict); ZSTD_cwksp_free(&cdict->workspace, cMem); if (!cdictInWorkspace) { - ZSTD_free(cdict, cMem); + ZSTD_customFree(cdict, cMem); } return 0; } @@ -3496,19 +4996,22 @@ const ZSTD_CDict* ZSTD_initStaticCDict( ZSTD_dictContentType_e dictContentType, ZSTD_compressionParameters cParams) { - size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0); + ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams); + /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */ + size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0); size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*)))) + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) + matchStateSize; ZSTD_CDict* cdict; + ZSTD_CCtx_params params; if ((size_t)workspace & 7) return NULL; /* 8-aligned */ { ZSTD_cwksp ws; - ZSTD_cwksp_init(&ws, workspace, workspaceSize); + ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc); cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict)); if (cdict == NULL) return NULL; ZSTD_cwksp_move(&cdict->workspace, &ws); @@ -3518,10 +5021,15 @@ const ZSTD_CDict* ZSTD_initStaticCDict( (unsigned)workspaceSize, (unsigned)neededSize, (unsigned)(workspaceSize < neededSize)); if (workspaceSize < neededSize) return NULL; + ZSTD_CCtxParams_init(¶ms, 0); + params.cParams = cParams; + params.useRowMatchFinder = useRowMatchFinder; + cdict->useRowMatchFinder = useRowMatchFinder; + if (ZSTD_isError( ZSTD_initCDict_internal(cdict, dict, dictSize, dictLoadMethod, dictContentType, - cParams) )) + params) )) return NULL; return cdict; @@ -3533,61 +5041,98 @@ ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict) return cdict->matchState.cParams; } -/* ZSTD_compressBegin_usingCDict_advanced() : - * cdict must be != NULL */ -size_t ZSTD_compressBegin_usingCDict_advanced( +/*! ZSTD_getDictID_fromCDict() : + * Provides the dictID of the dictionary loaded into `cdict`. + * If @return == 0, the dictionary is not conformant to Zstandard specification, or empty. + * Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */ +unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict) +{ + if (cdict==NULL) return 0; + return cdict->dictID; +} + +/* ZSTD_compressBegin_usingCDict_internal() : + * Implementation of various ZSTD_compressBegin_usingCDict* functions. + */ +static size_t ZSTD_compressBegin_usingCDict_internal( ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize) { - DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_advanced"); + ZSTD_CCtx_params cctxParams; + DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_internal"); RETURN_ERROR_IF(cdict==NULL, dictionary_wrong, "NULL pointer!"); - { ZSTD_CCtx_params params = cctx->requestedParams; + /* Initialize the cctxParams from the cdict */ + { + ZSTD_parameters params; + params.fParams = fParams; params.cParams = ( pledgedSrcSize < ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF || pledgedSrcSize < cdict->dictContentSize * ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN - || cdict->compressionLevel == 0 ) - && (params.attachDictPref != ZSTD_dictForceLoad) ? + || cdict->compressionLevel == 0 ) ? ZSTD_getCParamsFromCDict(cdict) : ZSTD_getCParams(cdict->compressionLevel, pledgedSrcSize, cdict->dictContentSize); - /* Increase window log to fit the entire dictionary and source if the - * source size is known. Limit the increase to 19, which is the - * window log for compression level 1 with the largest source size. - */ - if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) { - U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19); - U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1; - params.cParams.windowLog = MAX(params.cParams.windowLog, limitedSrcLog); - } - params.fParams = fParams; - return ZSTD_compressBegin_internal(cctx, - NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, - cdict, - ¶ms, pledgedSrcSize, - ZSTDb_not_buffered); + ZSTD_CCtxParams_init_internal(&cctxParams, ¶ms, cdict->compressionLevel); + } + /* Increase window log to fit the entire dictionary and source if the + * source size is known. Limit the increase to 19, which is the + * window log for compression level 1 with the largest source size. + */ + if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) { + U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19); + U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1; + cctxParams.cParams.windowLog = MAX(cctxParams.cParams.windowLog, limitedSrcLog); } + return ZSTD_compressBegin_internal(cctx, + NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, + cdict, + &cctxParams, pledgedSrcSize, + ZSTDb_not_buffered); +} + + +/* ZSTD_compressBegin_usingCDict_advanced() : + * This function is DEPRECATED. + * cdict must be != NULL */ +size_t ZSTD_compressBegin_usingCDict_advanced( + ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, + ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize) +{ + return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, pledgedSrcSize); } /* ZSTD_compressBegin_usingCDict() : - * pledgedSrcSize=0 means "unknown" - * if pledgedSrcSize>0, it will enable contentSizeFlag */ + * cdict must be != NULL */ size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict) { ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ }; - DEBUGLOG(4, "ZSTD_compressBegin_usingCDict : dictIDFlag == %u", !fParams.noDictIDFlag); - return ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN); + return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN); } -size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx, +/*! ZSTD_compress_usingCDict_internal(): + * Implementation of various ZSTD_compress_usingCDict* functions. + */ +static size_t ZSTD_compress_usingCDict_internal(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize, const ZSTD_CDict* cdict, ZSTD_frameParameters fParams) { - FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */ + FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */ return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize); } +/*! ZSTD_compress_usingCDict_advanced(): + * This function is DEPRECATED. + */ +size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx, + void* dst, size_t dstCapacity, + const void* src, size_t srcSize, + const ZSTD_CDict* cdict, ZSTD_frameParameters fParams) +{ + return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams); +} + /*! ZSTD_compress_usingCDict() : * Compression using a digested Dictionary. * Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times. @@ -3599,7 +5144,7 @@ size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict) { ZSTD_frameParameters const fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ }; - return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, fParams); + return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams); } @@ -3640,32 +5185,12 @@ size_t ZSTD_CStreamOutSize(void) return ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + 4 /* 32-bits hash */ ; } -static size_t ZSTD_resetCStream_internal(ZSTD_CStream* cctx, - const void* const dict, size_t const dictSize, ZSTD_dictContentType_e const dictContentType, - const ZSTD_CDict* const cdict, - ZSTD_CCtx_params params, unsigned long long const pledgedSrcSize) +static ZSTD_cParamMode_e ZSTD_getCParamMode(ZSTD_CDict const* cdict, ZSTD_CCtx_params const* params, U64 pledgedSrcSize) { - DEBUGLOG(4, "ZSTD_resetCStream_internal"); - /* Finalize the compression parameters */ - params.cParams = ZSTD_getCParamsFromCCtxParams(¶ms, pledgedSrcSize, dictSize); - /* params are supposed to be fully validated at this point */ - assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams))); - assert(!((dict) && (cdict))); /* either dict or cdict, not both */ - - FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx, - dict, dictSize, dictContentType, ZSTD_dtlm_fast, - cdict, - ¶ms, pledgedSrcSize, - ZSTDb_buffered) , ""); - - cctx->inToCompress = 0; - cctx->inBuffPos = 0; - cctx->inBuffTarget = cctx->blockSize - + (cctx->blockSize == pledgedSrcSize); /* for small input: avoid automatic flush on reaching end of block, since it would require to add a 3-bytes null block to end frame */ - cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0; - cctx->streamStage = zcss_load; - cctx->frameEnded = 0; - return 0; /* ready to go */ + if (cdict != NULL && ZSTD_shouldAttachDict(cdict, params, pledgedSrcSize)) + return ZSTD_cpm_attachDict; + else + return ZSTD_cpm_noAttachDict; } /* ZSTD_resetCStream(): @@ -3749,7 +5274,7 @@ size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs, FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , ""); FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , ""); FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , ""); - zcs->requestedParams = ZSTD_assignParamsToCCtxParams(&zcs->requestedParams, ¶ms); + ZSTD_CCtxParams_setZstdParams(&zcs->requestedParams, ¶ms); FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , ""); return 0; } @@ -3815,12 +5340,17 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs, /* check expectations */ DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%u", (unsigned)flushMode); - assert(zcs->inBuff != NULL); - assert(zcs->inBuffSize > 0); - assert(zcs->outBuff != NULL); - assert(zcs->outBuffSize > 0); + if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) { + assert(zcs->inBuff != NULL); + assert(zcs->inBuffSize > 0); + } + if (zcs->appliedParams.outBufferMode == ZSTD_bm_buffered) { + assert(zcs->outBuff != NULL); + assert(zcs->outBuffSize > 0); + } assert(output->pos <= output->size); assert(input->pos <= input->size); + assert((U32)flushMode <= (U32)ZSTD_e_end); while (someMoreWork) { switch(zcs->streamStage) @@ -3830,7 +5360,8 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs, case zcss_load: if ( (flushMode == ZSTD_e_end) - && ((size_t)(oend-op) >= ZSTD_compressBound(iend-ip)) /* enough dstCapacity */ + && ( (size_t)(oend-op) >= ZSTD_compressBound(iend-ip) /* Enough output space */ + || zcs->appliedParams.outBufferMode == ZSTD_bm_stable) /* OR we are allowed to return dstSizeTooSmall */ && (zcs->inBuffPos == 0) ) { /* shortcut to compression pass directly into output buffer */ size_t const cSize = ZSTD_compressEnd(zcs, @@ -3843,8 +5374,9 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs, ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only); someMoreWork = 0; break; } - /* complete loading into inBuffer */ - { size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos; + /* complete loading into inBuffer in buffered mode */ + if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) { + size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos; size_t const loaded = ZSTD_limitCopy( zcs->inBuff + zcs->inBuffPos, toLoad, ip, iend-ip); @@ -3864,31 +5396,49 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs, } /* compress current block (note : this stage cannot be stopped in the middle) */ DEBUGLOG(5, "stream compression stage (flushMode==%u)", flushMode); - { void* cDst; + { int const inputBuffered = (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered); + void* cDst; size_t cSize; - size_t const iSize = zcs->inBuffPos - zcs->inToCompress; size_t oSize = oend-op; - unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend); - if (oSize >= ZSTD_compressBound(iSize)) + size_t const iSize = inputBuffered + ? zcs->inBuffPos - zcs->inToCompress + : MIN((size_t)(iend - ip), zcs->blockSize); + if (oSize >= ZSTD_compressBound(iSize) || zcs->appliedParams.outBufferMode == ZSTD_bm_stable) cDst = op; /* compress into output buffer, to skip flush stage */ else cDst = zcs->outBuff, oSize = zcs->outBuffSize; - cSize = lastBlock ? - ZSTD_compressEnd(zcs, cDst, oSize, - zcs->inBuff + zcs->inToCompress, iSize) : - ZSTD_compressContinue(zcs, cDst, oSize, - zcs->inBuff + zcs->inToCompress, iSize); - FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed"); - zcs->frameEnded = lastBlock; - /* prepare next block */ - zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize; - if (zcs->inBuffTarget > zcs->inBuffSize) - zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize; - DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u", - (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize); - if (!lastBlock) - assert(zcs->inBuffTarget <= zcs->inBuffSize); - zcs->inToCompress = zcs->inBuffPos; + if (inputBuffered) { + unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend); + cSize = lastBlock ? + ZSTD_compressEnd(zcs, cDst, oSize, + zcs->inBuff + zcs->inToCompress, iSize) : + ZSTD_compressContinue(zcs, cDst, oSize, + zcs->inBuff + zcs->inToCompress, iSize); + FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed"); + zcs->frameEnded = lastBlock; + /* prepare next block */ + zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize; + if (zcs->inBuffTarget > zcs->inBuffSize) + zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize; + DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u", + (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize); + if (!lastBlock) + assert(zcs->inBuffTarget <= zcs->inBuffSize); + zcs->inToCompress = zcs->inBuffPos; + } else { + unsigned const lastBlock = (ip + iSize == iend); + assert(flushMode == ZSTD_e_end /* Already validated */); + cSize = lastBlock ? + ZSTD_compressEnd(zcs, cDst, oSize, ip, iSize) : + ZSTD_compressContinue(zcs, cDst, oSize, ip, iSize); + /* Consume the input prior to error checking to mirror buffered mode. */ + if (iSize > 0) + ip += iSize; + FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed"); + zcs->frameEnded = lastBlock; + if (lastBlock) + assert(ip == iend); + } if (cDst == op) { /* no need to flush */ op += cSize; if (zcs->frameEnded) { @@ -3902,9 +5452,10 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs, zcs->outBuffFlushedSize = 0; zcs->streamStage = zcss_flush; /* pass-through to flush stage */ } - /* fall-through */ + ZSTD_FALLTHROUGH; case zcss_flush: DEBUGLOG(5, "flush stage"); + assert(zcs->appliedParams.outBufferMode == ZSTD_bm_buffered); { size_t const toFlush = zcs->outBuffContentSize - zcs->outBuffFlushedSize; size_t const flushed = ZSTD_limitCopy(op, (size_t)(oend-op), zcs->outBuff + zcs->outBuffFlushedSize, toFlush); @@ -3959,6 +5510,126 @@ size_t ZSTD_compressStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output, ZSTD_inBuf return ZSTD_nextInputSizeHint_MTorST(zcs); } +/* After a compression call set the expected input/output buffer. + * This is validated at the start of the next compression call. + */ +static void ZSTD_setBufferExpectations(ZSTD_CCtx* cctx, ZSTD_outBuffer const* output, ZSTD_inBuffer const* input) +{ + if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) { + cctx->expectedInBuffer = *input; + } + if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) { + cctx->expectedOutBufferSize = output->size - output->pos; + } +} + +/* Validate that the input/output buffers match the expectations set by + * ZSTD_setBufferExpectations. + */ +static size_t ZSTD_checkBufferStability(ZSTD_CCtx const* cctx, + ZSTD_outBuffer const* output, + ZSTD_inBuffer const* input, + ZSTD_EndDirective endOp) +{ + if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) { + ZSTD_inBuffer const expect = cctx->expectedInBuffer; + if (expect.src != input->src || expect.pos != input->pos || expect.size != input->size) + RETURN_ERROR(srcBuffer_wrong, "ZSTD_c_stableInBuffer enabled but input differs!"); + if (endOp != ZSTD_e_end) + RETURN_ERROR(srcBuffer_wrong, "ZSTD_c_stableInBuffer can only be used with ZSTD_e_end!"); + } + if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) { + size_t const outBufferSize = output->size - output->pos; + if (cctx->expectedOutBufferSize != outBufferSize) + RETURN_ERROR(dstBuffer_wrong, "ZSTD_c_stableOutBuffer enabled but output size differs!"); + } + return 0; +} + +static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx, + ZSTD_EndDirective endOp, + size_t inSize) { + ZSTD_CCtx_params params = cctx->requestedParams; + ZSTD_prefixDict const prefixDict = cctx->prefixDict; + FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */ + ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict)); /* single usage */ + assert(prefixDict.dict==NULL || cctx->cdict==NULL); /* only one can be set */ + if (cctx->cdict && !cctx->localDict.cdict) { + /* Let the cdict's compression level take priority over the requested params. + * But do not take the cdict's compression level if the "cdict" is actually a localDict + * generated from ZSTD_initLocalDict(). + */ + params.compressionLevel = cctx->cdict->compressionLevel; + } + DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage"); + if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1; /* auto-fix pledgedSrcSize */ + { + size_t const dictSize = prefixDict.dict + ? prefixDict.dictSize + : (cctx->cdict ? cctx->cdict->dictContentSize : 0); + ZSTD_cParamMode_e const mode = ZSTD_getCParamMode(cctx->cdict, ¶ms, cctx->pledgedSrcSizePlusOne - 1); + params.cParams = ZSTD_getCParamsFromCCtxParams( + ¶ms, cctx->pledgedSrcSizePlusOne-1, + dictSize, mode); + } + + params.useBlockSplitter = ZSTD_resolveBlockSplitterMode(params.useBlockSplitter, ¶ms.cParams); + params.ldmParams.enableLdm = ZSTD_resolveEnableLdm(params.ldmParams.enableLdm, ¶ms.cParams); + params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, ¶ms.cParams); + +#ifdef ZSTD_MULTITHREAD + if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) { + params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */ + } + if (params.nbWorkers > 0) { +#if ZSTD_TRACE + cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0; +#endif + /* mt context creation */ + if (cctx->mtctx == NULL) { + DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u", + params.nbWorkers); + cctx->mtctx = ZSTDMT_createCCtx_advanced((U32)params.nbWorkers, cctx->customMem, cctx->pool); + RETURN_ERROR_IF(cctx->mtctx == NULL, memory_allocation, "NULL pointer!"); + } + /* mt compression */ + DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers); + FORWARD_IF_ERROR( ZSTDMT_initCStream_internal( + cctx->mtctx, + prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, + cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , ""); + cctx->dictID = cctx->cdict ? cctx->cdict->dictID : 0; + cctx->dictContentSize = cctx->cdict ? cctx->cdict->dictContentSize : prefixDict.dictSize; + cctx->consumedSrcSize = 0; + cctx->producedCSize = 0; + cctx->streamStage = zcss_load; + cctx->appliedParams = params; + } else +#endif + { U64 const pledgedSrcSize = cctx->pledgedSrcSizePlusOne - 1; + assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams))); + FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx, + prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, ZSTD_dtlm_fast, + cctx->cdict, + ¶ms, pledgedSrcSize, + ZSTDb_buffered) , ""); + assert(cctx->appliedParams.nbWorkers == 0); + cctx->inToCompress = 0; + cctx->inBuffPos = 0; + if (cctx->appliedParams.inBufferMode == ZSTD_bm_buffered) { + /* for small input: avoid automatic flush on reaching end of block, since + * it would require to add a 3-bytes null block to end frame + */ + cctx->inBuffTarget = cctx->blockSize + (cctx->blockSize == pledgedSrcSize); + } else { + cctx->inBuffTarget = 0; + } + cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0; + cctx->streamStage = zcss_load; + cctx->frameEnded = 0; + } + return 0; +} size_t ZSTD_compressStream2( ZSTD_CCtx* cctx, ZSTD_outBuffer* output, @@ -3967,82 +5638,69 @@ size_t ZSTD_compressStream2( ZSTD_CCtx* cctx, { DEBUGLOG(5, "ZSTD_compressStream2, endOp=%u ", (unsigned)endOp); /* check conditions */ - RETURN_ERROR_IF(output->pos > output->size, GENERIC, "invalid buffer"); - RETURN_ERROR_IF(input->pos > input->size, GENERIC, "invalid buffer"); - assert(cctx!=NULL); + RETURN_ERROR_IF(output->pos > output->size, dstSize_tooSmall, "invalid output buffer"); + RETURN_ERROR_IF(input->pos > input->size, srcSize_wrong, "invalid input buffer"); + RETURN_ERROR_IF((U32)endOp > (U32)ZSTD_e_end, parameter_outOfBound, "invalid endDirective"); + assert(cctx != NULL); /* transparent initialization stage */ if (cctx->streamStage == zcss_init) { - ZSTD_CCtx_params params = cctx->requestedParams; - ZSTD_prefixDict const prefixDict = cctx->prefixDict; - FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */ - memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict)); /* single usage */ - assert(prefixDict.dict==NULL || cctx->cdict==NULL); /* only one can be set */ - DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage"); - if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = input->size + 1; /* auto-fix pledgedSrcSize */ - params.cParams = ZSTD_getCParamsFromCCtxParams( - &cctx->requestedParams, cctx->pledgedSrcSizePlusOne-1, 0 /*dictSize*/); - - -#ifdef ZSTD_MULTITHREAD - if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) { - params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */ - } - if (params.nbWorkers > 0) { - /* mt context creation */ - if (cctx->mtctx == NULL) { - DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u", - params.nbWorkers); - cctx->mtctx = ZSTDMT_createCCtx_advanced((U32)params.nbWorkers, cctx->customMem); - RETURN_ERROR_IF(cctx->mtctx == NULL, memory_allocation, "NULL pointer!"); - } - /* mt compression */ - DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers); - FORWARD_IF_ERROR( ZSTDMT_initCStream_internal( - cctx->mtctx, - prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, - cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , ""); - cctx->streamStage = zcss_load; - cctx->appliedParams.nbWorkers = params.nbWorkers; - } else -#endif - { FORWARD_IF_ERROR( ZSTD_resetCStream_internal(cctx, - prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, - cctx->cdict, - params, cctx->pledgedSrcSizePlusOne-1) , ""); - assert(cctx->streamStage == zcss_load); - assert(cctx->appliedParams.nbWorkers == 0); - } } + FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, endOp, input->size), "CompressStream2 initialization failed"); + ZSTD_setBufferExpectations(cctx, output, input); /* Set initial buffer expectations now that we've initialized */ + } /* end of transparent initialization stage */ + FORWARD_IF_ERROR(ZSTD_checkBufferStability(cctx, output, input, endOp), "invalid buffers"); /* compression stage */ #ifdef ZSTD_MULTITHREAD if (cctx->appliedParams.nbWorkers > 0) { - int const forceMaxProgress = (endOp == ZSTD_e_flush || endOp == ZSTD_e_end); size_t flushMin; - assert(forceMaxProgress || endOp == ZSTD_e_continue /* Protection for a new flush type */); if (cctx->cParamsChanged) { ZSTDMT_updateCParams_whileCompressing(cctx->mtctx, &cctx->requestedParams); cctx->cParamsChanged = 0; } - do { + for (;;) { + size_t const ipos = input->pos; + size_t const opos = output->pos; flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp); + cctx->consumedSrcSize += (U64)(input->pos - ipos); + cctx->producedCSize += (U64)(output->pos - opos); if ( ZSTD_isError(flushMin) || (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */ + if (flushMin == 0) + ZSTD_CCtx_trace(cctx, 0); ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only); } FORWARD_IF_ERROR(flushMin, "ZSTDMT_compressStream_generic failed"); - } while (forceMaxProgress && flushMin != 0 && output->pos < output->size); + + if (endOp == ZSTD_e_continue) { + /* We only require some progress with ZSTD_e_continue, not maximal progress. + * We're done if we've consumed or produced any bytes, or either buffer is + * full. + */ + if (input->pos != ipos || output->pos != opos || input->pos == input->size || output->pos == output->size) + break; + } else { + assert(endOp == ZSTD_e_flush || endOp == ZSTD_e_end); + /* We require maximal progress. We're done when the flush is complete or the + * output buffer is full. + */ + if (flushMin == 0 || output->pos == output->size) + break; + } + } DEBUGLOG(5, "completed ZSTD_compressStream2 delegating to ZSTDMT_compressStream_generic"); /* Either we don't require maximum forward progress, we've finished the * flush, or we are out of output space. */ - assert(!forceMaxProgress || flushMin == 0 || output->pos == output->size); + assert(endOp == ZSTD_e_continue || flushMin == 0 || output->pos == output->size); + ZSTD_setBufferExpectations(cctx, output, input); return flushMin; } #endif FORWARD_IF_ERROR( ZSTD_compressStream_generic(cctx, output, input, endOp) , ""); DEBUGLOG(5, "completed ZSTD_compressStream2"); + ZSTD_setBufferExpectations(cctx, output, input); return cctx->outBuffContentSize - cctx->outBuffFlushedSize; /* remaining to flush */ } @@ -4065,14 +5723,22 @@ size_t ZSTD_compress2(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize) { + ZSTD_bufferMode_e const originalInBufferMode = cctx->requestedParams.inBufferMode; + ZSTD_bufferMode_e const originalOutBufferMode = cctx->requestedParams.outBufferMode; DEBUGLOG(4, "ZSTD_compress2 (srcSize=%u)", (unsigned)srcSize); ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only); + /* Enable stable input/output buffers. */ + cctx->requestedParams.inBufferMode = ZSTD_bm_stable; + cctx->requestedParams.outBufferMode = ZSTD_bm_stable; { size_t oPos = 0; size_t iPos = 0; size_t const result = ZSTD_compressStream2_simpleArgs(cctx, dst, dstCapacity, &oPos, src, srcSize, &iPos, ZSTD_e_end); + /* Reset to the original values. */ + cctx->requestedParams.inBufferMode = originalInBufferMode; + cctx->requestedParams.outBufferMode = originalOutBufferMode; FORWARD_IF_ERROR(result, "ZSTD_compressStream2_simpleArgs failed"); if (result != 0) { /* compression not completed, due to lack of output space */ assert(oPos == dstCapacity); @@ -4083,6 +5749,410 @@ size_t ZSTD_compress2(ZSTD_CCtx* cctx, } } +typedef struct { + U32 idx; /* Index in array of ZSTD_Sequence */ + U32 posInSequence; /* Position within sequence at idx */ + size_t posInSrc; /* Number of bytes given by sequences provided so far */ +} ZSTD_sequencePosition; + +/* ZSTD_validateSequence() : + * @offCode : is presumed to follow format required by ZSTD_storeSeq() + * @returns a ZSTD error code if sequence is not valid + */ +static size_t +ZSTD_validateSequence(U32 offCode, U32 matchLength, + size_t posInSrc, U32 windowLog, size_t dictSize) +{ + U32 const windowSize = 1 << windowLog; + /* posInSrc represents the amount of data the the decoder would decode up to this point. + * As long as the amount of data decoded is less than or equal to window size, offsets may be + * larger than the total length of output decoded in order to reference the dict, even larger than + * window size. After output surpasses windowSize, we're limited to windowSize offsets again. + */ + size_t const offsetBound = posInSrc > windowSize ? (size_t)windowSize : posInSrc + (size_t)dictSize; + RETURN_ERROR_IF(offCode > STORE_OFFSET(offsetBound), corruption_detected, "Offset too large!"); + RETURN_ERROR_IF(matchLength < MINMATCH, corruption_detected, "Matchlength too small"); + return 0; +} + +/* Returns an offset code, given a sequence's raw offset, the ongoing repcode array, and whether litLength == 0 */ +static U32 ZSTD_finalizeOffCode(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32 ll0) +{ + U32 offCode = STORE_OFFSET(rawOffset); + + if (!ll0 && rawOffset == rep[0]) { + offCode = STORE_REPCODE_1; + } else if (rawOffset == rep[1]) { + offCode = STORE_REPCODE(2 - ll0); + } else if (rawOffset == rep[2]) { + offCode = STORE_REPCODE(3 - ll0); + } else if (ll0 && rawOffset == rep[0] - 1) { + offCode = STORE_REPCODE_3; + } + return offCode; +} + +/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of + * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter. + */ +static size_t +ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx, + ZSTD_sequencePosition* seqPos, + const ZSTD_Sequence* const inSeqs, size_t inSeqsSize, + const void* src, size_t blockSize) +{ + U32 idx = seqPos->idx; + BYTE const* ip = (BYTE const*)(src); + const BYTE* const iend = ip + blockSize; + repcodes_t updatedRepcodes; + U32 dictSize; + + if (cctx->cdict) { + dictSize = (U32)cctx->cdict->dictContentSize; + } else if (cctx->prefixDict.dict) { + dictSize = (U32)cctx->prefixDict.dictSize; + } else { + dictSize = 0; + } + ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t)); + for (; (inSeqs[idx].matchLength != 0 || inSeqs[idx].offset != 0) && idx < inSeqsSize; ++idx) { + U32 const litLength = inSeqs[idx].litLength; + U32 const ll0 = (litLength == 0); + U32 const matchLength = inSeqs[idx].matchLength; + U32 const offCode = ZSTD_finalizeOffCode(inSeqs[idx].offset, updatedRepcodes.rep, ll0); + ZSTD_updateRep(updatedRepcodes.rep, offCode, ll0); + + DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offCode, matchLength, litLength); + if (cctx->appliedParams.validateSequences) { + seqPos->posInSrc += litLength + matchLength; + FORWARD_IF_ERROR(ZSTD_validateSequence(offCode, matchLength, seqPos->posInSrc, + cctx->appliedParams.cParams.windowLog, dictSize), + "Sequence validation failed"); + } + RETURN_ERROR_IF(idx - seqPos->idx > cctx->seqStore.maxNbSeq, memory_allocation, + "Not enough memory allocated. Try adjusting ZSTD_c_minMatch."); + ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offCode, matchLength); + ip += matchLength + litLength; + } + ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t)); + + if (inSeqs[idx].litLength) { + DEBUGLOG(6, "Storing last literals of size: %u", inSeqs[idx].litLength); + ZSTD_storeLastLiterals(&cctx->seqStore, ip, inSeqs[idx].litLength); + ip += inSeqs[idx].litLength; + seqPos->posInSrc += inSeqs[idx].litLength; + } + RETURN_ERROR_IF(ip != iend, corruption_detected, "Blocksize doesn't agree with block delimiter!"); + seqPos->idx = idx+1; + return 0; +} + +/* Returns the number of bytes to move the current read position back by. Only non-zero + * if we ended up splitting a sequence. Otherwise, it may return a ZSTD error if something + * went wrong. + * + * This function will attempt to scan through blockSize bytes represented by the sequences + * in inSeqs, storing any (partial) sequences. + * + * Occasionally, we may want to change the actual number of bytes we consumed from inSeqs to + * avoid splitting a match, or to avoid splitting a match such that it would produce a match + * smaller than MINMATCH. In this case, we return the number of bytes that we didn't read from this block. + */ +static size_t +ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos, + const ZSTD_Sequence* const inSeqs, size_t inSeqsSize, + const void* src, size_t blockSize) +{ + U32 idx = seqPos->idx; + U32 startPosInSequence = seqPos->posInSequence; + U32 endPosInSequence = seqPos->posInSequence + (U32)blockSize; + size_t dictSize; + BYTE const* ip = (BYTE const*)(src); + BYTE const* iend = ip + blockSize; /* May be adjusted if we decide to process fewer than blockSize bytes */ + repcodes_t updatedRepcodes; + U32 bytesAdjustment = 0; + U32 finalMatchSplit = 0; + + if (cctx->cdict) { + dictSize = cctx->cdict->dictContentSize; + } else if (cctx->prefixDict.dict) { + dictSize = cctx->prefixDict.dictSize; + } else { + dictSize = 0; + } + DEBUGLOG(5, "ZSTD_copySequencesToSeqStore: idx: %u PIS: %u blockSize: %zu", idx, startPosInSequence, blockSize); + DEBUGLOG(5, "Start seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength); + ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t)); + while (endPosInSequence && idx < inSeqsSize && !finalMatchSplit) { + const ZSTD_Sequence currSeq = inSeqs[idx]; + U32 litLength = currSeq.litLength; + U32 matchLength = currSeq.matchLength; + U32 const rawOffset = currSeq.offset; + U32 offCode; + + /* Modify the sequence depending on where endPosInSequence lies */ + if (endPosInSequence >= currSeq.litLength + currSeq.matchLength) { + if (startPosInSequence >= litLength) { + startPosInSequence -= litLength; + litLength = 0; + matchLength -= startPosInSequence; + } else { + litLength -= startPosInSequence; + } + /* Move to the next sequence */ + endPosInSequence -= currSeq.litLength + currSeq.matchLength; + startPosInSequence = 0; + idx++; + } else { + /* This is the final (partial) sequence we're adding from inSeqs, and endPosInSequence + does not reach the end of the match. So, we have to split the sequence */ + DEBUGLOG(6, "Require a split: diff: %u, idx: %u PIS: %u", + currSeq.litLength + currSeq.matchLength - endPosInSequence, idx, endPosInSequence); + if (endPosInSequence > litLength) { + U32 firstHalfMatchLength; + litLength = startPosInSequence >= litLength ? 0 : litLength - startPosInSequence; + firstHalfMatchLength = endPosInSequence - startPosInSequence - litLength; + if (matchLength > blockSize && firstHalfMatchLength >= cctx->appliedParams.cParams.minMatch) { + /* Only ever split the match if it is larger than the block size */ + U32 secondHalfMatchLength = currSeq.matchLength + currSeq.litLength - endPosInSequence; + if (secondHalfMatchLength < cctx->appliedParams.cParams.minMatch) { + /* Move the endPosInSequence backward so that it creates match of minMatch length */ + endPosInSequence -= cctx->appliedParams.cParams.minMatch - secondHalfMatchLength; + bytesAdjustment = cctx->appliedParams.cParams.minMatch - secondHalfMatchLength; + firstHalfMatchLength -= bytesAdjustment; + } + matchLength = firstHalfMatchLength; + /* Flag that we split the last match - after storing the sequence, exit the loop, + but keep the value of endPosInSequence */ + finalMatchSplit = 1; + } else { + /* Move the position in sequence backwards so that we don't split match, and break to store + * the last literals. We use the original currSeq.litLength as a marker for where endPosInSequence + * should go. We prefer to do this whenever it is not necessary to split the match, or if doing so + * would cause the first half of the match to be too small + */ + bytesAdjustment = endPosInSequence - currSeq.litLength; + endPosInSequence = currSeq.litLength; + break; + } + } else { + /* This sequence ends inside the literals, break to store the last literals */ + break; + } + } + /* Check if this offset can be represented with a repcode */ + { U32 const ll0 = (litLength == 0); + offCode = ZSTD_finalizeOffCode(rawOffset, updatedRepcodes.rep, ll0); + ZSTD_updateRep(updatedRepcodes.rep, offCode, ll0); + } + + if (cctx->appliedParams.validateSequences) { + seqPos->posInSrc += litLength + matchLength; + FORWARD_IF_ERROR(ZSTD_validateSequence(offCode, matchLength, seqPos->posInSrc, + cctx->appliedParams.cParams.windowLog, dictSize), + "Sequence validation failed"); + } + DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offCode, matchLength, litLength); + RETURN_ERROR_IF(idx - seqPos->idx > cctx->seqStore.maxNbSeq, memory_allocation, + "Not enough memory allocated. Try adjusting ZSTD_c_minMatch."); + ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offCode, matchLength); + ip += matchLength + litLength; + } + DEBUGLOG(5, "Ending seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength); + assert(idx == inSeqsSize || endPosInSequence <= inSeqs[idx].litLength + inSeqs[idx].matchLength); + seqPos->idx = idx; + seqPos->posInSequence = endPosInSequence; + ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t)); + + iend -= bytesAdjustment; + if (ip != iend) { + /* Store any last literals */ + U32 lastLLSize = (U32)(iend - ip); + assert(ip <= iend); + DEBUGLOG(6, "Storing last literals of size: %u", lastLLSize); + ZSTD_storeLastLiterals(&cctx->seqStore, ip, lastLLSize); + seqPos->posInSrc += lastLLSize; + } + + return bytesAdjustment; +} + +typedef size_t (*ZSTD_sequenceCopier) (ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos, + const ZSTD_Sequence* const inSeqs, size_t inSeqsSize, + const void* src, size_t blockSize); +static ZSTD_sequenceCopier ZSTD_selectSequenceCopier(ZSTD_sequenceFormat_e mode) +{ + ZSTD_sequenceCopier sequenceCopier = NULL; + assert(ZSTD_cParam_withinBounds(ZSTD_c_blockDelimiters, mode)); + if (mode == ZSTD_sf_explicitBlockDelimiters) { + return ZSTD_copySequencesToSeqStoreExplicitBlockDelim; + } else if (mode == ZSTD_sf_noBlockDelimiters) { + return ZSTD_copySequencesToSeqStoreNoBlockDelim; + } + assert(sequenceCopier != NULL); + return sequenceCopier; +} + +/* Compress, block-by-block, all of the sequences given. + * + * Returns the cumulative size of all compressed blocks (including their headers), + * otherwise a ZSTD error. + */ +static size_t +ZSTD_compressSequences_internal(ZSTD_CCtx* cctx, + void* dst, size_t dstCapacity, + const ZSTD_Sequence* inSeqs, size_t inSeqsSize, + const void* src, size_t srcSize) +{ + size_t cSize = 0; + U32 lastBlock; + size_t blockSize; + size_t compressedSeqsSize; + size_t remaining = srcSize; + ZSTD_sequencePosition seqPos = {0, 0, 0}; + + BYTE const* ip = (BYTE const*)src; + BYTE* op = (BYTE*)dst; + ZSTD_sequenceCopier const sequenceCopier = ZSTD_selectSequenceCopier(cctx->appliedParams.blockDelimiters); + + DEBUGLOG(4, "ZSTD_compressSequences_internal srcSize: %zu, inSeqsSize: %zu", srcSize, inSeqsSize); + /* Special case: empty frame */ + if (remaining == 0) { + U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1); + RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "No room for empty frame block header"); + MEM_writeLE32(op, cBlockHeader24); + op += ZSTD_blockHeaderSize; + dstCapacity -= ZSTD_blockHeaderSize; + cSize += ZSTD_blockHeaderSize; + } + + while (remaining) { + size_t cBlockSize; + size_t additionalByteAdjustment; + lastBlock = remaining <= cctx->blockSize; + blockSize = lastBlock ? (U32)remaining : (U32)cctx->blockSize; + ZSTD_resetSeqStore(&cctx->seqStore); + DEBUGLOG(4, "Working on new block. Blocksize: %zu", blockSize); + + additionalByteAdjustment = sequenceCopier(cctx, &seqPos, inSeqs, inSeqsSize, ip, blockSize); + FORWARD_IF_ERROR(additionalByteAdjustment, "Bad sequence copy"); + blockSize -= additionalByteAdjustment; + + /* If blocks are too small, emit as a nocompress block */ + if (blockSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) { + cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock); + FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed"); + DEBUGLOG(4, "Block too small, writing out nocompress block: cSize: %zu", cBlockSize); + cSize += cBlockSize; + ip += blockSize; + op += cBlockSize; + remaining -= blockSize; + dstCapacity -= cBlockSize; + continue; + } + + compressedSeqsSize = ZSTD_entropyCompressSeqStore(&cctx->seqStore, + &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy, + &cctx->appliedParams, + op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize, + blockSize, + cctx->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */, + cctx->bmi2); + FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed"); + DEBUGLOG(4, "Compressed sequences size: %zu", compressedSeqsSize); + + if (!cctx->isFirstBlock && + ZSTD_maybeRLE(&cctx->seqStore) && + ZSTD_isRLE((BYTE const*)src, srcSize)) { + /* We don't want to emit our first block as a RLE even if it qualifies because + * doing so will cause the decoder (cli only) to throw a "should consume all input error." + * This is only an issue for zstd <= v1.4.3 + */ + compressedSeqsSize = 1; + } + + if (compressedSeqsSize == 0) { + /* ZSTD_noCompressBlock writes the block header as well */ + cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock); + FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed"); + DEBUGLOG(4, "Writing out nocompress block, size: %zu", cBlockSize); + } else if (compressedSeqsSize == 1) { + cBlockSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, blockSize, lastBlock); + FORWARD_IF_ERROR(cBlockSize, "RLE compress block failed"); + DEBUGLOG(4, "Writing out RLE block, size: %zu", cBlockSize); + } else { + U32 cBlockHeader; + /* Error checking and repcodes update */ + ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState); + if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid) + cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check; + + /* Write block header into beginning of block*/ + cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3); + MEM_writeLE24(op, cBlockHeader); + cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize; + DEBUGLOG(4, "Writing out compressed block, size: %zu", cBlockSize); + } + + cSize += cBlockSize; + DEBUGLOG(4, "cSize running total: %zu", cSize); + + if (lastBlock) { + break; + } else { + ip += blockSize; + op += cBlockSize; + remaining -= blockSize; + dstCapacity -= cBlockSize; + cctx->isFirstBlock = 0; + } + } + + return cSize; +} + +size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstCapacity, + const ZSTD_Sequence* inSeqs, size_t inSeqsSize, + const void* src, size_t srcSize) +{ + BYTE* op = (BYTE*)dst; + size_t cSize = 0; + size_t compressedBlocksSize = 0; + size_t frameHeaderSize = 0; + + /* Transparent initialization stage, same as compressStream2() */ + DEBUGLOG(3, "ZSTD_compressSequences()"); + assert(cctx != NULL); + FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, srcSize), "CCtx initialization failed"); + /* Begin writing output, starting with frame header */ + frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity, &cctx->appliedParams, srcSize, cctx->dictID); + op += frameHeaderSize; + dstCapacity -= frameHeaderSize; + cSize += frameHeaderSize; + if (cctx->appliedParams.fParams.checksumFlag && srcSize) { + XXH64_update(&cctx->xxhState, src, srcSize); + } + /* cSize includes block header size and compressed sequences size */ + compressedBlocksSize = ZSTD_compressSequences_internal(cctx, + op, dstCapacity, + inSeqs, inSeqsSize, + src, srcSize); + FORWARD_IF_ERROR(compressedBlocksSize, "Compressing blocks failed!"); + cSize += compressedBlocksSize; + dstCapacity -= compressedBlocksSize; + + if (cctx->appliedParams.fParams.checksumFlag) { + U32 const checksum = (U32) XXH64_digest(&cctx->xxhState); + RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for checksum"); + DEBUGLOG(4, "Write checksum : %08X", (unsigned)checksum); + MEM_writeLE32((char*)dst + cSize, checksum); + cSize += 4; + } + + DEBUGLOG(3, "Final compressed size: %zu", cSize); + return cSize; +} + /*====== Finalize ======*/ /*! ZSTD_flushStream() : @@ -4111,137 +6181,116 @@ size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output) /*-===== Pre-defined compression levels =====-*/ +#include "clevels.h" -#define ZSTD_MAX_CLEVEL 22 int ZSTD_maxCLevel(void) { return ZSTD_MAX_CLEVEL; } int ZSTD_minCLevel(void) { return (int)-ZSTD_TARGETLENGTH_MAX; } +int ZSTD_defaultCLevel(void) { return ZSTD_CLEVEL_DEFAULT; } -static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = { -{ /* "default" - for any srcSize > 256 KB */ - /* W, C, H, S, L, TL, strat */ - { 19, 12, 13, 1, 6, 1, ZSTD_fast }, /* base for negative levels */ - { 19, 13, 14, 1, 7, 0, ZSTD_fast }, /* level 1 */ - { 20, 15, 16, 1, 6, 0, ZSTD_fast }, /* level 2 */ - { 21, 16, 17, 1, 5, 0, ZSTD_dfast }, /* level 3 */ - { 21, 18, 18, 1, 5, 0, ZSTD_dfast }, /* level 4 */ - { 21, 18, 19, 2, 5, 2, ZSTD_greedy }, /* level 5 */ - { 21, 19, 19, 3, 5, 4, ZSTD_greedy }, /* level 6 */ - { 21, 19, 19, 3, 5, 8, ZSTD_lazy }, /* level 7 */ - { 21, 19, 19, 3, 5, 16, ZSTD_lazy2 }, /* level 8 */ - { 21, 19, 20, 4, 5, 16, ZSTD_lazy2 }, /* level 9 */ - { 22, 20, 21, 4, 5, 16, ZSTD_lazy2 }, /* level 10 */ - { 22, 21, 22, 4, 5, 16, ZSTD_lazy2 }, /* level 11 */ - { 22, 21, 22, 5, 5, 16, ZSTD_lazy2 }, /* level 12 */ - { 22, 21, 22, 5, 5, 32, ZSTD_btlazy2 }, /* level 13 */ - { 22, 22, 23, 5, 5, 32, ZSTD_btlazy2 }, /* level 14 */ - { 22, 23, 23, 6, 5, 32, ZSTD_btlazy2 }, /* level 15 */ - { 22, 22, 22, 5, 5, 48, ZSTD_btopt }, /* level 16 */ - { 23, 23, 22, 5, 4, 64, ZSTD_btopt }, /* level 17 */ - { 23, 23, 22, 6, 3, 64, ZSTD_btultra }, /* level 18 */ - { 23, 24, 22, 7, 3,256, ZSTD_btultra2}, /* level 19 */ - { 25, 25, 23, 7, 3,256, ZSTD_btultra2}, /* level 20 */ - { 26, 26, 24, 7, 3,512, ZSTD_btultra2}, /* level 21 */ - { 27, 27, 25, 9, 3,999, ZSTD_btultra2}, /* level 22 */ -}, -{ /* for srcSize <= 256 KB */ - /* W, C, H, S, L, T, strat */ - { 18, 12, 13, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ - { 18, 13, 14, 1, 6, 0, ZSTD_fast }, /* level 1 */ - { 18, 14, 14, 1, 5, 0, ZSTD_dfast }, /* level 2 */ - { 18, 16, 16, 1, 4, 0, ZSTD_dfast }, /* level 3 */ - { 18, 16, 17, 2, 5, 2, ZSTD_greedy }, /* level 4.*/ - { 18, 18, 18, 3, 5, 2, ZSTD_greedy }, /* level 5.*/ - { 18, 18, 19, 3, 5, 4, ZSTD_lazy }, /* level 6.*/ - { 18, 18, 19, 4, 4, 4, ZSTD_lazy }, /* level 7 */ - { 18, 18, 19, 4, 4, 8, ZSTD_lazy2 }, /* level 8 */ - { 18, 18, 19, 5, 4, 8, ZSTD_lazy2 }, /* level 9 */ - { 18, 18, 19, 6, 4, 8, ZSTD_lazy2 }, /* level 10 */ - { 18, 18, 19, 5, 4, 12, ZSTD_btlazy2 }, /* level 11.*/ - { 18, 19, 19, 7, 4, 12, ZSTD_btlazy2 }, /* level 12.*/ - { 18, 18, 19, 4, 4, 16, ZSTD_btopt }, /* level 13 */ - { 18, 18, 19, 4, 3, 32, ZSTD_btopt }, /* level 14.*/ - { 18, 18, 19, 6, 3,128, ZSTD_btopt }, /* level 15.*/ - { 18, 19, 19, 6, 3,128, ZSTD_btultra }, /* level 16.*/ - { 18, 19, 19, 8, 3,256, ZSTD_btultra }, /* level 17.*/ - { 18, 19, 19, 6, 3,128, ZSTD_btultra2}, /* level 18.*/ - { 18, 19, 19, 8, 3,256, ZSTD_btultra2}, /* level 19.*/ - { 18, 19, 19, 10, 3,512, ZSTD_btultra2}, /* level 20.*/ - { 18, 19, 19, 12, 3,512, ZSTD_btultra2}, /* level 21.*/ - { 18, 19, 19, 13, 3,999, ZSTD_btultra2}, /* level 22.*/ -}, -{ /* for srcSize <= 128 KB */ - /* W, C, H, S, L, T, strat */ - { 17, 12, 12, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ - { 17, 12, 13, 1, 6, 0, ZSTD_fast }, /* level 1 */ - { 17, 13, 15, 1, 5, 0, ZSTD_fast }, /* level 2 */ - { 17, 15, 16, 2, 5, 0, ZSTD_dfast }, /* level 3 */ - { 17, 17, 17, 2, 4, 0, ZSTD_dfast }, /* level 4 */ - { 17, 16, 17, 3, 4, 2, ZSTD_greedy }, /* level 5 */ - { 17, 17, 17, 3, 4, 4, ZSTD_lazy }, /* level 6 */ - { 17, 17, 17, 3, 4, 8, ZSTD_lazy2 }, /* level 7 */ - { 17, 17, 17, 4, 4, 8, ZSTD_lazy2 }, /* level 8 */ - { 17, 17, 17, 5, 4, 8, ZSTD_lazy2 }, /* level 9 */ - { 17, 17, 17, 6, 4, 8, ZSTD_lazy2 }, /* level 10 */ - { 17, 17, 17, 5, 4, 8, ZSTD_btlazy2 }, /* level 11 */ - { 17, 18, 17, 7, 4, 12, ZSTD_btlazy2 }, /* level 12 */ - { 17, 18, 17, 3, 4, 12, ZSTD_btopt }, /* level 13.*/ - { 17, 18, 17, 4, 3, 32, ZSTD_btopt }, /* level 14.*/ - { 17, 18, 17, 6, 3,256, ZSTD_btopt }, /* level 15.*/ - { 17, 18, 17, 6, 3,128, ZSTD_btultra }, /* level 16.*/ - { 17, 18, 17, 8, 3,256, ZSTD_btultra }, /* level 17.*/ - { 17, 18, 17, 10, 3,512, ZSTD_btultra }, /* level 18.*/ - { 17, 18, 17, 5, 3,256, ZSTD_btultra2}, /* level 19.*/ - { 17, 18, 17, 7, 3,512, ZSTD_btultra2}, /* level 20.*/ - { 17, 18, 17, 9, 3,512, ZSTD_btultra2}, /* level 21.*/ - { 17, 18, 17, 11, 3,999, ZSTD_btultra2}, /* level 22.*/ -}, -{ /* for srcSize <= 16 KB */ - /* W, C, H, S, L, T, strat */ - { 14, 12, 13, 1, 5, 1, ZSTD_fast }, /* base for negative levels */ - { 14, 14, 15, 1, 5, 0, ZSTD_fast }, /* level 1 */ - { 14, 14, 15, 1, 4, 0, ZSTD_fast }, /* level 2 */ - { 14, 14, 15, 2, 4, 0, ZSTD_dfast }, /* level 3 */ - { 14, 14, 14, 4, 4, 2, ZSTD_greedy }, /* level 4 */ - { 14, 14, 14, 3, 4, 4, ZSTD_lazy }, /* level 5.*/ - { 14, 14, 14, 4, 4, 8, ZSTD_lazy2 }, /* level 6 */ - { 14, 14, 14, 6, 4, 8, ZSTD_lazy2 }, /* level 7 */ - { 14, 14, 14, 8, 4, 8, ZSTD_lazy2 }, /* level 8.*/ - { 14, 15, 14, 5, 4, 8, ZSTD_btlazy2 }, /* level 9.*/ - { 14, 15, 14, 9, 4, 8, ZSTD_btlazy2 }, /* level 10.*/ - { 14, 15, 14, 3, 4, 12, ZSTD_btopt }, /* level 11.*/ - { 14, 15, 14, 4, 3, 24, ZSTD_btopt }, /* level 12.*/ - { 14, 15, 14, 5, 3, 32, ZSTD_btultra }, /* level 13.*/ - { 14, 15, 15, 6, 3, 64, ZSTD_btultra }, /* level 14.*/ - { 14, 15, 15, 7, 3,256, ZSTD_btultra }, /* level 15.*/ - { 14, 15, 15, 5, 3, 48, ZSTD_btultra2}, /* level 16.*/ - { 14, 15, 15, 6, 3,128, ZSTD_btultra2}, /* level 17.*/ - { 14, 15, 15, 7, 3,256, ZSTD_btultra2}, /* level 18.*/ - { 14, 15, 15, 8, 3,256, ZSTD_btultra2}, /* level 19.*/ - { 14, 15, 15, 8, 3,512, ZSTD_btultra2}, /* level 20.*/ - { 14, 15, 15, 9, 3,512, ZSTD_btultra2}, /* level 21.*/ - { 14, 15, 15, 10, 3,999, ZSTD_btultra2}, /* level 22.*/ -}, -}; +static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(int const compressionLevel, size_t const dictSize) +{ + ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, 0, dictSize, ZSTD_cpm_createCDict); + switch (cParams.strategy) { + case ZSTD_fast: + case ZSTD_dfast: + break; + case ZSTD_greedy: + case ZSTD_lazy: + case ZSTD_lazy2: + cParams.hashLog += ZSTD_LAZY_DDSS_BUCKET_LOG; + break; + case ZSTD_btlazy2: + case ZSTD_btopt: + case ZSTD_btultra: + case ZSTD_btultra2: + break; + } + return cParams; +} + +static int ZSTD_dedicatedDictSearch_isSupported( + ZSTD_compressionParameters const* cParams) +{ + return (cParams->strategy >= ZSTD_greedy) + && (cParams->strategy <= ZSTD_lazy2) + && (cParams->hashLog > cParams->chainLog) + && (cParams->chainLog <= 24); +} + +/** + * Reverses the adjustment applied to cparams when enabling dedicated dict + * search. This is used to recover the params set to be used in the working + * context. (Otherwise, those tables would also grow.) + */ +static void ZSTD_dedicatedDictSearch_revertCParams( + ZSTD_compressionParameters* cParams) { + switch (cParams->strategy) { + case ZSTD_fast: + case ZSTD_dfast: + break; + case ZSTD_greedy: + case ZSTD_lazy: + case ZSTD_lazy2: + cParams->hashLog -= ZSTD_LAZY_DDSS_BUCKET_LOG; + if (cParams->hashLog < ZSTD_HASHLOG_MIN) { + cParams->hashLog = ZSTD_HASHLOG_MIN; + } + break; + case ZSTD_btlazy2: + case ZSTD_btopt: + case ZSTD_btultra: + case ZSTD_btultra2: + break; + } +} + +static U64 ZSTD_getCParamRowSize(U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) +{ + switch (mode) { + case ZSTD_cpm_unknown: + case ZSTD_cpm_noAttachDict: + case ZSTD_cpm_createCDict: + break; + case ZSTD_cpm_attachDict: + dictSize = 0; + break; + default: + assert(0); + break; + } + { int const unknown = srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN; + size_t const addedSize = unknown && dictSize > 0 ? 500 : 0; + return unknown && dictSize == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : srcSizeHint+dictSize+addedSize; + } +} /*! ZSTD_getCParams_internal() : * @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize. * Note: srcSizeHint 0 means 0, use ZSTD_CONTENTSIZE_UNKNOWN for unknown. - * Use dictSize == 0 for unknown or unused. */ -static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) + * Use dictSize == 0 for unknown or unused. + * Note: `mode` controls how we treat the `dictSize`. See docs for `ZSTD_cParamMode_e`. */ +static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) { - int const unknown = srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN; - size_t const addedSize = unknown && dictSize > 0 ? 500 : 0; - U64 const rSize = unknown && dictSize == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : srcSizeHint+dictSize+addedSize; + U64 const rSize = ZSTD_getCParamRowSize(srcSizeHint, dictSize, mode); U32 const tableID = (rSize <= 256 KB) + (rSize <= 128 KB) + (rSize <= 16 KB); - int row = compressionLevel; + int row; DEBUGLOG(5, "ZSTD_getCParams_internal (cLevel=%i)", compressionLevel); + + /* row */ if (compressionLevel == 0) row = ZSTD_CLEVEL_DEFAULT; /* 0 == default */ - if (compressionLevel < 0) row = 0; /* entry 0 is baseline for fast mode */ - if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL; + else if (compressionLevel < 0) row = 0; /* entry 0 is baseline for fast mode */ + else if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL; + else row = compressionLevel; + { ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row]; - if (compressionLevel < 0) cp.targetLength = (unsigned)(-compressionLevel); /* acceleration factor */ + DEBUGLOG(5, "ZSTD_getCParams_internal selected tableID: %u row: %u strat: %u", tableID, row, (U32)cp.strategy); + /* acceleration factor */ + if (compressionLevel < 0) { + int const clampedCompressionLevel = MAX(ZSTD_minCLevel(), compressionLevel); + cp.targetLength = (unsigned)(-clampedCompressionLevel); + } /* refine parameters based on srcSize & dictSize */ - return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize); + return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize, mode); } } @@ -4251,18 +6300,18 @@ static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) { if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN; - return ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize); + return ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown); } /*! ZSTD_getParams() : * same idea as ZSTD_getCParams() * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`). * Fields of `ZSTD_frameParameters` are set to default values */ -static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) { +static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) { ZSTD_parameters params; - ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize); + ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, mode); DEBUGLOG(5, "ZSTD_getParams (cLevel=%i)", compressionLevel); - memset(¶ms, 0, sizeof(params)); + ZSTD_memset(¶ms, 0, sizeof(params)); params.cParams = cParams; params.fParams.contentSizeFlag = 1; return params; @@ -4274,5 +6323,5 @@ static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned lo * Fields of `ZSTD_frameParameters` are set to default values */ ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) { if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN; - return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize); + return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown); } diff --git a/thirdparty/zstd/compress/zstd_compress_internal.h b/thirdparty/zstd/compress/zstd_compress_internal.h index db73f6ce21..c406e794bd 100644 --- a/thirdparty/zstd/compress/zstd_compress_internal.h +++ b/thirdparty/zstd/compress/zstd_compress_internal.h @@ -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 @@ -28,7 +28,6 @@ extern "C" { #endif - /*-************************************* * Constants ***************************************/ @@ -64,7 +63,7 @@ typedef struct { } ZSTD_localDict; typedef struct { - U32 CTable[HUF_CTABLE_SIZE_U32(255)]; + HUF_CElt CTable[HUF_CTABLE_SIZE_ST(255)]; HUF_repeat repeatMode; } ZSTD_hufCTables_t; @@ -82,12 +81,76 @@ typedef struct { ZSTD_fseCTables_t fse; } ZSTD_entropyCTables_t; +/*********************************************** +* Entropy buffer statistics structs and funcs * +***********************************************/ +/** ZSTD_hufCTablesMetadata_t : + * Stores Literals Block Type for a super-block in hType, and + * huffman tree description in hufDesBuffer. + * hufDesSize refers to the size of huffman tree description in bytes. + * This metadata is populated in ZSTD_buildBlockEntropyStats_literals() */ typedef struct { - U32 off; - U32 len; + symbolEncodingType_e hType; + BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE]; + size_t hufDesSize; +} ZSTD_hufCTablesMetadata_t; + +/** ZSTD_fseCTablesMetadata_t : + * Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and + * fse tables in fseTablesBuffer. + * fseTablesSize refers to the size of fse tables in bytes. + * This metadata is populated in ZSTD_buildBlockEntropyStats_sequences() */ +typedef struct { + symbolEncodingType_e llType; + symbolEncodingType_e ofType; + symbolEncodingType_e mlType; + BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE]; + size_t fseTablesSize; + size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */ +} ZSTD_fseCTablesMetadata_t; + +typedef struct { + ZSTD_hufCTablesMetadata_t hufMetadata; + ZSTD_fseCTablesMetadata_t fseMetadata; +} ZSTD_entropyCTablesMetadata_t; + +/** ZSTD_buildBlockEntropyStats() : + * Builds entropy for the block. + * @return : 0 on success or error code */ +size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr, + const ZSTD_entropyCTables_t* prevEntropy, + ZSTD_entropyCTables_t* nextEntropy, + const ZSTD_CCtx_params* cctxParams, + ZSTD_entropyCTablesMetadata_t* entropyMetadata, + void* workspace, size_t wkspSize); + +/********************************* +* Compression internals structs * +*********************************/ + +typedef struct { + U32 off; /* Offset sumtype code for the match, using ZSTD_storeSeq() format */ + U32 len; /* Raw length of match */ } ZSTD_match_t; typedef struct { + U32 offset; /* Offset of sequence */ + U32 litLength; /* Length of literals prior to match */ + U32 matchLength; /* Raw length of match */ +} rawSeq; + +typedef struct { + rawSeq* seq; /* The start of the sequences */ + size_t pos; /* The index in seq where reading stopped. pos <= size. */ + size_t posInSequence; /* The position within the sequence at seq[pos] where reading + stopped. posInSequence <= seq[pos].litLength + seq[pos].matchLength */ + size_t size; /* The number of sequences. <= capacity. */ + size_t capacity; /* The capacity starting from `seq` pointer */ +} rawSeqStore_t; + +UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0}; + +typedef struct { int price; U32 off; U32 mlen; @@ -116,7 +179,7 @@ typedef struct { U32 offCodeSumBasePrice; /* to compare to log2(offreq) */ ZSTD_OptPrice_e priceType; /* prices can be determined dynamically, or follow a pre-defined cost structure */ const ZSTD_entropyCTables_t* symbolCosts; /* pre-calculated dictionary statistics */ - ZSTD_literalCompressionMode_e literalCompressionMode; + ZSTD_paramSwitch_e literalCompressionMode; } optState_t; typedef struct { @@ -125,14 +188,23 @@ typedef struct { } ZSTD_compressedBlockState_t; typedef struct { - BYTE const* nextSrc; /* next block here to continue on current prefix */ - BYTE const* base; /* All regular indexes relative to this position */ - BYTE const* dictBase; /* extDict indexes relative to this position */ - U32 dictLimit; /* below that point, need extDict */ - U32 lowLimit; /* below that point, no more valid data */ + BYTE const* nextSrc; /* next block here to continue on current prefix */ + BYTE const* base; /* All regular indexes relative to this position */ + BYTE const* dictBase; /* extDict indexes relative to this position */ + U32 dictLimit; /* below that point, need extDict */ + U32 lowLimit; /* below that point, no more valid data */ + U32 nbOverflowCorrections; /* Number of times overflow correction has run since + * ZSTD_window_init(). Useful for debugging coredumps + * and for ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY. + */ } ZSTD_window_t; +#define ZSTD_WINDOW_START_INDEX 2 + typedef struct ZSTD_matchState_t ZSTD_matchState_t; + +#define ZSTD_ROW_HASH_CACHE_SIZE 8 /* Size of prefetching hash cache for row-based matchfinder */ + struct ZSTD_matchState_t { ZSTD_window_t window; /* State for window round buffer management */ U32 loadedDictEnd; /* index of end of dictionary, within context's referential. @@ -144,12 +216,24 @@ struct ZSTD_matchState_t { */ U32 nextToUpdate; /* index from which to continue table update */ U32 hashLog3; /* dispatch table for matches of len==3 : larger == faster, more memory */ + + U32 rowHashLog; /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/ + U16* tagTable; /* For row-based matchFinder: A row-based table containing the hashes and head index. */ + U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */ + U32* hashTable; U32* hashTable3; U32* chainTable; + + U32 forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */ + + int dedicatedDictSearch; /* Indicates whether this matchState is using the + * dedicated dictionary search structure. + */ optState_t opt; /* optimal parser state */ const ZSTD_matchState_t* dictMatchState; ZSTD_compressionParameters cParams; + const rawSeqStore_t* ldmSeqStore; }; typedef struct { @@ -164,16 +248,25 @@ typedef struct { } ldmEntry_t; typedef struct { + BYTE const* split; + U32 hash; + U32 checksum; + ldmEntry_t* bucket; +} ldmMatchCandidate_t; + +#define LDM_BATCH_SIZE 64 + +typedef struct { ZSTD_window_t window; /* State for the window round buffer management */ ldmEntry_t* hashTable; U32 loadedDictEnd; BYTE* bucketOffsets; /* Next position in bucket to insert entry */ - U64 hashPower; /* Used to compute the rolling hash. - * Depends on ldmParams.minMatchLength */ + size_t splitIndices[LDM_BATCH_SIZE]; + ldmMatchCandidate_t matchCandidates[LDM_BATCH_SIZE]; } ldmState_t; typedef struct { - U32 enableLdm; /* 1 if enable long distance matching */ + ZSTD_paramSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */ U32 hashLog; /* Log size of hashTable */ U32 bucketSizeLog; /* Log bucket size for collision resolution, at most 8 */ U32 minMatchLength; /* Minimum match length */ @@ -182,19 +275,6 @@ typedef struct { } ldmParams_t; typedef struct { - U32 offset; - U32 litLength; - U32 matchLength; -} rawSeq; - -typedef struct { - rawSeq* seq; /* The start of the sequences */ - size_t pos; /* The position where reading stopped. <= size. */ - size_t size; /* The number of sequences. <= capacity. */ - size_t capacity; /* The capacity starting from `seq` pointer */ -} rawSeqStore_t; - -typedef struct { int collectSequences; ZSTD_Sequence* seqStart; size_t seqIndex; @@ -217,7 +297,7 @@ struct ZSTD_CCtx_params_s { * There is no guarantee that hint is close to actual source size */ ZSTD_dictAttachPref_e attachDictPref; - ZSTD_literalCompressionMode_e literalCompressionMode; + ZSTD_paramSwitch_e literalCompressionMode; /* Multithreading: used to pass parameters to mtctx */ int nbWorkers; @@ -228,17 +308,68 @@ struct ZSTD_CCtx_params_s { /* Long distance matching parameters */ ldmParams_t ldmParams; + /* Dedicated dict search algorithm trigger */ + int enableDedicatedDictSearch; + + /* Input/output buffer modes */ + ZSTD_bufferMode_e inBufferMode; + ZSTD_bufferMode_e outBufferMode; + + /* Sequence compression API */ + ZSTD_sequenceFormat_e blockDelimiters; + int validateSequences; + + /* Block splitting */ + ZSTD_paramSwitch_e useBlockSplitter; + + /* Param for deciding whether to use row-based matchfinder */ + ZSTD_paramSwitch_e useRowMatchFinder; + + /* Always load a dictionary in ext-dict mode (not prefix mode)? */ + int deterministicRefPrefix; + /* Internal use, for createCCtxParams() and freeCCtxParams() only */ ZSTD_customMem customMem; }; /* typedef'd to ZSTD_CCtx_params within "zstd.h" */ +#define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2)) +#define ENTROPY_WORKSPACE_SIZE (HUF_WORKSPACE_SIZE + COMPRESS_SEQUENCES_WORKSPACE_SIZE) + +/** + * Indicates whether this compression proceeds directly from user-provided + * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or + * whether the context needs to buffer the input/output (ZSTDb_buffered). + */ +typedef enum { + ZSTDb_not_buffered, + ZSTDb_buffered +} ZSTD_buffered_policy_e; + +/** + * Struct that contains all elements of block splitter that should be allocated + * in a wksp. + */ +#define ZSTD_MAX_NB_BLOCK_SPLITS 196 +typedef struct { + seqStore_t fullSeqStoreChunk; + seqStore_t firstHalfSeqStore; + seqStore_t secondHalfSeqStore; + seqStore_t currSeqStore; + seqStore_t nextSeqStore; + + U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS]; + ZSTD_entropyCTablesMetadata_t entropyMetadata; +} ZSTD_blockSplitCtx; + struct ZSTD_CCtx_s { ZSTD_compressionStage_e stage; int cParamsChanged; /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */ int bmi2; /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */ ZSTD_CCtx_params requestedParams; ZSTD_CCtx_params appliedParams; + ZSTD_CCtx_params simpleApiParams; /* Param storage used by the simple API - not sticky. Must only be used in top-level simple API functions for storage. */ U32 dictID; + size_t dictContentSize; ZSTD_cwksp workspace; /* manages buffer for dynamic allocations */ size_t blockSize; @@ -247,6 +378,7 @@ struct ZSTD_CCtx_s { unsigned long long producedCSize; XXH64_state_t xxhState; ZSTD_customMem customMem; + ZSTD_threadPool* pool; size_t staticSize; SeqCollector seqCollector; int isFirstBlock; @@ -258,7 +390,10 @@ struct ZSTD_CCtx_s { size_t maxNbLdmSequences; rawSeqStore_t externSeqStore; /* Mutable reference to external sequences */ ZSTD_blockState_t blockState; - U32* entropyWorkspace; /* entropy workspace of HUF_WORKSPACE_SIZE bytes */ + U32* entropyWorkspace; /* entropy workspace of ENTROPY_WORKSPACE_SIZE bytes */ + + /* Whether we are streaming or not */ + ZSTD_buffered_policy_e bufferedPolicy; /* streaming */ char* inBuff; @@ -273,6 +408,10 @@ struct ZSTD_CCtx_s { ZSTD_cStreamStage streamStage; U32 frameEnded; + /* Stable in/out buffer verification */ + ZSTD_inBuffer expectedInBuffer; + size_t expectedOutBufferSize; + /* Dictionary */ ZSTD_localDict localDict; const ZSTD_CDict* cdict; @@ -282,17 +421,49 @@ struct ZSTD_CCtx_s { #ifdef ZSTD_MULTITHREAD ZSTDMT_CCtx* mtctx; #endif + + /* Tracing */ +#if ZSTD_TRACE + ZSTD_TraceCtx traceCtx; +#endif + + /* Workspace for block splitter */ + ZSTD_blockSplitCtx blockSplitCtx; }; typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e; -typedef enum { ZSTD_noDict = 0, ZSTD_extDict = 1, ZSTD_dictMatchState = 2 } ZSTD_dictMode_e; - +typedef enum { + ZSTD_noDict = 0, + ZSTD_extDict = 1, + ZSTD_dictMatchState = 2, + ZSTD_dedicatedDictSearch = 3 +} ZSTD_dictMode_e; + +typedef enum { + ZSTD_cpm_noAttachDict = 0, /* Compression with ZSTD_noDict or ZSTD_extDict. + * In this mode we use both the srcSize and the dictSize + * when selecting and adjusting parameters. + */ + ZSTD_cpm_attachDict = 1, /* Compression with ZSTD_dictMatchState or ZSTD_dedicatedDictSearch. + * In this mode we only take the srcSize into account when selecting + * and adjusting parameters. + */ + ZSTD_cpm_createCDict = 2, /* Creating a CDict. + * In this mode we take both the source size and the dictionary size + * into account when selecting and adjusting the parameters. + */ + ZSTD_cpm_unknown = 3, /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams. + * We don't know what these parameters are for. We default to the legacy + * behavior of taking both the source size and the dict size into account + * when selecting and adjusting parameters. + */ +} ZSTD_cParamMode_e; typedef size_t (*ZSTD_blockCompressor) ( ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); -ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode); +ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode); MEM_STATIC U32 ZSTD_LLcode(U32 litLength) @@ -326,31 +497,6 @@ MEM_STATIC U32 ZSTD_MLcode(U32 mlBase) return (mlBase > 127) ? ZSTD_highbit32(mlBase) + ML_deltaCode : ML_Code[mlBase]; } -typedef struct repcodes_s { - U32 rep[3]; -} repcodes_t; - -MEM_STATIC repcodes_t ZSTD_updateRep(U32 const rep[3], U32 const offset, U32 const ll0) -{ - repcodes_t newReps; - if (offset >= ZSTD_REP_NUM) { /* full offset */ - newReps.rep[2] = rep[1]; - newReps.rep[1] = rep[0]; - newReps.rep[0] = offset - ZSTD_REP_MOVE; - } else { /* repcode */ - U32 const repCode = offset + ll0; - if (repCode > 0) { /* note : if repCode==0, no change */ - U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode]; - newReps.rep[2] = (repCode >= 2) ? rep[1] : rep[2]; - newReps.rep[1] = rep[0]; - newReps.rep[0] = currentOffset; - } else { /* repCode == 0 */ - memcpy(&newReps, rep, sizeof(newReps)); - } - } - return newReps; -} - /* ZSTD_cParam_withinBounds: * @return 1 if value is within cParam bounds, * 0 otherwise */ @@ -372,7 +518,7 @@ MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const voi RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity, dstSize_tooSmall, "dst buf too small for uncompressed block"); MEM_writeLE24(dst, cBlockHeader24); - memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize); + ZSTD_memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize); return ZSTD_blockHeaderSize + srcSize; } @@ -399,17 +545,17 @@ MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat) return (srcSize >> minlog) + 2; } -MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParams) +MEM_STATIC int ZSTD_literalsCompressionIsDisabled(const ZSTD_CCtx_params* cctxParams) { switch (cctxParams->literalCompressionMode) { - case ZSTD_lcm_huffman: + case ZSTD_ps_enable: return 0; - case ZSTD_lcm_uncompressed: + case ZSTD_ps_disable: return 1; default: assert(0 /* impossible: pre-validated */); - /* fall-through */ - case ZSTD_lcm_auto: + ZSTD_FALLTHROUGH; + case ZSTD_ps_auto: return (cctxParams->cParams.strategy == ZSTD_fast) && (cctxParams->cParams.targetLength > 0); } } @@ -419,7 +565,9 @@ MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParam * Only called when the sequence ends past ilimit_w, so it only needs to be optimized for single * large copies. */ -static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w) { +static void +ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w) +{ assert(iend > ilimit_w); if (ip <= ilimit_w) { ZSTD_wildcopy(op, ip, ilimit_w - ip, ZSTD_no_overlap); @@ -429,14 +577,30 @@ static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const ie while (ip < iend) *op++ = *ip++; } +#define ZSTD_REP_MOVE (ZSTD_REP_NUM-1) +#define STORE_REPCODE_1 STORE_REPCODE(1) +#define STORE_REPCODE_2 STORE_REPCODE(2) +#define STORE_REPCODE_3 STORE_REPCODE(3) +#define STORE_REPCODE(r) (assert((r)>=1), assert((r)<=3), (r)-1) +#define STORE_OFFSET(o) (assert((o)>0), o + ZSTD_REP_MOVE) +#define STORED_IS_OFFSET(o) ((o) > ZSTD_REP_MOVE) +#define STORED_IS_REPCODE(o) ((o) <= ZSTD_REP_MOVE) +#define STORED_OFFSET(o) (assert(STORED_IS_OFFSET(o)), (o)-ZSTD_REP_MOVE) +#define STORED_REPCODE(o) (assert(STORED_IS_REPCODE(o)), (o)+1) /* returns ID 1,2,3 */ +#define STORED_TO_OFFBASE(o) ((o)+1) +#define OFFBASE_TO_STORED(o) ((o)-1) + /*! ZSTD_storeSeq() : - * Store a sequence (litlen, litPtr, offCode and mlBase) into seqStore_t. - * `offCode` : distance to match + ZSTD_REP_MOVE (values <= ZSTD_REP_MOVE are repCodes). - * `mlBase` : matchLength - MINMATCH + * Store a sequence (litlen, litPtr, offCode and matchLength) into seqStore_t. + * @offBase_minus1 : Users should use employ macros STORE_REPCODE_X and STORE_OFFSET(). + * @matchLength : must be >= MINMATCH * Allowed to overread literals up to litLimit. */ -HINT_INLINE UNUSED_ATTR -void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* literals, const BYTE* litLimit, U32 offCode, size_t mlBase) +HINT_INLINE UNUSED_ATTR void +ZSTD_storeSeq(seqStore_t* seqStorePtr, + size_t litLength, const BYTE* literals, const BYTE* litLimit, + U32 offBase_minus1, + size_t matchLength) { BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH; BYTE const* const litEnd = literals + litLength; @@ -445,7 +609,7 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera if (g_start==NULL) g_start = (const BYTE*)literals; /* note : index only works for compression within a single segment */ { U32 const pos = (U32)((const BYTE*)literals - g_start); DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offCode%7u", - pos, (U32)litLength, (U32)mlBase+MINMATCH, (U32)offCode); + pos, (U32)litLength, (U32)matchLength, (U32)offBase_minus1); } #endif assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq); @@ -469,26 +633,66 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera /* literal Length */ if (litLength>0xFFFF) { - assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */ - seqStorePtr->longLengthID = 1; + assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */ + seqStorePtr->longLengthType = ZSTD_llt_literalLength; seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart); } seqStorePtr->sequences[0].litLength = (U16)litLength; /* match offset */ - seqStorePtr->sequences[0].offset = offCode + 1; + seqStorePtr->sequences[0].offBase = STORED_TO_OFFBASE(offBase_minus1); /* match Length */ - if (mlBase>0xFFFF) { - assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */ - seqStorePtr->longLengthID = 2; - seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart); + assert(matchLength >= MINMATCH); + { size_t const mlBase = matchLength - MINMATCH; + if (mlBase>0xFFFF) { + assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */ + seqStorePtr->longLengthType = ZSTD_llt_matchLength; + seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart); + } + seqStorePtr->sequences[0].mlBase = (U16)mlBase; } - seqStorePtr->sequences[0].matchLength = (U16)mlBase; seqStorePtr->sequences++; } +/* ZSTD_updateRep() : + * updates in-place @rep (array of repeat offsets) + * @offBase_minus1 : sum-type, with same numeric representation as ZSTD_storeSeq() + */ +MEM_STATIC void +ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0) +{ + if (STORED_IS_OFFSET(offBase_minus1)) { /* full offset */ + rep[2] = rep[1]; + rep[1] = rep[0]; + rep[0] = STORED_OFFSET(offBase_minus1); + } else { /* repcode */ + U32 const repCode = STORED_REPCODE(offBase_minus1) - 1 + ll0; + if (repCode > 0) { /* note : if repCode==0, no change */ + U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode]; + rep[2] = (repCode >= 2) ? rep[1] : rep[2]; + rep[1] = rep[0]; + rep[0] = currentOffset; + } else { /* repCode == 0 */ + /* nothing to do */ + } + } +} + +typedef struct repcodes_s { + U32 rep[3]; +} repcodes_t; + +MEM_STATIC repcodes_t +ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0) +{ + repcodes_t newReps; + ZSTD_memcpy(&newReps, rep, sizeof(newReps)); + ZSTD_updateRep(newReps.rep, offBase_minus1, ll0); + return newReps; +} + /*-************************************* * Match length counter @@ -498,8 +702,18 @@ static unsigned ZSTD_NbCommonBytes (size_t val) if (MEM_isLittleEndian()) { if (MEM_64bits()) { # if defined(_MSC_VER) && defined(_WIN64) - unsigned long r = 0; - return _BitScanForward64( &r, (U64)val ) ? (unsigned)(r >> 3) : 0; +# if STATIC_BMI2 + return _tzcnt_u64(val) >> 3; +# else + if (val != 0) { + unsigned long r; + _BitScanForward64(&r, (U64)val); + return (unsigned)(r >> 3); + } else { + /* Should not reach this code path */ + __assume(0); + } +# endif # elif defined(__GNUC__) && (__GNUC__ >= 4) return (__builtin_ctzll((U64)val) >> 3); # else @@ -515,8 +729,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val) # endif } else { /* 32 bits */ # if defined(_MSC_VER) - unsigned long r=0; - return _BitScanForward( &r, (U32)val ) ? (unsigned)(r >> 3) : 0; + if (val != 0) { + unsigned long r; + _BitScanForward(&r, (U32)val); + return (unsigned)(r >> 3); + } else { + /* Should not reach this code path */ + __assume(0); + } # elif defined(__GNUC__) && (__GNUC__ >= 3) return (__builtin_ctz((U32)val) >> 3); # else @@ -530,8 +750,18 @@ static unsigned ZSTD_NbCommonBytes (size_t val) } else { /* Big Endian CPU */ if (MEM_64bits()) { # if defined(_MSC_VER) && defined(_WIN64) - unsigned long r = 0; - return _BitScanReverse64( &r, val ) ? (unsigned)(r >> 3) : 0; +# if STATIC_BMI2 + return _lzcnt_u64(val) >> 3; +# else + if (val != 0) { + unsigned long r; + _BitScanReverse64(&r, (U64)val); + return (unsigned)(r >> 3); + } else { + /* Should not reach this code path */ + __assume(0); + } +# endif # elif defined(__GNUC__) && (__GNUC__ >= 4) return (__builtin_clzll(val) >> 3); # else @@ -544,8 +774,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val) # endif } else { /* 32 bits */ # if defined(_MSC_VER) - unsigned long r = 0; - return _BitScanReverse( &r, (unsigned long)val ) ? (unsigned)(r >> 3) : 0; + if (val != 0) { + unsigned long r; + _BitScanReverse(&r, (unsigned long)val); + return (unsigned)(r >> 3); + } else { + /* Should not reach this code path */ + __assume(0); + } # elif defined(__GNUC__) && (__GNUC__ >= 3) return (__builtin_clz((U32)val) >> 3); # else @@ -626,7 +862,8 @@ static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL; static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; } static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); } -MEM_STATIC size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls) +MEM_STATIC FORCE_INLINE_ATTR +size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls) { switch(mls) { @@ -723,6 +960,13 @@ MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window) window->dictLimit = end; } +MEM_STATIC U32 ZSTD_window_isEmpty(ZSTD_window_t const window) +{ + return window.dictLimit == ZSTD_WINDOW_START_INDEX && + window.lowLimit == ZSTD_WINDOW_START_INDEX && + (window.nextSrc - window.base) == ZSTD_WINDOW_START_INDEX; +} + /** * ZSTD_window_hasExtDict(): * Returns non-zero if the window has a non-empty extDict. @@ -742,20 +986,76 @@ MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_matchState_t *ms) return ZSTD_window_hasExtDict(ms->window) ? ZSTD_extDict : ms->dictMatchState != NULL ? - ZSTD_dictMatchState : + (ms->dictMatchState->dedicatedDictSearch ? ZSTD_dedicatedDictSearch : ZSTD_dictMatchState) : ZSTD_noDict; } +/* Defining this macro to non-zero tells zstd to run the overflow correction + * code much more frequently. This is very inefficient, and should only be + * used for tests and fuzzers. + */ +#ifndef ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY +# ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION +# define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 1 +# else +# define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 0 +# endif +#endif + +/** + * ZSTD_window_canOverflowCorrect(): + * Returns non-zero if the indices are large enough for overflow correction + * to work correctly without impacting compression ratio. + */ +MEM_STATIC U32 ZSTD_window_canOverflowCorrect(ZSTD_window_t const window, + U32 cycleLog, + U32 maxDist, + U32 loadedDictEnd, + void const* src) +{ + U32 const cycleSize = 1u << cycleLog; + U32 const curr = (U32)((BYTE const*)src - window.base); + U32 const minIndexToOverflowCorrect = cycleSize + + MAX(maxDist, cycleSize) + + ZSTD_WINDOW_START_INDEX; + + /* Adjust the min index to backoff the overflow correction frequency, + * so we don't waste too much CPU in overflow correction. If this + * computation overflows we don't really care, we just need to make + * sure it is at least minIndexToOverflowCorrect. + */ + U32 const adjustment = window.nbOverflowCorrections + 1; + U32 const adjustedIndex = MAX(minIndexToOverflowCorrect * adjustment, + minIndexToOverflowCorrect); + U32 const indexLargeEnough = curr > adjustedIndex; + + /* Only overflow correct early if the dictionary is invalidated already, + * so we don't hurt compression ratio. + */ + U32 const dictionaryInvalidated = curr > maxDist + loadedDictEnd; + + return indexLargeEnough && dictionaryInvalidated; +} + /** * ZSTD_window_needOverflowCorrection(): * Returns non-zero if the indices are getting too large and need overflow * protection. */ MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window, + U32 cycleLog, + U32 maxDist, + U32 loadedDictEnd, + void const* src, void const* srcEnd) { - U32 const current = (U32)((BYTE const*)srcEnd - window.base); - return current > ZSTD_CURRENT_MAX; + U32 const curr = (U32)((BYTE const*)srcEnd - window.base); + if (ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) { + if (ZSTD_window_canOverflowCorrect(window, cycleLog, maxDist, loadedDictEnd, src)) { + return 1; + } + } + return curr > ZSTD_CURRENT_MAX; } /** @@ -766,7 +1066,6 @@ MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window, * * The least significant cycleLog bits of the indices must remain the same, * which may be 0. Every index up to maxDist in the past must be valid. - * NOTE: (maxDist & cycleMask) must be zero. */ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog, U32 maxDist, void const* src) @@ -790,32 +1089,52 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog, * 3. (cctx->lowLimit + 1<<windowLog) < 1<<32: * windowLog <= 31 ==> 3<<29 + 1<<windowLog < 7<<29 < 1<<32. */ - U32 const cycleMask = (1U << cycleLog) - 1; - U32 const current = (U32)((BYTE const*)src - window->base); - U32 const currentCycle0 = current & cycleMask; - /* Exclude zero so that newCurrent - maxDist >= 1. */ - U32 const currentCycle1 = currentCycle0 == 0 ? (1U << cycleLog) : currentCycle0; - U32 const newCurrent = currentCycle1 + maxDist; - U32 const correction = current - newCurrent; - assert((maxDist & cycleMask) == 0); - assert(current > newCurrent); - /* Loose bound, should be around 1<<29 (see above) */ - assert(correction > 1<<28); + U32 const cycleSize = 1u << cycleLog; + U32 const cycleMask = cycleSize - 1; + U32 const curr = (U32)((BYTE const*)src - window->base); + U32 const currentCycle = curr & cycleMask; + /* Ensure newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX. */ + U32 const currentCycleCorrection = currentCycle < ZSTD_WINDOW_START_INDEX + ? MAX(cycleSize, ZSTD_WINDOW_START_INDEX) + : 0; + U32 const newCurrent = currentCycle + + currentCycleCorrection + + MAX(maxDist, cycleSize); + U32 const correction = curr - newCurrent; + /* maxDist must be a power of two so that: + * (newCurrent & cycleMask) == (curr & cycleMask) + * This is required to not corrupt the chains / binary tree. + */ + assert((maxDist & (maxDist - 1)) == 0); + assert((curr & cycleMask) == (newCurrent & cycleMask)); + assert(curr > newCurrent); + if (!ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) { + /* Loose bound, should be around 1<<29 (see above) */ + assert(correction > 1<<28); + } window->base += correction; window->dictBase += correction; - if (window->lowLimit <= correction) window->lowLimit = 1; - else window->lowLimit -= correction; - if (window->dictLimit <= correction) window->dictLimit = 1; - else window->dictLimit -= correction; + if (window->lowLimit < correction + ZSTD_WINDOW_START_INDEX) { + window->lowLimit = ZSTD_WINDOW_START_INDEX; + } else { + window->lowLimit -= correction; + } + if (window->dictLimit < correction + ZSTD_WINDOW_START_INDEX) { + window->dictLimit = ZSTD_WINDOW_START_INDEX; + } else { + window->dictLimit -= correction; + } /* Ensure we can still reference the full window. */ assert(newCurrent >= maxDist); - assert(newCurrent - maxDist >= 1); + assert(newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX); /* Ensure that lowLimit and dictLimit didn't underflow. */ assert(window->lowLimit <= newCurrent); assert(window->dictLimit <= newCurrent); + ++window->nbOverflowCorrections; + DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction, window->lowLimit); return correction; @@ -919,12 +1238,14 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window, } MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) { - memset(window, 0, sizeof(*window)); - window->base = (BYTE const*)""; - window->dictBase = (BYTE const*)""; - window->dictLimit = 1; /* start from 1, so that 1st position is valid */ - window->lowLimit = 1; /* it ensures first and later CCtx usages compress the same */ - window->nextSrc = window->base + 1; /* see issue #1241 */ + ZSTD_memset(window, 0, sizeof(*window)); + window->base = (BYTE const*)" "; + window->dictBase = (BYTE const*)" "; + ZSTD_STATIC_ASSERT(ZSTD_DUBT_UNSORTED_MARK < ZSTD_WINDOW_START_INDEX); /* Start above ZSTD_DUBT_UNSORTED_MARK */ + window->dictLimit = ZSTD_WINDOW_START_INDEX; /* start from >0, so that 1st position is valid */ + window->lowLimit = ZSTD_WINDOW_START_INDEX; /* it ensures first and later CCtx usages compress the same */ + window->nextSrc = window->base + ZSTD_WINDOW_START_INDEX; /* see issue #1241 */ + window->nbOverflowCorrections = 0; } /** @@ -935,7 +1256,8 @@ MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) { * Returns non-zero if the segment is contiguous. */ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window, - void const* src, size_t srcSize) + void const* src, size_t srcSize, + int forceNonContiguous) { BYTE const* const ip = (BYTE const*)src; U32 contiguous = 1; @@ -945,7 +1267,7 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window, assert(window->base != NULL); assert(window->dictBase != NULL); /* Check if blocks follow each other */ - if (src != window->nextSrc) { + if (src != window->nextSrc || forceNonContiguous) { /* not contiguous */ size_t const distanceFromBase = (size_t)(window->nextSrc - window->base); DEBUGLOG(5, "Non contiguous blocks, new segment starts at %u", window->dictLimit); @@ -973,25 +1295,32 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window, /** * Returns the lowest allowed match index. It may either be in the ext-dict or the prefix. */ -MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog) +MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog) { - U32 const maxDistance = 1U << windowLog; - U32 const lowestValid = ms->window.lowLimit; - U32 const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid; - U32 const isDictionary = (ms->loadedDictEnd != 0); - U32 const matchLowest = isDictionary ? lowestValid : withinWindow; + U32 const maxDistance = 1U << windowLog; + U32 const lowestValid = ms->window.lowLimit; + U32 const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid; + U32 const isDictionary = (ms->loadedDictEnd != 0); + /* When using a dictionary the entire dictionary is valid if a single byte of the dictionary + * is within the window. We invalidate the dictionary (and set loadedDictEnd to 0) when it isn't + * valid for the entire block. So this check is sufficient to find the lowest valid match index. + */ + U32 const matchLowest = isDictionary ? lowestValid : withinWindow; return matchLowest; } /** * Returns the lowest allowed match index in the prefix. */ -MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog) +MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog) { U32 const maxDistance = 1U << windowLog; U32 const lowestValid = ms->window.dictLimit; - U32 const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid; + U32 const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid; U32 const isDictionary = (ms->loadedDictEnd != 0); + /* When computing the lowest prefix index we need to take the dictionary into account to handle + * the edge case where the dictionary and the source are contiguous in memory. + */ U32 const matchLowest = isDictionary ? lowestValid : withinWindow; return matchLowest; } @@ -1045,7 +1374,6 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max) * assumptions : magic number supposed already checked * and dictSize >= 8 */ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace, - short* offcodeNCount, unsigned* offcodeMaxValue, const void* const dict, size_t dictSize); void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs); @@ -1061,7 +1389,7 @@ void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs); * Note: srcSizeHint == 0 means 0! */ ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams( - const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize); + const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode); /*! ZSTD_initCStream_internal() : * Private use only. Init streaming operation. @@ -1122,4 +1450,9 @@ size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSe * condition for correct operation : hashLog > 1 */ U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat); +/** ZSTD_CCtx_trace() : + * Trace the end of a compression call. + */ +void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize); + #endif /* ZSTD_COMPRESS_H */ diff --git a/thirdparty/zstd/compress/zstd_compress_literals.c b/thirdparty/zstd/compress/zstd_compress_literals.c index 17e7168d89..52b0a8059a 100644 --- a/thirdparty/zstd/compress/zstd_compress_literals.c +++ b/thirdparty/zstd/compress/zstd_compress_literals.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 @@ -15,7 +15,7 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize) { - BYTE* const ostart = (BYTE* const)dst; + BYTE* const ostart = (BYTE*)dst; U32 const flSize = 1 + (srcSize>31) + (srcSize>4095); RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, ""); @@ -35,14 +35,14 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, assert(0); } - memcpy(ostart + flSize, src, srcSize); + ZSTD_memcpy(ostart + flSize, src, srcSize); DEBUGLOG(5, "Raw literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize)); return srcSize + flSize; } size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize) { - BYTE* const ostart = (BYTE* const)dst; + BYTE* const ostart = (BYTE*)dst; U32 const flSize = 1 + (srcSize>31) + (srcSize>4095); (void)dstCapacity; /* dstCapacity already guaranteed to be >=4, hence large enough */ @@ -73,7 +73,8 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf, void* dst, size_t dstCapacity, const void* src, size_t srcSize, void* entropyWorkspace, size_t entropyWorkspaceSize, - const int bmi2) + const int bmi2, + unsigned suspectUncompressible) { size_t const minGain = ZSTD_minGain(srcSize, strategy); size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB); @@ -86,7 +87,7 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf, disableLiteralCompression, (U32)srcSize); /* Prepare nextEntropy assuming reusing the existing table */ - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); if (disableLiteralCompression) return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize); @@ -105,11 +106,11 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf, HUF_compress1X_repeat( ostart+lhSize, dstCapacity-lhSize, src, srcSize, HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize, - (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2) : + (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible) : HUF_compress4X_repeat( ostart+lhSize, dstCapacity-lhSize, src, srcSize, HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize, - (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2); + (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible); if (repeat != HUF_repeat_none) { /* reused the existing table */ DEBUGLOG(5, "Reusing previous huffman table"); @@ -117,12 +118,12 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf, } } - if ((cLitSize==0) | (cLitSize >= srcSize - minGain) | ERR_isError(cLitSize)) { - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) { + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize); } if (cLitSize==1) { - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); + ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize); } diff --git a/thirdparty/zstd/compress/zstd_compress_literals.h b/thirdparty/zstd/compress/zstd_compress_literals.h index 8b08705743..9775fb97cb 100644 --- a/thirdparty/zstd/compress/zstd_compress_literals.h +++ b/thirdparty/zstd/compress/zstd_compress_literals.h @@ -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 @@ -18,12 +18,14 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize); +/* If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf, ZSTD_hufCTables_t* nextHuf, ZSTD_strategy strategy, int disableLiteralCompression, void* dst, size_t dstCapacity, const void* src, size_t srcSize, void* entropyWorkspace, size_t entropyWorkspaceSize, - const int bmi2); + const int bmi2, + unsigned suspectUncompressible); #endif /* ZSTD_COMPRESS_LITERALS_H */ diff --git a/thirdparty/zstd/compress/zstd_compress_sequences.c b/thirdparty/zstd/compress/zstd_compress_sequences.c index f9f8097c83..f1e40af2ea 100644 --- a/thirdparty/zstd/compress/zstd_compress_sequences.c +++ b/thirdparty/zstd/compress/zstd_compress_sequences.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 @@ -51,6 +51,19 @@ static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) { } /** + * Returns true if we should use ncount=-1 else we should + * use ncount=1 for low probability symbols instead. + */ +static unsigned ZSTD_useLowProbCount(size_t const nbSeq) +{ + /* Heuristic: This should cover most blocks <= 16K and + * start to fade out after 16K to about 32K depending on + * comprssibility. + */ + return nbSeq >= 2048; +} + +/** * Returns the cost in bytes of encoding the normalized count header. * Returns an error if any of the helper functions return an error. */ @@ -60,7 +73,7 @@ static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max, BYTE wksp[FSE_NCOUNTBOUND]; S16 norm[MaxSeq + 1]; const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max); - FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max), ""); + FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), ""); return FSE_writeNCount(wksp, sizeof(wksp), norm, max, tableLog); } @@ -72,6 +85,8 @@ static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t { unsigned cost = 0; unsigned s; + + assert(total > 0); for (s = 0; s <= max; ++s) { unsigned norm = (unsigned)((256 * count[s]) / total); if (count[s] != 0 && norm == 0) @@ -219,6 +234,11 @@ ZSTD_selectEncodingType( return set_compressed; } +typedef struct { + S16 norm[MaxSeq + 1]; + U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)]; +} ZSTD_BuildCTableWksp; + size_t ZSTD_buildCTable(void* dst, size_t dstCapacity, FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type, @@ -239,13 +259,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity, *op = codeTable[0]; return 1; case set_repeat: - memcpy(nextCTable, prevCTable, prevCTableSize); + ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize); return 0; case set_basic: FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), ""); /* note : could be pre-calculated */ return 0; case set_compressed: { - S16 norm[MaxSeq + 1]; + ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace; size_t nbSeq_1 = nbSeq; const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max); if (count[codeTable[nbSeq-1]] > 1) { @@ -253,10 +273,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity, nbSeq_1--; } assert(nbSeq_1 > 1); - FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq_1, max), ""); - { size_t const NCountSize = FSE_writeNCount(op, oend - op, norm, max, tableLog); /* overflow protected */ + assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp)); + (void)entropyWorkspaceSize; + FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed"); + assert(oend >= op); + { size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog); /* overflow protected */ FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed"); - FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, norm, max, tableLog, entropyWorkspace, entropyWorkspaceSize), ""); + FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed"); return NCountSize; } } @@ -290,19 +313,19 @@ ZSTD_encodeSequences_body( FSE_initCState2(&stateLitLength, CTable_LitLength, llCodeTable[nbSeq-1]); BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]); if (MEM_32bits()) BIT_flushBits(&blockStream); - BIT_addBits(&blockStream, sequences[nbSeq-1].matchLength, ML_bits[mlCodeTable[nbSeq-1]]); + BIT_addBits(&blockStream, sequences[nbSeq-1].mlBase, ML_bits[mlCodeTable[nbSeq-1]]); if (MEM_32bits()) BIT_flushBits(&blockStream); if (longOffsets) { U32 const ofBits = ofCodeTable[nbSeq-1]; unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1); if (extraBits) { - BIT_addBits(&blockStream, sequences[nbSeq-1].offset, extraBits); + BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, extraBits); BIT_flushBits(&blockStream); } - BIT_addBits(&blockStream, sequences[nbSeq-1].offset >> extraBits, + BIT_addBits(&blockStream, sequences[nbSeq-1].offBase >> extraBits, ofBits - extraBits); } else { - BIT_addBits(&blockStream, sequences[nbSeq-1].offset, ofCodeTable[nbSeq-1]); + BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, ofCodeTable[nbSeq-1]); } BIT_flushBits(&blockStream); @@ -316,8 +339,8 @@ ZSTD_encodeSequences_body( U32 const mlBits = ML_bits[mlCode]; DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u", (unsigned)sequences[n].litLength, - (unsigned)sequences[n].matchLength + MINMATCH, - (unsigned)sequences[n].offset); + (unsigned)sequences[n].mlBase + MINMATCH, + (unsigned)sequences[n].offBase); /* 32b*/ /* 64b*/ /* (7)*/ /* (7)*/ FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode); /* 15 */ /* 15 */ @@ -328,18 +351,18 @@ ZSTD_encodeSequences_body( BIT_flushBits(&blockStream); /* (7)*/ BIT_addBits(&blockStream, sequences[n].litLength, llBits); if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream); - BIT_addBits(&blockStream, sequences[n].matchLength, mlBits); + BIT_addBits(&blockStream, sequences[n].mlBase, mlBits); if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream); if (longOffsets) { unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1); if (extraBits) { - BIT_addBits(&blockStream, sequences[n].offset, extraBits); + BIT_addBits(&blockStream, sequences[n].offBase, extraBits); BIT_flushBits(&blockStream); /* (7)*/ } - BIT_addBits(&blockStream, sequences[n].offset >> extraBits, + BIT_addBits(&blockStream, sequences[n].offBase >> extraBits, ofBits - extraBits); /* 31 */ } else { - BIT_addBits(&blockStream, sequences[n].offset, ofBits); /* 31 */ + BIT_addBits(&blockStream, sequences[n].offBase, ofBits); /* 31 */ } BIT_flushBits(&blockStream); /* (7)*/ DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr)); @@ -376,7 +399,7 @@ ZSTD_encodeSequences_default( #if DYNAMIC_BMI2 -static TARGET_ATTRIBUTE("bmi2") size_t +static BMI2_TARGET_ATTRIBUTE size_t ZSTD_encodeSequences_bmi2( void* dst, size_t dstCapacity, FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable, diff --git a/thirdparty/zstd/compress/zstd_compress_sequences.h b/thirdparty/zstd/compress/zstd_compress_sequences.h index 68c6f9a5ac..7991364c2f 100644 --- a/thirdparty/zstd/compress/zstd_compress_sequences.h +++ b/thirdparty/zstd/compress/zstd_compress_sequences.h @@ -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 diff --git a/thirdparty/zstd/compress/zstd_compress_superblock.c b/thirdparty/zstd/compress/zstd_compress_superblock.c index b693866c0a..10e3378577 100644 --- a/thirdparty/zstd/compress/zstd_compress_superblock.c +++ b/thirdparty/zstd/compress/zstd_compress_superblock.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 @@ -15,288 +15,10 @@ #include "../common/zstd_internal.h" /* ZSTD_getSequenceLength */ #include "hist.h" /* HIST_countFast_wksp */ -#include "zstd_compress_internal.h" +#include "zstd_compress_internal.h" /* ZSTD_[huf|fse|entropy]CTablesMetadata_t */ #include "zstd_compress_sequences.h" #include "zstd_compress_literals.h" -/*-************************************* -* Superblock entropy buffer structs -***************************************/ -/** ZSTD_hufCTablesMetadata_t : - * Stores Literals Block Type for a super-block in hType, and - * huffman tree description in hufDesBuffer. - * hufDesSize refers to the size of huffman tree description in bytes. - * This metadata is populated in ZSTD_buildSuperBlockEntropy_literal() */ -typedef struct { - symbolEncodingType_e hType; - BYTE hufDesBuffer[500]; /* TODO give name to this value */ - size_t hufDesSize; -} ZSTD_hufCTablesMetadata_t; - -/** ZSTD_fseCTablesMetadata_t : - * Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and - * fse tables in fseTablesBuffer. - * fseTablesSize refers to the size of fse tables in bytes. - * This metadata is populated in ZSTD_buildSuperBlockEntropy_sequences() */ -typedef struct { - symbolEncodingType_e llType; - symbolEncodingType_e ofType; - symbolEncodingType_e mlType; - BYTE fseTablesBuffer[500]; /* TODO give name to this value */ - size_t fseTablesSize; - size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_compressSubBlock_sequences() */ -} ZSTD_fseCTablesMetadata_t; - -typedef struct { - ZSTD_hufCTablesMetadata_t hufMetadata; - ZSTD_fseCTablesMetadata_t fseMetadata; -} ZSTD_entropyCTablesMetadata_t; - - -/** ZSTD_buildSuperBlockEntropy_literal() : - * Builds entropy for the super-block literals. - * Stores literals block type (raw, rle, compressed, repeat) and - * huffman description table to hufMetadata. - * @return : size of huffman description table or error code */ -static size_t ZSTD_buildSuperBlockEntropy_literal(void* const src, size_t srcSize, - const ZSTD_hufCTables_t* prevHuf, - ZSTD_hufCTables_t* nextHuf, - ZSTD_hufCTablesMetadata_t* hufMetadata, - const int disableLiteralsCompression, - void* workspace, size_t wkspSize) -{ - BYTE* const wkspStart = (BYTE*)workspace; - BYTE* const wkspEnd = wkspStart + wkspSize; - BYTE* const countWkspStart = wkspStart; - unsigned* const countWksp = (unsigned*)workspace; - const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned); - BYTE* const nodeWksp = countWkspStart + countWkspSize; - const size_t nodeWkspSize = wkspEnd-nodeWksp; - unsigned maxSymbolValue = 255; - unsigned huffLog = HUF_TABLELOG_DEFAULT; - HUF_repeat repeat = prevHuf->repeatMode; - - DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_literal (srcSize=%zu)", srcSize); - - /* Prepare nextEntropy assuming reusing the existing table */ - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); - - if (disableLiteralsCompression) { - DEBUGLOG(5, "set_basic - disabled"); - hufMetadata->hType = set_basic; - return 0; - } - - /* small ? don't even attempt compression (speed opt) */ -# define COMPRESS_LITERALS_SIZE_MIN 63 - { size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN; - if (srcSize <= minLitSize) { - DEBUGLOG(5, "set_basic - too small"); - hufMetadata->hType = set_basic; - return 0; - } - } - - /* Scan input and build symbol stats */ - { size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize); - FORWARD_IF_ERROR(largest, "HIST_count_wksp failed"); - if (largest == srcSize) { - DEBUGLOG(5, "set_rle"); - hufMetadata->hType = set_rle; - return 0; - } - if (largest <= (srcSize >> 7)+4) { - DEBUGLOG(5, "set_basic - no gain"); - hufMetadata->hType = set_basic; - return 0; - } - } - - /* Validate the previous Huffman table */ - if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) { - repeat = HUF_repeat_none; - } - - /* Build Huffman Tree */ - memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable)); - huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue); - { size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp, - maxSymbolValue, huffLog, - nodeWksp, nodeWkspSize); - FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp"); - huffLog = (U32)maxBits; - { /* Build and write the CTable */ - size_t const newCSize = HUF_estimateCompressedSize( - (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue); - size_t const hSize = HUF_writeCTable( - hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer), - (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog); - /* Check against repeating the previous CTable */ - if (repeat != HUF_repeat_none) { - size_t const oldCSize = HUF_estimateCompressedSize( - (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue); - if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) { - DEBUGLOG(5, "set_repeat - smaller"); - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); - hufMetadata->hType = set_repeat; - return 0; - } - } - if (newCSize + hSize >= srcSize) { - DEBUGLOG(5, "set_basic - no gains"); - memcpy(nextHuf, prevHuf, sizeof(*prevHuf)); - hufMetadata->hType = set_basic; - return 0; - } - DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize); - hufMetadata->hType = set_compressed; - nextHuf->repeatMode = HUF_repeat_check; - return hSize; - } - } -} - -/** ZSTD_buildSuperBlockEntropy_sequences() : - * Builds entropy for the super-block sequences. - * Stores symbol compression modes and fse table to fseMetadata. - * @return : size of fse tables or error code */ -static size_t ZSTD_buildSuperBlockEntropy_sequences(seqStore_t* seqStorePtr, - const ZSTD_fseCTables_t* prevEntropy, - ZSTD_fseCTables_t* nextEntropy, - const ZSTD_CCtx_params* cctxParams, - ZSTD_fseCTablesMetadata_t* fseMetadata, - void* workspace, size_t wkspSize) -{ - BYTE* const wkspStart = (BYTE*)workspace; - BYTE* const wkspEnd = wkspStart + wkspSize; - BYTE* const countWkspStart = wkspStart; - unsigned* const countWksp = (unsigned*)workspace; - const size_t countWkspSize = (MaxSeq + 1) * sizeof(unsigned); - BYTE* const cTableWksp = countWkspStart + countWkspSize; - const size_t cTableWkspSize = wkspEnd-cTableWksp; - ZSTD_strategy const strategy = cctxParams->cParams.strategy; - FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable; - FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable; - FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable; - const BYTE* const ofCodeTable = seqStorePtr->ofCode; - const BYTE* const llCodeTable = seqStorePtr->llCode; - const BYTE* const mlCodeTable = seqStorePtr->mlCode; - size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart; - BYTE* const ostart = fseMetadata->fseTablesBuffer; - BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer); - BYTE* op = ostart; - - assert(cTableWkspSize >= (1 << MaxFSELog) * sizeof(FSE_FUNCTION_TYPE)); - DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_sequences (nbSeq=%zu)", nbSeq); - memset(workspace, 0, wkspSize); - - fseMetadata->lastCountSize = 0; - /* convert length/distances into codes */ - ZSTD_seqToCodes(seqStorePtr); - /* build CTable for Literal Lengths */ - { U32 LLtype; - unsigned max = MaxLL; - size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, llCodeTable, nbSeq, workspace, wkspSize); /* can't fail */ - DEBUGLOG(5, "Building LL table"); - nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode; - LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode, - countWksp, max, mostFrequent, nbSeq, - LLFSELog, prevEntropy->litlengthCTable, - LL_defaultNorm, LL_defaultNormLog, - ZSTD_defaultAllowed, strategy); - assert(set_basic < set_compressed && set_rle < set_compressed); - assert(!(LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype, - countWksp, max, llCodeTable, nbSeq, LL_defaultNorm, LL_defaultNormLog, MaxLL, - prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable), - cTableWksp, cTableWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed"); - if (LLtype == set_compressed) - fseMetadata->lastCountSize = countSize; - op += countSize; - fseMetadata->llType = (symbolEncodingType_e) LLtype; - } } - /* build CTable for Offsets */ - { U32 Offtype; - unsigned max = MaxOff; - size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, ofCodeTable, nbSeq, workspace, wkspSize); /* can't fail */ - /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */ - ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed; - DEBUGLOG(5, "Building OF table"); - nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode; - Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode, - countWksp, max, mostFrequent, nbSeq, - OffFSELog, prevEntropy->offcodeCTable, - OF_defaultNorm, OF_defaultNormLog, - defaultPolicy, strategy); - assert(!(Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype, - countWksp, max, ofCodeTable, nbSeq, OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff, - prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable), - cTableWksp, cTableWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed"); - if (Offtype == set_compressed) - fseMetadata->lastCountSize = countSize; - op += countSize; - fseMetadata->ofType = (symbolEncodingType_e) Offtype; - } } - /* build CTable for MatchLengths */ - { U32 MLtype; - unsigned max = MaxML; - size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, mlCodeTable, nbSeq, workspace, wkspSize); /* can't fail */ - DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op)); - nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode; - MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode, - countWksp, max, mostFrequent, nbSeq, - MLFSELog, prevEntropy->matchlengthCTable, - ML_defaultNorm, ML_defaultNormLog, - ZSTD_defaultAllowed, strategy); - assert(!(MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */ - { size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype, - countWksp, max, mlCodeTable, nbSeq, ML_defaultNorm, ML_defaultNormLog, MaxML, - prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable), - cTableWksp, cTableWkspSize); - FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed"); - if (MLtype == set_compressed) - fseMetadata->lastCountSize = countSize; - op += countSize; - fseMetadata->mlType = (symbolEncodingType_e) MLtype; - } } - assert((size_t) (op-ostart) <= sizeof(fseMetadata->fseTablesBuffer)); - return op-ostart; -} - - -/** ZSTD_buildSuperBlockEntropy() : - * Builds entropy for the super-block. - * @return : 0 on success or error code */ -static size_t -ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr, - const ZSTD_entropyCTables_t* prevEntropy, - ZSTD_entropyCTables_t* nextEntropy, - const ZSTD_CCtx_params* cctxParams, - ZSTD_entropyCTablesMetadata_t* entropyMetadata, - void* workspace, size_t wkspSize) -{ - size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart; - DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy"); - entropyMetadata->hufMetadata.hufDesSize = - ZSTD_buildSuperBlockEntropy_literal(seqStorePtr->litStart, litSize, - &prevEntropy->huf, &nextEntropy->huf, - &entropyMetadata->hufMetadata, - ZSTD_disableLiteralsCompression(cctxParams), - workspace, wkspSize); - FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildSuperBlockEntropy_literal failed"); - entropyMetadata->fseMetadata.fseTablesSize = - ZSTD_buildSuperBlockEntropy_sequences(seqStorePtr, - &prevEntropy->fse, &nextEntropy->fse, - cctxParams, - &entropyMetadata->fseMetadata, - workspace, wkspSize); - FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildSuperBlockEntropy_sequences failed"); - return 0; -} - /** ZSTD_compressSubBlock_literal() : * Compresses literals section for a sub-block. * When we have to write the Huffman table we will sometimes choose a header @@ -304,7 +26,7 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr, * before we know the table size + compressed size, so we have a bound on the * table size. If we guessed incorrectly, we fall back to uncompressed literals. * - * We write the header when writeEntropy=1 and set entropyWrriten=1 when we succeeded + * We write the header when writeEntropy=1 and set entropyWritten=1 when we succeeded * in writing the header, otherwise it is set to 0. * * hufMetadata->hType has literals block type info. @@ -348,7 +70,7 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable, assert(hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat); if (writeEntropy && hufMetadata->hType == set_compressed) { - memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize); + ZSTD_memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize); op += hufMetadata->hufDesSize; cLitSize += hufMetadata->hufDesSize; DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize); @@ -410,6 +132,7 @@ static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef* const seqDef* sp = sstart; size_t matchLengthSum = 0; size_t litLengthSum = 0; + (void)(litLengthSum); /* suppress unused variable warning on some environments */ while (send-sp > 0) { ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp); litLengthSum += seqLen.litLength; @@ -474,7 +197,7 @@ static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables const U32 MLtype = fseMetadata->mlType; DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (fseTablesSize=%zu)", fseMetadata->fseTablesSize); *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2)); - memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize); + ZSTD_memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize); op += fseMetadata->fseTablesSize; } else { const U32 repeat = set_repeat; @@ -602,8 +325,8 @@ static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t lit static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type, const BYTE* codeTable, unsigned maxCode, size_t nbSeq, const FSE_CTable* fseCTable, - const U32* additionalBits, - short const* defaultNorm, U32 defaultNormLog, + const U8* additionalBits, + short const* defaultNorm, U32 defaultNormLog, U32 defaultMax, void* workspace, size_t wkspSize) { unsigned* const countWksp = (unsigned*)workspace; @@ -615,7 +338,11 @@ static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type, HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize); /* can't fail */ if (type == set_basic) { - cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max); + /* We selected this encoding type, so it must be valid. */ + assert(max <= defaultMax); + cSymbolTypeSizeEstimateInBits = max <= defaultMax + ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max) + : ERROR(GENERIC); } else if (type == set_rle) { cSymbolTypeSizeEstimateInBits = 0; } else if (type == set_compressed || type == set_repeat) { @@ -639,19 +366,20 @@ static size_t ZSTD_estimateSubBlockSize_sequences(const BYTE* ofCodeTable, void* workspace, size_t wkspSize, int writeEntropy) { - size_t sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */ + size_t const sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */ size_t cSeqSizeEstimate = 0; + if (nbSeq == 0) return sequencesSectionHeaderSize; cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, MaxOff, nbSeq, fseTables->offcodeCTable, NULL, - OF_defaultNorm, OF_defaultNormLog, + OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff, workspace, wkspSize); cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->llType, llCodeTable, MaxLL, nbSeq, fseTables->litlengthCTable, LL_bits, - LL_defaultNorm, LL_defaultNormLog, + LL_defaultNorm, LL_defaultNormLog, MaxLL, workspace, wkspSize); cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, MaxML, nbSeq, fseTables->matchlengthCTable, ML_bits, - ML_defaultNorm, ML_defaultNormLog, + ML_defaultNorm, ML_defaultNormLog, MaxML, workspace, wkspSize); if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize; return cSeqSizeEstimate + sequencesSectionHeaderSize; @@ -747,7 +475,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr, /* I think there is an optimization opportunity here. * Calling ZSTD_estimateSubBlockSize for every sequence can be wasteful * since it recalculates estimate from scratch. - * For example, it would recount literal distribution and symbol codes everytime. + * For example, it would recount literal distribution and symbol codes every time. */ cBlockSizeEstimate = ZSTD_estimateSubBlockSize(lp, litSize, ofCodePtr, llCodePtr, mlCodePtr, seqCount, &nextCBlock->entropy, entropyMetadata, @@ -790,7 +518,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr, } while (!lastSequence); if (writeLitEntropy) { DEBUGLOG(5, "ZSTD_compressSubBlock_multi has literal entropy tables unwritten"); - memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf)); + ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf)); } if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) { /* If we haven't written our entropy tables, then we've violated our contract and @@ -809,11 +537,11 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr, if (sp < send) { seqDef const* seq; repcodes_t rep; - memcpy(&rep, prevCBlock->rep, sizeof(rep)); + ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep)); for (seq = sstart; seq < sp; ++seq) { - rep = ZSTD_updateRep(rep.rep, seq->offset - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0); + ZSTD_updateRep(rep.rep, seq->offBase - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0); } - memcpy(nextCBlock->rep, &rep, sizeof(rep)); + ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep)); } } DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed"); @@ -826,12 +554,12 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc, unsigned lastBlock) { ZSTD_entropyCTablesMetadata_t entropyMetadata; - FORWARD_IF_ERROR(ZSTD_buildSuperBlockEntropy(&zc->seqStore, + FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore, &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy, &zc->appliedParams, &entropyMetadata, - zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */), ""); + zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), ""); return ZSTD_compressSubBlock_multi(&zc->seqStore, zc->blockState.prevCBlock, @@ -841,5 +569,5 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc, dst, dstCapacity, src, srcSize, zc->bmi2, lastBlock, - zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */); + zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */); } diff --git a/thirdparty/zstd/compress/zstd_compress_superblock.h b/thirdparty/zstd/compress/zstd_compress_superblock.h index 07f4cb1dc6..176f9b106f 100644 --- a/thirdparty/zstd/compress/zstd_compress_superblock.h +++ b/thirdparty/zstd/compress/zstd_compress_superblock.h @@ -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 diff --git a/thirdparty/zstd/compress/zstd_cwksp.h b/thirdparty/zstd/compress/zstd_cwksp.h index a25c9263b7..dc3f40c80c 100644 --- a/thirdparty/zstd/compress/zstd_cwksp.h +++ b/thirdparty/zstd/compress/zstd_cwksp.h @@ -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 @@ -35,6 +35,10 @@ extern "C" { #define ZSTD_CWKSP_ASAN_REDZONE_SIZE 128 #endif + +/* Set our tables and aligneds to align by 64 bytes */ +#define ZSTD_CWKSP_ALIGNMENT_BYTES 64 + /*-************************************* * Structures ***************************************/ @@ -45,6 +49,16 @@ typedef enum { } ZSTD_cwksp_alloc_phase_e; /** + * Used to describe whether the workspace is statically allocated (and will not + * necessarily ever be freed), or if it's dynamically allocated and we can + * expect a well-formed caller to free this. + */ +typedef enum { + ZSTD_cwksp_dynamic_alloc, + ZSTD_cwksp_static_alloc +} ZSTD_cwksp_static_alloc_e; + +/** * Zstd fits all its internal datastructures into a single continuous buffer, * so that it only needs to perform a single OS allocation (or so that a buffer * can be provided to it and it can perform no allocations at all). This buffer @@ -92,7 +106,7 @@ typedef enum { * * - Static objects: this is optionally the enclosing ZSTD_CCtx or ZSTD_CDict, * so that literally everything fits in a single buffer. Note: if present, - * this must be the first object in the workspace, since ZSTD_free{CCtx, + * this must be the first object in the workspace, since ZSTD_customFree{CCtx, * CDict}() rely on a pointer comparison to see whether one or two frees are * required. * @@ -107,10 +121,11 @@ typedef enum { * - Tables: these are any of several different datastructures (hash tables, * chain tables, binary trees) that all respect a common format: they are * uint32_t arrays, all of whose values are between 0 and (nextSrc - base). - * Their sizes depend on the cparams. + * Their sizes depend on the cparams. These tables are 64-byte aligned. * * - Aligned: these buffers are used for various purposes that require 4 byte - * alignment, but don't require any initialization before they're used. + * alignment, but don't require any initialization before they're used. These + * buffers are each aligned to 64 bytes. * * - Buffers: these buffers are used for various purposes that don't require * any alignment or initialization before they're used. This means they can @@ -123,8 +138,7 @@ typedef enum { * * 1. Objects * 2. Buffers - * 3. Aligned - * 4. Tables + * 3. Aligned/Tables * * Attempts to reserve objects of different types out of order will fail. */ @@ -137,9 +151,10 @@ typedef struct { void* tableValidEnd; void* allocStart; - int allocFailed; + BYTE allocFailed; int workspaceOversizedDuration; ZSTD_cwksp_alloc_phase_e phase; + ZSTD_cwksp_static_alloc_e isStatic; } ZSTD_cwksp; /*-************************************* @@ -176,82 +191,166 @@ MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t const align) { * Since tables aren't currently redzoned, you don't need to call through this * to figure out how much space you need for the matchState tables. Everything * else is though. + * + * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned_alloc_size(). */ MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) { -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) + if (size == 0) + return 0; +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) return size + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE; #else return size; #endif } -MEM_STATIC void ZSTD_cwksp_internal_advance_phase( - ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase) { +/** + * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes. + * Used to determine the number of bytes required for a given "aligned". + */ +MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) { + return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, ZSTD_CWKSP_ALIGNMENT_BYTES)); +} + +/** + * Returns the amount of additional space the cwksp must allocate + * for internal purposes (currently only alignment). + */ +MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) { + /* For alignment, the wksp will always allocate an additional n_1=[1, 64] bytes + * to align the beginning of tables section, as well as another n_2=[0, 63] bytes + * to align the beginning of the aligned section. + * + * n_1 + n_2 == 64 bytes if the cwksp is freshly allocated, due to tables and + * aligneds being sized in multiples of 64 bytes. + */ + size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES; + return slackSpace; +} + + +/** + * Return the number of additional bytes required to align a pointer to the given number of bytes. + * alignBytes must be a power of two. + */ +MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) { + size_t const alignBytesMask = alignBytes - 1; + size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask; + assert((alignBytes & alignBytesMask) == 0); + assert(bytes != ZSTD_CWKSP_ALIGNMENT_BYTES); + return bytes; +} + +/** + * Internal function. Do not use directly. + * Reserves the given number of bytes within the aligned/buffer segment of the wksp, + * which counts from the end of the wksp (as opposed to the object/table segment). + * + * Returns a pointer to the beginning of that space. + */ +MEM_STATIC void* +ZSTD_cwksp_reserve_internal_buffer_space(ZSTD_cwksp* ws, size_t const bytes) +{ + void* const alloc = (BYTE*)ws->allocStart - bytes; + void* const bottom = ws->tableEnd; + DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining", + alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes); + ZSTD_cwksp_assert_internal_consistency(ws); + assert(alloc >= bottom); + if (alloc < bottom) { + DEBUGLOG(4, "cwksp: alloc failed!"); + ws->allocFailed = 1; + return NULL; + } + /* the area is reserved from the end of wksp. + * If it overlaps with tableValidEnd, it voids guarantees on values' range */ + if (alloc < ws->tableValidEnd) { + ws->tableValidEnd = alloc; + } + ws->allocStart = alloc; + return alloc; +} + +/** + * Moves the cwksp to the next phase, and does any necessary allocations. + * cwksp initialization must necessarily go through each phase in order. + * Returns a 0 on success, or zstd error + */ +MEM_STATIC size_t +ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase) +{ assert(phase >= ws->phase); if (phase > ws->phase) { + /* Going from allocating objects to allocating buffers */ if (ws->phase < ZSTD_cwksp_alloc_buffers && phase >= ZSTD_cwksp_alloc_buffers) { ws->tableValidEnd = ws->objectEnd; } + + /* Going from allocating buffers to allocating aligneds/tables */ if (ws->phase < ZSTD_cwksp_alloc_aligned && phase >= ZSTD_cwksp_alloc_aligned) { - /* If unaligned allocations down from a too-large top have left us - * unaligned, we need to realign our alloc ptr. Technically, this - * can consume space that is unaccounted for in the neededSpace - * calculation. However, I believe this can only happen when the - * workspace is too large, and specifically when it is too large - * by a larger margin than the space that will be consumed. */ - /* TODO: cleaner, compiler warning friendly way to do this??? */ - ws->allocStart = (BYTE*)ws->allocStart - ((size_t)ws->allocStart & (sizeof(U32)-1)); - if (ws->allocStart < ws->tableValidEnd) { - ws->tableValidEnd = ws->allocStart; + { /* Align the start of the "aligned" to 64 bytes. Use [1, 64] bytes. */ + size_t const bytesToAlign = + ZSTD_CWKSP_ALIGNMENT_BYTES - ZSTD_cwksp_bytes_to_align_ptr(ws->allocStart, ZSTD_CWKSP_ALIGNMENT_BYTES); + DEBUGLOG(5, "reserving aligned alignment addtl space: %zu", bytesToAlign); + ZSTD_STATIC_ASSERT((ZSTD_CWKSP_ALIGNMENT_BYTES & (ZSTD_CWKSP_ALIGNMENT_BYTES - 1)) == 0); /* power of 2 */ + RETURN_ERROR_IF(!ZSTD_cwksp_reserve_internal_buffer_space(ws, bytesToAlign), + memory_allocation, "aligned phase - alignment initial allocation failed!"); } - } + { /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */ + void* const alloc = ws->objectEnd; + size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES); + void* const objectEnd = (BYTE*)alloc + bytesToAlign; + DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign); + RETURN_ERROR_IF(objectEnd > ws->workspaceEnd, memory_allocation, + "table phase - alignment initial allocation failed!"); + ws->objectEnd = objectEnd; + ws->tableEnd = objectEnd; /* table area starts being empty */ + if (ws->tableValidEnd < ws->tableEnd) { + ws->tableValidEnd = ws->tableEnd; + } } } ws->phase = phase; + ZSTD_cwksp_assert_internal_consistency(ws); } + return 0; } /** * Returns whether this object/buffer/etc was allocated in this workspace. */ -MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr) { +MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr) +{ return (ptr != NULL) && (ws->workspace <= ptr) && (ptr <= ws->workspaceEnd); } /** * Internal function. Do not use directly. */ -MEM_STATIC void* ZSTD_cwksp_reserve_internal( - ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase) { +MEM_STATIC void* +ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase) +{ void* alloc; - void* bottom = ws->tableEnd; - ZSTD_cwksp_internal_advance_phase(ws, phase); - alloc = (BYTE *)ws->allocStart - bytes; + if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase)) || bytes == 0) { + return NULL; + } -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) /* over-reserve space */ - alloc = (BYTE *)alloc - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE; + bytes += 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE; #endif - DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining", - alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes); - ZSTD_cwksp_assert_internal_consistency(ws); - assert(alloc >= bottom); - if (alloc < bottom) { - DEBUGLOG(4, "cwksp: alloc failed!"); - ws->allocFailed = 1; - return NULL; - } - if (alloc < ws->tableValidEnd) { - ws->tableValidEnd = alloc; - } - ws->allocStart = alloc; + alloc = ZSTD_cwksp_reserve_internal_buffer_space(ws, bytes); -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on * either size. */ - alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE; - __asan_unpoison_memory_region(alloc, bytes); + if (alloc) { + alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE; + if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) { + __asan_unpoison_memory_region(alloc, bytes); + } + } #endif return alloc; @@ -260,33 +359,44 @@ MEM_STATIC void* ZSTD_cwksp_reserve_internal( /** * Reserves and returns unaligned memory. */ -MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes) { +MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes) +{ return (BYTE*)ZSTD_cwksp_reserve_internal(ws, bytes, ZSTD_cwksp_alloc_buffers); } /** - * Reserves and returns memory sized on and aligned on sizeof(unsigned). + * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes). */ -MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes) { - assert((bytes & (sizeof(U32)-1)) == 0); - return ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, sizeof(U32)), ZSTD_cwksp_alloc_aligned); +MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes) +{ + void* ptr = ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES), + ZSTD_cwksp_alloc_aligned); + assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0); + return ptr; } /** - * Aligned on sizeof(unsigned). These buffers have the special property that + * Aligned on 64 bytes. These buffers have the special property that * their values remain constrained, allowing us to re-use them without * memset()-ing them. */ -MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) { +MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) +{ const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned; - void* alloc = ws->tableEnd; - void* end = (BYTE *)alloc + bytes; - void* top = ws->allocStart; + void* alloc; + void* end; + void* top; + + if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) { + return NULL; + } + alloc = ws->tableEnd; + end = (BYTE *)alloc + bytes; + top = ws->allocStart; DEBUGLOG(5, "cwksp: reserving %p table %zd bytes, %zd bytes remaining", alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes); assert((bytes & (sizeof(U32)-1)) == 0); - ZSTD_cwksp_internal_advance_phase(ws, phase); ZSTD_cwksp_assert_internal_consistency(ws); assert(end <= top); if (end > top) { @@ -296,35 +406,41 @@ MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) { } ws->tableEnd = end; -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) - __asan_unpoison_memory_region(alloc, bytes); +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) + if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) { + __asan_unpoison_memory_region(alloc, bytes); + } #endif + assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0); + assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0); return alloc; } /** * Aligned on sizeof(void*). + * Note : should happen only once, at workspace first initialization */ -MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) { - size_t roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*)); +MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) +{ + size_t const roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*)); void* alloc = ws->objectEnd; void* end = (BYTE*)alloc + roundedBytes; -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) /* over-reserve space */ end = (BYTE *)end + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE; #endif - DEBUGLOG(5, + DEBUGLOG(4, "cwksp: reserving %p object %zd bytes (rounded to %zd), %zd bytes remaining", alloc, bytes, roundedBytes, ZSTD_cwksp_available_space(ws) - roundedBytes); - assert(((size_t)alloc & (sizeof(void*)-1)) == 0); - assert((bytes & (sizeof(void*)-1)) == 0); + assert((size_t)alloc % ZSTD_ALIGNOF(void*) == 0); + assert(bytes % ZSTD_ALIGNOF(void*) == 0); ZSTD_cwksp_assert_internal_consistency(ws); /* we must be in the first phase, no advance is possible */ if (ws->phase != ZSTD_cwksp_alloc_objects || end > ws->workspaceEnd) { - DEBUGLOG(4, "cwksp: object alloc failed!"); + DEBUGLOG(3, "cwksp: object alloc failed!"); ws->allocFailed = 1; return NULL; } @@ -332,20 +448,23 @@ MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) { ws->tableEnd = end; ws->tableValidEnd = end; -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on * either size. */ - alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE; - __asan_unpoison_memory_region(alloc, bytes); + alloc = (BYTE*)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE; + if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) { + __asan_unpoison_memory_region(alloc, bytes); + } #endif return alloc; } -MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws) { +MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws) +{ DEBUGLOG(4, "cwksp: ZSTD_cwksp_mark_tables_dirty"); -#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) +#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) /* To validate that the table re-use logic is sound, and that we don't * access table space that we haven't cleaned, we re-"poison" the table * space every time we mark it dirty. */ @@ -380,7 +499,7 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) { assert(ws->tableValidEnd >= ws->objectEnd); assert(ws->tableValidEnd <= ws->allocStart); if (ws->tableValidEnd < ws->tableEnd) { - memset(ws->tableValidEnd, 0, (BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd); + ZSTD_memset(ws->tableValidEnd, 0, (BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd); } ZSTD_cwksp_mark_tables_clean(ws); } @@ -392,8 +511,12 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) { MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) { DEBUGLOG(4, "cwksp: clearing tables!"); -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) - { +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) + /* We don't do this when the workspace is statically allocated, because + * when that is the case, we have no capability to hook into the end of the + * workspace's lifecycle to unpoison the memory. + */ + if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) { size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd; __asan_poison_memory_region(ws->objectEnd, size); } @@ -410,7 +533,7 @@ MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) { MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) { DEBUGLOG(4, "cwksp: clearing!"); -#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) +#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE) /* To validate that the context re-use logic is sound, and that we don't * access stuff that this compression hasn't initialized, we re-"poison" * the workspace (or at least the non-static, non-table parts of it) @@ -421,8 +544,12 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) { } #endif -#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) - { +#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE) + /* We don't do this when the workspace is statically allocated, because + * when that is the case, we have no capability to hook into the end of the + * workspace's lifecycle to unpoison the memory. + */ + if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) { size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->objectEnd; __asan_poison_memory_region(ws->objectEnd, size); } @@ -442,7 +569,7 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) { * Any existing values in the workspace are ignored (the previously managed * buffer, if present, must be separately freed). */ -MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size) { +MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size, ZSTD_cwksp_static_alloc_e isStatic) { DEBUGLOG(4, "cwksp: init'ing workspace with %zd bytes", size); assert(((size_t)start & (sizeof(void*)-1)) == 0); /* ensure correct alignment */ ws->workspace = start; @@ -450,39 +577,45 @@ MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size) { ws->objectEnd = ws->workspace; ws->tableValidEnd = ws->objectEnd; ws->phase = ZSTD_cwksp_alloc_objects; + ws->isStatic = isStatic; ZSTD_cwksp_clear(ws); ws->workspaceOversizedDuration = 0; ZSTD_cwksp_assert_internal_consistency(ws); } MEM_STATIC size_t ZSTD_cwksp_create(ZSTD_cwksp* ws, size_t size, ZSTD_customMem customMem) { - void* workspace = ZSTD_malloc(size, customMem); + void* workspace = ZSTD_customMalloc(size, customMem); DEBUGLOG(4, "cwksp: creating new workspace with %zd bytes", size); RETURN_ERROR_IF(workspace == NULL, memory_allocation, "NULL pointer!"); - ZSTD_cwksp_init(ws, workspace, size); + ZSTD_cwksp_init(ws, workspace, size, ZSTD_cwksp_dynamic_alloc); return 0; } MEM_STATIC void ZSTD_cwksp_free(ZSTD_cwksp* ws, ZSTD_customMem customMem) { void *ptr = ws->workspace; DEBUGLOG(4, "cwksp: freeing workspace"); - memset(ws, 0, sizeof(ZSTD_cwksp)); - ZSTD_free(ptr, customMem); + ZSTD_memset(ws, 0, sizeof(ZSTD_cwksp)); + ZSTD_customFree(ptr, customMem); } /** * Moves the management of a workspace from one cwksp to another. The src cwksp - * is left in an invalid state (src must be re-init()'ed before its used again). + * is left in an invalid state (src must be re-init()'ed before it's used again). */ MEM_STATIC void ZSTD_cwksp_move(ZSTD_cwksp* dst, ZSTD_cwksp* src) { *dst = *src; - memset(src, 0, sizeof(ZSTD_cwksp)); + ZSTD_memset(src, 0, sizeof(ZSTD_cwksp)); } MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) { return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace); } +MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) { + return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace) + + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart); +} + MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) { return ws->allocFailed; } @@ -491,6 +624,24 @@ MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) { * Functions Checking Free Space ***************************************/ +/* ZSTD_alignmentSpaceWithinBounds() : + * Returns if the estimated space needed for a wksp is within an acceptable limit of the + * actual amount of space used. + */ +MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp* const ws, + size_t const estimatedSpace, int resizedWorkspace) { + if (resizedWorkspace) { + /* Resized/newly allocated wksp should have exact bounds */ + return ZSTD_cwksp_used(ws) == estimatedSpace; + } else { + /* Due to alignment, when reusing a workspace, we can actually consume 63 fewer or more bytes + * than estimatedSpace. See the comments in zstd_cwksp.h for details. + */ + return (ZSTD_cwksp_used(ws) >= estimatedSpace - 63) && (ZSTD_cwksp_used(ws) <= estimatedSpace + 63); + } +} + + MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws) { return (size_t)((BYTE*)ws->allocStart - (BYTE*)ws->tableEnd); } diff --git a/thirdparty/zstd/compress/zstd_double_fast.c b/thirdparty/zstd/compress/zstd_double_fast.c index 27eed66cfe..76933dea26 100644 --- a/thirdparty/zstd/compress/zstd_double_fast.c +++ b/thirdparty/zstd/compress/zstd_double_fast.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 @@ -31,15 +31,15 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, * is empty. */ for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) { - U32 const current = (U32)(ip - base); + U32 const curr = (U32)(ip - base); U32 i; for (i = 0; i < fastHashFillStep; ++i) { size_t const smHash = ZSTD_hashPtr(ip + i, hBitsS, mls); size_t const lgHash = ZSTD_hashPtr(ip + i, hBitsL, 8); if (i == 0) - hashSmall[smHash] = current + i; + hashSmall[smHash] = curr + i; if (i == 0 || hashLarge[lgHash] == 0) - hashLarge[lgHash] = current + i; + hashLarge[lgHash] = curr + i; /* Only load extra positions for ZSTD_dtlm_full */ if (dtlm == ZSTD_dtlm_fast) break; @@ -48,10 +48,9 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, FORCE_INLINE_TEMPLATE -size_t ZSTD_compressBlock_doubleFast_generic( +size_t ZSTD_compressBlock_doubleFast_noDict_generic( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], - void const* src, size_t srcSize, - U32 const mls /* template */, ZSTD_dictMode_e const dictMode) + void const* src, size_t srcSize, U32 const mls /* template */) { ZSTD_compressionParameters const* cParams = &ms->cParams; U32* const hashLong = ms->hashTable; @@ -60,7 +59,6 @@ size_t ZSTD_compressBlock_doubleFast_generic( const U32 hBitsS = cParams->chainLog; const BYTE* const base = ms->window.base; const BYTE* const istart = (const BYTE*)src; - const BYTE* ip = istart; const BYTE* anchor = istart; const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); /* presumes that, if there is a dictionary, it must be using Attach mode */ @@ -71,55 +69,239 @@ size_t ZSTD_compressBlock_doubleFast_generic( U32 offset_1=rep[0], offset_2=rep[1]; U32 offsetSaved = 0; - const ZSTD_matchState_t* const dms = ms->dictMatchState; - const ZSTD_compressionParameters* const dictCParams = - dictMode == ZSTD_dictMatchState ? - &dms->cParams : NULL; - const U32* const dictHashLong = dictMode == ZSTD_dictMatchState ? - dms->hashTable : NULL; - const U32* const dictHashSmall = dictMode == ZSTD_dictMatchState ? - dms->chainTable : NULL; - const U32 dictStartIndex = dictMode == ZSTD_dictMatchState ? - dms->window.dictLimit : 0; - const BYTE* const dictBase = dictMode == ZSTD_dictMatchState ? - dms->window.base : NULL; - const BYTE* const dictStart = dictMode == ZSTD_dictMatchState ? - dictBase + dictStartIndex : NULL; - const BYTE* const dictEnd = dictMode == ZSTD_dictMatchState ? - dms->window.nextSrc : NULL; - const U32 dictIndexDelta = dictMode == ZSTD_dictMatchState ? - prefixLowestIndex - (U32)(dictEnd - dictBase) : - 0; - const U32 dictHBitsL = dictMode == ZSTD_dictMatchState ? - dictCParams->hashLog : hBitsL; - const U32 dictHBitsS = dictMode == ZSTD_dictMatchState ? - dictCParams->chainLog : hBitsS; - const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictStart)); + size_t mLength; + U32 offset; + U32 curr; - DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_generic"); + /* how many positions to search before increasing step size */ + const size_t kStepIncr = 1 << kSearchStrength; + /* the position at which to increment the step size if no match is found */ + const BYTE* nextStep; + size_t step; /* the current step size */ - assert(dictMode == ZSTD_noDict || dictMode == ZSTD_dictMatchState); + size_t hl0; /* the long hash at ip */ + size_t hl1; /* the long hash at ip1 */ - /* if a dictionary is attached, it must be within window range */ - if (dictMode == ZSTD_dictMatchState) { - assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex); - } + U32 idxl0; /* the long match index for ip */ + U32 idxl1; /* the long match index for ip1 */ + + const BYTE* matchl0; /* the long match for ip */ + const BYTE* matchs0; /* the short match for ip */ + const BYTE* matchl1; /* the long match for ip1 */ + + const BYTE* ip = istart; /* the current position */ + const BYTE* ip1; /* the next position */ + + DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic"); /* init */ - ip += (dictAndPrefixLength == 0); - if (dictMode == ZSTD_noDict) { + ip += ((ip - prefixLowest) == 0); + { U32 const current = (U32)(ip - base); U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog); U32 const maxRep = current - windowLow; if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0; if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0; } - if (dictMode == ZSTD_dictMatchState) { - /* dictMatchState repCode checks don't currently handle repCode == 0 - * disabling. */ - assert(offset_1 <= dictAndPrefixLength); - assert(offset_2 <= dictAndPrefixLength); + + /* Outer Loop: one iteration per match found and stored */ + while (1) { + step = 1; + nextStep = ip + kStepIncr; + ip1 = ip + step; + + if (ip1 > ilimit) { + goto _cleanup; + } + + hl0 = ZSTD_hashPtr(ip, hBitsL, 8); + idxl0 = hashLong[hl0]; + matchl0 = base + idxl0; + + /* Inner Loop: one iteration per search / position */ + do { + const size_t hs0 = ZSTD_hashPtr(ip, hBitsS, mls); + const U32 idxs0 = hashSmall[hs0]; + curr = (U32)(ip-base); + matchs0 = base + idxs0; + + hashLong[hl0] = hashSmall[hs0] = curr; /* update hash tables */ + + /* check noDict repcode */ + if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) { + mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4; + ip++; + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength); + goto _match_stored; + } + + hl1 = ZSTD_hashPtr(ip1, hBitsL, 8); + + if (idxl0 > prefixLowestIndex) { + /* check prefix long match */ + if (MEM_read64(matchl0) == MEM_read64(ip)) { + mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8; + offset = (U32)(ip-matchl0); + while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */ + goto _match_found; + } + } + + idxl1 = hashLong[hl1]; + matchl1 = base + idxl1; + + if (idxs0 > prefixLowestIndex) { + /* check prefix short match */ + if (MEM_read32(matchs0) == MEM_read32(ip)) { + goto _search_next_long; + } + } + + if (ip1 >= nextStep) { + PREFETCH_L1(ip1 + 64); + PREFETCH_L1(ip1 + 128); + step++; + nextStep += kStepIncr; + } + ip = ip1; + ip1 += step; + + hl0 = hl1; + idxl0 = idxl1; + matchl0 = matchl1; + #if defined(__aarch64__) + PREFETCH_L1(ip+256); + #endif + } while (ip1 <= ilimit); + +_cleanup: + /* save reps for next block */ + rep[0] = offset_1 ? offset_1 : offsetSaved; + rep[1] = offset_2 ? offset_2 : offsetSaved; + + /* Return the last literals size */ + return (size_t)(iend - anchor); + +_search_next_long: + + /* check prefix long +1 match */ + if (idxl1 > prefixLowestIndex) { + if (MEM_read64(matchl1) == MEM_read64(ip1)) { + ip = ip1; + mLength = ZSTD_count(ip+8, matchl1+8, iend) + 8; + offset = (U32)(ip-matchl1); + while (((ip>anchor) & (matchl1>prefixLowest)) && (ip[-1] == matchl1[-1])) { ip--; matchl1--; mLength++; } /* catch up */ + goto _match_found; + } + } + + /* if no long +1 match, explore the short match we found */ + mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4; + offset = (U32)(ip - matchs0); + while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* catch up */ + + /* fall-through */ + +_match_found: /* requires ip, offset, mLength */ + offset_2 = offset_1; + offset_1 = offset; + + if (step < 4) { + /* It is unsafe to write this value back to the hashtable when ip1 is + * greater than or equal to the new ip we will have after we're done + * processing this match. Rather than perform that test directly + * (ip1 >= ip + mLength), which costs speed in practice, we do a simpler + * more predictable test. The minmatch even if we take a short match is + * 4 bytes, so as long as step, the distance between ip and ip1 + * (initially) is less than 4, we know ip1 < new ip. */ + hashLong[hl1] = (U32)(ip1 - base); + } + + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); + +_match_stored: + /* match found */ + ip += mLength; + anchor = ip; + + if (ip <= ilimit) { + /* Complementary insertion */ + /* done after iLimit test, as candidates could be > iend-8 */ + { U32 const indexToInsert = curr+2; + hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert; + hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base); + hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert; + hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base); + } + + /* check immediate repcode */ + while ( (ip <= ilimit) + && ( (offset_2>0) + & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) { + /* store sequence */ + size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4; + U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; /* swap offset_2 <=> offset_1 */ + hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base); + hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base); + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, rLength); + ip += rLength; + anchor = ip; + continue; /* faster when present ... (?) */ + } + } } +} + + +FORCE_INLINE_TEMPLATE +size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize, + U32 const mls /* template */) +{ + ZSTD_compressionParameters const* cParams = &ms->cParams; + U32* const hashLong = ms->hashTable; + const U32 hBitsL = cParams->hashLog; + U32* const hashSmall = ms->chainTable; + const U32 hBitsS = cParams->chainLog; + const BYTE* const base = ms->window.base; + const BYTE* const istart = (const BYTE*)src; + const BYTE* ip = istart; + const BYTE* anchor = istart; + const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); + /* presumes that, if there is a dictionary, it must be using Attach mode */ + const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog); + const BYTE* const prefixLowest = base + prefixLowestIndex; + const BYTE* const iend = istart + srcSize; + const BYTE* const ilimit = iend - HASH_READ_SIZE; + U32 offset_1=rep[0], offset_2=rep[1]; + U32 offsetSaved = 0; + + const ZSTD_matchState_t* const dms = ms->dictMatchState; + const ZSTD_compressionParameters* const dictCParams = &dms->cParams; + const U32* const dictHashLong = dms->hashTable; + const U32* const dictHashSmall = dms->chainTable; + const U32 dictStartIndex = dms->window.dictLimit; + const BYTE* const dictBase = dms->window.base; + const BYTE* const dictStart = dictBase + dictStartIndex; + const BYTE* const dictEnd = dms->window.nextSrc; + const U32 dictIndexDelta = prefixLowestIndex - (U32)(dictEnd - dictBase); + const U32 dictHBitsL = dictCParams->hashLog; + const U32 dictHBitsS = dictCParams->chainLog; + const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictStart)); + + DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic"); + + /* if a dictionary is attached, it must be within window range */ + assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex); + + /* init */ + ip += (dictAndPrefixLength == 0); + + /* dictMatchState repCode checks don't currently handle repCode == 0 + * disabling. */ + assert(offset_1 <= dictAndPrefixLength); + assert(offset_2 <= dictAndPrefixLength); /* Main Search Loop */ while (ip < ilimit) { /* < instead of <=, because repcode check at (ip+1) */ @@ -129,35 +311,24 @@ size_t ZSTD_compressBlock_doubleFast_generic( size_t const h = ZSTD_hashPtr(ip, hBitsS, mls); size_t const dictHL = ZSTD_hashPtr(ip, dictHBitsL, 8); size_t const dictHS = ZSTD_hashPtr(ip, dictHBitsS, mls); - U32 const current = (U32)(ip-base); + U32 const curr = (U32)(ip-base); U32 const matchIndexL = hashLong[h2]; U32 matchIndexS = hashSmall[h]; const BYTE* matchLong = base + matchIndexL; const BYTE* match = base + matchIndexS; - const U32 repIndex = current + 1 - offset_1; - const BYTE* repMatch = (dictMode == ZSTD_dictMatchState - && repIndex < prefixLowestIndex) ? + const U32 repIndex = curr + 1 - offset_1; + const BYTE* repMatch = (repIndex < prefixLowestIndex) ? dictBase + (repIndex - dictIndexDelta) : base + repIndex; - hashLong[h2] = hashSmall[h] = current; /* update hash tables */ + hashLong[h2] = hashSmall[h] = curr; /* update hash tables */ - /* check dictMatchState repcode */ - if (dictMode == ZSTD_dictMatchState - && ((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */) + /* check repcode */ + if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */) && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) { const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend; mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4; ip++; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH); - goto _match_stored; - } - - /* check noDict repcode */ - if ( dictMode == ZSTD_noDict - && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) { - mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4; - ip++; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength); goto _match_stored; } @@ -169,7 +340,7 @@ size_t ZSTD_compressBlock_doubleFast_generic( while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */ goto _match_found; } - } else if (dictMode == ZSTD_dictMatchState) { + } else { /* check dictMatchState long match */ U32 const dictMatchIndexL = dictHashLong[dictHL]; const BYTE* dictMatchL = dictBase + dictMatchIndexL; @@ -177,7 +348,7 @@ size_t ZSTD_compressBlock_doubleFast_generic( if (dictMatchL > dictStart && MEM_read64(dictMatchL) == MEM_read64(ip)) { mLength = ZSTD_count_2segments(ip+8, dictMatchL+8, iend, dictEnd, prefixLowest) + 8; - offset = (U32)(current - dictMatchIndexL - dictIndexDelta); + offset = (U32)(curr - dictMatchIndexL - dictIndexDelta); while (((ip>anchor) & (dictMatchL>dictStart)) && (ip[-1] == dictMatchL[-1])) { ip--; dictMatchL--; mLength++; } /* catch up */ goto _match_found; } } @@ -187,7 +358,7 @@ size_t ZSTD_compressBlock_doubleFast_generic( if (MEM_read32(match) == MEM_read32(ip)) { goto _search_next_long; } - } else if (dictMode == ZSTD_dictMatchState) { + } else { /* check dictMatchState short match */ U32 const dictMatchIndexS = dictHashSmall[dictHS]; match = dictBase + dictMatchIndexS; @@ -209,7 +380,7 @@ _search_next_long: size_t const dictHLNext = ZSTD_hashPtr(ip+1, dictHBitsL, 8); U32 const matchIndexL3 = hashLong[hl3]; const BYTE* matchL3 = base + matchIndexL3; - hashLong[hl3] = current + 1; + hashLong[hl3] = curr + 1; /* check prefix long +1 match */ if (matchIndexL3 > prefixLowestIndex) { @@ -220,7 +391,7 @@ _search_next_long: while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */ goto _match_found; } - } else if (dictMode == ZSTD_dictMatchState) { + } else { /* check dict long +1 match */ U32 const dictMatchIndexL3 = dictHashLong[dictHLNext]; const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3; @@ -228,15 +399,15 @@ _search_next_long: if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) { mLength = ZSTD_count_2segments(ip+1+8, dictMatchL3+8, iend, dictEnd, prefixLowest) + 8; ip++; - offset = (U32)(current + 1 - dictMatchIndexL3 - dictIndexDelta); + offset = (U32)(curr + 1 - dictMatchIndexL3 - dictIndexDelta); while (((ip>anchor) & (dictMatchL3>dictStart)) && (ip[-1] == dictMatchL3[-1])) { ip--; dictMatchL3--; mLength++; } /* catch up */ goto _match_found; } } } /* if no long +1 match, explore the short match we found */ - if (dictMode == ZSTD_dictMatchState && matchIndexS < prefixLowestIndex) { + if (matchIndexS < prefixLowestIndex) { mLength = ZSTD_count_2segments(ip+4, match+4, iend, dictEnd, prefixLowest) + 4; - offset = (U32)(current - matchIndexS); + offset = (U32)(curr - matchIndexS); while (((ip>anchor) & (match>dictStart)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */ } else { mLength = ZSTD_count(ip+4, match+4, iend) + 4; @@ -244,13 +415,11 @@ _search_next_long: while (((ip>anchor) & (match>prefixLowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */ } - /* fall-through */ - _match_found: offset_2 = offset_1; offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); _match_stored: /* match found */ @@ -260,7 +429,7 @@ _match_stored: if (ip <= ilimit) { /* Complementary insertion */ /* done after iLimit test, as candidates could be > iend-8 */ - { U32 const indexToInsert = current+2; + { U32 const indexToInsert = curr+2; hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert; hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base); hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert; @@ -268,43 +437,27 @@ _match_stored: } /* check immediate repcode */ - if (dictMode == ZSTD_dictMatchState) { - while (ip <= ilimit) { - U32 const current2 = (U32)(ip-base); - U32 const repIndex2 = current2 - offset_2; - const BYTE* repMatch2 = dictMode == ZSTD_dictMatchState - && repIndex2 < prefixLowestIndex ? - dictBase + repIndex2 - dictIndexDelta : - base + repIndex2; - if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */) - && (MEM_read32(repMatch2) == MEM_read32(ip)) ) { - const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend; - size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4; - U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH); - hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2; - hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2; - ip += repLength2; - anchor = ip; - continue; - } - break; - } } - - if (dictMode == ZSTD_noDict) { - while ( (ip <= ilimit) - && ( (offset_2>0) - & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) { - /* store sequence */ - size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4; - U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; /* swap offset_2 <=> offset_1 */ - hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base); - hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base); - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, rLength-MINMATCH); - ip += rLength; + while (ip <= ilimit) { + U32 const current2 = (U32)(ip-base); + U32 const repIndex2 = current2 - offset_2; + const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ? + dictBase + repIndex2 - dictIndexDelta : + base + repIndex2; + if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */) + && (MEM_read32(repMatch2) == MEM_read32(ip)) ) { + const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend; + size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4; + U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */ + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2); + hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2; + hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2; + ip += repLength2; anchor = ip; - continue; /* faster when present ... (?) */ - } } } + continue; + } + break; + } + } } /* while (ip < ilimit) */ /* save reps for next block */ @@ -315,6 +468,24 @@ _match_stored: return (size_t)(iend - anchor); } +#define ZSTD_GEN_DFAST_FN(dictMode, mls) \ + static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls( \ + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], \ + void const* src, size_t srcSize) \ + { \ + return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \ + } + +ZSTD_GEN_DFAST_FN(noDict, 4) +ZSTD_GEN_DFAST_FN(noDict, 5) +ZSTD_GEN_DFAST_FN(noDict, 6) +ZSTD_GEN_DFAST_FN(noDict, 7) + +ZSTD_GEN_DFAST_FN(dictMatchState, 4) +ZSTD_GEN_DFAST_FN(dictMatchState, 5) +ZSTD_GEN_DFAST_FN(dictMatchState, 6) +ZSTD_GEN_DFAST_FN(dictMatchState, 7) + size_t ZSTD_compressBlock_doubleFast( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -325,13 +496,13 @@ size_t ZSTD_compressBlock_doubleFast( { default: /* includes case 3 */ case 4 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_noDict); + return ZSTD_compressBlock_doubleFast_noDict_4(ms, seqStore, rep, src, srcSize); case 5 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_noDict); + return ZSTD_compressBlock_doubleFast_noDict_5(ms, seqStore, rep, src, srcSize); case 6 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_noDict); + return ZSTD_compressBlock_doubleFast_noDict_6(ms, seqStore, rep, src, srcSize); case 7 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_noDict); + return ZSTD_compressBlock_doubleFast_noDict_7(ms, seqStore, rep, src, srcSize); } } @@ -345,13 +516,13 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState( { default: /* includes case 3 */ case 4 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_dictMatchState); + return ZSTD_compressBlock_doubleFast_dictMatchState_4(ms, seqStore, rep, src, srcSize); case 5 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_dictMatchState); + return ZSTD_compressBlock_doubleFast_dictMatchState_5(ms, seqStore, rep, src, srcSize); case 6 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_dictMatchState); + return ZSTD_compressBlock_doubleFast_dictMatchState_6(ms, seqStore, rep, src, srcSize); case 7 : - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_dictMatchState); + return ZSTD_compressBlock_doubleFast_dictMatchState_7(ms, seqStore, rep, src, srcSize); } } @@ -387,7 +558,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( /* if extDict is invalidated due to maxDistance, switch to "regular" variant */ if (prefixStartIndex == dictStartIndex) - return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, mls, ZSTD_noDict); + return ZSTD_compressBlock_doubleFast(ms, seqStore, rep, src, srcSize); /* Search Loop */ while (ip < ilimit) { /* < instead of <=, because (ip+1) */ @@ -401,31 +572,31 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( const BYTE* const matchLongBase = matchLongIndex < prefixStartIndex ? dictBase : base; const BYTE* matchLong = matchLongBase + matchLongIndex; - const U32 current = (U32)(ip-base); - const U32 repIndex = current + 1 - offset_1; /* offset_1 expected <= current +1 */ + const U32 curr = (U32)(ip-base); + const U32 repIndex = curr + 1 - offset_1; /* offset_1 expected <= curr +1 */ const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; size_t mLength; - hashSmall[hSmall] = hashLong[hLong] = current; /* update hash table */ + hashSmall[hSmall] = hashLong[hLong] = curr; /* update hash table */ if ((((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex doesn't overlap dict + prefix */ - & (repIndex > dictStartIndex)) + & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */ && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) { const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend; mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4; ip++; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength); } else { if ((matchLongIndex > dictStartIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) { const BYTE* const matchEnd = matchLongIndex < prefixStartIndex ? dictEnd : iend; const BYTE* const lowMatchPtr = matchLongIndex < prefixStartIndex ? dictStart : prefixStart; U32 offset; mLength = ZSTD_count_2segments(ip+8, matchLong+8, iend, matchEnd, prefixStart) + 8; - offset = current - matchLongIndex; + offset = curr - matchLongIndex; while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */ offset_2 = offset_1; offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); } else if ((matchIndex > dictStartIndex) && (MEM_read32(match) == MEM_read32(ip))) { size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8); @@ -433,24 +604,24 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( const BYTE* const match3Base = matchIndex3 < prefixStartIndex ? dictBase : base; const BYTE* match3 = match3Base + matchIndex3; U32 offset; - hashLong[h3] = current + 1; + hashLong[h3] = curr + 1; if ( (matchIndex3 > dictStartIndex) && (MEM_read64(match3) == MEM_read64(ip+1)) ) { const BYTE* const matchEnd = matchIndex3 < prefixStartIndex ? dictEnd : iend; const BYTE* const lowMatchPtr = matchIndex3 < prefixStartIndex ? dictStart : prefixStart; mLength = ZSTD_count_2segments(ip+9, match3+8, iend, matchEnd, prefixStart) + 8; ip++; - offset = current+1 - matchIndex3; + offset = curr+1 - matchIndex3; while (((ip>anchor) & (match3>lowMatchPtr)) && (ip[-1] == match3[-1])) { ip--; match3--; mLength++; } /* catch up */ } else { const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend; const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart; mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4; - offset = current - matchIndex; + offset = curr - matchIndex; while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */ } offset_2 = offset_1; offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); } else { ip += ((ip-anchor) >> kSearchStrength) + 1; @@ -464,7 +635,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( if (ip <= ilimit) { /* Complementary insertion */ /* done after iLimit test, as candidates could be > iend-8 */ - { U32 const indexToInsert = current+2; + { U32 const indexToInsert = curr+2; hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert; hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base); hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert; @@ -477,12 +648,12 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( U32 const repIndex2 = current2 - offset_2; const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2; if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) /* intentional overflow : ensure repIndex2 doesn't overlap dict + prefix */ - & (repIndex2 > dictStartIndex)) + & (offset_2 <= current2 - dictStartIndex)) && (MEM_read32(repMatch2) == MEM_read32(ip)) ) { const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend; size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4; U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH); + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2); hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2; hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2; ip += repLength2; @@ -500,6 +671,10 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic( return (size_t)(iend - anchor); } +ZSTD_GEN_DFAST_FN(extDict, 4) +ZSTD_GEN_DFAST_FN(extDict, 5) +ZSTD_GEN_DFAST_FN(extDict, 6) +ZSTD_GEN_DFAST_FN(extDict, 7) size_t ZSTD_compressBlock_doubleFast_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -510,12 +685,12 @@ size_t ZSTD_compressBlock_doubleFast_extDict( { default: /* includes case 3 */ case 4 : - return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 4); + return ZSTD_compressBlock_doubleFast_extDict_4(ms, seqStore, rep, src, srcSize); case 5 : - return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 5); + return ZSTD_compressBlock_doubleFast_extDict_5(ms, seqStore, rep, src, srcSize); case 6 : - return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 6); + return ZSTD_compressBlock_doubleFast_extDict_6(ms, seqStore, rep, src, srcSize); case 7 : - return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 7); + return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize); } } diff --git a/thirdparty/zstd/compress/zstd_double_fast.h b/thirdparty/zstd/compress/zstd_double_fast.h index 14d944d69b..e16b7b03a3 100644 --- a/thirdparty/zstd/compress/zstd_double_fast.h +++ b/thirdparty/zstd/compress/zstd_double_fast.h @@ -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 diff --git a/thirdparty/zstd/compress/zstd_fast.c b/thirdparty/zstd/compress/zstd_fast.c index 85a3a7a91e..802fc31579 100644 --- a/thirdparty/zstd/compress/zstd_fast.c +++ b/thirdparty/zstd/compress/zstd_fast.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 @@ -29,159 +29,308 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms, * Insert the other positions if their hash entry is empty. */ for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) { - U32 const current = (U32)(ip - base); + U32 const curr = (U32)(ip - base); size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls); - hashTable[hash0] = current; + hashTable[hash0] = curr; if (dtlm == ZSTD_dtlm_fast) continue; /* Only load extra positions for ZSTD_dtlm_full */ { U32 p; for (p = 1; p < fastHashFillStep; ++p) { size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls); if (hashTable[hash] == 0) { /* not yet filled */ - hashTable[hash] = current + p; + hashTable[hash] = curr + p; } } } } } +/** + * If you squint hard enough (and ignore repcodes), the search operation at any + * given position is broken into 4 stages: + * + * 1. Hash (map position to hash value via input read) + * 2. Lookup (map hash val to index via hashtable read) + * 3. Load (map index to value at that position via input read) + * 4. Compare + * + * Each of these steps involves a memory read at an address which is computed + * from the previous step. This means these steps must be sequenced and their + * latencies are cumulative. + * + * Rather than do 1->2->3->4 sequentially for a single position before moving + * onto the next, this implementation interleaves these operations across the + * next few positions: + * + * R = Repcode Read & Compare + * H = Hash + * T = Table Lookup + * M = Match Read & Compare + * + * Pos | Time --> + * ----+------------------- + * N | ... M + * N+1 | ... TM + * N+2 | R H T M + * N+3 | H TM + * N+4 | R H T M + * N+5 | H ... + * N+6 | R ... + * + * This is very much analogous to the pipelining of execution in a CPU. And just + * like a CPU, we have to dump the pipeline when we find a match (i.e., take a + * branch). + * + * When this happens, we throw away our current state, and do the following prep + * to re-enter the loop: + * + * Pos | Time --> + * ----+------------------- + * N | H T + * N+1 | H + * + * This is also the work we do at the beginning to enter the loop initially. + */ FORCE_INLINE_TEMPLATE size_t -ZSTD_compressBlock_fast_generic( +ZSTD_compressBlock_fast_noDict_generic( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize, - U32 const mls) + U32 const mls, U32 const hasStep) { const ZSTD_compressionParameters* const cParams = &ms->cParams; U32* const hashTable = ms->hashTable; U32 const hlog = cParams->hashLog; /* support stepSize of 0 */ - size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1; + size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2; const BYTE* const base = ms->window.base; const BYTE* const istart = (const BYTE*)src; - /* We check ip0 (ip + 0) and ip1 (ip + 1) each loop */ - const BYTE* ip0 = istart; - const BYTE* ip1; - const BYTE* anchor = istart; const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); const U32 prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog); const BYTE* const prefixStart = base + prefixStartIndex; const BYTE* const iend = istart + srcSize; const BYTE* const ilimit = iend - HASH_READ_SIZE; - U32 offset_1=rep[0], offset_2=rep[1]; + + const BYTE* anchor = istart; + const BYTE* ip0 = istart; + const BYTE* ip1; + const BYTE* ip2; + const BYTE* ip3; + U32 current0; + + U32 rep_offset1 = rep[0]; + U32 rep_offset2 = rep[1]; U32 offsetSaved = 0; - /* init */ + size_t hash0; /* hash for ip0 */ + size_t hash1; /* hash for ip1 */ + U32 idx; /* match idx for ip0 */ + U32 mval; /* src value at match idx */ + + U32 offcode; + const BYTE* match0; + size_t mLength; + + /* ip0 and ip1 are always adjacent. The targetLength skipping and + * uncompressibility acceleration is applied to every other position, + * matching the behavior of #1562. step therefore represents the gap + * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */ + size_t step; + const BYTE* nextStep; + const size_t kStepIncr = (1 << (kSearchStrength - 1)); + DEBUGLOG(5, "ZSTD_compressBlock_fast_generic"); ip0 += (ip0 == prefixStart); + { U32 const curr = (U32)(ip0 - base); + U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog); + U32 const maxRep = curr - windowLow; + if (rep_offset2 > maxRep) offsetSaved = rep_offset2, rep_offset2 = 0; + if (rep_offset1 > maxRep) offsetSaved = rep_offset1, rep_offset1 = 0; + } + + /* start each op */ +_start: /* Requires: ip0 */ + + step = stepSize; + nextStep = ip0 + kStepIncr; + + /* calculate positions, ip0 - anchor == 0, so we skip step calc */ ip1 = ip0 + 1; - { U32 const current = (U32)(ip0 - base); - U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog); - U32 const maxRep = current - windowLow; - if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0; - if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0; + ip2 = ip0 + step; + ip3 = ip2 + 1; + + if (ip3 >= ilimit) { + goto _cleanup; } - /* Main Search Loop */ -#ifdef __INTEL_COMPILER - /* From intel 'The vector pragma indicates that the loop should be - * vectorized if it is legal to do so'. Can be used together with - * #pragma ivdep (but have opted to exclude that because intel - * warns against using it).*/ - #pragma vector always -#endif - while (ip1 < ilimit) { /* < instead of <=, because check at ip0+2 */ - size_t mLength; - BYTE const* ip2 = ip0 + 2; - size_t const h0 = ZSTD_hashPtr(ip0, hlog, mls); - U32 const val0 = MEM_read32(ip0); - size_t const h1 = ZSTD_hashPtr(ip1, hlog, mls); - U32 const val1 = MEM_read32(ip1); - U32 const current0 = (U32)(ip0-base); - U32 const current1 = (U32)(ip1-base); - U32 const matchIndex0 = hashTable[h0]; - U32 const matchIndex1 = hashTable[h1]; - BYTE const* repMatch = ip2 - offset_1; - const BYTE* match0 = base + matchIndex0; - const BYTE* match1 = base + matchIndex1; - U32 offcode; - -#if defined(__aarch64__) - PREFETCH_L1(ip0+256); -#endif - - hashTable[h0] = current0; /* update hash table */ - hashTable[h1] = current1; /* update hash table */ - - assert(ip0 + 1 == ip1); - - if ((offset_1 > 0) & (MEM_read32(repMatch) == MEM_read32(ip2))) { - mLength = (ip2[-1] == repMatch[-1]) ? 1 : 0; - ip0 = ip2 - mLength; - match0 = repMatch - mLength; + hash0 = ZSTD_hashPtr(ip0, hlog, mls); + hash1 = ZSTD_hashPtr(ip1, hlog, mls); + + idx = hashTable[hash0]; + + do { + /* load repcode match for ip[2]*/ + const U32 rval = MEM_read32(ip2 - rep_offset1); + + /* write back hash table entry */ + current0 = (U32)(ip0 - base); + hashTable[hash0] = current0; + + /* check repcode at ip[2] */ + if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) { + ip0 = ip2; + match0 = ip0 - rep_offset1; + mLength = ip0[-1] == match0[-1]; + ip0 -= mLength; + match0 -= mLength; + offcode = STORE_REPCODE_1; mLength += 4; - offcode = 0; goto _match; } - if ((matchIndex0 > prefixStartIndex) && MEM_read32(match0) == val0) { - /* found a regular match */ - goto _offset; + + /* load match for ip[0] */ + if (idx >= prefixStartIndex) { + mval = MEM_read32(base + idx); + } else { + mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */ } - if ((matchIndex1 > prefixStartIndex) && MEM_read32(match1) == val1) { - /* found a regular match after one literal */ - ip0 = ip1; - match0 = match1; + + /* check match at ip[0] */ + if (MEM_read32(ip0) == mval) { + /* found a match! */ goto _offset; } - { size_t const step = ((size_t)(ip0-anchor) >> (kSearchStrength - 1)) + stepSize; - assert(step >= 2); - ip0 += step; - ip1 += step; - continue; + + /* lookup ip[1] */ + idx = hashTable[hash1]; + + /* hash ip[2] */ + hash0 = hash1; + hash1 = ZSTD_hashPtr(ip2, hlog, mls); + + /* advance to next positions */ + ip0 = ip1; + ip1 = ip2; + ip2 = ip3; + + /* write back hash table entry */ + current0 = (U32)(ip0 - base); + hashTable[hash0] = current0; + + /* load match for ip[0] */ + if (idx >= prefixStartIndex) { + mval = MEM_read32(base + idx); + } else { + mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */ } -_offset: /* Requires: ip0, match0 */ - /* Compute the offset code */ - offset_2 = offset_1; - offset_1 = (U32)(ip0-match0); - offcode = offset_1 + ZSTD_REP_MOVE; - mLength = 4; - /* Count the backwards match length */ - while (((ip0>anchor) & (match0>prefixStart)) - && (ip0[-1] == match0[-1])) { ip0--; match0--; mLength++; } /* catch up */ -_match: /* Requires: ip0, match0, offcode */ - /* Count the forward length */ - mLength += ZSTD_count(ip0+mLength, match0+mLength, iend); - ZSTD_storeSeq(seqStore, (size_t)(ip0-anchor), anchor, iend, offcode, mLength-MINMATCH); - /* match found */ - ip0 += mLength; - anchor = ip0; + /* check match at ip[0] */ + if (MEM_read32(ip0) == mval) { + /* found a match! */ + goto _offset; + } - if (ip0 <= ilimit) { - /* Fill Table */ - assert(base+current0+2 > istart); /* check base overflow */ - hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */ - hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base); - - if (offset_2 > 0) { /* offset_2==0 means offset_2 is invalidated */ - while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - offset_2)) ) { - /* store sequence */ - size_t const rLength = ZSTD_count(ip0+4, ip0+4-offset_2, iend) + 4; - { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */ - hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base); - ip0 += rLength; - ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, 0 /*offCode*/, rLength-MINMATCH); - anchor = ip0; - continue; /* faster when present (confirmed on gcc-8) ... (?) */ - } } } - ip1 = ip0 + 1; - } + /* lookup ip[1] */ + idx = hashTable[hash1]; + + /* hash ip[2] */ + hash0 = hash1; + hash1 = ZSTD_hashPtr(ip2, hlog, mls); + + /* advance to next positions */ + ip0 = ip1; + ip1 = ip2; + ip2 = ip0 + step; + ip3 = ip1 + step; + + /* calculate step */ + if (ip2 >= nextStep) { + step++; + PREFETCH_L1(ip1 + 64); + PREFETCH_L1(ip1 + 128); + nextStep += kStepIncr; + } + } while (ip3 < ilimit); + +_cleanup: + /* Note that there are probably still a couple positions we could search. + * However, it seems to be a meaningful performance hit to try to search + * them. So let's not. */ /* save reps for next block */ - rep[0] = offset_1 ? offset_1 : offsetSaved; - rep[1] = offset_2 ? offset_2 : offsetSaved; + rep[0] = rep_offset1 ? rep_offset1 : offsetSaved; + rep[1] = rep_offset2 ? rep_offset2 : offsetSaved; /* Return the last literals size */ return (size_t)(iend - anchor); + +_offset: /* Requires: ip0, idx */ + + /* Compute the offset code. */ + match0 = base + idx; + rep_offset2 = rep_offset1; + rep_offset1 = (U32)(ip0-match0); + offcode = STORE_OFFSET(rep_offset1); + mLength = 4; + + /* Count the backwards match length. */ + while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) { + ip0--; + match0--; + mLength++; + } + +_match: /* Requires: ip0, match0, offcode */ + + /* Count the forward length. */ + mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend); + + ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength); + + ip0 += mLength; + anchor = ip0; + + /* write next hash table entry */ + if (ip1 < ip0) { + hashTable[hash1] = (U32)(ip1 - base); + } + + /* Fill table and check for immediate repcode. */ + if (ip0 <= ilimit) { + /* Fill Table */ + assert(base+current0+2 > istart); /* check base overflow */ + hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */ + hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base); + + if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */ + while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) { + /* store sequence */ + size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4; + { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */ + hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base); + ip0 += rLength; + ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, STORE_REPCODE_1, rLength); + anchor = ip0; + continue; /* faster when present (confirmed on gcc-8) ... (?) */ + } } } + + goto _start; } +#define ZSTD_GEN_FAST_FN(dictMode, mls, step) \ + static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step( \ + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], \ + void const* src, size_t srcSize) \ + { \ + return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \ + } + +ZSTD_GEN_FAST_FN(noDict, 4, 1) +ZSTD_GEN_FAST_FN(noDict, 5, 1) +ZSTD_GEN_FAST_FN(noDict, 6, 1) +ZSTD_GEN_FAST_FN(noDict, 7, 1) + +ZSTD_GEN_FAST_FN(noDict, 4, 0) +ZSTD_GEN_FAST_FN(noDict, 5, 0) +ZSTD_GEN_FAST_FN(noDict, 6, 0) +ZSTD_GEN_FAST_FN(noDict, 7, 0) size_t ZSTD_compressBlock_fast( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -189,24 +338,40 @@ size_t ZSTD_compressBlock_fast( { U32 const mls = ms->cParams.minMatch; assert(ms->dictMatchState == NULL); - switch(mls) - { - default: /* includes case 3 */ - case 4 : - return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 4); - case 5 : - return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 5); - case 6 : - return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 6); - case 7 : - return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 7); + if (ms->cParams.targetLength > 1) { + switch(mls) + { + default: /* includes case 3 */ + case 4 : + return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize); + case 5 : + return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize); + case 6 : + return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize); + case 7 : + return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize); + } + } else { + switch(mls) + { + default: /* includes case 3 */ + case 4 : + return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize); + case 5 : + return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize); + case 6 : + return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize); + case 7 : + return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize); + } + } } FORCE_INLINE_TEMPLATE size_t ZSTD_compressBlock_fast_dictMatchState_generic( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], - void const* src, size_t srcSize, U32 const mls) + void const* src, size_t srcSize, U32 const mls, U32 const hasStep) { const ZSTD_compressionParameters* const cParams = &ms->cParams; U32* const hashTable = ms->hashTable; @@ -242,7 +407,9 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( assert(endIndex - prefixStartIndex <= maxDistance); (void)maxDistance; (void)endIndex; /* these variables are not used when assert() is disabled */ - /* ensure there will be no no underflow + (void)hasStep; /* not currently specialized on whether it's accelerated */ + + /* ensure there will be no underflow * when translating a dict index into a local index */ assert(prefixStartIndex >= (U32)(dictEnd - dictBase)); @@ -258,21 +425,21 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( while (ip < ilimit) { /* < instead of <=, because repcode check at (ip+1) */ size_t mLength; size_t const h = ZSTD_hashPtr(ip, hlog, mls); - U32 const current = (U32)(ip-base); + U32 const curr = (U32)(ip-base); U32 const matchIndex = hashTable[h]; const BYTE* match = base + matchIndex; - const U32 repIndex = current + 1 - offset_1; + const U32 repIndex = curr + 1 - offset_1; const BYTE* repMatch = (repIndex < prefixStartIndex) ? dictBase + (repIndex - dictIndexDelta) : base + repIndex; - hashTable[h] = current; /* update hash table */ + hashTable[h] = curr; /* update hash table */ if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */ && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) { const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend; mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4; ip++; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength); } else if ( (matchIndex <= prefixStartIndex) ) { size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls); U32 const dictMatchIndex = dictHashTable[dictHash]; @@ -284,7 +451,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( continue; } else { /* found a dict match */ - U32 const offset = (U32)(current-dictMatchIndex-dictIndexDelta); + U32 const offset = (U32)(curr-dictMatchIndex-dictIndexDelta); mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4; while (((ip>anchor) & (dictMatch>dictStart)) && (ip[-1] == dictMatch[-1])) { @@ -292,7 +459,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( } /* catch up */ offset_2 = offset_1; offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); } } else if (MEM_read32(match) != MEM_read32(ip)) { /* it's not a match, and we're not going to check the dictionary */ @@ -307,7 +474,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */ offset_2 = offset_1; offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); } /* match found */ @@ -316,8 +483,8 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( if (ip <= ilimit) { /* Fill Table */ - assert(base+current+2 > istart); /* check base overflow */ - hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2; /* here because current+2 could be > iend-8 */ + assert(base+curr+2 > istart); /* check base overflow */ + hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2; /* here because curr+2 could be > iend-8 */ hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base); /* check immediate repcode */ @@ -332,7 +499,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend; size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4; U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH); + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2); hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2; ip += repLength2; anchor = ip; @@ -351,6 +518,12 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( return (size_t)(iend - anchor); } + +ZSTD_GEN_FAST_FN(dictMatchState, 4, 0) +ZSTD_GEN_FAST_FN(dictMatchState, 5, 0) +ZSTD_GEN_FAST_FN(dictMatchState, 6, 0) +ZSTD_GEN_FAST_FN(dictMatchState, 7, 0) + size_t ZSTD_compressBlock_fast_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize) @@ -361,20 +534,20 @@ size_t ZSTD_compressBlock_fast_dictMatchState( { default: /* includes case 3 */ case 4 : - return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 4); + return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize); case 5 : - return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 5); + return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize); case 6 : - return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 6); + return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize); case 7 : - return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 7); + return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize); } } static size_t ZSTD_compressBlock_fast_extDict_generic( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], - void const* src, size_t srcSize, U32 const mls) + void const* src, size_t srcSize, U32 const mls, U32 const hasStep) { const ZSTD_compressionParameters* const cParams = &ms->cParams; U32* const hashTable = ms->hashTable; @@ -398,11 +571,13 @@ static size_t ZSTD_compressBlock_fast_extDict_generic( const BYTE* const ilimit = iend - 8; U32 offset_1=rep[0], offset_2=rep[1]; + (void)hasStep; /* not currently specialized on whether it's accelerated */ + DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1); /* switch to "regular" variant if extDict is invalidated due to maxDistance */ if (prefixStartIndex == dictStartIndex) - return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, mls); + return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize); /* Search Loop */ while (ip < ilimit) { /* < instead of <=, because (ip+1) */ @@ -410,20 +585,20 @@ static size_t ZSTD_compressBlock_fast_extDict_generic( const U32 matchIndex = hashTable[h]; const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base; const BYTE* match = matchBase + matchIndex; - const U32 current = (U32)(ip-base); - const U32 repIndex = current + 1 - offset_1; + const U32 curr = (U32)(ip-base); + const U32 repIndex = curr + 1 - offset_1; const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; - hashTable[h] = current; /* update hash table */ - DEBUGLOG(7, "offset_1 = %u , current = %u", offset_1, current); - assert(offset_1 <= current +1); /* check repIndex */ + hashTable[h] = curr; /* update hash table */ + DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr); - if ( (((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > dictStartIndex)) + if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ + & (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */ && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) { const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend; size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4; ip++; - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, rLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, rLength); ip += rLength; anchor = ip; } else { @@ -435,30 +610,30 @@ static size_t ZSTD_compressBlock_fast_extDict_generic( } { const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend; const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart; - U32 const offset = current - matchIndex; + U32 const offset = curr - matchIndex; size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4; while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */ offset_2 = offset_1; offset_1 = offset; /* update offset history */ - ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH); + ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength); ip += mLength; anchor = ip; } } if (ip <= ilimit) { /* Fill Table */ - hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2; + hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2; hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base); /* check immediate repcode */ while (ip <= ilimit) { U32 const current2 = (U32)(ip-base); U32 const repIndex2 = current2 - offset_2; const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2; - if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (repIndex2 > dictStartIndex)) /* intentional overflow */ + if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 <= curr - dictStartIndex)) /* intentional overflow */ && (MEM_read32(repMatch2) == MEM_read32(ip)) ) { const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend; size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4; { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; } /* swap offset_2 <=> offset_1 */ - ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, 0 /*offcode*/, repLength2-MINMATCH); + ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, STORE_REPCODE_1, repLength2); hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2; ip += repLength2; anchor = ip; @@ -475,6 +650,10 @@ static size_t ZSTD_compressBlock_fast_extDict_generic( return (size_t)(iend - anchor); } +ZSTD_GEN_FAST_FN(extDict, 4, 0) +ZSTD_GEN_FAST_FN(extDict, 5, 0) +ZSTD_GEN_FAST_FN(extDict, 6, 0) +ZSTD_GEN_FAST_FN(extDict, 7, 0) size_t ZSTD_compressBlock_fast_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -485,12 +664,12 @@ size_t ZSTD_compressBlock_fast_extDict( { default: /* includes case 3 */ case 4 : - return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 4); + return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize); case 5 : - return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 5); + return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize); case 6 : - return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 6); + return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize); case 7 : - return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 7); + return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize); } } diff --git a/thirdparty/zstd/compress/zstd_fast.h b/thirdparty/zstd/compress/zstd_fast.h index cf6aaa8e67..0d4a0c1090 100644 --- a/thirdparty/zstd/compress/zstd_fast.h +++ b/thirdparty/zstd/compress/zstd_fast.h @@ -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 diff --git a/thirdparty/zstd/compress/zstd_lazy.c b/thirdparty/zstd/compress/zstd_lazy.c index 4cf5c88b53..2e38dcb46d 100644 --- a/thirdparty/zstd/compress/zstd_lazy.c +++ b/thirdparty/zstd/compress/zstd_lazy.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 @@ -58,11 +58,11 @@ ZSTD_updateDUBT(ZSTD_matchState_t* ms, /** ZSTD_insertDUBT1() : * sort one already inserted but unsorted position - * assumption : current >= btlow == (current - btmask) + * assumption : curr >= btlow == (curr - btmask) * doesn't fail */ static void -ZSTD_insertDUBT1(ZSTD_matchState_t* ms, - U32 current, const BYTE* inputEnd, +ZSTD_insertDUBT1(const ZSTD_matchState_t* ms, + U32 curr, const BYTE* inputEnd, U32 nbCompares, U32 btLow, const ZSTD_dictMode_e dictMode) { @@ -74,41 +74,41 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms, const BYTE* const base = ms->window.base; const BYTE* const dictBase = ms->window.dictBase; const U32 dictLimit = ms->window.dictLimit; - const BYTE* const ip = (current>=dictLimit) ? base + current : dictBase + current; - const BYTE* const iend = (current>=dictLimit) ? inputEnd : dictBase + dictLimit; + const BYTE* const ip = (curr>=dictLimit) ? base + curr : dictBase + curr; + const BYTE* const iend = (curr>=dictLimit) ? inputEnd : dictBase + dictLimit; const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const prefixStart = base + dictLimit; const BYTE* match; - U32* smallerPtr = bt + 2*(current&btMask); + U32* smallerPtr = bt + 2*(curr&btMask); U32* largerPtr = smallerPtr + 1; U32 matchIndex = *smallerPtr; /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */ U32 dummy32; /* to be nullified at the end */ U32 const windowValid = ms->window.lowLimit; U32 const maxDistance = 1U << cParams->windowLog; - U32 const windowLow = (current - windowValid > maxDistance) ? current - maxDistance : windowValid; + U32 const windowLow = (curr - windowValid > maxDistance) ? curr - maxDistance : windowValid; DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)", - current, dictLimit, windowLow); - assert(current >= btLow); + curr, dictLimit, windowLow); + assert(curr >= btLow); assert(ip < iend); /* condition for ZSTD_count */ - while (nbCompares-- && (matchIndex > windowLow)) { + for (; nbCompares && (matchIndex > windowLow); --nbCompares) { U32* const nextPtr = bt + 2*(matchIndex & btMask); size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ - assert(matchIndex < current); + assert(matchIndex < curr); /* note : all candidates are now supposed sorted, * but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK * when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */ if ( (dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit) /* both in current segment*/ - || (current < dictLimit) /* both in extDict */) { + || (curr < dictLimit) /* both in extDict */) { const BYTE* const mBase = ( (dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit)) ? base : dictBase; assert( (matchIndex+matchLength >= dictLimit) /* might be wrong if extDict is incorrectly set to 0 */ - || (current < dictLimit) ); + || (curr < dictLimit) ); match = mBase + matchIndex; matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend); } else { @@ -119,7 +119,7 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms, } DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ", - current, matchIndex, (U32)matchLength); + curr, matchIndex, (U32)matchLength); if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */ break; /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */ @@ -151,7 +151,7 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms, static size_t ZSTD_DUBT_findBetterDictMatch ( - ZSTD_matchState_t* ms, + const ZSTD_matchState_t* ms, const BYTE* const ip, const BYTE* const iend, size_t* offsetPtr, size_t bestLength, @@ -168,7 +168,7 @@ ZSTD_DUBT_findBetterDictMatch ( const BYTE* const base = ms->window.base; const BYTE* const prefixStart = base + ms->window.dictLimit; - U32 const current = (U32)(ip-base); + U32 const curr = (U32)(ip-base); const BYTE* const dictBase = dms->window.base; const BYTE* const dictEnd = dms->window.nextSrc; U32 const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base); @@ -185,7 +185,7 @@ ZSTD_DUBT_findBetterDictMatch ( (void)dictMode; assert(dictMode == ZSTD_dictMatchState); - while (nbCompares-- && (dictMatchIndex > dictLowLimit)) { + for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) { U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask); size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ const BYTE* match = dictBase + dictMatchIndex; @@ -195,10 +195,10 @@ ZSTD_DUBT_findBetterDictMatch ( if (matchLength > bestLength) { U32 matchIndex = dictMatchIndex + dictIndexDelta; - if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) { + if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) { DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)", - current, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, ZSTD_REP_MOVE + current - matchIndex, dictMatchIndex, matchIndex); - bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex; + curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, STORE_OFFSET(curr - matchIndex), dictMatchIndex, matchIndex); + bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex); } if (ip+matchLength == iend) { /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */ break; /* drop, to guarantee consistency (miss a little bit of compression) */ @@ -218,9 +218,9 @@ ZSTD_DUBT_findBetterDictMatch ( } if (bestLength >= MINMATCH) { - U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex; + U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex; DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)", - current, (U32)bestLength, (U32)*offsetPtr, mIndex); + curr, (U32)bestLength, (U32)*offsetPtr, mIndex); } return bestLength; @@ -241,13 +241,13 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, U32 matchIndex = hashTable[h]; const BYTE* const base = ms->window.base; - U32 const current = (U32)(ip-base); - U32 const windowLow = ZSTD_getLowestMatchIndex(ms, current, cParams->windowLog); + U32 const curr = (U32)(ip-base); + U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog); U32* const bt = ms->chainTable; U32 const btLog = cParams->chainLog - 1; U32 const btMask = (1 << btLog) - 1; - U32 const btLow = (btMask >= current) ? 0 : current - btMask; + U32 const btLow = (btMask >= curr) ? 0 : curr - btMask; U32 const unsortLimit = MAX(btLow, windowLow); U32* nextCandidate = bt + 2*(matchIndex&btMask); @@ -256,8 +256,9 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, U32 nbCandidates = nbCompares; U32 previousCandidate = 0; - DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", current); + DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", curr); assert(ip <= iend-8); /* required for h calculation */ + assert(dictMode != ZSTD_dedicatedDictSearch); /* reach end of unsorted candidates list */ while ( (matchIndex > unsortLimit) @@ -299,16 +300,16 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, const U32 dictLimit = ms->window.dictLimit; const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const prefixStart = base + dictLimit; - U32* smallerPtr = bt + 2*(current&btMask); - U32* largerPtr = bt + 2*(current&btMask) + 1; - U32 matchEndIdx = current + 8 + 1; + U32* smallerPtr = bt + 2*(curr&btMask); + U32* largerPtr = bt + 2*(curr&btMask) + 1; + U32 matchEndIdx = curr + 8 + 1; U32 dummy32; /* to be nullified at the end */ size_t bestLength = 0; matchIndex = hashTable[h]; - hashTable[h] = current; /* Update Hash Table */ + hashTable[h] = curr; /* Update Hash Table */ - while (nbCompares-- && (matchIndex > windowLow)) { + for (; nbCompares && (matchIndex > windowLow); --nbCompares) { U32* const nextPtr = bt + 2*(matchIndex & btMask); size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ const BYTE* match; @@ -326,8 +327,8 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, if (matchLength > bestLength) { if (matchLength > matchEndIdx - matchIndex) matchEndIdx = matchIndex + (U32)matchLength; - if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) - bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex; + if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) + bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex); if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */ if (dictMode == ZSTD_dictMatchState) { nbCompares = 0; /* in addition to avoiding checking any @@ -356,6 +357,7 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, *smallerPtr = *largerPtr = 0; + assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */ if (dictMode == ZSTD_dictMatchState && nbCompares) { bestLength = ZSTD_DUBT_findBetterDictMatch( ms, ip, iend, @@ -363,12 +365,12 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms, mls, dictMode); } - assert(matchEndIdx > current+8); /* ensure nextToUpdate is increased */ + assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */ ms->nextToUpdate = matchEndIdx - 8; /* skip repetitive patterns */ if (bestLength >= MINMATCH) { - U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex; + U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex; DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)", - current, (U32)bestLength, (U32)*offsetPtr, mIndex); + curr, (U32)bestLength, (U32)*offsetPtr, mIndex); } return bestLength; } @@ -389,56 +391,222 @@ ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms, return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode); } +/*********************************** +* Dedicated dict search +***********************************/ -static size_t -ZSTD_BtFindBestMatch_selectMLS ( ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) +void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip) { - switch(ms->cParams.minMatch) + const BYTE* const base = ms->window.base; + U32 const target = (U32)(ip - base); + U32* const hashTable = ms->hashTable; + U32* const chainTable = ms->chainTable; + U32 const chainSize = 1 << ms->cParams.chainLog; + U32 idx = ms->nextToUpdate; + U32 const minChain = chainSize < target - idx ? target - chainSize : idx; + U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG; + U32 const cacheSize = bucketSize - 1; + U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize; + U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts; + + /* We know the hashtable is oversized by a factor of `bucketSize`. + * We are going to temporarily pretend `bucketSize == 1`, keeping only a + * single entry. We will use the rest of the space to construct a temporary + * chaintable. + */ + U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG; + U32* const tmpHashTable = hashTable; + U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog); + U32 const tmpChainSize = (U32)((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog; + U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx; + U32 hashIdx; + + assert(ms->cParams.chainLog <= 24); + assert(ms->cParams.hashLog > ms->cParams.chainLog); + assert(idx != 0); + assert(tmpMinChain <= minChain); + + /* fill conventional hash table and conventional chain table */ + for ( ; idx < target; idx++) { + U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch); + if (idx >= tmpMinChain) { + tmpChainTable[idx - tmpMinChain] = hashTable[h]; + } + tmpHashTable[h] = idx; + } + + /* sort chains into ddss chain table */ { - default : /* includes case 3 */ - case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict); - case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict); - case 7 : - case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict); + U32 chainPos = 0; + for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) { + U32 count; + U32 countBeyondMinChain = 0; + U32 i = tmpHashTable[hashIdx]; + for (count = 0; i >= tmpMinChain && count < cacheSize; count++) { + /* skip through the chain to the first position that won't be + * in the hash cache bucket */ + if (i < minChain) { + countBeyondMinChain++; + } + i = tmpChainTable[i - tmpMinChain]; + } + if (count == cacheSize) { + for (count = 0; count < chainLimit;) { + if (i < minChain) { + if (!i || ++countBeyondMinChain > cacheSize) { + /* only allow pulling `cacheSize` number of entries + * into the cache or chainTable beyond `minChain`, + * to replace the entries pulled out of the + * chainTable into the cache. This lets us reach + * back further without increasing the total number + * of entries in the chainTable, guaranteeing the + * DDSS chain table will fit into the space + * allocated for the regular one. */ + break; + } + } + chainTable[chainPos++] = i; + count++; + if (i < tmpMinChain) { + break; + } + i = tmpChainTable[i - tmpMinChain]; + } + } else { + count = 0; + } + if (count) { + tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count; + } else { + tmpHashTable[hashIdx] = 0; + } + } + assert(chainPos <= chainSize); /* I believe this is guaranteed... */ + } + + /* move chain pointers into the last entry of each hash bucket */ + for (hashIdx = (1 << hashLog); hashIdx; ) { + U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG; + U32 const chainPackedPointer = tmpHashTable[hashIdx]; + U32 i; + for (i = 0; i < cacheSize; i++) { + hashTable[bucketIdx + i] = 0; + } + hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer; + } + + /* fill the buckets of the hash table */ + for (idx = ms->nextToUpdate; idx < target; idx++) { + U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch) + << ZSTD_LAZY_DDSS_BUCKET_LOG; + U32 i; + /* Shift hash cache down 1. */ + for (i = cacheSize - 1; i; i--) + hashTable[h + i] = hashTable[h + i - 1]; + hashTable[h] = idx; } + + ms->nextToUpdate = target; } +/* Returns the longest match length found in the dedicated dict search structure. + * If none are longer than the argument ml, then ml will be returned. + */ +FORCE_INLINE_TEMPLATE +size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts, + const ZSTD_matchState_t* const dms, + const BYTE* const ip, const BYTE* const iLimit, + const BYTE* const prefixStart, const U32 curr, + const U32 dictLimit, const size_t ddsIdx) { + const U32 ddsLowestIndex = dms->window.dictLimit; + const BYTE* const ddsBase = dms->window.base; + const BYTE* const ddsEnd = dms->window.nextSrc; + const U32 ddsSize = (U32)(ddsEnd - ddsBase); + const U32 ddsIndexDelta = dictLimit - ddsSize; + const U32 bucketSize = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG); + const U32 bucketLimit = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1; + U32 ddsAttempt; + U32 matchIndex; + + for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) { + PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]); + } -static size_t ZSTD_BtFindBestMatch_dictMatchState_selectMLS ( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) -{ - switch(ms->cParams.minMatch) { - default : /* includes case 3 */ - case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState); - case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState); - case 7 : - case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState); + U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; + U32 const chainIndex = chainPackedPointer >> 8; + + PREFETCH_L1(&dms->chainTable[chainIndex]); } -} + for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) { + size_t currentMl=0; + const BYTE* match; + matchIndex = dms->hashTable[ddsIdx + ddsAttempt]; + match = ddsBase + matchIndex; + + if (!matchIndex) { + return ml; + } + + /* guaranteed by table construction */ + (void)ddsLowestIndex; + assert(matchIndex >= ddsLowestIndex); + assert(match+4 <= ddsEnd); + if (MEM_read32(match) == MEM_read32(ip)) { + /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; + } + + /* save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta)); + if (ip+currentMl == iLimit) { + /* best possible, avoids read overflow on next attempt */ + return ml; + } + } + } -static size_t ZSTD_BtFindBestMatch_extDict_selectMLS ( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) -{ - switch(ms->cParams.minMatch) { - default : /* includes case 3 */ - case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict); - case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict); - case 7 : - case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict); + U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1]; + U32 chainIndex = chainPackedPointer >> 8; + U32 const chainLength = chainPackedPointer & 0xFF; + U32 const chainAttempts = nbAttempts - ddsAttempt; + U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts; + U32 chainAttempt; + + for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) { + PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]); + } + + for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) { + size_t currentMl=0; + const BYTE* match; + matchIndex = dms->chainTable[chainIndex]; + match = ddsBase + matchIndex; + + /* guaranteed by table construction */ + assert(matchIndex >= ddsLowestIndex); + assert(match+4 <= ddsEnd); + if (MEM_read32(match) == MEM_read32(ip)) { + /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4; + } + + /* save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta)); + if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ + } + } } + return ml; } - /* ********************************* * Hash Chain ***********************************/ @@ -446,7 +614,7 @@ static size_t ZSTD_BtFindBestMatch_extDict_selectMLS ( /* Update chains up to ip (excluded) Assumption : always within prefix (i.e. not within extDict) */ -static U32 ZSTD_insertAndFindFirstIndex_internal( +FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal( ZSTD_matchState_t* ms, const ZSTD_compressionParameters* const cParams, const BYTE* ip, U32 const mls) @@ -475,10 +643,9 @@ U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) { return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch); } - /* inlining is important to hardwire a hot branch (template emulation) */ FORCE_INLINE_TEMPLATE -size_t ZSTD_HcFindBestMatch_generic ( +size_t ZSTD_HcFindBestMatch( ZSTD_matchState_t* ms, const BYTE* const ip, const BYTE* const iLimit, size_t* offsetPtr, @@ -493,20 +660,33 @@ size_t ZSTD_HcFindBestMatch_generic ( const U32 dictLimit = ms->window.dictLimit; const BYTE* const prefixStart = base + dictLimit; const BYTE* const dictEnd = dictBase + dictLimit; - const U32 current = (U32)(ip-base); + const U32 curr = (U32)(ip-base); const U32 maxDistance = 1U << cParams->windowLog; const U32 lowestValid = ms->window.lowLimit; - const U32 withinMaxDistance = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid; + const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid; const U32 isDictionary = (ms->loadedDictEnd != 0); const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance; - const U32 minChain = current > chainSize ? current - chainSize : 0; + const U32 minChain = curr > chainSize ? curr - chainSize : 0; U32 nbAttempts = 1U << cParams->searchLog; size_t ml=4-1; + const ZSTD_matchState_t* const dms = ms->dictMatchState; + const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch + ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0; + const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch + ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0; + + U32 matchIndex; + + if (dictMode == ZSTD_dedicatedDictSearch) { + const U32* entry = &dms->hashTable[ddsIdx]; + PREFETCH_L1(entry); + } + /* HC4 match finder */ - U32 matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls); + matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls); - for ( ; (matchIndex>lowLimit) & (nbAttempts>0) ; nbAttempts--) { + for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) { size_t currentMl=0; if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) { const BYTE* const match = base + matchIndex; @@ -523,7 +703,7 @@ size_t ZSTD_HcFindBestMatch_generic ( /* save best solution */ if (currentMl > ml) { ml = currentMl; - *offsetPtr = current - matchIndex + ZSTD_REP_MOVE; + *offsetPtr = STORE_OFFSET(curr - matchIndex); if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ } @@ -531,8 +711,11 @@ size_t ZSTD_HcFindBestMatch_generic ( matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask); } - if (dictMode == ZSTD_dictMatchState) { - const ZSTD_matchState_t* const dms = ms->dictMatchState; + assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */ + if (dictMode == ZSTD_dedicatedDictSearch) { + ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms, + ip, iLimit, prefixStart, curr, dictLimit, ddsIdx); + } else if (dictMode == ZSTD_dictMatchState) { const U32* const dmsChainTable = dms->chainTable; const U32 dmsChainSize = (1 << dms->cParams.chainLog); const U32 dmsChainMask = dmsChainSize - 1; @@ -545,7 +728,7 @@ size_t ZSTD_HcFindBestMatch_generic ( matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)]; - for ( ; (matchIndex>dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) { + for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) { size_t currentMl=0; const BYTE* const match = dmsBase + matchIndex; assert(match+4 <= dmsEnd); @@ -555,11 +738,13 @@ size_t ZSTD_HcFindBestMatch_generic ( /* save best solution */ if (currentMl > ml) { ml = currentMl; - *offsetPtr = current - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE; + assert(curr > matchIndex + dmsIndexDelta); + *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta)); if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ } if (matchIndex <= dmsMinChain) break; + matchIndex = dmsChainTable[matchIndex & dmsChainMask]; } } @@ -567,59 +752,725 @@ size_t ZSTD_HcFindBestMatch_generic ( return ml; } +/* ********************************* +* (SIMD) Row-based matchfinder +***********************************/ +/* Constants for row-based hash */ +#define ZSTD_ROW_HASH_TAG_OFFSET 16 /* byte offset of hashes in the match state's tagTable from the beginning of a row */ +#define ZSTD_ROW_HASH_TAG_BITS 8 /* nb bits to use for the tag */ +#define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1) +#define ZSTD_ROW_HASH_MAX_ENTRIES 64 /* absolute maximum number of entries per row, for all configurations */ -FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS ( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) +#define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1) + +typedef U64 ZSTD_VecMask; /* Clarifies when we are interacting with a U64 representing a mask of matches */ + +/* ZSTD_VecMask_next(): + * Starting from the LSB, returns the idx of the next non-zero bit. + * Basically counting the nb of trailing zeroes. + */ +static U32 ZSTD_VecMask_next(ZSTD_VecMask val) { + assert(val != 0); +# if defined(_MSC_VER) && defined(_WIN64) + if (val != 0) { + unsigned long r; + _BitScanForward64(&r, val); + return (U32)(r); + } else { + /* Should not reach this code path */ + __assume(0); + } +# elif (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4)))) + if (sizeof(size_t) == 4) { + U32 mostSignificantWord = (U32)(val >> 32); + U32 leastSignificantWord = (U32)val; + if (leastSignificantWord == 0) { + return 32 + (U32)__builtin_ctz(mostSignificantWord); + } else { + return (U32)__builtin_ctz(leastSignificantWord); + } + } else { + return (U32)__builtin_ctzll(val); + } +# else + /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count + * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer + */ + val = ~val & (val - 1ULL); /* Lowest set bit mask */ + val = val - ((val >> 1) & 0x5555555555555555); + val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL); + return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56); +# endif +} + +/* ZSTD_rotateRight_*(): + * Rotates a bitfield to the right by "count" bits. + * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts + */ +FORCE_INLINE_TEMPLATE +U64 ZSTD_rotateRight_U64(U64 const value, U32 count) { + assert(count < 64); + count &= 0x3F; /* for fickle pattern recognition */ + return (value >> count) | (U64)(value << ((0U - count) & 0x3F)); +} + +FORCE_INLINE_TEMPLATE +U32 ZSTD_rotateRight_U32(U32 const value, U32 count) { + assert(count < 32); + count &= 0x1F; /* for fickle pattern recognition */ + return (value >> count) | (U32)(value << ((0U - count) & 0x1F)); +} + +FORCE_INLINE_TEMPLATE +U16 ZSTD_rotateRight_U16(U16 const value, U32 count) { + assert(count < 16); + count &= 0x0F; /* for fickle pattern recognition */ + return (value >> count) | (U16)(value << ((0U - count) & 0x0F)); +} + +/* ZSTD_row_nextIndex(): + * Returns the next index to insert at within a tagTable row, and updates the "head" + * value to reflect the update. Essentially cycles backwards from [0, {entries per row}) + */ +FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) { + U32 const next = (*tagRow - 1) & rowMask; + *tagRow = (BYTE)next; + return next; +} + +/* ZSTD_isAligned(): + * Checks that a pointer is aligned to "align" bytes which must be a power of 2. + */ +MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) { + assert((align & (align - 1)) == 0); + return (((size_t)ptr) & (align - 1)) == 0; +} + +/* ZSTD_row_prefetch(): + * Performs prefetching for the hashTable and tagTable at a given row. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) { + PREFETCH_L1(hashTable + relRow); + if (rowLog >= 5) { + PREFETCH_L1(hashTable + relRow + 16); + /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */ + } + PREFETCH_L1(tagTable + relRow); + if (rowLog == 6) { + PREFETCH_L1(tagTable + relRow + 32); + } + assert(rowLog == 4 || rowLog == 5 || rowLog == 6); + assert(ZSTD_isAligned(hashTable + relRow, 64)); /* prefetched hash row always 64-byte aligned */ + assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */ +} + +/* ZSTD_row_fillHashCache(): + * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries, + * but not beyond iLimit. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base, + U32 const rowLog, U32 const mls, + U32 idx, const BYTE* const iLimit) { - switch(ms->cParams.minMatch) - { - default : /* includes case 3 */ - case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict); - case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict); - case 7 : - case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict); + U32 const* const hashTable = ms->hashTable; + U16 const* const tagTable = ms->tagTable; + U32 const hashLog = ms->rowHashLog; + U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1); + U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch); + + for (; idx < lim; ++idx) { + U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + ZSTD_row_prefetch(hashTable, tagTable, row, rowLog); + ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash; } + + DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1], + ms->hashCache[2], ms->hashCache[3], ms->hashCache[4], + ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]); } +/* ZSTD_row_nextCachedHash(): + * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at + * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable. + */ +FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable, + U16 const* tagTable, BYTE const* base, + U32 idx, U32 const hashLog, + U32 const rowLog, U32 const mls) +{ + U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + ZSTD_row_prefetch(hashTable, tagTable, row, rowLog); + { U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK]; + cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash; + return hash; + } +} -static size_t ZSTD_HcFindBestMatch_dictMatchState_selectMLS ( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) +/* ZSTD_row_update_internalImpl(): + * Updates the hash table with positions starting from updateStartIdx until updateEndIdx. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms, + U32 updateStartIdx, U32 const updateEndIdx, + U32 const mls, U32 const rowLog, + U32 const rowMask, U32 const useCache) { - switch(ms->cParams.minMatch) - { - default : /* includes case 3 */ - case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState); - case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState); - case 7 : - case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState); + U32* const hashTable = ms->hashTable; + U16* const tagTable = ms->tagTable; + U32 const hashLog = ms->rowHashLog; + const BYTE* const base = ms->window.base; + + DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx); + for (; updateStartIdx < updateEndIdx; ++updateStartIdx) { + U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls) + : (U32)ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + U32* const row = hashTable + relRow; + BYTE* tagRow = (BYTE*)(tagTable + relRow); /* Though tagTable is laid out as a table of U16, each tag is only 1 byte. + Explicit cast allows us to get exact desired position within each row */ + U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask); + + assert(hash == ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls)); + ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK; + row[pos] = updateStartIdx; + } +} + +/* ZSTD_row_update_internal(): + * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate. + * Skips sections of long matches as is necessary. + */ +FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip, + U32 const mls, U32 const rowLog, + U32 const rowMask, U32 const useCache) +{ + U32 idx = ms->nextToUpdate; + const BYTE* const base = ms->window.base; + const U32 target = (U32)(ip - base); + const U32 kSkipThreshold = 384; + const U32 kMaxMatchStartPositionsToUpdate = 96; + const U32 kMaxMatchEndPositionsToUpdate = 32; + + if (useCache) { + /* Only skip positions when using hash cache, i.e. + * if we are loading a dict, don't skip anything. + * If we decide to skip, then we only update a set number + * of positions at the beginning and end of the match. + */ + if (UNLIKELY(target - idx > kSkipThreshold)) { + U32 const bound = idx + kMaxMatchStartPositionsToUpdate; + ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache); + idx = target - kMaxMatchEndPositionsToUpdate; + ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1); + } } + assert(target >= idx); + ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache); + ms->nextToUpdate = target; +} + +/* ZSTD_row_update(): + * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary + * processing. + */ +void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) { + const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6); + const U32 rowMask = (1u << rowLog) - 1; + const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */); + + DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog); + ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */); } +#if defined(ZSTD_ARCH_X86_SSE2) +FORCE_INLINE_TEMPLATE ZSTD_VecMask +ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head) +{ + const __m128i comparisonMask = _mm_set1_epi8((char)tag); + int matches[4] = {0}; + int i; + assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4); + for (i=0; i<nbChunks; i++) { + const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i)); + const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask); + matches[i] = _mm_movemask_epi8(equalMask); + } + if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head); + if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head); + assert(nbChunks == 4); + return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head); +} +#endif -FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS ( +/* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches + * the hash at the nth position in a row of the tagTable. + * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield + * to match up with the actual layout of the entries within the hashTable */ +FORCE_INLINE_TEMPLATE ZSTD_VecMask +ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries) +{ + const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET; + assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64); + assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES); + +#if defined(ZSTD_ARCH_X86_SSE2) + + return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, head); + +#else /* SW or NEON-LE */ + +# if defined(ZSTD_ARCH_ARM_NEON) + /* This NEON path only works for little endian - otherwise use SWAR below */ + if (MEM_isLittleEndian()) { + if (rowEntries == 16) { + const uint8x16_t chunk = vld1q_u8(src); + const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag))); + const uint16x8_t t0 = vshlq_n_u16(equalMask, 7); + const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14)); + const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14)); + const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28)); + const U16 hi = (U16)vgetq_lane_u8(t3, 8); + const U16 lo = (U16)vgetq_lane_u8(t3, 0); + return ZSTD_rotateRight_U16((hi << 8) | lo, head); + } else if (rowEntries == 32) { + const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src); + const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]); + const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]); + const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag)); + const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag)); + const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0)); + const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1)); + const uint8x8_t t0 = vreinterpret_u8_s8(pack0); + const uint8x8_t t1 = vreinterpret_u8_s8(pack1); + const uint8x8_t t2 = vsri_n_u8(t1, t0, 2); + const uint8x8x2_t t3 = vuzp_u8(t2, t0); + const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4); + const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0); + return ZSTD_rotateRight_U32(matches, head); + } else { /* rowEntries == 64 */ + const uint8x16x4_t chunk = vld4q_u8(src); + const uint8x16_t dup = vdupq_n_u8(tag); + const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup); + const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup); + const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup); + const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup); + + const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1); + const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1); + const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2); + const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4); + const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4); + const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0); + return ZSTD_rotateRight_U64(matches, head); + } + } +# endif /* ZSTD_ARCH_ARM_NEON */ + /* SWAR */ + { const size_t chunkSize = sizeof(size_t); + const size_t shiftAmount = ((chunkSize * 8) - chunkSize); + const size_t xFF = ~((size_t)0); + const size_t x01 = xFF / 0xFF; + const size_t x80 = x01 << 7; + const size_t splatChar = tag * x01; + ZSTD_VecMask matches = 0; + int i = rowEntries - chunkSize; + assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8)); + if (MEM_isLittleEndian()) { /* runtime check so have two loops */ + const size_t extractMagic = (xFF / 0x7F) >> chunkSize; + do { + size_t chunk = MEM_readST(&src[i]); + chunk ^= splatChar; + chunk = (((chunk | x80) - x01) | chunk) & x80; + matches <<= chunkSize; + matches |= (chunk * extractMagic) >> shiftAmount; + i -= chunkSize; + } while (i >= 0); + } else { /* big endian: reverse bits during extraction */ + const size_t msb = xFF ^ (xFF >> 1); + const size_t extractMagic = (msb / 0x1FF) | msb; + do { + size_t chunk = MEM_readST(&src[i]); + chunk ^= splatChar; + chunk = (((chunk | x80) - x01) | chunk) & x80; + matches <<= chunkSize; + matches |= ((chunk >> 7) * extractMagic) >> shiftAmount; + i -= chunkSize; + } while (i >= 0); + } + matches = ~matches; + if (rowEntries == 16) { + return ZSTD_rotateRight_U16((U16)matches, head); + } else if (rowEntries == 32) { + return ZSTD_rotateRight_U32((U32)matches, head); + } else { + return ZSTD_rotateRight_U64((U64)matches, head); + } + } +#endif +} + +/* The high-level approach of the SIMD row based match finder is as follows: + * - Figure out where to insert the new entry: + * - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag" + * - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines + * which row to insert into. + * - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can + * be considered as a circular buffer with a "head" index that resides in the tagTable. + * - Also insert the "tag" into the equivalent row and position in the tagTable. + * - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry. + * The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively, + * for alignment/performance reasons, leaving some bytes unused. + * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and + * generate a bitfield that we can cycle through to check the collisions in the hash table. + * - Pick the longest match. + */ +FORCE_INLINE_TEMPLATE +size_t ZSTD_RowFindBestMatch( ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* const iLimit, - size_t* offsetPtr) + const BYTE* const ip, const BYTE* const iLimit, + size_t* offsetPtr, + const U32 mls, const ZSTD_dictMode_e dictMode, + const U32 rowLog) { - switch(ms->cParams.minMatch) - { - default : /* includes case 3 */ - case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict); - case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict); - case 7 : - case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict); + U32* const hashTable = ms->hashTable; + U16* const tagTable = ms->tagTable; + U32* const hashCache = ms->hashCache; + const U32 hashLog = ms->rowHashLog; + const ZSTD_compressionParameters* const cParams = &ms->cParams; + const BYTE* const base = ms->window.base; + const BYTE* const dictBase = ms->window.dictBase; + const U32 dictLimit = ms->window.dictLimit; + const BYTE* const prefixStart = base + dictLimit; + const BYTE* const dictEnd = dictBase + dictLimit; + const U32 curr = (U32)(ip-base); + const U32 maxDistance = 1U << cParams->windowLog; + const U32 lowestValid = ms->window.lowLimit; + const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid; + const U32 isDictionary = (ms->loadedDictEnd != 0); + const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance; + const U32 rowEntries = (1U << rowLog); + const U32 rowMask = rowEntries - 1; + const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */ + U32 nbAttempts = 1U << cappedSearchLog; + size_t ml=4-1; + + /* DMS/DDS variables that may be referenced laster */ + const ZSTD_matchState_t* const dms = ms->dictMatchState; + + /* Initialize the following variables to satisfy static analyzer */ + size_t ddsIdx = 0; + U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */ + U32 dmsTag = 0; + U32* dmsRow = NULL; + BYTE* dmsTagRow = NULL; + + if (dictMode == ZSTD_dedicatedDictSearch) { + const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG; + { /* Prefetch DDS hashtable entry */ + ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG; + PREFETCH_L1(&dms->hashTable[ddsIdx]); + } + ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0; + } + + if (dictMode == ZSTD_dictMatchState) { + /* Prefetch DMS rows */ + U32* const dmsHashTable = dms->hashTable; + U16* const dmsTagTable = dms->tagTable; + U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls); + U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK; + dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow); + dmsRow = dmsHashTable + dmsRelRow; + ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog); + } + + /* Update the hashTable and tagTable up to (but not including) ip */ + ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */); + { /* Get the hash for ip, compute the appropriate row */ + U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls); + U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog; + U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK; + U32* const row = hashTable + relRow; + BYTE* tagRow = (BYTE*)(tagTable + relRow); + U32 const head = *tagRow & rowMask; + U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES]; + size_t numMatches = 0; + size_t currMatch = 0; + ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries); + + /* Cycle through the matches and prefetch */ + for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) { + U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask; + U32 const matchIndex = row[matchPos]; + assert(numMatches < rowEntries); + if (matchIndex < lowLimit) + break; + if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) { + PREFETCH_L1(base + matchIndex); + } else { + PREFETCH_L1(dictBase + matchIndex); + } + matchBuffer[numMatches++] = matchIndex; + } + + /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop + in ZSTD_row_update_internal() at the next search. */ + { + U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask); + tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag; + row[pos] = ms->nextToUpdate++; + } + + /* Return the longest match */ + for (; currMatch < numMatches; ++currMatch) { + U32 const matchIndex = matchBuffer[currMatch]; + size_t currentMl=0; + assert(matchIndex < curr); + assert(matchIndex >= lowLimit); + + if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) { + const BYTE* const match = base + matchIndex; + assert(matchIndex >= dictLimit); /* ensures this is true if dictMode != ZSTD_extDict */ + if (match[ml] == ip[ml]) /* potentially better */ + currentMl = ZSTD_count(ip, match, iLimit); + } else { + const BYTE* const match = dictBase + matchIndex; + assert(match+4 <= dictEnd); + if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */ + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4; + } + + /* Save best solution */ + if (currentMl > ml) { + ml = currentMl; + *offsetPtr = STORE_OFFSET(curr - matchIndex); + if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */ + } + } } + + assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */ + if (dictMode == ZSTD_dedicatedDictSearch) { + ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms, + ip, iLimit, prefixStart, curr, dictLimit, ddsIdx); + } else if (dictMode == ZSTD_dictMatchState) { + /* TODO: Measure and potentially add prefetching to DMS */ + const U32 dmsLowestIndex = dms->window.dictLimit; + const BYTE* const dmsBase = dms->window.base; + const BYTE* const dmsEnd = dms->window.nextSrc; + const U32 dmsSize = (U32)(dmsEnd - dmsBase); + const U32 dmsIndexDelta = dictLimit - dmsSize; + + { U32 const head = *dmsTagRow & rowMask; + U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES]; + size_t numMatches = 0; + size_t currMatch = 0; + ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries); + + for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) { + U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask; + U32 const matchIndex = dmsRow[matchPos]; + if (matchIndex < dmsLowestIndex) + break; + PREFETCH_L1(dmsBase + matchIndex); + matchBuffer[numMatches++] = matchIndex; + } + + /* Return the longest match */ + for (; currMatch < numMatches; ++currMatch) { + U32 const matchIndex = matchBuffer[currMatch]; + size_t currentMl=0; + assert(matchIndex >= dmsLowestIndex); + assert(matchIndex < curr); + + { const BYTE* const match = dmsBase + matchIndex; + assert(match+4 <= dmsEnd); + if (MEM_read32(match) == MEM_read32(ip)) + currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4; + } + + if (currentMl > ml) { + ml = currentMl; + assert(curr > matchIndex + dmsIndexDelta); + *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta)); + if (ip+currentMl == iLimit) break; + } + } + } + } + return ml; } +typedef size_t (*searchMax_f)( + ZSTD_matchState_t* ms, + const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr); + +/** + * This struct contains the functions necessary for lazy to search. + * Currently, that is only searchMax. However, it is still valuable to have the + * VTable because this makes it easier to add more functions to the VTable later. + * + * TODO: The start of the search function involves loading and calculating a + * bunch of constants from the ZSTD_matchState_t. These computations could be + * done in an initialization function, and saved somewhere in the match state. + * Then we could pass a pointer to the saved state instead of the match state, + * and avoid duplicate computations. + * + * TODO: Move the match re-winding into searchMax. This improves compression + * ratio, and unlocks further simplifications with the next TODO. + * + * TODO: Try moving the repcode search into searchMax. After the re-winding + * and repcode search are in searchMax, there is no more logic in the match + * finder loop that requires knowledge about the dictMode. So we should be + * able to avoid force inlining it, and we can join the extDict loop with + * the single segment loop. It should go in searchMax instead of its own + * function to avoid having multiple virtual function calls per search. + */ +typedef struct { + searchMax_f searchMax; +} ZSTD_LazyVTable; + +#define GEN_ZSTD_BT_VTABLE(dictMode, mls) \ + static size_t ZSTD_BtFindBestMatch_##dictMode##_##mls( \ + ZSTD_matchState_t* ms, \ + const BYTE* ip, const BYTE* const iLimit, \ + size_t* offsetPtr) \ + { \ + assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \ + return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \ + } \ + static const ZSTD_LazyVTable ZSTD_BtVTable_##dictMode##_##mls = { \ + ZSTD_BtFindBestMatch_##dictMode##_##mls \ + }; + +#define GEN_ZSTD_HC_VTABLE(dictMode, mls) \ + static size_t ZSTD_HcFindBestMatch_##dictMode##_##mls( \ + ZSTD_matchState_t* ms, \ + const BYTE* ip, const BYTE* const iLimit, \ + size_t* offsetPtr) \ + { \ + assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \ + return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \ + } \ + static const ZSTD_LazyVTable ZSTD_HcVTable_##dictMode##_##mls = { \ + ZSTD_HcFindBestMatch_##dictMode##_##mls \ + }; + +#define GEN_ZSTD_ROW_VTABLE(dictMode, mls, rowLog) \ + static size_t ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog( \ + ZSTD_matchState_t* ms, \ + const BYTE* ip, const BYTE* const iLimit, \ + size_t* offsetPtr) \ + { \ + assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \ + assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog); \ + return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \ + } \ + static const ZSTD_LazyVTable ZSTD_RowVTable_##dictMode##_##mls##_##rowLog = { \ + ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog \ + }; + +#define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \ + X(dictMode, mls, 4) \ + X(dictMode, mls, 5) \ + X(dictMode, mls, 6) + +#define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \ + ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4) \ + ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5) \ + ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6) + +#define ZSTD_FOR_EACH_MLS(X, dictMode) \ + X(dictMode, 4) \ + X(dictMode, 5) \ + X(dictMode, 6) + +#define ZSTD_FOR_EACH_DICT_MODE(X, ...) \ + X(__VA_ARGS__, noDict) \ + X(__VA_ARGS__, extDict) \ + X(__VA_ARGS__, dictMatchState) \ + X(__VA_ARGS__, dedicatedDictSearch) + +/* Generate Row VTables for each combination of (dictMode, mls, rowLog) */ +ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_VTABLE) +/* Generate Binary Tree VTables for each combination of (dictMode, mls) */ +ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_VTABLE) +/* Generate Hash Chain VTables for each combination of (dictMode, mls) */ +ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_VTABLE) + +#define GEN_ZSTD_BT_VTABLE_ARRAY(dictMode) \ + { \ + &ZSTD_BtVTable_##dictMode##_4, \ + &ZSTD_BtVTable_##dictMode##_5, \ + &ZSTD_BtVTable_##dictMode##_6 \ + } + +#define GEN_ZSTD_HC_VTABLE_ARRAY(dictMode) \ + { \ + &ZSTD_HcVTable_##dictMode##_4, \ + &ZSTD_HcVTable_##dictMode##_5, \ + &ZSTD_HcVTable_##dictMode##_6 \ + } + +#define GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, mls) \ + { \ + &ZSTD_RowVTable_##dictMode##_##mls##_4, \ + &ZSTD_RowVTable_##dictMode##_##mls##_5, \ + &ZSTD_RowVTable_##dictMode##_##mls##_6 \ + } + +#define GEN_ZSTD_ROW_VTABLE_ARRAY(dictMode) \ + { \ + GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 4), \ + GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 5), \ + GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 6) \ + } + +#define GEN_ZSTD_VTABLE_ARRAY(X) \ + { \ + X(noDict), \ + X(extDict), \ + X(dictMatchState), \ + X(dedicatedDictSearch) \ + } + /* ******************************* * Common parser - lazy strategy *********************************/ -typedef enum { search_hashChain, search_binaryTree } searchMethod_e; +typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e; + +/** + * This table is indexed first by the four ZSTD_dictMode_e values, and then + * by the two searchMethod_e values. NULLs are placed for configurations + * that should never occur (extDict modes go to the other implementation + * below and there is no DDSS for binary tree search yet). + */ + +static ZSTD_LazyVTable const* +ZSTD_selectLazyVTable(ZSTD_matchState_t const* ms, searchMethod_e searchMethod, ZSTD_dictMode_e dictMode) +{ + /* Fill the Hc/Bt VTable arrays with the right functions for the (dictMode, mls) combination. */ + ZSTD_LazyVTable const* const hcVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_HC_VTABLE_ARRAY); + ZSTD_LazyVTable const* const btVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_BT_VTABLE_ARRAY); + /* Fill the Row VTable array with the right functions for the (dictMode, mls, rowLog) combination. */ + ZSTD_LazyVTable const* const rowVTables[4][3][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_ROW_VTABLE_ARRAY); + + U32 const mls = MAX(4, MIN(6, ms->cParams.minMatch)); + U32 const rowLog = MAX(4, MIN(6, ms->cParams.searchLog)); + switch (searchMethod) { + case search_hashChain: + return hcVTables[dictMode][mls - 4]; + case search_binaryTree: + return btVTables[dictMode][mls - 4]; + case search_rowHash: + return rowVTables[dictMode][mls - 4][rowLog - 4]; + default: + return NULL; + } +} FORCE_INLINE_TEMPLATE size_t ZSTD_compressBlock_lazy_generic( @@ -633,53 +1484,52 @@ ZSTD_compressBlock_lazy_generic( const BYTE* ip = istart; const BYTE* anchor = istart; const BYTE* const iend = istart + srcSize; - const BYTE* const ilimit = iend - 8; + const BYTE* const ilimit = (searchMethod == search_rowHash) ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8; const BYTE* const base = ms->window.base; const U32 prefixLowestIndex = ms->window.dictLimit; const BYTE* const prefixLowest = base + prefixLowestIndex; - typedef size_t (*searchMax_f)( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr); - searchMax_f const searchMax = dictMode == ZSTD_dictMatchState ? - (searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_dictMatchState_selectMLS - : ZSTD_HcFindBestMatch_dictMatchState_selectMLS) : - (searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_selectMLS - : ZSTD_HcFindBestMatch_selectMLS); + searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, dictMode)->searchMax; U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0; + const int isDMS = dictMode == ZSTD_dictMatchState; + const int isDDS = dictMode == ZSTD_dedicatedDictSearch; + const int isDxS = isDMS || isDDS; const ZSTD_matchState_t* const dms = ms->dictMatchState; - const U32 dictLowestIndex = dictMode == ZSTD_dictMatchState ? - dms->window.dictLimit : 0; - const BYTE* const dictBase = dictMode == ZSTD_dictMatchState ? - dms->window.base : NULL; - const BYTE* const dictLowest = dictMode == ZSTD_dictMatchState ? - dictBase + dictLowestIndex : NULL; - const BYTE* const dictEnd = dictMode == ZSTD_dictMatchState ? - dms->window.nextSrc : NULL; - const U32 dictIndexDelta = dictMode == ZSTD_dictMatchState ? + const U32 dictLowestIndex = isDxS ? dms->window.dictLimit : 0; + const BYTE* const dictBase = isDxS ? dms->window.base : NULL; + const BYTE* const dictLowest = isDxS ? dictBase + dictLowestIndex : NULL; + const BYTE* const dictEnd = isDxS ? dms->window.nextSrc : NULL; + const U32 dictIndexDelta = isDxS ? prefixLowestIndex - (U32)(dictEnd - dictBase) : 0; const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest)); - DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (U32)dictMode); + assert(searchMax != NULL); - /* init */ + DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod); ip += (dictAndPrefixLength == 0); if (dictMode == ZSTD_noDict) { - U32 const current = (U32)(ip - base); - U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, ms->cParams.windowLog); - U32 const maxRep = current - windowLow; + U32 const curr = (U32)(ip - base); + U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog); + U32 const maxRep = curr - windowLow; if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0; if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0; } - if (dictMode == ZSTD_dictMatchState) { + if (isDxS) { /* dictMatchState repCode checks don't currently handle repCode == 0 * disabling. */ assert(offset_1 <= dictAndPrefixLength); assert(offset_2 <= dictAndPrefixLength); } + if (searchMethod == search_rowHash) { + const U32 rowLog = MAX(4, MIN(6, ms->cParams.searchLog)); + ZSTD_row_fillHashCache(ms, base, rowLog, + MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */), + ms->nextToUpdate, ilimit); + } + /* Match Loop */ #if defined(__GNUC__) && defined(__x86_64__) /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the @@ -689,13 +1539,14 @@ ZSTD_compressBlock_lazy_generic( #endif while (ip < ilimit) { size_t matchLength=0; - size_t offset=0; + size_t offcode=STORE_REPCODE_1; const BYTE* start=ip+1; + DEBUGLOG(7, "search baseline (depth 0)"); /* check repCode */ - if (dictMode == ZSTD_dictMatchState) { + if (isDxS) { const U32 repIndex = (U32)(ip - base) + 1 - offset_1; - const BYTE* repMatch = (dictMode == ZSTD_dictMatchState + const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch) && repIndex < prefixLowestIndex) ? dictBase + (repIndex - dictIndexDelta) : base + repIndex; @@ -716,7 +1567,7 @@ ZSTD_compressBlock_lazy_generic( { size_t offsetFound = 999999999; size_t const ml2 = searchMax(ms, ip, iend, &offsetFound); if (ml2 > matchLength) - matchLength = ml2, start = ip, offset=offsetFound; + matchLength = ml2, start = ip, offcode=offsetFound; } if (matchLength < 4) { @@ -727,16 +1578,17 @@ ZSTD_compressBlock_lazy_generic( /* let's try to find a better solution */ if (depth>=1) while (ip<ilimit) { + DEBUGLOG(7, "search depth 1"); ip ++; if ( (dictMode == ZSTD_noDict) - && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) { + && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) { size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4; int const gain2 = (int)(mlRep * 3); - int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((mlRep >= 4) && (gain2 > gain1)) - matchLength = mlRep, offset = 0, start = ip; + matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip; } - if (dictMode == ZSTD_dictMatchState) { + if (isDxS) { const U32 repIndex = (U32)(ip - base) - offset_1; const BYTE* repMatch = repIndex < prefixLowestIndex ? dictBase + (repIndex - dictIndexDelta) : @@ -746,32 +1598,33 @@ ZSTD_compressBlock_lazy_generic( const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend; size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4; int const gain2 = (int)(mlRep * 3); - int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((mlRep >= 4) && (gain2 > gain1)) - matchLength = mlRep, offset = 0, start = ip; + matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip; } } { size_t offset2=999999999; size_t const ml2 = searchMax(ms, ip, iend, &offset2); - int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */ - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4); + int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2))); /* raw approx */ + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4); if ((ml2 >= 4) && (gain2 > gain1)) { - matchLength = ml2, offset = offset2, start = ip; + matchLength = ml2, offcode = offset2, start = ip; continue; /* search a better one */ } } /* let's find an even better one */ if ((depth==2) && (ip<ilimit)) { + DEBUGLOG(7, "search depth 2"); ip ++; if ( (dictMode == ZSTD_noDict) - && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) { + && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) { size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4; int const gain2 = (int)(mlRep * 4); - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((mlRep >= 4) && (gain2 > gain1)) - matchLength = mlRep, offset = 0, start = ip; + matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip; } - if (dictMode == ZSTD_dictMatchState) { + if (isDxS) { const U32 repIndex = (U32)(ip - base) - offset_1; const BYTE* repMatch = repIndex < prefixLowestIndex ? dictBase + (repIndex - dictIndexDelta) : @@ -781,64 +1634,62 @@ ZSTD_compressBlock_lazy_generic( const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend; size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4; int const gain2 = (int)(mlRep * 4); - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((mlRep >= 4) && (gain2 > gain1)) - matchLength = mlRep, offset = 0, start = ip; + matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip; } } { size_t offset2=999999999; size_t const ml2 = searchMax(ms, ip, iend, &offset2); - int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */ - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7); + int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2))); /* raw approx */ + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7); if ((ml2 >= 4) && (gain2 > gain1)) { - matchLength = ml2, offset = offset2, start = ip; + matchLength = ml2, offcode = offset2, start = ip; continue; } } } break; /* nothing found : store previous solution */ } /* NOTE: - * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior. - * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which - * overflows the pointer, which is undefined behavior. + * Pay attention that `start[-value]` can lead to strange undefined behavior + * notably if `value` is unsigned, resulting in a large positive `-value`. */ /* catch up */ - if (offset) { + if (STORED_IS_OFFSET(offcode)) { if (dictMode == ZSTD_noDict) { - while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest)) - && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) ) /* only search for offset within prefix */ + while ( ((start > anchor) & (start - STORED_OFFSET(offcode) > prefixLowest)) + && (start[-1] == (start-STORED_OFFSET(offcode))[-1]) ) /* only search for offset within prefix */ { start--; matchLength++; } } - if (dictMode == ZSTD_dictMatchState) { - U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE)); + if (isDxS) { + U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode)); const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex; const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest; while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */ } - offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE); + offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode); } /* store sequence */ _storeSequence: - { size_t const litLength = start - anchor; - ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH); + { size_t const litLength = (size_t)(start - anchor); + ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength); anchor = ip = start + matchLength; } /* check immediate repcode */ - if (dictMode == ZSTD_dictMatchState) { + if (isDxS) { while (ip <= ilimit) { U32 const current2 = (U32)(ip-base); U32 const repIndex = current2 - offset_2; - const BYTE* repMatch = dictMode == ZSTD_dictMatchState - && repIndex < prefixLowestIndex ? + const BYTE* repMatch = repIndex < prefixLowestIndex ? dictBase - dictIndexDelta + repIndex : base + repIndex; if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */) && (MEM_read32(repMatch) == MEM_read32(ip)) ) { const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend; matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4; - offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset_2 <=> offset_1 */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH); + offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap offset_2 <=> offset_1 */ + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength); ip += matchLength; anchor = ip; continue; @@ -852,8 +1703,8 @@ _storeSequence: && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) { /* store sequence */ matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4; - offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH); + offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap repcodes */ + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength); ip += matchLength; anchor = ip; continue; /* faster when present ... (?) */ @@ -925,6 +1776,92 @@ size_t ZSTD_compressBlock_greedy_dictMatchState( } +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch); +} + +/* Row-based matchfinder */ +size_t ZSTD_compressBlock_lazy2_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_lazy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_greedy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict); +} + +size_t ZSTD_compressBlock_lazy2_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState); +} + +size_t ZSTD_compressBlock_lazy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState); +} + +size_t ZSTD_compressBlock_greedy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState); +} + + +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch); +} + +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch); +} + FORCE_INLINE_TEMPLATE size_t ZSTD_compressBlock_lazy_extDict_generic( ZSTD_matchState_t* ms, seqStore_t* seqStore, @@ -936,7 +1873,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( const BYTE* ip = istart; const BYTE* anchor = istart; const BYTE* const iend = istart + srcSize; - const BYTE* const ilimit = iend - 8; + const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8; const BYTE* const base = ms->window.base; const U32 dictLimit = ms->window.dictLimit; const BYTE* const prefixStart = base + dictLimit; @@ -944,18 +1881,20 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const dictStart = dictBase + ms->window.lowLimit; const U32 windowLog = ms->cParams.windowLog; + const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5; - typedef size_t (*searchMax_f)( - ZSTD_matchState_t* ms, - const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr); - searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS; - + searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, ZSTD_extDict)->searchMax; U32 offset_1 = rep[0], offset_2 = rep[1]; - DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic"); + DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod); /* init */ ip += (ip == prefixStart); + if (searchMethod == search_rowHash) { + ZSTD_row_fillHashCache(ms, base, rowLog, + MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */), + ms->nextToUpdate, ilimit); + } /* Match Loop */ #if defined(__GNUC__) && defined(__x86_64__) @@ -966,16 +1905,17 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( #endif while (ip < ilimit) { size_t matchLength=0; - size_t offset=0; + size_t offcode=STORE_REPCODE_1; const BYTE* start=ip+1; - U32 current = (U32)(ip-base); + U32 curr = (U32)(ip-base); /* check repCode */ - { const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current+1, windowLog); - const U32 repIndex = (U32)(current+1 - offset_1); + { const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog); + const U32 repIndex = (U32)(curr+1 - offset_1); const BYTE* const repBase = repIndex < dictLimit ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; - if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow)) /* intentional overflow */ + if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */ + & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */ if (MEM_read32(ip+1) == MEM_read32(repMatch)) { /* repcode detected we should take it */ const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend; @@ -987,10 +1927,10 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( { size_t offsetFound = 999999999; size_t const ml2 = searchMax(ms, ip, iend, &offsetFound); if (ml2 > matchLength) - matchLength = ml2, start = ip, offset=offsetFound; + matchLength = ml2, start = ip, offcode=offsetFound; } - if (matchLength < 4) { + if (matchLength < 4) { ip += ((ip-anchor) >> kSearchStrength) + 1; /* jump faster over incompressible sections */ continue; } @@ -999,80 +1939,82 @@ size_t ZSTD_compressBlock_lazy_extDict_generic( if (depth>=1) while (ip<ilimit) { ip ++; - current++; + curr++; /* check repCode */ - if (offset) { - const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current, windowLog); - const U32 repIndex = (U32)(current - offset_1); + if (offcode) { + const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog); + const U32 repIndex = (U32)(curr - offset_1); const BYTE* const repBase = repIndex < dictLimit ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; - if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow)) /* intentional overflow */ + if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */ + & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */ if (MEM_read32(ip) == MEM_read32(repMatch)) { /* repcode detected */ const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend; size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4; int const gain2 = (int)(repLength * 3); - int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((repLength >= 4) && (gain2 > gain1)) - matchLength = repLength, offset = 0, start = ip; + matchLength = repLength, offcode = STORE_REPCODE_1, start = ip; } } /* search match, depth 1 */ { size_t offset2=999999999; size_t const ml2 = searchMax(ms, ip, iend, &offset2); - int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */ - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4); + int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2))); /* raw approx */ + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4); if ((ml2 >= 4) && (gain2 > gain1)) { - matchLength = ml2, offset = offset2, start = ip; + matchLength = ml2, offcode = offset2, start = ip; continue; /* search a better one */ } } /* let's find an even better one */ if ((depth==2) && (ip<ilimit)) { ip ++; - current++; + curr++; /* check repCode */ - if (offset) { - const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current, windowLog); - const U32 repIndex = (U32)(current - offset_1); + if (offcode) { + const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog); + const U32 repIndex = (U32)(curr - offset_1); const BYTE* const repBase = repIndex < dictLimit ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; - if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow)) /* intentional overflow */ + if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */ + & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */ if (MEM_read32(ip) == MEM_read32(repMatch)) { /* repcode detected */ const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend; size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4; int const gain2 = (int)(repLength * 4); - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1); + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1); if ((repLength >= 4) && (gain2 > gain1)) - matchLength = repLength, offset = 0, start = ip; + matchLength = repLength, offcode = STORE_REPCODE_1, start = ip; } } /* search match, depth 2 */ { size_t offset2=999999999; size_t const ml2 = searchMax(ms, ip, iend, &offset2); - int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */ - int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7); + int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2))); /* raw approx */ + int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7); if ((ml2 >= 4) && (gain2 > gain1)) { - matchLength = ml2, offset = offset2, start = ip; + matchLength = ml2, offcode = offset2, start = ip; continue; } } } break; /* nothing found : store previous solution */ } /* catch up */ - if (offset) { - U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE)); + if (STORED_IS_OFFSET(offcode)) { + U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode)); const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex; const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart; while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */ - offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE); + offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode); } /* store sequence */ _storeSequence: - { size_t const litLength = start - anchor; - ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH); + { size_t const litLength = (size_t)(start - anchor); + ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength); anchor = ip = start + matchLength; } @@ -1083,13 +2025,14 @@ _storeSequence: const U32 repIndex = repCurrent - offset_2; const BYTE* const repBase = repIndex < dictLimit ? dictBase : base; const BYTE* const repMatch = repBase + repIndex; - if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow)) /* intentional overflow */ + if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */ + & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */ if (MEM_read32(ip) == MEM_read32(repMatch)) { /* repcode detected we should take it */ const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend; matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4; - offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset history */ - ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH); + offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap offset history */ + ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength); ip += matchLength; anchor = ip; continue; /* faster when present ... (?) */ @@ -1136,3 +2079,26 @@ size_t ZSTD_compressBlock_btlazy2_extDict( { return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2); } + +size_t ZSTD_compressBlock_greedy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0); +} + +size_t ZSTD_compressBlock_lazy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) + +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1); +} + +size_t ZSTD_compressBlock_lazy2_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize) + +{ + return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2); +} diff --git a/thirdparty/zstd/compress/zstd_lazy.h b/thirdparty/zstd/compress/zstd_lazy.h index 581936f03b..150f7b390b 100644 --- a/thirdparty/zstd/compress/zstd_lazy.h +++ b/thirdparty/zstd/compress/zstd_lazy.h @@ -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 @@ -17,7 +17,18 @@ extern "C" { #include "zstd_compress_internal.h" +/** + * Dedicated Dictionary Search Structure bucket log. In the + * ZSTD_dedicatedDictSearch mode, the hashTable has + * 2 ** ZSTD_LAZY_DDSS_BUCKET_LOG entries in each bucket, rather than just + * one. + */ +#define ZSTD_LAZY_DDSS_BUCKET_LOG 2 + U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip); +void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip); + +void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip); void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue); /*! used in ZSTD_reduceIndex(). preemptively increase value of ZSTD_DUBT_UNSORTED_MARK */ @@ -33,6 +44,15 @@ size_t ZSTD_compressBlock_lazy( size_t ZSTD_compressBlock_greedy( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_btlazy2_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -46,6 +66,34 @@ size_t ZSTD_compressBlock_lazy_dictMatchState( size_t ZSTD_compressBlock_greedy_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_dictMatchState_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); + +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_greedy_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], @@ -56,9 +104,19 @@ size_t ZSTD_compressBlock_lazy_extDict( size_t ZSTD_compressBlock_lazy2_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); +size_t ZSTD_compressBlock_greedy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); +size_t ZSTD_compressBlock_lazy2_extDict_row( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + void const* src, size_t srcSize); size_t ZSTD_compressBlock_btlazy2_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); + #if defined (__cplusplus) } diff --git a/thirdparty/zstd/compress/zstd_ldm.c b/thirdparty/zstd/compress/zstd_ldm.c index 8c47948358..f662b2546e 100644 --- a/thirdparty/zstd/compress/zstd_ldm.c +++ b/thirdparty/zstd/compress/zstd_ldm.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 @@ -11,13 +11,126 @@ #include "zstd_ldm.h" #include "../common/debug.h" +#include "../common/xxhash.h" #include "zstd_fast.h" /* ZSTD_fillHashTable() */ #include "zstd_double_fast.h" /* ZSTD_fillDoubleHashTable() */ +#include "zstd_ldm_geartab.h" #define LDM_BUCKET_SIZE_LOG 3 #define LDM_MIN_MATCH_LENGTH 64 #define LDM_HASH_RLOG 7 -#define LDM_HASH_CHAR_OFFSET 10 + +typedef struct { + U64 rolling; + U64 stopMask; +} ldmRollingHashState_t; + +/** ZSTD_ldm_gear_init(): + * + * Initializes the rolling hash state such that it will honor the + * settings in params. */ +static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params) +{ + unsigned maxBitsInMask = MIN(params->minMatchLength, 64); + unsigned hashRateLog = params->hashRateLog; + + state->rolling = ~(U32)0; + + /* The choice of the splitting criterion is subject to two conditions: + * 1. it has to trigger on average every 2^(hashRateLog) bytes; + * 2. ideally, it has to depend on a window of minMatchLength bytes. + * + * In the gear hash algorithm, bit n depends on the last n bytes; + * so in order to obtain a good quality splitting criterion it is + * preferable to use bits with high weight. + * + * To match condition 1 we use a mask with hashRateLog bits set + * and, because of the previous remark, we make sure these bits + * have the highest possible weight while still respecting + * condition 2. + */ + if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) { + state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog); + } else { + /* In this degenerate case we simply honor the hash rate. */ + state->stopMask = ((U64)1 << hashRateLog) - 1; + } +} + +/** ZSTD_ldm_gear_reset() + * Feeds [data, data + minMatchLength) into the hash without registering any + * splits. This effectively resets the hash state. This is used when skipping + * over data, either at the beginning of a block, or skipping sections. + */ +static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state, + BYTE const* data, size_t minMatchLength) +{ + U64 hash = state->rolling; + size_t n = 0; + +#define GEAR_ITER_ONCE() do { \ + hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \ + n += 1; \ + } while (0) + while (n + 3 < minMatchLength) { + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + } + while (n < minMatchLength) { + GEAR_ITER_ONCE(); + } +#undef GEAR_ITER_ONCE +} + +/** ZSTD_ldm_gear_feed(): + * + * Registers in the splits array all the split points found in the first + * size bytes following the data pointer. This function terminates when + * either all the data has been processed or LDM_BATCH_SIZE splits are + * present in the splits array. + * + * Precondition: The splits array must not be full. + * Returns: The number of bytes processed. */ +static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state, + BYTE const* data, size_t size, + size_t* splits, unsigned* numSplits) +{ + size_t n; + U64 hash, mask; + + hash = state->rolling; + mask = state->stopMask; + n = 0; + +#define GEAR_ITER_ONCE() do { \ + hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \ + n += 1; \ + if (UNLIKELY((hash & mask) == 0)) { \ + splits[*numSplits] = n; \ + *numSplits += 1; \ + if (*numSplits == LDM_BATCH_SIZE) \ + goto done; \ + } \ + } while (0) + + while (n + 3 < size) { + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + GEAR_ITER_ONCE(); + } + while (n < size) { + GEAR_ITER_ONCE(); + } + +#undef GEAR_ITER_ONCE + +done: + state->rolling = hash; + return n; +} void ZSTD_ldm_adjustParameters(ldmParams_t* params, ZSTD_compressionParameters const* cParams) @@ -27,13 +140,6 @@ void ZSTD_ldm_adjustParameters(ldmParams_t* params, DEBUGLOG(4, "ZSTD_ldm_adjustParameters"); if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG; if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH; - if (cParams->strategy >= ZSTD_btopt) { - /* Get out of the way of the optimal parser */ - U32 const minMatch = MAX(cParams->targetLength, params->minMatchLength); - assert(minMatch >= ZSTD_LDM_MINMATCH_MIN); - assert(minMatch <= ZSTD_LDM_MINMATCH_MAX); - params->minMatchLength = minMatch; - } if (params->hashLog == 0) { params->hashLog = MAX(ZSTD_HASHLOG_MIN, params->windowLog - LDM_HASH_RLOG); assert(params->hashLog <= ZSTD_HASHLOG_MAX); @@ -53,47 +159,12 @@ size_t ZSTD_ldm_getTableSize(ldmParams_t params) size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog); size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize) + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t)); - return params.enableLdm ? totalSize : 0; + return params.enableLdm == ZSTD_ps_enable ? totalSize : 0; } size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize) { - return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0; -} - -/** ZSTD_ldm_getSmallHash() : - * numBits should be <= 32 - * If numBits==0, returns 0. - * @return : the most significant numBits of value. */ -static U32 ZSTD_ldm_getSmallHash(U64 value, U32 numBits) -{ - assert(numBits <= 32); - return numBits == 0 ? 0 : (U32)(value >> (64 - numBits)); -} - -/** ZSTD_ldm_getChecksum() : - * numBitsToDiscard should be <= 32 - * @return : the next most significant 32 bits after numBitsToDiscard */ -static U32 ZSTD_ldm_getChecksum(U64 hash, U32 numBitsToDiscard) -{ - assert(numBitsToDiscard <= 32); - return (hash >> (64 - 32 - numBitsToDiscard)) & 0xFFFFFFFF; -} - -/** ZSTD_ldm_getTag() ; - * Given the hash, returns the most significant numTagBits bits - * after (32 + hbits) bits. - * - * If there are not enough bits remaining, return the last - * numTagBits bits. */ -static U32 ZSTD_ldm_getTag(U64 hash, U32 hbits, U32 numTagBits) -{ - assert(numTagBits < 32 && hbits <= 32); - if (32 - hbits < numTagBits) { - return hash & (((U32)1 << numTagBits) - 1); - } else { - return (hash >> (32 - hbits - numTagBits)) & (((U32)1 << numTagBits) - 1); - } + return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0; } /** ZSTD_ldm_getBucket() : @@ -110,38 +181,12 @@ static void ZSTD_ldm_insertEntry(ldmState_t* ldmState, size_t const hash, const ldmEntry_t entry, ldmParams_t const ldmParams) { - BYTE* const bucketOffsets = ldmState->bucketOffsets; - *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + bucketOffsets[hash]) = entry; - bucketOffsets[hash]++; - bucketOffsets[hash] &= ((U32)1 << ldmParams.bucketSizeLog) - 1; -} + BYTE* const pOffset = ldmState->bucketOffsets + hash; + unsigned const offset = *pOffset; + + *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry; + *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1)); -/** ZSTD_ldm_makeEntryAndInsertByTag() : - * - * Gets the small hash, checksum, and tag from the rollingHash. - * - * If the tag matches (1 << ldmParams.hashRateLog)-1, then - * creates an ldmEntry from the offset, and inserts it into the hash table. - * - * hBits is the length of the small hash, which is the most significant hBits - * of rollingHash. The checksum is the next 32 most significant bits, followed - * by ldmParams.hashRateLog bits that make up the tag. */ -static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState, - U64 const rollingHash, - U32 const hBits, - U32 const offset, - ldmParams_t const ldmParams) -{ - U32 const tag = ZSTD_ldm_getTag(rollingHash, hBits, ldmParams.hashRateLog); - U32 const tagMask = ((U32)1 << ldmParams.hashRateLog) - 1; - if (tag == tagMask) { - U32 const hash = ZSTD_ldm_getSmallHash(rollingHash, hBits); - U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits); - ldmEntry_t entry; - entry.offset = offset; - entry.checksum = checksum; - ZSTD_ldm_insertEntry(ldmState, hash, entry, ldmParams); - } } /** ZSTD_ldm_countBackwardsMatch() : @@ -150,10 +195,10 @@ static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState, * We count only bytes where pMatch >= pBase and pIn >= pAnchor. */ static size_t ZSTD_ldm_countBackwardsMatch( const BYTE* pIn, const BYTE* pAnchor, - const BYTE* pMatch, const BYTE* pBase) + const BYTE* pMatch, const BYTE* pMatchBase) { size_t matchLength = 0; - while (pIn > pAnchor && pMatch > pBase && pIn[-1] == pMatch[-1]) { + while (pIn > pAnchor && pMatch > pMatchBase && pIn[-1] == pMatch[-1]) { pIn--; pMatch--; matchLength++; @@ -161,6 +206,27 @@ static size_t ZSTD_ldm_countBackwardsMatch( return matchLength; } +/** ZSTD_ldm_countBackwardsMatch_2segments() : + * Returns the number of bytes that match backwards from pMatch, + * even with the backwards match spanning 2 different segments. + * + * On reaching `pMatchBase`, start counting from mEnd */ +static size_t ZSTD_ldm_countBackwardsMatch_2segments( + const BYTE* pIn, const BYTE* pAnchor, + const BYTE* pMatch, const BYTE* pMatchBase, + const BYTE* pExtDictStart, const BYTE* pExtDictEnd) +{ + size_t matchLength = ZSTD_ldm_countBackwardsMatch(pIn, pAnchor, pMatch, pMatchBase); + if (pMatch - matchLength != pMatchBase || pMatchBase == pExtDictStart) { + /* If backwards match is entirely in the extDict or prefix, immediately return */ + return matchLength; + } + DEBUGLOG(7, "ZSTD_ldm_countBackwardsMatch_2segments: found 2-parts backwards match (length in prefix==%zu)", matchLength); + matchLength += ZSTD_ldm_countBackwardsMatch(pIn - matchLength, pAnchor, pExtDictEnd, pExtDictStart); + DEBUGLOG(7, "final backwards match length = %zu", matchLength); + return matchLength; +} + /** ZSTD_ldm_fillFastTables() : * * Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies. @@ -198,43 +264,42 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms, return 0; } -/** ZSTD_ldm_fillLdmHashTable() : - * - * Fills hashTable from (lastHashed + 1) to iend (non-inclusive). - * lastHash is the rolling hash that corresponds to lastHashed. - * - * Returns the rolling hash corresponding to position iend-1. */ -static U64 ZSTD_ldm_fillLdmHashTable(ldmState_t* state, - U64 lastHash, const BYTE* lastHashed, - const BYTE* iend, const BYTE* base, - U32 hBits, ldmParams_t const ldmParams) -{ - U64 rollingHash = lastHash; - const BYTE* cur = lastHashed + 1; - - while (cur < iend) { - rollingHash = ZSTD_rollingHash_rotate(rollingHash, cur[-1], - cur[ldmParams.minMatchLength-1], - state->hashPower); - ZSTD_ldm_makeEntryAndInsertByTag(state, - rollingHash, hBits, - (U32)(cur - base), ldmParams); - ++cur; - } - return rollingHash; -} - void ZSTD_ldm_fillHashTable( - ldmState_t* state, const BYTE* ip, + ldmState_t* ldmState, const BYTE* ip, const BYTE* iend, ldmParams_t const* params) { + U32 const minMatchLength = params->minMatchLength; + U32 const hBits = params->hashLog - params->bucketSizeLog; + BYTE const* const base = ldmState->window.base; + BYTE const* const istart = ip; + ldmRollingHashState_t hashState; + size_t* const splits = ldmState->splitIndices; + unsigned numSplits; + DEBUGLOG(5, "ZSTD_ldm_fillHashTable"); - if ((size_t)(iend - ip) >= params->minMatchLength) { - U64 startingHash = ZSTD_rollingHash_compute(ip, params->minMatchLength); - ZSTD_ldm_fillLdmHashTable( - state, startingHash, ip, iend - params->minMatchLength, state->window.base, - params->hashLog - params->bucketSizeLog, - *params); + + ZSTD_ldm_gear_init(&hashState, params); + while (ip < iend) { + size_t hashed; + unsigned n; + + numSplits = 0; + hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits); + + for (n = 0; n < numSplits; n++) { + if (ip + splits[n] >= istart + minMatchLength) { + BYTE const* const split = ip + splits[n] - minMatchLength; + U64 const xxhash = XXH64(split, minMatchLength, 0); + U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1)); + ldmEntry_t entry; + + entry.offset = (U32)(split - base); + entry.checksum = (U32)(xxhash >> 32); + ZSTD_ldm_insertEntry(ldmState, hash, entry, *params); + } + } + + ip += hashed; } } @@ -246,10 +311,10 @@ void ZSTD_ldm_fillHashTable( * (after a long match, only update tables a limited amount). */ static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor) { - U32 const current = (U32)(anchor - ms->window.base); - if (current > ms->nextToUpdate + 1024) { + U32 const curr = (U32)(anchor - ms->window.base); + if (curr > ms->nextToUpdate + 1024) { ms->nextToUpdate = - current - MIN(512, current - ms->nextToUpdate - 1024); + curr - MIN(512, curr - ms->nextToUpdate - 1024); } } @@ -260,11 +325,8 @@ static size_t ZSTD_ldm_generateSequences_internal( /* LDM parameters */ int const extDict = ZSTD_window_hasExtDict(ldmState->window); U32 const minMatchLength = params->minMatchLength; - U64 const hashPower = ldmState->hashPower; + U32 const entsPerBucket = 1U << params->bucketSizeLog; U32 const hBits = params->hashLog - params->bucketSizeLog; - U32 const ldmBucketSize = 1U << params->bucketSizeLog; - U32 const hashRateLog = params->hashRateLog; - U32 const ldmTagMask = (1U << params->hashRateLog) - 1; /* Prefix and extDict parameters */ U32 const dictLimit = ldmState->window.dictLimit; U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit; @@ -276,45 +338,69 @@ static size_t ZSTD_ldm_generateSequences_internal( /* Input bounds */ BYTE const* const istart = (BYTE const*)src; BYTE const* const iend = istart + srcSize; - BYTE const* const ilimit = iend - MAX(minMatchLength, HASH_READ_SIZE); + BYTE const* const ilimit = iend - HASH_READ_SIZE; /* Input positions */ BYTE const* anchor = istart; BYTE const* ip = istart; - /* Rolling hash */ - BYTE const* lastHashed = NULL; - U64 rollingHash = 0; - - while (ip <= ilimit) { - size_t mLength; - U32 const current = (U32)(ip - base); - size_t forwardMatchLength = 0, backwardMatchLength = 0; - ldmEntry_t* bestEntry = NULL; - if (ip != istart) { - rollingHash = ZSTD_rollingHash_rotate(rollingHash, lastHashed[0], - lastHashed[minMatchLength], - hashPower); - } else { - rollingHash = ZSTD_rollingHash_compute(ip, minMatchLength); + /* Rolling hash state */ + ldmRollingHashState_t hashState; + /* Arrays for staged-processing */ + size_t* const splits = ldmState->splitIndices; + ldmMatchCandidate_t* const candidates = ldmState->matchCandidates; + unsigned numSplits; + + if (srcSize < minMatchLength) + return iend - anchor; + + /* Initialize the rolling hash state with the first minMatchLength bytes */ + ZSTD_ldm_gear_init(&hashState, params); + ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength); + ip += minMatchLength; + + while (ip < ilimit) { + size_t hashed; + unsigned n; + + numSplits = 0; + hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip, + splits, &numSplits); + + for (n = 0; n < numSplits; n++) { + BYTE const* const split = ip + splits[n] - minMatchLength; + U64 const xxhash = XXH64(split, minMatchLength, 0); + U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1)); + + candidates[n].split = split; + candidates[n].hash = hash; + candidates[n].checksum = (U32)(xxhash >> 32); + candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params); + PREFETCH_L1(candidates[n].bucket); } - lastHashed = ip; - /* Do not insert and do not look for a match */ - if (ZSTD_ldm_getTag(rollingHash, hBits, hashRateLog) != ldmTagMask) { - ip++; - continue; - } + for (n = 0; n < numSplits; n++) { + size_t forwardMatchLength = 0, backwardMatchLength = 0, + bestMatchLength = 0, mLength; + U32 offset; + BYTE const* const split = candidates[n].split; + U32 const checksum = candidates[n].checksum; + U32 const hash = candidates[n].hash; + ldmEntry_t* const bucket = candidates[n].bucket; + ldmEntry_t const* cur; + ldmEntry_t const* bestEntry = NULL; + ldmEntry_t newEntry; + + newEntry.offset = (U32)(split - base); + newEntry.checksum = checksum; + + /* If a split point would generate a sequence overlapping with + * the previous one, we merely register it in the hash table and + * move on */ + if (split < anchor) { + ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params); + continue; + } - /* Get the best entry and compute the match lengths */ - { - ldmEntry_t* const bucket = - ZSTD_ldm_getBucket(ldmState, - ZSTD_ldm_getSmallHash(rollingHash, hBits), - *params); - ldmEntry_t* cur; - size_t bestMatchLength = 0; - U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits); - - for (cur = bucket; cur < bucket + ldmBucketSize; ++cur) { + for (cur = bucket; cur < bucket + entsPerBucket; cur++) { size_t curForwardMatchLength, curBackwardMatchLength, curTotalMatchLength; if (cur->checksum != checksum || cur->offset <= lowestIndex) { @@ -328,30 +414,23 @@ static size_t ZSTD_ldm_generateSequences_internal( cur->offset < dictLimit ? dictEnd : iend; BYTE const* const lowMatchPtr = cur->offset < dictLimit ? dictStart : lowPrefixPtr; - - curForwardMatchLength = ZSTD_count_2segments( - ip, pMatch, iend, - matchEnd, lowPrefixPtr); + curForwardMatchLength = + ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr); if (curForwardMatchLength < minMatchLength) { continue; } - curBackwardMatchLength = - ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch, - lowMatchPtr); - curTotalMatchLength = curForwardMatchLength + - curBackwardMatchLength; + curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments( + split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd); } else { /* !extDict */ BYTE const* const pMatch = base + cur->offset; - curForwardMatchLength = ZSTD_count(ip, pMatch, iend); + curForwardMatchLength = ZSTD_count(split, pMatch, iend); if (curForwardMatchLength < minMatchLength) { continue; } curBackwardMatchLength = - ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch, - lowPrefixPtr); - curTotalMatchLength = curForwardMatchLength + - curBackwardMatchLength; + ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr); } + curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength; if (curTotalMatchLength > bestMatchLength) { bestMatchLength = curTotalMatchLength; @@ -360,57 +439,54 @@ static size_t ZSTD_ldm_generateSequences_internal( bestEntry = cur; } } - } - - /* No match found -- continue searching */ - if (bestEntry == NULL) { - ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, - hBits, current, - *params); - ip++; - continue; - } - /* Match found */ - mLength = forwardMatchLength + backwardMatchLength; - ip -= backwardMatchLength; + /* No match found -- insert an entry into the hash table + * and process the next candidate match */ + if (bestEntry == NULL) { + ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params); + continue; + } - { - /* Store the sequence: - * ip = current - backwardMatchLength - * The match is at (bestEntry->offset - backwardMatchLength) - */ - U32 const matchIndex = bestEntry->offset; - U32 const offset = current - matchIndex; - rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size; - - /* Out of sequence storage */ - if (rawSeqStore->size == rawSeqStore->capacity) - return ERROR(dstSize_tooSmall); - seq->litLength = (U32)(ip - anchor); - seq->matchLength = (U32)mLength; - seq->offset = offset; - rawSeqStore->size++; - } + /* Match found */ + offset = (U32)(split - base) - bestEntry->offset; + mLength = forwardMatchLength + backwardMatchLength; + { + rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size; + + /* Out of sequence storage */ + if (rawSeqStore->size == rawSeqStore->capacity) + return ERROR(dstSize_tooSmall); + seq->litLength = (U32)(split - backwardMatchLength - anchor); + seq->matchLength = (U32)mLength; + seq->offset = offset; + rawSeqStore->size++; + } - /* Insert the current entry into the hash table */ - ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, hBits, - (U32)(lastHashed - base), - *params); + /* Insert the current entry into the hash table --- it must be + * done after the previous block to avoid clobbering bestEntry */ + ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params); - assert(ip + backwardMatchLength == lastHashed); + anchor = split + forwardMatchLength; - /* Fill the hash table from lastHashed+1 to ip+mLength*/ - /* Heuristic: don't need to fill the entire table at end of block */ - if (ip + mLength <= ilimit) { - rollingHash = ZSTD_ldm_fillLdmHashTable( - ldmState, rollingHash, lastHashed, - ip + mLength, base, hBits, *params); - lastHashed = ip + mLength - 1; + /* If we find a match that ends after the data that we've hashed + * then we have a repeating, overlapping, pattern. E.g. all zeros. + * If one repetition of the pattern matches our `stopMask` then all + * repetitions will. We don't need to insert them all into out table, + * only the first one. So skip over overlapping matches. + * This is a major speed boost (20x) for compressing a single byte + * repeated, when that byte ends up in the table. + */ + if (anchor > ip + hashed) { + ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength); + /* Continue the outer loop at anchor (ip + hashed == anchor). */ + ip = anchor - hashed; + break; + } } - ip += mLength; - anchor = ip; + + ip += hashed; } + return iend - anchor; } @@ -459,7 +535,7 @@ size_t ZSTD_ldm_generateSequences( assert(chunkStart < iend); /* 1. Perform overflow correction if necessary. */ - if (ZSTD_window_needOverflowCorrection(ldmState->window, chunkEnd)) { + if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) { U32 const ldmHSize = 1U << params->hashLog; U32 const correction = ZSTD_window_correctOverflow( &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart); @@ -503,7 +579,9 @@ size_t ZSTD_ldm_generateSequences( return 0; } -void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch) { +void +ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch) +{ while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) { rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos; if (srcSize <= seq->litLength) { @@ -562,14 +640,32 @@ static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore, return sequence; } +void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) { + U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes); + while (currPos && rawSeqStore->pos < rawSeqStore->size) { + rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos]; + if (currPos >= currSeq.litLength + currSeq.matchLength) { + currPos -= currSeq.litLength + currSeq.matchLength; + rawSeqStore->pos++; + } else { + rawSeqStore->posInSequence = currPos; + break; + } + } + if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) { + rawSeqStore->posInSequence = 0; + } +} + size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + ZSTD_paramSwitch_e useRowMatchFinder, void const* src, size_t srcSize) { const ZSTD_compressionParameters* const cParams = &ms->cParams; unsigned const minMatch = cParams->minMatch; ZSTD_blockCompressor const blockCompressor = - ZSTD_selectBlockCompressor(cParams->strategy, ZSTD_matchState_dictMode(ms)); + ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms)); /* Input bounds */ BYTE const* const istart = (BYTE const*)src; BYTE const* const iend = istart + srcSize; @@ -577,9 +673,18 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, BYTE const* ip = istart; DEBUGLOG(5, "ZSTD_ldm_blockCompress: srcSize=%zu", srcSize); + /* If using opt parser, use LDMs only as candidates rather than always accepting them */ + if (cParams->strategy >= ZSTD_btopt) { + size_t lastLLSize; + ms->ldmSeqStore = rawSeqStore; + lastLLSize = blockCompressor(ms, seqStore, rep, src, srcSize); + ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore, srcSize); + return lastLLSize; + } + assert(rawSeqStore->pos <= rawSeqStore->size); assert(rawSeqStore->size <= rawSeqStore->capacity); - /* Loop through each sequence and apply the block compressor to the lits */ + /* Loop through each sequence and apply the block compressor to the literals */ while (rawSeqStore->pos < rawSeqStore->size && ip < iend) { /* maybeSplitSequence updates rawSeqStore->pos */ rawSeq const sequence = maybeSplitSequence(rawSeqStore, @@ -606,8 +711,8 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, rep[0] = sequence.offset; /* Store the sequence */ ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend, - sequence.offset + ZSTD_REP_MOVE, - sequence.matchLength - MINMATCH); + STORE_OFFSET(sequence.offset), + sequence.matchLength); ip += sequence.matchLength; } } diff --git a/thirdparty/zstd/compress/zstd_ldm.h b/thirdparty/zstd/compress/zstd_ldm.h index 229ea05a9e..4e68dbf52e 100644 --- a/thirdparty/zstd/compress/zstd_ldm.h +++ b/thirdparty/zstd/compress/zstd_ldm.h @@ -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 @@ -66,6 +66,7 @@ size_t ZSTD_ldm_generateSequences( */ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + ZSTD_paramSwitch_e useRowMatchFinder, void const* src, size_t srcSize); /** @@ -73,11 +74,17 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore, * * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`. * Avoids emitting matches less than `minMatch` bytes. - * Must be called for data with is not passed to ZSTD_ldm_blockCompress(). + * Must be called for data that is not passed to ZSTD_ldm_blockCompress(). */ void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch); +/* ZSTD_ldm_skipRawSeqStoreBytes(): + * Moves forward in rawSeqStore by nbBytes, updating fields 'pos' and 'posInSequence'. + * Not to be used in conjunction with ZSTD_ldm_skipSequences(). + * Must be called for data with is not passed to ZSTD_ldm_blockCompress(). + */ +void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes); /** ZSTD_ldm_getTableSize() : * Estimate the space needed for long distance matching tables or 0 if LDM is diff --git a/thirdparty/zstd/compress/zstd_ldm_geartab.h b/thirdparty/zstd/compress/zstd_ldm_geartab.h new file mode 100644 index 0000000000..647f865be2 --- /dev/null +++ b/thirdparty/zstd/compress/zstd_ldm_geartab.h @@ -0,0 +1,106 @@ +/* + * Copyright (c) Yann Collet, Facebook, Inc. + * 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). + * You may select, at your option, one of the above-listed licenses. + */ + +#ifndef ZSTD_LDM_GEARTAB_H +#define ZSTD_LDM_GEARTAB_H + +#include "../common/compiler.h" /* UNUSED_ATTR */ +#include "../common/mem.h" /* U64 */ + +static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = { + 0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc, + 0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05, + 0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e, + 0x9c8528f65badeaca, 0x86563706e2097529, 0x2902475fa375d889, + 0xafb32a9739a5ebe6, 0xce2714da3883e639, 0x21eaf821722e69e, + 0x37b628620b628, 0x49a8d455d88caf5, 0x8556d711e6958140, + 0x4f7ae74fc605c1f, 0x829f0c3468bd3a20, 0x4ffdc885c625179e, + 0x8473de048a3daf1b, 0x51008822b05646b2, 0x69d75d12b2d1cc5f, + 0x8c9d4a19159154bc, 0xc3cc10f4abbd4003, 0xd06ddc1cecb97391, + 0xbe48e6e7ed80302e, 0x3481db31cee03547, 0xacc3f67cdaa1d210, + 0x65cb771d8c7f96cc, 0x8eb27177055723dd, 0xc789950d44cd94be, + 0x934feadc3700b12b, 0x5e485f11edbdf182, 0x1e2e2a46fd64767a, + 0x2969ca71d82efa7c, 0x9d46e9935ebbba2e, 0xe056b67e05e6822b, + 0x94d73f55739d03a0, 0xcd7010bdb69b5a03, 0x455ef9fcd79b82f4, + 0x869cb54a8749c161, 0x38d1a4fa6185d225, 0xb475166f94bbe9bb, + 0xa4143548720959f1, 0x7aed4780ba6b26ba, 0xd0ce264439e02312, + 0x84366d746078d508, 0xa8ce973c72ed17be, 0x21c323a29a430b01, + 0x9962d617e3af80ee, 0xab0ce91d9c8cf75b, 0x530e8ee6d19a4dbc, + 0x2ef68c0cf53f5d72, 0xc03a681640a85506, 0x496e4e9f9c310967, + 0x78580472b59b14a0, 0x273824c23b388577, 0x66bf923ad45cb553, + 0x47ae1a5a2492ba86, 0x35e304569e229659, 0x4765182a46870b6f, + 0x6cbab625e9099412, 0xddac9a2e598522c1, 0x7172086e666624f2, + 0xdf5003ca503b7837, 0x88c0c1db78563d09, 0x58d51865acfc289d, + 0x177671aec65224f1, 0xfb79d8a241e967d7, 0x2be1e101cad9a49a, + 0x6625682f6e29186b, 0x399553457ac06e50, 0x35dffb4c23abb74, + 0x429db2591f54aade, 0xc52802a8037d1009, 0x6acb27381f0b25f3, + 0xf45e2551ee4f823b, 0x8b0ea2d99580c2f7, 0x3bed519cbcb4e1e1, + 0xff452823dbb010a, 0x9d42ed614f3dd267, 0x5b9313c06257c57b, + 0xa114b8008b5e1442, 0xc1fe311c11c13d4b, 0x66e8763ea34c5568, + 0x8b982af1c262f05d, 0xee8876faaa75fbb7, 0x8a62a4d0d172bb2a, + 0xc13d94a3b7449a97, 0x6dbbba9dc15d037c, 0xc786101f1d92e0f1, + 0xd78681a907a0b79b, 0xf61aaf2962c9abb9, 0x2cfd16fcd3cb7ad9, + 0x868c5b6744624d21, 0x25e650899c74ddd7, 0xba042af4a7c37463, + 0x4eb1a539465a3eca, 0xbe09dbf03b05d5ca, 0x774e5a362b5472ba, + 0x47a1221229d183cd, 0x504b0ca18ef5a2df, 0xdffbdfbde2456eb9, + 0x46cd2b2fbee34634, 0xf2aef8fe819d98c3, 0x357f5276d4599d61, + 0x24a5483879c453e3, 0x88026889192b4b9, 0x28da96671782dbec, + 0x4ef37c40588e9aaa, 0x8837b90651bc9fb3, 0xc164f741d3f0e5d6, + 0xbc135a0a704b70ba, 0x69cd868f7622ada, 0xbc37ba89e0b9c0ab, + 0x47c14a01323552f6, 0x4f00794bacee98bb, 0x7107de7d637a69d5, + 0x88af793bb6f2255e, 0xf3c6466b8799b598, 0xc288c616aa7f3b59, + 0x81ca63cf42fca3fd, 0x88d85ace36a2674b, 0xd056bd3792389e7, + 0xe55c396c4e9dd32d, 0xbefb504571e6c0a6, 0x96ab32115e91e8cc, + 0xbf8acb18de8f38d1, 0x66dae58801672606, 0x833b6017872317fb, + 0xb87c16f2d1c92864, 0xdb766a74e58b669c, 0x89659f85c61417be, + 0xc8daad856011ea0c, 0x76a4b565b6fe7eae, 0xa469d085f6237312, + 0xaaf0365683a3e96c, 0x4dbb746f8424f7b8, 0x638755af4e4acc1, + 0x3d7807f5bde64486, 0x17be6d8f5bbb7639, 0x903f0cd44dc35dc, + 0x67b672eafdf1196c, 0xa676ff93ed4c82f1, 0x521d1004c5053d9d, + 0x37ba9ad09ccc9202, 0x84e54d297aacfb51, 0xa0b4b776a143445, + 0x820d471e20b348e, 0x1874383cb83d46dc, 0x97edeec7a1efe11c, + 0xb330e50b1bdc42aa, 0x1dd91955ce70e032, 0xa514cdb88f2939d5, + 0x2791233fd90db9d3, 0x7b670a4cc50f7a9b, 0x77c07d2a05c6dfa5, + 0xe3778b6646d0a6fa, 0xb39c8eda47b56749, 0x933ed448addbef28, + 0xaf846af6ab7d0bf4, 0xe5af208eb666e49, 0x5e6622f73534cd6a, + 0x297daeca42ef5b6e, 0x862daef3d35539a6, 0xe68722498f8e1ea9, + 0x981c53093dc0d572, 0xfa09b0bfbf86fbf5, 0x30b1e96166219f15, + 0x70e7d466bdc4fb83, 0x5a66736e35f2a8e9, 0xcddb59d2b7c1baef, + 0xd6c7d247d26d8996, 0xea4e39eac8de1ba3, 0x539c8bb19fa3aff2, + 0x9f90e4c5fd508d8, 0xa34e5956fbaf3385, 0x2e2f8e151d3ef375, + 0x173691e9b83faec1, 0xb85a8d56bf016379, 0x8382381267408ae3, + 0xb90f901bbdc0096d, 0x7c6ad32933bcec65, 0x76bb5e2f2c8ad595, + 0x390f851a6cf46d28, 0xc3e6064da1c2da72, 0xc52a0c101cfa5389, + 0xd78eaf84a3fbc530, 0x3781b9e2288b997e, 0x73c2f6dea83d05c4, + 0x4228e364c5b5ed7, 0x9d7a3edf0da43911, 0x8edcfeda24686756, + 0x5e7667a7b7a9b3a1, 0x4c4f389fa143791d, 0xb08bc1023da7cddc, + 0x7ab4be3ae529b1cc, 0x754e6132dbe74ff9, 0x71635442a839df45, + 0x2f6fb1643fbe52de, 0x961e0a42cf7a8177, 0xf3b45d83d89ef2ea, + 0xee3de4cf4a6e3e9b, 0xcd6848542c3295e7, 0xe4cee1664c78662f, + 0x9947548b474c68c4, 0x25d73777a5ed8b0b, 0xc915b1d636b7fc, + 0x21c2ba75d9b0d2da, 0x5f6b5dcf608a64a1, 0xdcf333255ff9570c, + 0x633b922418ced4ee, 0xc136dde0b004b34a, 0x58cc83b05d4b2f5a, + 0x5eb424dda28e42d2, 0x62df47369739cd98, 0xb4e0b42485e4ce17, + 0x16e1f0c1f9a8d1e7, 0x8ec3916707560ebf, 0x62ba6e2df2cc9db3, + 0xcbf9f4ff77d83a16, 0x78d9d7d07d2bbcc4, 0xef554ce1e02c41f4, + 0x8d7581127eccf94d, 0xa9b53336cb3c8a05, 0x38c42c0bf45c4f91, + 0x640893cdf4488863, 0x80ec34bc575ea568, 0x39f324f5b48eaa40, + 0xe9d9ed1f8eff527f, 0x9224fc058cc5a214, 0xbaba00b04cfe7741, + 0x309a9f120fcf52af, 0xa558f3ec65626212, 0x424bec8b7adabe2f, + 0x41622513a6aea433, 0xb88da2d5324ca798, 0xd287733b245528a4, + 0x9a44697e6d68aec3, 0x7b1093be2f49bb28, 0x50bbec632e3d8aad, + 0x6cd90723e1ea8283, 0x897b9e7431b02bf3, 0x219efdcb338a7047, + 0x3b0311f0a27c0656, 0xdb17bf91c0db96e7, 0x8cd4fd6b4e85a5b2, + 0xfab071054ba6409d, 0x40d6fe831fa9dfd9, 0xaf358debad7d791e, + 0xeb8d0e25a65e3e58, 0xbbcbd3df14e08580, 0xcf751f27ecdab2b, + 0x2b4da14f2613d8f4 +}; + +#endif /* ZSTD_LDM_GEARTAB_H */ diff --git a/thirdparty/zstd/compress/zstd_opt.c b/thirdparty/zstd/compress/zstd_opt.c index 36fff050cf..1b1ddad428 100644 --- a/thirdparty/zstd/compress/zstd_opt.c +++ b/thirdparty/zstd/compress/zstd_opt.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2020, Przemyslaw Skibinski, Yann Collet, Facebook, Inc. + * Copyright (c) Przemyslaw Skibinski, Yann Collet, Facebook, Inc. * All rights reserved. * * This source code is licensed under both the BSD-style license (found in the @@ -14,7 +14,6 @@ #define ZSTD_LITFREQ_ADD 2 /* scaling factor for litFreq, so that frequencies adapt faster to new stats */ -#define ZSTD_FREQ_DIV 4 /* log factor when using previous stats to init next stats */ #define ZSTD_MAX_PRICE (1<<30) #define ZSTD_PREDEF_THRESHOLD 1024 /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */ @@ -24,11 +23,11 @@ * Price functions for optimal parser ***************************************/ -#if 0 /* approximation at bit level */ +#if 0 /* approximation at bit level (for tests) */ # define BITCOST_ACCURACY 0 # define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY) -# define WEIGHT(stat) ((void)opt, ZSTD_bitWeight(stat)) -#elif 0 /* fractional bit accuracy */ +# define WEIGHT(stat, opt) ((void)opt, ZSTD_bitWeight(stat)) +#elif 0 /* fractional bit accuracy (for tests) */ # define BITCOST_ACCURACY 8 # define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY) # define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat)) @@ -66,7 +65,7 @@ MEM_STATIC double ZSTD_fCost(U32 price) static int ZSTD_compressedLiterals(optState_t const* const optPtr) { - return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed; + return optPtr->literalCompressionMode != ZSTD_ps_disable; } static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel) @@ -79,25 +78,46 @@ static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel) } -/* ZSTD_downscaleStat() : - * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus) - * return the resulting sum of elements */ -static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus) +static U32 sum_u32(const unsigned table[], size_t nbElts) +{ + size_t n; + U32 total = 0; + for (n=0; n<nbElts; n++) { + total += table[n]; + } + return total; +} + +static U32 ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift) { U32 s, sum=0; - DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1); - assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31); + DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)", (unsigned)lastEltIndex+1, (unsigned)shift); + assert(shift < 30); for (s=0; s<lastEltIndex+1; s++) { - table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus)); + table[s] = 1 + (table[s] >> shift); sum += table[s]; } return sum; } +/* ZSTD_scaleStats() : + * reduce all elements in table is sum too large + * return the resulting sum of elements */ +static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget) +{ + U32 const prevsum = sum_u32(table, lastEltIndex+1); + U32 const factor = prevsum >> logTarget; + DEBUGLOG(5, "ZSTD_scaleStats (nbElts=%u, target=%u)", (unsigned)lastEltIndex+1, (unsigned)logTarget); + assert(logTarget < 30); + if (factor <= 1) return prevsum; + return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor)); +} + /* ZSTD_rescaleFreqs() : * if first block (detected by optPtr->litLengthSum == 0) : init statistics * take hints from dictionary if there is one - * or init from zero, using src for literals stats, or flat 1 for match symbols + * and init from zero if there is none, + * using src for literals stats, and baseline stats for sequence symbols * otherwise downscale existing stats, to be used as seed for next block. */ static void @@ -126,7 +146,7 @@ ZSTD_rescaleFreqs(optState_t* const optPtr, optPtr->litSum = 0; for (lit=0; lit<=MaxLit; lit++) { U32 const scaleLog = 11; /* scale to 2K */ - U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit); + U32 const bitCost = HUF_getNbBitsFromCTable(optPtr->symbolCosts->huf.CTable, lit); assert(bitCost <= scaleLog); optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/; optPtr->litSum += optPtr->litFreq[lit]; @@ -174,14 +194,19 @@ ZSTD_rescaleFreqs(optState_t* const optPtr, if (compressedLiterals) { unsigned lit = MaxLit; HIST_count_simple(optPtr->litFreq, &lit, src, srcSize); /* use raw first block to init statistics */ - optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1); + optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8); } - { unsigned ll; - for (ll=0; ll<=MaxLL; ll++) - optPtr->litLengthFreq[ll] = 1; + { unsigned const baseLLfreqs[MaxLL+1] = { + 4, 2, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1 + }; + ZSTD_memcpy(optPtr->litLengthFreq, baseLLfreqs, sizeof(baseLLfreqs)); + optPtr->litLengthSum = sum_u32(baseLLfreqs, MaxLL+1); } - optPtr->litLengthSum = MaxLL+1; { unsigned ml; for (ml=0; ml<=MaxML; ml++) @@ -189,21 +214,26 @@ ZSTD_rescaleFreqs(optState_t* const optPtr, } optPtr->matchLengthSum = MaxML+1; - { unsigned of; - for (of=0; of<=MaxOff; of++) - optPtr->offCodeFreq[of] = 1; + { unsigned const baseOFCfreqs[MaxOff+1] = { + 6, 2, 1, 1, 2, 3, 4, 4, + 4, 3, 2, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1 + }; + ZSTD_memcpy(optPtr->offCodeFreq, baseOFCfreqs, sizeof(baseOFCfreqs)); + optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1); } - optPtr->offCodeSum = MaxOff+1; + } } else { /* new block : re-use previous statistics, scaled down */ if (compressedLiterals) - optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1); - optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0); - optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0); - optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0); + optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12); + optPtr->litLengthSum = ZSTD_scaleStats(optPtr->litLengthFreq, MaxLL, 11); + optPtr->matchLengthSum = ZSTD_scaleStats(optPtr->matchLengthFreq, MaxML, 11); + optPtr->offCodeSum = ZSTD_scaleStats(optPtr->offCodeFreq, MaxOff, 11); } ZSTD_setBasePrices(optPtr, optLevel); @@ -239,7 +269,16 @@ static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength, * cost of literalLength symbol */ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel) { - if (optPtr->priceType == zop_predef) return WEIGHT(litLength, optLevel); + assert(litLength <= ZSTD_BLOCKSIZE_MAX); + if (optPtr->priceType == zop_predef) + return WEIGHT(litLength, optLevel); + /* We can't compute the litLength price for sizes >= ZSTD_BLOCKSIZE_MAX + * because it isn't representable in the zstd format. So instead just + * call it 1 bit more than ZSTD_BLOCKSIZE_MAX - 1. In this case the block + * would be all literals. + */ + if (litLength == ZSTD_BLOCKSIZE_MAX) + return BITCOST_MULTIPLIER + ZSTD_litLengthPrice(ZSTD_BLOCKSIZE_MAX - 1, optPtr, optLevel); /* dynamic statistics */ { U32 const llCode = ZSTD_LLcode(litLength); @@ -252,15 +291,17 @@ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optP /* ZSTD_getMatchPrice() : * Provides the cost of the match part (offset + matchLength) of a sequence * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence. - * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */ + * @offcode : expects a scale where 0,1,2 are repcodes 1-3, and 3+ are real_offsets+2 + * @optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) + */ FORCE_INLINE_TEMPLATE U32 -ZSTD_getMatchPrice(U32 const offset, +ZSTD_getMatchPrice(U32 const offcode, U32 const matchLength, const optState_t* const optPtr, int const optLevel) { U32 price; - U32 const offCode = ZSTD_highbit32(offset+1); + U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offcode)); U32 const mlBase = matchLength - MINMATCH; assert(matchLength >= MINMATCH); @@ -303,8 +344,8 @@ static void ZSTD_updateStats(optState_t* const optPtr, optPtr->litLengthSum++; } - /* match offset code (0-2=>repCode; 3+=>offset+2) */ - { U32 const offCode = ZSTD_highbit32(offsetCode+1); + /* offset code : expected to follow storeSeq() numeric representation */ + { U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offsetCode)); assert(offCode <= MaxOff); optPtr->offCodeFreq[offCode]++; optPtr->offCodeSum++; @@ -338,7 +379,7 @@ MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length) /* Update hashTable3 up to ip (excluded) Assumption : always within prefix (i.e. not within extDict) */ -static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms, +static U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_matchState_t* ms, U32* nextToUpdate3, const BYTE* const ip) { @@ -364,11 +405,13 @@ static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms, * Binary Tree search ***************************************/ /** ZSTD_insertBt1() : add one or multiple positions to tree. - * ip : assumed <= iend-8 . + * @param ip assumed <= iend-8 . + * @param target The target of ZSTD_updateTree_internal() - we are filling to this position * @return : nb of positions added */ static U32 ZSTD_insertBt1( - ZSTD_matchState_t* ms, + const ZSTD_matchState_t* ms, const BYTE* const ip, const BYTE* const iend, + U32 const target, U32 const mls, const int extDict) { const ZSTD_compressionParameters* const cParams = &ms->cParams; @@ -386,32 +429,36 @@ static U32 ZSTD_insertBt1( const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const prefixStart = base + dictLimit; const BYTE* match; - const U32 current = (U32)(ip-base); - const U32 btLow = btMask >= current ? 0 : current - btMask; - U32* smallerPtr = bt + 2*(current&btMask); + const U32 curr = (U32)(ip-base); + const U32 btLow = btMask >= curr ? 0 : curr - btMask; + U32* smallerPtr = bt + 2*(curr&btMask); U32* largerPtr = smallerPtr + 1; U32 dummy32; /* to be nullified at the end */ - U32 const windowLow = ms->window.lowLimit; - U32 matchEndIdx = current+8+1; + /* windowLow is based on target because + * we only need positions that will be in the window at the end of the tree update. + */ + U32 const windowLow = ZSTD_getLowestMatchIndex(ms, target, cParams->windowLog); + U32 matchEndIdx = curr+8+1; size_t bestLength = 8; U32 nbCompares = 1U << cParams->searchLog; #ifdef ZSTD_C_PREDICT - U32 predictedSmall = *(bt + 2*((current-1)&btMask) + 0); - U32 predictedLarge = *(bt + 2*((current-1)&btMask) + 1); + U32 predictedSmall = *(bt + 2*((curr-1)&btMask) + 0); + U32 predictedLarge = *(bt + 2*((curr-1)&btMask) + 1); predictedSmall += (predictedSmall>0); predictedLarge += (predictedLarge>0); #endif /* ZSTD_C_PREDICT */ - DEBUGLOG(8, "ZSTD_insertBt1 (%u)", current); + DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr); + assert(curr <= target); assert(ip <= iend-8); /* required for h calculation */ - hashTable[h] = current; /* Update Hash Table */ + hashTable[h] = curr; /* Update Hash Table */ assert(windowLow > 0); - while (nbCompares-- && (matchIndex >= windowLow)) { + for (; nbCompares && (matchIndex >= windowLow); --nbCompares) { U32* const nextPtr = bt + 2*(matchIndex & btMask); size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ - assert(matchIndex < current); + assert(matchIndex < curr); #ifdef ZSTD_C_PREDICT /* note : can create issues when hlog small <= 11 */ const U32* predictPtr = bt + 2*((matchIndex-1) & btMask); /* written this way, as bt is a roll buffer */ @@ -474,8 +521,8 @@ static U32 ZSTD_insertBt1( *smallerPtr = *largerPtr = 0; { U32 positions = 0; if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384)); /* speed optimization */ - assert(matchEndIdx > current + 8); - return MAX(positions, matchEndIdx - (current + 8)); + assert(matchEndIdx > curr + 8); + return MAX(positions, matchEndIdx - (curr + 8)); } } @@ -492,7 +539,7 @@ void ZSTD_updateTree_internal( idx, target, dictMode); while(idx < target) { - U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict); + U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, target, mls, dictMode == ZSTD_extDict); assert(idx < (U32)(idx + forward)); idx += forward; } @@ -519,7 +566,7 @@ U32 ZSTD_insertBtAndGetAllMatches ( const ZSTD_compressionParameters* const cParams = &ms->cParams; U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1); const BYTE* const base = ms->window.base; - U32 const current = (U32)(ip-base); + U32 const curr = (U32)(ip-base); U32 const hashLog = cParams->hashLog; U32 const minMatch = (mls==3) ? 3 : 4; U32* const hashTable = ms->hashTable; @@ -533,12 +580,12 @@ U32 ZSTD_insertBtAndGetAllMatches ( U32 const dictLimit = ms->window.dictLimit; const BYTE* const dictEnd = dictBase + dictLimit; const BYTE* const prefixStart = base + dictLimit; - U32 const btLow = (btMask >= current) ? 0 : current - btMask; - U32 const windowLow = ZSTD_getLowestMatchIndex(ms, current, cParams->windowLog); + U32 const btLow = (btMask >= curr) ? 0 : curr - btMask; + U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog); U32 const matchLow = windowLow ? windowLow : 1; - U32* smallerPtr = bt + 2*(current&btMask); - U32* largerPtr = bt + 2*(current&btMask) + 1; - U32 matchEndIdx = current+8+1; /* farthest referenced position of any match => detects repetitive patterns */ + U32* smallerPtr = bt + 2*(curr&btMask); + U32* largerPtr = bt + 2*(curr&btMask) + 1; + U32 matchEndIdx = curr+8+1; /* farthest referenced position of any match => detects repetitive patterns */ U32 dummy32; /* to be nullified at the end */ U32 mnum = 0; U32 nbCompares = 1U << cParams->searchLog; @@ -557,7 +604,7 @@ U32 ZSTD_insertBtAndGetAllMatches ( U32 const dmsBtLow = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit; size_t bestLength = lengthToBeat-1; - DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", current); + DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", curr); /* check repCode */ assert(ll0 <= 1); /* necessarily 1 or 0 */ @@ -565,29 +612,29 @@ U32 ZSTD_insertBtAndGetAllMatches ( U32 repCode; for (repCode = ll0; repCode < lastR; repCode++) { U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode]; - U32 const repIndex = current - repOffset; + U32 const repIndex = curr - repOffset; U32 repLen = 0; - assert(current >= dictLimit); - if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < current-dictLimit) { /* equivalent to `current > repIndex >= dictLimit` */ + assert(curr >= dictLimit); + if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < curr-dictLimit) { /* equivalent to `curr > repIndex >= dictLimit` */ /* We must validate the repcode offset because when we're using a dictionary the * valid offset range shrinks when the dictionary goes out of bounds. */ if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) { repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch; } - } else { /* repIndex < dictLimit || repIndex >= current */ + } else { /* repIndex < dictLimit || repIndex >= curr */ const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ? dmsBase + repIndex - dmsIndexDelta : dictBase + repIndex; - assert(current >= windowLow); + assert(curr >= windowLow); if ( dictMode == ZSTD_extDict - && ( ((repOffset-1) /*intentional overflow*/ < current - windowLow) /* equivalent to `current > repIndex >= windowLow` */ + && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow) /* equivalent to `curr > repIndex >= windowLow` */ & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */) && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) { repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch; } if (dictMode == ZSTD_dictMatchState - && ( ((repOffset-1) /*intentional overflow*/ < current - (dmsLowLimit + dmsIndexDelta)) /* equivalent to `current > repIndex >= dmsLowLimit` */ + && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta)) /* equivalent to `curr > repIndex >= dmsLowLimit` */ & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */ && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) { repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch; @@ -597,7 +644,7 @@ U32 ZSTD_insertBtAndGetAllMatches ( DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u", repCode, ll0, repOffset, repLen); bestLength = repLen; - matches[mnum].off = repCode - ll0; + matches[mnum].off = STORE_REPCODE(repCode - ll0 + 1); /* expect value between 1 and 3 */ matches[mnum].len = (U32)repLen; mnum++; if ( (repLen > sufficient_len) @@ -609,7 +656,7 @@ U32 ZSTD_insertBtAndGetAllMatches ( if ((mls == 3) /*static*/ && (bestLength < mls)) { U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip); if ((matchIndex3 >= matchLow) - & (current - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) { + & (curr - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) { size_t mlen; if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) { const BYTE* const match = base + matchIndex3; @@ -624,26 +671,26 @@ U32 ZSTD_insertBtAndGetAllMatches ( DEBUGLOG(8, "found small match with hlog3, of length %u", (U32)mlen); bestLength = mlen; - assert(current > matchIndex3); + assert(curr > matchIndex3); assert(mnum==0); /* no prior solution */ - matches[0].off = (current - matchIndex3) + ZSTD_REP_MOVE; + matches[0].off = STORE_OFFSET(curr - matchIndex3); matches[0].len = (U32)mlen; mnum = 1; if ( (mlen > sufficient_len) | (ip+mlen == iLimit) ) { /* best possible length */ - ms->nextToUpdate = current+1; /* skip insertion */ + ms->nextToUpdate = curr+1; /* skip insertion */ return 1; } } } /* no dictMatchState lookup: dicts don't have a populated HC3 table */ - } + } /* if (mls == 3) */ - hashTable[h] = current; /* Update Hash Table */ + hashTable[h] = curr; /* Update Hash Table */ - while (nbCompares-- && (matchIndex >= matchLow)) { + for (; nbCompares && (matchIndex >= matchLow); --nbCompares) { U32* const nextPtr = bt + 2*(matchIndex & btMask); const BYTE* match; size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ - assert(current > matchIndex); + assert(curr > matchIndex); if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) { assert(matchIndex+matchLength >= dictLimit); /* ensure the condition is correct when !extDict */ @@ -660,20 +707,19 @@ U32 ZSTD_insertBtAndGetAllMatches ( if (matchLength > bestLength) { DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)", - (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE); + (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex)); assert(matchEndIdx > matchIndex); if (matchLength > matchEndIdx - matchIndex) matchEndIdx = matchIndex + (U32)matchLength; bestLength = matchLength; - matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE; + matches[mnum].off = STORE_OFFSET(curr - matchIndex); matches[mnum].len = (U32)matchLength; mnum++; if ( (matchLength > ZSTD_OPT_NUM) | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) { if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */ break; /* drop, to preserve bt consistency (miss a little bit of compression) */ - } - } + } } if (match[matchLength] < ip[matchLength]) { /* match smaller than current */ @@ -692,12 +738,13 @@ U32 ZSTD_insertBtAndGetAllMatches ( *smallerPtr = *largerPtr = 0; + assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */ if (dictMode == ZSTD_dictMatchState && nbCompares) { size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls); U32 dictMatchIndex = dms->hashTable[dmsH]; const U32* const dmsBt = dms->chainTable; commonLengthSmaller = commonLengthLarger = 0; - while (nbCompares-- && (dictMatchIndex > dmsLowLimit)) { + for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) { const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask); size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ const BYTE* match = dmsBase + dictMatchIndex; @@ -708,18 +755,17 @@ U32 ZSTD_insertBtAndGetAllMatches ( if (matchLength > bestLength) { matchIndex = dictMatchIndex + dmsIndexDelta; DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)", - (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE); + (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex)); if (matchLength > matchEndIdx - matchIndex) matchEndIdx = matchIndex + (U32)matchLength; bestLength = matchLength; - matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE; + matches[mnum].off = STORE_OFFSET(curr - matchIndex); matches[mnum].len = (U32)matchLength; mnum++; if ( (matchLength > ZSTD_OPT_NUM) | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) { break; /* drop, to guarantee consistency (miss a little bit of compression) */ - } - } + } } if (dictMatchIndex <= dmsBtLow) { break; } /* beyond tree size, stop the search */ if (match[matchLength] < ip[matchLength]) { @@ -729,47 +775,242 @@ U32 ZSTD_insertBtAndGetAllMatches ( /* match is larger than current */ commonLengthLarger = matchLength; dictMatchIndex = nextPtr[0]; - } - } - } + } } } /* if (dictMode == ZSTD_dictMatchState) */ - assert(matchEndIdx > current+8); + assert(matchEndIdx > curr+8); ms->nextToUpdate = matchEndIdx - 8; /* skip repetitive patterns */ return mnum; } +typedef U32 (*ZSTD_getAllMatchesFn)( + ZSTD_match_t*, + ZSTD_matchState_t*, + U32*, + const BYTE*, + const BYTE*, + const U32 rep[ZSTD_REP_NUM], + U32 const ll0, + U32 const lengthToBeat); + +FORCE_INLINE_TEMPLATE U32 ZSTD_btGetAllMatches_internal( + ZSTD_match_t* matches, + ZSTD_matchState_t* ms, + U32* nextToUpdate3, + const BYTE* ip, + const BYTE* const iHighLimit, + const U32 rep[ZSTD_REP_NUM], + U32 const ll0, + U32 const lengthToBeat, + const ZSTD_dictMode_e dictMode, + const U32 mls) +{ + assert(BOUNDED(3, ms->cParams.minMatch, 6) == mls); + DEBUGLOG(8, "ZSTD_BtGetAllMatches(dictMode=%d, mls=%u)", (int)dictMode, mls); + if (ip < ms->window.base + ms->nextToUpdate) + return 0; /* skipped area */ + ZSTD_updateTree_internal(ms, ip, iHighLimit, mls, dictMode); + return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, mls); +} + +#define ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls) ZSTD_btGetAllMatches_##dictMode##_##mls + +#define GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, mls) \ + static U32 ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls)( \ + ZSTD_match_t* matches, \ + ZSTD_matchState_t* ms, \ + U32* nextToUpdate3, \ + const BYTE* ip, \ + const BYTE* const iHighLimit, \ + const U32 rep[ZSTD_REP_NUM], \ + U32 const ll0, \ + U32 const lengthToBeat) \ + { \ + return ZSTD_btGetAllMatches_internal( \ + matches, ms, nextToUpdate3, ip, iHighLimit, \ + rep, ll0, lengthToBeat, ZSTD_##dictMode, mls); \ + } -FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches ( - ZSTD_match_t* matches, /* store result (match found, increasing size) in this table */ - ZSTD_matchState_t* ms, - U32* nextToUpdate3, - const BYTE* ip, const BYTE* const iHighLimit, const ZSTD_dictMode_e dictMode, - const U32 rep[ZSTD_REP_NUM], - U32 const ll0, - U32 const lengthToBeat) +#define GEN_ZSTD_BT_GET_ALL_MATCHES(dictMode) \ + GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 3) \ + GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 4) \ + GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 5) \ + GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 6) + +GEN_ZSTD_BT_GET_ALL_MATCHES(noDict) +GEN_ZSTD_BT_GET_ALL_MATCHES(extDict) +GEN_ZSTD_BT_GET_ALL_MATCHES(dictMatchState) + +#define ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMode) \ + { \ + ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 3), \ + ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 4), \ + ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 5), \ + ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 6) \ + } + +static ZSTD_getAllMatchesFn +ZSTD_selectBtGetAllMatches(ZSTD_matchState_t const* ms, ZSTD_dictMode_e const dictMode) { - const ZSTD_compressionParameters* const cParams = &ms->cParams; - U32 const matchLengthSearch = cParams->minMatch; - DEBUGLOG(8, "ZSTD_BtGetAllMatches"); - if (ip < ms->window.base + ms->nextToUpdate) return 0; /* skipped area */ - ZSTD_updateTree_internal(ms, ip, iHighLimit, matchLengthSearch, dictMode); - switch(matchLengthSearch) - { - case 3 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 3); - default : - case 4 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 4); - case 5 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 5); - case 7 : - case 6 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 6); + ZSTD_getAllMatchesFn const getAllMatchesFns[3][4] = { + ZSTD_BT_GET_ALL_MATCHES_ARRAY(noDict), + ZSTD_BT_GET_ALL_MATCHES_ARRAY(extDict), + ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMatchState) + }; + U32 const mls = BOUNDED(3, ms->cParams.minMatch, 6); + assert((U32)dictMode < 3); + assert(mls - 3 < 4); + return getAllMatchesFns[(int)dictMode][mls - 3]; +} + +/************************* +* LDM helper functions * +*************************/ + +/* Struct containing info needed to make decision about ldm inclusion */ +typedef struct { + rawSeqStore_t seqStore; /* External match candidates store for this block */ + U32 startPosInBlock; /* Start position of the current match candidate */ + U32 endPosInBlock; /* End position of the current match candidate */ + U32 offset; /* Offset of the match candidate */ +} ZSTD_optLdm_t; + +/* ZSTD_optLdm_skipRawSeqStoreBytes(): + * Moves forward in @rawSeqStore by @nbBytes, + * which will update the fields 'pos' and 'posInSequence'. + */ +static void ZSTD_optLdm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) +{ + U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes); + while (currPos && rawSeqStore->pos < rawSeqStore->size) { + rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos]; + if (currPos >= currSeq.litLength + currSeq.matchLength) { + currPos -= currSeq.litLength + currSeq.matchLength; + rawSeqStore->pos++; + } else { + rawSeqStore->posInSequence = currPos; + break; + } + } + if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) { + rawSeqStore->posInSequence = 0; } } +/* ZSTD_opt_getNextMatchAndUpdateSeqStore(): + * Calculates the beginning and end of the next match in the current block. + * Updates 'pos' and 'posInSequence' of the ldmSeqStore. + */ +static void +ZSTD_opt_getNextMatchAndUpdateSeqStore(ZSTD_optLdm_t* optLdm, U32 currPosInBlock, + U32 blockBytesRemaining) +{ + rawSeq currSeq; + U32 currBlockEndPos; + U32 literalsBytesRemaining; + U32 matchBytesRemaining; + + /* Setting match end position to MAX to ensure we never use an LDM during this block */ + if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) { + optLdm->startPosInBlock = UINT_MAX; + optLdm->endPosInBlock = UINT_MAX; + return; + } + /* Calculate appropriate bytes left in matchLength and litLength + * after adjusting based on ldmSeqStore->posInSequence */ + currSeq = optLdm->seqStore.seq[optLdm->seqStore.pos]; + assert(optLdm->seqStore.posInSequence <= currSeq.litLength + currSeq.matchLength); + currBlockEndPos = currPosInBlock + blockBytesRemaining; + literalsBytesRemaining = (optLdm->seqStore.posInSequence < currSeq.litLength) ? + currSeq.litLength - (U32)optLdm->seqStore.posInSequence : + 0; + matchBytesRemaining = (literalsBytesRemaining == 0) ? + currSeq.matchLength - ((U32)optLdm->seqStore.posInSequence - currSeq.litLength) : + currSeq.matchLength; + + /* If there are more literal bytes than bytes remaining in block, no ldm is possible */ + if (literalsBytesRemaining >= blockBytesRemaining) { + optLdm->startPosInBlock = UINT_MAX; + optLdm->endPosInBlock = UINT_MAX; + ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, blockBytesRemaining); + return; + } + + /* Matches may be < MINMATCH by this process. In that case, we will reject them + when we are deciding whether or not to add the ldm */ + optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining; + optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining; + optLdm->offset = currSeq.offset; + + if (optLdm->endPosInBlock > currBlockEndPos) { + /* Match ends after the block ends, we can't use the whole match */ + optLdm->endPosInBlock = currBlockEndPos; + ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, currBlockEndPos - currPosInBlock); + } else { + /* Consume nb of bytes equal to size of sequence left */ + ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, literalsBytesRemaining + matchBytesRemaining); + } +} + +/* ZSTD_optLdm_maybeAddMatch(): + * Adds a match if it's long enough, + * based on it's 'matchStartPosInBlock' and 'matchEndPosInBlock', + * into 'matches'. Maintains the correct ordering of 'matches'. + */ +static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches, + const ZSTD_optLdm_t* optLdm, U32 currPosInBlock) +{ + U32 const posDiff = currPosInBlock - optLdm->startPosInBlock; + /* Note: ZSTD_match_t actually contains offCode and matchLength (before subtracting MINMATCH) */ + U32 const candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff; + + /* Ensure that current block position is not outside of the match */ + if (currPosInBlock < optLdm->startPosInBlock + || currPosInBlock >= optLdm->endPosInBlock + || candidateMatchLength < MINMATCH) { + return; + } + + if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) { + U32 const candidateOffCode = STORE_OFFSET(optLdm->offset); + DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offCode: %u matchLength %u) at block position=%u", + candidateOffCode, candidateMatchLength, currPosInBlock); + matches[*nbMatches].len = candidateMatchLength; + matches[*nbMatches].off = candidateOffCode; + (*nbMatches)++; + } +} + +/* ZSTD_optLdm_processMatchCandidate(): + * Wrapper function to update ldm seq store and call ldm functions as necessary. + */ +static void +ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm, + ZSTD_match_t* matches, U32* nbMatches, + U32 currPosInBlock, U32 remainingBytes) +{ + if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) { + return; + } + + if (currPosInBlock >= optLdm->endPosInBlock) { + if (currPosInBlock > optLdm->endPosInBlock) { + /* The position at which ZSTD_optLdm_processMatchCandidate() is called is not necessarily + * at the end of a match from the ldm seq store, and will often be some bytes + * over beyond matchEndPosInBlock. As such, we need to correct for these "overshoots" + */ + U32 const posOvershoot = currPosInBlock - optLdm->endPosInBlock; + ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot); + } + ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes); + } + ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock); +} + /*-******************************* * Optimal parser *********************************/ - static U32 ZSTD_totalLen(ZSTD_optimal_t sol) { return sol.litlen + sol.mlen; @@ -810,6 +1051,8 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, const BYTE* const prefixStart = base + ms->window.dictLimit; const ZSTD_compressionParameters* const cParams = &ms->cParams; + ZSTD_getAllMatchesFn getAllMatches = ZSTD_selectBtGetAllMatches(ms, dictMode); + U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1); U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4; U32 nextToUpdate3 = ms->nextToUpdate; @@ -817,6 +1060,11 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, ZSTD_optimal_t* const opt = optStatePtr->priceTable; ZSTD_match_t* const matches = optStatePtr->matchTable; ZSTD_optimal_t lastSequence; + ZSTD_optLdm_t optLdm; + + optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore; + optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0; + ZSTD_opt_getNextMatchAndUpdateSeqStore(&optLdm, (U32)(ip-istart), (U32)(iend-ip)); /* init */ DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u", @@ -832,7 +1080,9 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, /* find first match */ { U32 const litlen = (U32)(ip - anchor); U32 const ll0 = !litlen; - U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch); + U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch); + ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches, + (U32)(ip-istart), (U32)(iend - ip)); if (!nbMatches) { ip++; continue; } /* initialize opt[0] */ @@ -844,18 +1094,18 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, * in every price. We include the literal length to avoid negative * prices when we subtract the previous literal length. */ - opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel); + opt[0].price = (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel); /* large match -> immediate encoding */ { U32 const maxML = matches[nbMatches-1].len; - U32 const maxOffset = matches[nbMatches-1].off; + U32 const maxOffcode = matches[nbMatches-1].off; DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series", - nbMatches, maxML, maxOffset, (U32)(ip-prefixStart)); + nbMatches, maxML, maxOffcode, (U32)(ip-prefixStart)); if (maxML > sufficient_len) { lastSequence.litlen = litlen; lastSequence.mlen = maxML; - lastSequence.off = maxOffset; + lastSequence.off = maxOffcode; DEBUGLOG(6, "large match (%u>%u), immediate encoding", maxML, sufficient_len); cur = 0; @@ -864,24 +1114,25 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, } } /* set prices for first matches starting position == 0 */ - { U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel); + assert(opt[0].price >= 0); + { U32 const literalsPrice = (U32)opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel); U32 pos; U32 matchNb; for (pos = 1; pos < minMatch; pos++) { opt[pos].price = ZSTD_MAX_PRICE; /* mlen, litlen and price will be fixed during forward scanning */ } for (matchNb = 0; matchNb < nbMatches; matchNb++) { - U32 const offset = matches[matchNb].off; + U32 const offcode = matches[matchNb].off; U32 const end = matches[matchNb].len; for ( ; pos <= end ; pos++ ) { - U32 const matchPrice = ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel); + U32 const matchPrice = ZSTD_getMatchPrice(offcode, pos, optStatePtr, optLevel); U32 const sequencePrice = literalsPrice + matchPrice; DEBUGLOG(7, "rPos:%u => set initial price : %.2f", pos, ZSTD_fCost(sequencePrice)); opt[pos].mlen = pos; - opt[pos].off = offset; + opt[pos].off = offcode; opt[pos].litlen = litlen; - opt[pos].price = sequencePrice; + opt[pos].price = (int)sequencePrice; } } last_pos = pos-1; } @@ -896,9 +1147,9 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, /* Fix current position with one literal if cheaper */ { U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1; int const price = opt[cur-1].price - + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel) - + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel) - - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel); + + (int)ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel) + + (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel) + - (int)ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel); assert(price < 1000000000); /* overflow check */ if (price <= opt[cur].price) { DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)", @@ -924,10 +1175,10 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, assert(cur >= opt[cur].mlen); if (opt[cur].mlen != 0) { U32 const prev = cur - opt[cur].mlen; - repcodes_t newReps = ZSTD_updateRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0); - memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t)); + repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0); + ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t)); } else { - memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t)); + ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t)); } /* last match must start at a minimum distance of 8 from oend */ @@ -941,12 +1192,17 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, continue; /* skip unpromising positions; about ~+6% speed, -0.01 ratio */ } + assert(opt[cur].price >= 0); { U32 const ll0 = (opt[cur].mlen != 0); U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0; - U32 const previousPrice = opt[cur].price; + U32 const previousPrice = (U32)opt[cur].price; U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel); - U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch); + U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch); U32 matchNb; + + ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches, + (U32)(inr-istart), (U32)(iend-inr)); + if (!nbMatches) { DEBUGLOG(7, "rPos:%u : no match found", cur); continue; @@ -979,7 +1235,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, for (mlen = lastML; mlen >= startML; mlen--) { /* scan downward */ U32 const pos = cur + mlen; - int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel); + int const price = (int)basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel); if ((pos > last_pos) || (price < opt[pos].price)) { DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)", @@ -1009,10 +1265,10 @@ _shortestPath: /* cur, last_pos, best_mlen, best_off have to be set */ * update them while traversing the sequences. */ if (lastSequence.mlen != 0) { - repcodes_t reps = ZSTD_updateRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0); - memcpy(rep, &reps, sizeof(reps)); + repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0); + ZSTD_memcpy(rep, &reps, sizeof(reps)); } else { - memcpy(rep, opt[cur].rep, sizeof(repcodes_t)); + ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t)); } { U32 const storeEnd = cur + 1; @@ -1053,7 +1309,7 @@ _shortestPath: /* cur, last_pos, best_mlen, best_off have to be set */ assert(anchor + llen <= iend); ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen); - ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen-MINMATCH); + ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen); anchor += advance; ip = anchor; } } @@ -1065,38 +1321,30 @@ _shortestPath: /* cur, last_pos, best_mlen, best_off have to be set */ return (size_t)(iend - anchor); } +static size_t ZSTD_compressBlock_opt0( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode) +{ + return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /* optLevel */, dictMode); +} + +static size_t ZSTD_compressBlock_opt2( + ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], + const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode) +{ + return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /* optLevel */, dictMode); +} size_t ZSTD_compressBlock_btopt( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { DEBUGLOG(5, "ZSTD_compressBlock_btopt"); - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict); + return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_noDict); } -/* used in 2-pass strategy */ -static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus) -{ - U32 s, sum=0; - assert(ZSTD_FREQ_DIV+bonus >= 0); - for (s=0; s<lastEltIndex+1; s++) { - table[s] <<= ZSTD_FREQ_DIV+bonus; - table[s]--; - sum += table[s]; - } - return sum; -} -/* used in 2-pass strategy */ -MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr) -{ - if (ZSTD_compressedLiterals(optPtr)) - optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0); - optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0); - optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0); - optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0); -} /* ZSTD_initStats_ultra(): * make a first compression pass, just to seed stats with more accurate starting values. @@ -1110,7 +1358,7 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms, const void* src, size_t srcSize) { U32 tmpRep[ZSTD_REP_NUM]; /* updated rep codes will sink here */ - memcpy(tmpRep, rep, sizeof(tmpRep)); + ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep)); DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize); assert(ms->opt.litLengthSum == 0); /* first block */ @@ -1118,7 +1366,7 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms, assert(ms->window.dictLimit == ms->window.lowLimit); /* no dictionary */ assert(ms->window.dictLimit - ms->nextToUpdate <= 1); /* no prefix (note: intentional overflow, defined as 2-complement) */ - ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); /* generate stats into ms->opt*/ + ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict); /* generate stats into ms->opt*/ /* invalidate first scan from history */ ZSTD_resetSeqStore(seqStore); @@ -1127,8 +1375,6 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms, ms->window.lowLimit = ms->window.dictLimit; ms->nextToUpdate = ms->window.dictLimit; - /* re-inforce weight of collected statistics */ - ZSTD_upscaleStats(&ms->opt); } size_t ZSTD_compressBlock_btultra( @@ -1136,14 +1382,14 @@ size_t ZSTD_compressBlock_btultra( const void* src, size_t srcSize) { DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize); - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); + return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict); } size_t ZSTD_compressBlock_btultra2( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { - U32 const current = (U32)((const BYTE*)src - ms->window.base); + U32 const curr = (U32)((const BYTE*)src - ms->window.base); DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize); /* 2-pass strategy: @@ -1158,41 +1404,41 @@ size_t ZSTD_compressBlock_btultra2( if ( (ms->opt.litLengthSum==0) /* first block */ && (seqStore->sequences == seqStore->sequencesStart) /* no ldm */ && (ms->window.dictLimit == ms->window.lowLimit) /* no dictionary */ - && (current == ms->window.dictLimit) /* start of frame, nothing already loaded nor skipped */ + && (curr == ms->window.dictLimit) /* start of frame, nothing already loaded nor skipped */ && (srcSize > ZSTD_PREDEF_THRESHOLD) ) { ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize); } - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); + return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict); } size_t ZSTD_compressBlock_btopt_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState); + return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState); } size_t ZSTD_compressBlock_btultra_dictMatchState( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState); + return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState); } size_t ZSTD_compressBlock_btopt_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict); + return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict); } size_t ZSTD_compressBlock_btultra_extDict( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], const void* src, size_t srcSize) { - return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict); + return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_extDict); } /* note : no btultra2 variant for extDict nor dictMatchState, diff --git a/thirdparty/zstd/compress/zstd_opt.h b/thirdparty/zstd/compress/zstd_opt.h index 9aba8a9018..627255f53d 100644 --- a/thirdparty/zstd/compress/zstd_opt.h +++ b/thirdparty/zstd/compress/zstd_opt.h @@ -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 diff --git a/thirdparty/zstd/compress/zstdmt_compress.c b/thirdparty/zstd/compress/zstdmt_compress.c index 1e3c8fdbee..6bc14b035e 100644 --- a/thirdparty/zstd/compress/zstdmt_compress.c +++ b/thirdparty/zstd/compress/zstdmt_compress.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 @@ -20,8 +20,7 @@ /* ====== Dependencies ====== */ -#include <string.h> /* memcpy, memset */ -#include <limits.h> /* INT_MAX, UINT_MAX */ +#include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset, INT_MAX, UINT_MAX */ #include "../common/mem.h" /* MEM_STATIC */ #include "../common/pool.h" /* threadpool */ #include "../common/threading.h" /* mutex */ @@ -103,14 +102,13 @@ typedef struct ZSTDMT_bufferPool_s { buffer_t bTable[1]; /* variable size */ } ZSTDMT_bufferPool; -static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem) +static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned maxNbBuffers, ZSTD_customMem cMem) { - unsigned const maxNbBuffers = 2*nbWorkers + 3; - ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc( + ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_customCalloc( sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem); if (bufPool==NULL) return NULL; if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) { - ZSTD_free(bufPool, cMem); + ZSTD_customFree(bufPool, cMem); return NULL; } bufPool->bufferSize = 64 KB; @@ -127,10 +125,10 @@ static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool) if (!bufPool) return; /* compatibility with free on NULL */ for (u=0; u<bufPool->totalBuffers; u++) { DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->bTable[u].start); - ZSTD_free(bufPool->bTable[u].start, bufPool->cMem); + ZSTD_customFree(bufPool->bTable[u].start, bufPool->cMem); } ZSTD_pthread_mutex_destroy(&bufPool->poolMutex); - ZSTD_free(bufPool, bufPool->cMem); + ZSTD_customFree(bufPool, bufPool->cMem); } /* only works at initialization, not during compression */ @@ -161,9 +159,8 @@ static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const } -static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, U32 nbWorkers) +static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, unsigned maxNbBuffers) { - unsigned const maxNbBuffers = 2*nbWorkers + 3; if (srcBufPool==NULL) return NULL; if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */ return srcBufPool; @@ -172,7 +169,7 @@ static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, size_t const bSize = srcBufPool->bufferSize; /* forward parameters */ ZSTDMT_bufferPool* newBufPool; ZSTDMT_freeBufferPool(srcBufPool); - newBufPool = ZSTDMT_createBufferPool(nbWorkers, cMem); + newBufPool = ZSTDMT_createBufferPool(maxNbBuffers, cMem); if (newBufPool==NULL) return newBufPool; ZSTDMT_setBufferSize(newBufPool, bSize); return newBufPool; @@ -201,13 +198,13 @@ static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool) } /* size conditions not respected : scratch this buffer, create new one */ DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing"); - ZSTD_free(buf.start, bufPool->cMem); + ZSTD_customFree(buf.start, bufPool->cMem); } ZSTD_pthread_mutex_unlock(&bufPool->poolMutex); /* create new buffer */ DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer"); { buffer_t buffer; - void* const start = ZSTD_malloc(bSize, bufPool->cMem); + void* const start = ZSTD_customMalloc(bSize, bufPool->cMem); buffer.start = start; /* note : start can be NULL if malloc fails ! */ buffer.capacity = (start==NULL) ? 0 : bSize; if (start==NULL) { @@ -229,13 +226,13 @@ static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer) { size_t const bSize = bufPool->bufferSize; if (buffer.capacity < bSize) { - void* const start = ZSTD_malloc(bSize, bufPool->cMem); + void* const start = ZSTD_customMalloc(bSize, bufPool->cMem); buffer_t newBuffer; newBuffer.start = start; newBuffer.capacity = start == NULL ? 0 : bSize; if (start != NULL) { assert(newBuffer.capacity >= buffer.capacity); - memcpy(newBuffer.start, buffer.start, buffer.capacity); + ZSTD_memcpy(newBuffer.start, buffer.start, buffer.capacity); DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize); return newBuffer; } @@ -261,13 +258,21 @@ static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf) ZSTD_pthread_mutex_unlock(&bufPool->poolMutex); /* Reached bufferPool capacity (should not happen) */ DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing "); - ZSTD_free(buf.start, bufPool->cMem); + ZSTD_customFree(buf.start, bufPool->cMem); } +/* We need 2 output buffers per worker since each dstBuff must be flushed after it is released. + * The 3 additional buffers are as follows: + * 1 buffer for input loading + * 1 buffer for "next input" when submitting current one + * 1 buffer stuck in queue */ +#define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) 2*nbWorkers + 3 -/* ===== Seq Pool Wrapper ====== */ +/* After a worker releases its rawSeqStore, it is immediately ready for reuse. + * So we only need one seq buffer per worker. */ +#define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) nbWorkers -static rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0}; +/* ===== Seq Pool Wrapper ====== */ typedef ZSTDMT_bufferPool ZSTDMT_seqPool; @@ -278,7 +283,7 @@ static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool) static rawSeqStore_t bufferToSeq(buffer_t buffer) { - rawSeqStore_t seq = {NULL, 0, 0, 0}; + rawSeqStore_t seq = kNullRawSeqStore; seq.seq = (rawSeq*)buffer.start; seq.capacity = buffer.capacity / sizeof(rawSeq); return seq; @@ -319,7 +324,7 @@ static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq) static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem) { - ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(nbWorkers, cMem); + ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(SEQ_POOL_MAX_NB_BUFFERS(nbWorkers), cMem); if (seqPool == NULL) return NULL; ZSTDMT_setNbSeq(seqPool, 0); return seqPool; @@ -332,7 +337,7 @@ static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool) static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers) { - return ZSTDMT_expandBufferPool(pool, nbWorkers); + return ZSTDMT_expandBufferPool(pool, SEQ_POOL_MAX_NB_BUFFERS(nbWorkers)); } @@ -354,7 +359,7 @@ static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool) for (cid=0; cid<pool->totalCCtx; cid++) ZSTD_freeCCtx(pool->cctx[cid]); /* note : compatible with free on NULL */ ZSTD_pthread_mutex_destroy(&pool->poolMutex); - ZSTD_free(pool, pool->cMem); + ZSTD_customFree(pool, pool->cMem); } /* ZSTDMT_createCCtxPool() : @@ -362,12 +367,12 @@ static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool) static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers, ZSTD_customMem cMem) { - ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc( + ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_customCalloc( sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem); assert(nbWorkers > 0); if (!cctxPool) return NULL; if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) { - ZSTD_free(cctxPool, cMem); + ZSTD_customFree(cctxPool, cMem); return NULL; } cctxPool->cMem = cMem; @@ -470,54 +475,52 @@ ZSTDMT_serialState_reset(serialState_t* serialState, ZSTD_dictContentType_e dictContentType) { /* Adjust parameters */ - if (params.ldmParams.enableLdm) { + if (params.ldmParams.enableLdm == ZSTD_ps_enable) { DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10); ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams); assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog); assert(params.ldmParams.hashRateLog < 32); - serialState->ldmState.hashPower = - ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength); } else { - memset(¶ms.ldmParams, 0, sizeof(params.ldmParams)); + ZSTD_memset(¶ms.ldmParams, 0, sizeof(params.ldmParams)); } serialState->nextJobID = 0; if (params.fParams.checksumFlag) XXH64_reset(&serialState->xxhState, 0); - if (params.ldmParams.enableLdm) { + if (params.ldmParams.enableLdm == ZSTD_ps_enable) { ZSTD_customMem cMem = params.customMem; unsigned const hashLog = params.ldmParams.hashLog; size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t); unsigned const bucketLog = params.ldmParams.hashLog - params.ldmParams.bucketSizeLog; - size_t const bucketSize = (size_t)1 << bucketLog; unsigned const prevBucketLog = serialState->params.ldmParams.hashLog - serialState->params.ldmParams.bucketSizeLog; + size_t const numBuckets = (size_t)1 << bucketLog; /* Size the seq pool tables */ ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize)); /* Reset the window */ ZSTD_window_init(&serialState->ldmState.window); /* Resize tables and output space if necessary. */ if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) { - ZSTD_free(serialState->ldmState.hashTable, cMem); - serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_malloc(hashSize, cMem); + ZSTD_customFree(serialState->ldmState.hashTable, cMem); + serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_customMalloc(hashSize, cMem); } if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) { - ZSTD_free(serialState->ldmState.bucketOffsets, cMem); - serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_malloc(bucketSize, cMem); + ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem); + serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(numBuckets, cMem); } if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets) return 1; /* Zero the tables */ - memset(serialState->ldmState.hashTable, 0, hashSize); - memset(serialState->ldmState.bucketOffsets, 0, bucketSize); + ZSTD_memset(serialState->ldmState.hashTable, 0, hashSize); + ZSTD_memset(serialState->ldmState.bucketOffsets, 0, numBuckets); /* Update window state and fill hash table with dict */ serialState->ldmState.loadedDictEnd = 0; if (dictSize > 0) { if (dictContentType == ZSTD_dct_rawContent) { BYTE const* const dictEnd = (const BYTE*)dict + dictSize; - ZSTD_window_update(&serialState->ldmState.window, dict, dictSize); + ZSTD_window_update(&serialState->ldmState.window, dict, dictSize, /* forceNonContiguous */ 0); ZSTD_ldm_fillHashTable(&serialState->ldmState, (const BYTE*)dict, dictEnd, ¶ms.ldmParams); serialState->ldmState.loadedDictEnd = params.forceWindow ? 0 : (U32)(dictEnd - serialState->ldmState.window.base); } else { @@ -537,7 +540,7 @@ ZSTDMT_serialState_reset(serialState_t* serialState, static int ZSTDMT_serialState_init(serialState_t* serialState) { int initError = 0; - memset(serialState, 0, sizeof(*serialState)); + ZSTD_memset(serialState, 0, sizeof(*serialState)); initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL); initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL); initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL); @@ -552,8 +555,8 @@ static void ZSTDMT_serialState_free(serialState_t* serialState) ZSTD_pthread_cond_destroy(&serialState->cond); ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex); ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond); - ZSTD_free(serialState->ldmState.hashTable, cMem); - ZSTD_free(serialState->ldmState.bucketOffsets, cMem); + ZSTD_customFree(serialState->ldmState.hashTable, cMem); + ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem); } static void ZSTDMT_serialState_update(serialState_t* serialState, @@ -569,12 +572,12 @@ static void ZSTDMT_serialState_update(serialState_t* serialState, /* A future job may error and skip our job */ if (serialState->nextJobID == jobID) { /* It is now our turn, do any processing necessary */ - if (serialState->params.ldmParams.enableLdm) { + if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) { size_t error; assert(seqStore.seq != NULL && seqStore.pos == 0 && seqStore.size == 0 && seqStore.capacity > 0); assert(src.size <= serialState->params.jobSize); - ZSTD_window_update(&serialState->ldmState.window, src.start, src.size); + ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0); error = ZSTD_ldm_generateSequences( &serialState->ldmState, &seqStore, &serialState->params.ldmParams, src.start, src.size); @@ -599,7 +602,7 @@ static void ZSTDMT_serialState_update(serialState_t* serialState, if (seqStore.size > 0) { size_t const err = ZSTD_referenceExternalSequences( jobCCtx, seqStore.seq, seqStore.size); - assert(serialState->params.ldmParams.enableLdm); + assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable); assert(!ZSTD_isError(err)); (void)err; } @@ -677,7 +680,7 @@ static void ZSTDMT_compressionJob(void* jobDescription) if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation)); job->dstBuff = dstBuff; /* this value can be read in ZSTDMT_flush, when it copies the whole job */ } - if (jobParams.ldmParams.enableLdm && rawSeqStore.seq == NULL) + if (jobParams.ldmParams.enableLdm == ZSTD_ps_enable && rawSeqStore.seq == NULL) JOB_ERROR(ERROR(memory_allocation)); /* Don't compute the checksum for chunks, since we compute it externally, @@ -685,7 +688,9 @@ static void ZSTDMT_compressionJob(void* jobDescription) */ if (job->jobID != 0) jobParams.fParams.checksumFlag = 0; /* Don't run LDM for the chunks, since we handle it externally */ - jobParams.ldmParams.enableLdm = 0; + jobParams.ldmParams.enableLdm = ZSTD_ps_disable; + /* Correct nbWorkers to 0. */ + jobParams.nbWorkers = 0; /* init */ @@ -698,6 +703,10 @@ static void ZSTDMT_compressionJob(void* jobDescription) { size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob); if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError); } + if (!job->firstJob) { + size_t const err = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_deterministicRefPrefix, 0); + if (ZSTD_isError(err)) JOB_ERROR(err); + } { size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */ ZSTD_dtlm_fast, @@ -753,6 +762,13 @@ static void ZSTDMT_compressionJob(void* jobDescription) if (ZSTD_isError(cSize)) JOB_ERROR(cSize); lastCBlockSize = cSize; } } + if (!job->firstJob) { + /* Double check that we don't have an ext-dict, because then our + * repcode invalidation doesn't work. + */ + assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window)); + } + ZSTD_CCtx_trace(cctx, 0); _endJob: ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize); @@ -799,6 +815,15 @@ typedef struct { static const roundBuff_t kNullRoundBuff = {NULL, 0, 0}; #define RSYNC_LENGTH 32 +/* Don't create chunks smaller than the zstd block size. + * This stops us from regressing compression ratio too much, + * and ensures our output fits in ZSTD_compressBound(). + * + * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then + * ZSTD_COMPRESSBOUND() will need to be updated. + */ +#define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX +#define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG) typedef struct { U64 hash; @@ -820,7 +845,6 @@ struct ZSTDMT_CCtx_s { roundBuff_t roundBuff; serialState_t serial; rsyncState_t rsync; - unsigned singleBlockingThread; unsigned jobIDMask; unsigned doneJobID; unsigned nextJobID; @@ -832,6 +856,7 @@ struct ZSTDMT_CCtx_s { ZSTD_customMem cMem; ZSTD_CDict* cdictLocal; const ZSTD_CDict* cdict; + unsigned providedFactory: 1; }; static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem) @@ -842,7 +867,7 @@ static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZS ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex); ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond); } - ZSTD_free(jobTable, cMem); + ZSTD_customFree(jobTable, cMem); } /* ZSTDMT_allocJobsTable() @@ -854,7 +879,7 @@ static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_custom U32 const nbJobs = 1 << nbJobsLog2; U32 jobNb; ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*) - ZSTD_calloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem); + ZSTD_customCalloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem); int initError = 0; if (jobTable==NULL) return NULL; *nbJobsPtr = nbJobs; @@ -885,12 +910,12 @@ static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) { /* ZSTDMT_CCtxParam_setNbWorkers(): * Internal use only */ -size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers) +static size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers) { return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers); } -MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem) +MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool) { ZSTDMT_CCtx* mtctx; U32 nbJobs = nbWorkers + 2; @@ -903,16 +928,23 @@ MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, /* invalid custom allocator */ return NULL; - mtctx = (ZSTDMT_CCtx*) ZSTD_calloc(sizeof(ZSTDMT_CCtx), cMem); + mtctx = (ZSTDMT_CCtx*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtx), cMem); if (!mtctx) return NULL; ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers); mtctx->cMem = cMem; mtctx->allJobsCompleted = 1; - mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem); + if (pool != NULL) { + mtctx->factory = pool; + mtctx->providedFactory = 1; + } + else { + mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem); + mtctx->providedFactory = 0; + } mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem); assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */ mtctx->jobIDMask = nbJobs - 1; - mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem); + mtctx->bufPool = ZSTDMT_createBufferPool(BUF_POOL_MAX_NB_BUFFERS(nbWorkers), cMem); mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem); mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem); initError = ZSTDMT_serialState_init(&mtctx->serial); @@ -925,22 +957,18 @@ MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, return mtctx; } -ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem) +ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool) { #ifdef ZSTD_MULTITHREAD - return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem); + return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem, pool); #else (void)nbWorkers; (void)cMem; + (void)pool; return NULL; #endif } -ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers) -{ - return ZSTDMT_createCCtx_advanced(nbWorkers, ZSTD_defaultCMem); -} - /* ZSTDMT_releaseAllJobResources() : * note : ensure all workers are killed first ! */ @@ -957,7 +985,7 @@ static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx) ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff); /* Clear the job description, but keep the mutex/cond */ - memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID])); + ZSTD_memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID])); mtctx->jobs[jobID].job_mutex = mutex; mtctx->jobs[jobID].job_cond = cond; } @@ -984,7 +1012,8 @@ static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx) size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx) { if (mtctx==NULL) return 0; /* compatible with free on NULL */ - POOL_free(mtctx->factory); /* stop and free worker threads */ + if (!mtctx->providedFactory) + POOL_free(mtctx->factory); /* stop and free worker threads */ ZSTDMT_releaseAllJobResources(mtctx); /* release job resources into pools first */ ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem); ZSTDMT_freeBufferPool(mtctx->bufPool); @@ -993,8 +1022,8 @@ size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx) ZSTDMT_serialState_free(&mtctx->serial); ZSTD_freeCDict(mtctx->cdictLocal); if (mtctx->roundBuff.buffer) - ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem); - ZSTD_free(mtctx, mtctx->cMem); + ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem); + ZSTD_customFree(mtctx, mtctx->cMem); return 0; } @@ -1011,65 +1040,6 @@ size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx) + mtctx->roundBuff.capacity; } -/* Internal only */ -size_t -ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params, - ZSTDMT_parameter parameter, - int value) -{ - DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter"); - switch(parameter) - { - case ZSTDMT_p_jobSize : - DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter : set jobSize to %i", value); - return ZSTD_CCtxParams_setParameter(params, ZSTD_c_jobSize, value); - case ZSTDMT_p_overlapLog : - DEBUGLOG(4, "ZSTDMT_p_overlapLog : %i", value); - return ZSTD_CCtxParams_setParameter(params, ZSTD_c_overlapLog, value); - case ZSTDMT_p_rsyncable : - DEBUGLOG(4, "ZSTD_p_rsyncable : %i", value); - return ZSTD_CCtxParams_setParameter(params, ZSTD_c_rsyncable, value); - default : - return ERROR(parameter_unsupported); - } -} - -size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int value) -{ - DEBUGLOG(4, "ZSTDMT_setMTCtxParameter"); - return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value); -} - -size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int* value) -{ - switch (parameter) { - case ZSTDMT_p_jobSize: - return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_jobSize, value); - case ZSTDMT_p_overlapLog: - return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_overlapLog, value); - case ZSTDMT_p_rsyncable: - return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_rsyncable, value); - default: - return ERROR(parameter_unsupported); - } -} - -/* Sets parameters relevant to the compression job, - * initializing others to default values. */ -static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(const ZSTD_CCtx_params* params) -{ - ZSTD_CCtx_params jobParams = *params; - /* Clear parameters related to multithreading */ - jobParams.forceWindow = 0; - jobParams.nbWorkers = 0; - jobParams.jobSize = 0; - jobParams.overlapLog = 0; - jobParams.rsyncable = 0; - memset(&jobParams.ldmParams, 0, sizeof(ldmParams_t)); - memset(&jobParams.customMem, 0, sizeof(ZSTD_customMem)); - return jobParams; -} - /* ZSTDMT_resize() : * @return : error code if fails, 0 on success */ @@ -1077,7 +1047,7 @@ static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers) { if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation); FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) , ""); - mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, nbWorkers); + mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, BUF_POOL_MAX_NB_BUFFERS(nbWorkers)); if (mtctx->bufPool == NULL) return ERROR(memory_allocation); mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers); if (mtctx->cctxPool == NULL) return ERROR(memory_allocation); @@ -1098,7 +1068,7 @@ void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_p DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)", compressionLevel); mtctx->params.compressionLevel = compressionLevel; - { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0); + { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict); cParams.windowLog = saved_wlog; mtctx->params.cParams = cParams; } @@ -1182,11 +1152,11 @@ size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx) static unsigned ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params* params) { unsigned jobLog; - if (params->ldmParams.enableLdm) { + if (params->ldmParams.enableLdm == ZSTD_ps_enable) { /* In Long Range Mode, the windowLog is typically oversized. * In which case, it's preferable to determine the jobSize - * based on chainLog instead. */ - jobLog = MAX(21, params->cParams.chainLog + 4); + * based on cycleLog instead. */ + jobLog = MAX(21, ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy) + 3); } else { jobLog = MAX(20, params->cParams.windowLog + 2); } @@ -1226,7 +1196,7 @@ static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params) int const overlapRLog = 9 - ZSTDMT_overlapLog(params->overlapLog, params->cParams.strategy); int ovLog = (overlapRLog >= 8) ? 0 : (params->cParams.windowLog - overlapRLog); assert(0 <= overlapRLog && overlapRLog <= 8); - if (params->ldmParams.enableLdm) { + if (params->ldmParams.enableLdm == ZSTD_ps_enable) { /* In Long Range Mode, the windowLog is typically oversized. * In which case, it's preferable to determine the jobSize * based on chainLog instead. @@ -1240,174 +1210,6 @@ static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params) return (ovLog==0) ? 0 : (size_t)1 << ovLog; } -static unsigned -ZSTDMT_computeNbJobs(const ZSTD_CCtx_params* params, size_t srcSize, unsigned nbWorkers) -{ - assert(nbWorkers>0); - { size_t const jobSizeTarget = (size_t)1 << ZSTDMT_computeTargetJobLog(params); - size_t const jobMaxSize = jobSizeTarget << 2; - size_t const passSizeMax = jobMaxSize * nbWorkers; - unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1; - unsigned const nbJobsLarge = multiplier * nbWorkers; - unsigned const nbJobsMax = (unsigned)(srcSize / jobSizeTarget) + 1; - unsigned const nbJobsSmall = MIN(nbJobsMax, nbWorkers); - return (multiplier>1) ? nbJobsLarge : nbJobsSmall; -} } - -/* ZSTDMT_compress_advanced_internal() : - * This is a blocking function : it will only give back control to caller after finishing its compression job. - */ -static size_t -ZSTDMT_compress_advanced_internal( - ZSTDMT_CCtx* mtctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - const ZSTD_CDict* cdict, - ZSTD_CCtx_params params) -{ - ZSTD_CCtx_params const jobParams = ZSTDMT_initJobCCtxParams(¶ms); - size_t const overlapSize = ZSTDMT_computeOverlapSize(¶ms); - unsigned const nbJobs = ZSTDMT_computeNbJobs(¶ms, srcSize, params.nbWorkers); - size_t const proposedJobSize = (srcSize + (nbJobs-1)) / nbJobs; - size_t const avgJobSize = (((proposedJobSize-1) & 0x1FFFF) < 0x7FFF) ? proposedJobSize + 0xFFFF : proposedJobSize; /* avoid too small last block */ - const char* const srcStart = (const char*)src; - size_t remainingSrcSize = srcSize; - unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbJobs : (unsigned)(dstCapacity / ZSTD_compressBound(avgJobSize)); /* presumes avgJobSize >= 256 KB, which should be the case */ - size_t frameStartPos = 0, dstBufferPos = 0; - assert(jobParams.nbWorkers == 0); - assert(mtctx->cctxPool->totalCCtx == params.nbWorkers); - - params.jobSize = (U32)avgJobSize; - DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: nbJobs=%2u (rawSize=%u bytes; fixedSize=%u) ", - nbJobs, (U32)proposedJobSize, (U32)avgJobSize); - - if ((nbJobs==1) | (params.nbWorkers<=1)) { /* fallback to single-thread mode : this is a blocking invocation anyway */ - ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0]; - DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: fallback to single-thread mode"); - if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, jobParams.fParams); - return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, NULL, 0, &jobParams); - } - - assert(avgJobSize >= 256 KB); /* condition for ZSTD_compressBound(A) + ZSTD_compressBound(B) <= ZSTD_compressBound(A+B), required to compress directly into Dst (no additional buffer) */ - ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(avgJobSize) ); - /* LDM doesn't even try to load the dictionary in single-ingestion mode */ - if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, avgJobSize, NULL, 0, ZSTD_dct_auto)) - return ERROR(memory_allocation); - - FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbJobs) , ""); /* only expands if necessary */ - - { unsigned u; - for (u=0; u<nbJobs; u++) { - size_t const jobSize = MIN(remainingSrcSize, avgJobSize); - size_t const dstBufferCapacity = ZSTD_compressBound(jobSize); - buffer_t const dstAsBuffer = { (char*)dst + dstBufferPos, dstBufferCapacity }; - buffer_t const dstBuffer = u < compressWithinDst ? dstAsBuffer : g_nullBuffer; - size_t dictSize = u ? overlapSize : 0; - - mtctx->jobs[u].prefix.start = srcStart + frameStartPos - dictSize; - mtctx->jobs[u].prefix.size = dictSize; - mtctx->jobs[u].src.start = srcStart + frameStartPos; - mtctx->jobs[u].src.size = jobSize; assert(jobSize > 0); /* avoid job.src.size == 0 */ - mtctx->jobs[u].consumed = 0; - mtctx->jobs[u].cSize = 0; - mtctx->jobs[u].cdict = (u==0) ? cdict : NULL; - mtctx->jobs[u].fullFrameSize = srcSize; - mtctx->jobs[u].params = jobParams; - /* do not calculate checksum within sections, but write it in header for first section */ - mtctx->jobs[u].dstBuff = dstBuffer; - mtctx->jobs[u].cctxPool = mtctx->cctxPool; - mtctx->jobs[u].bufPool = mtctx->bufPool; - mtctx->jobs[u].seqPool = mtctx->seqPool; - mtctx->jobs[u].serial = &mtctx->serial; - mtctx->jobs[u].jobID = u; - mtctx->jobs[u].firstJob = (u==0); - mtctx->jobs[u].lastJob = (u==nbJobs-1); - - DEBUGLOG(5, "ZSTDMT_compress_advanced_internal: posting job %u (%u bytes)", u, (U32)jobSize); - DEBUG_PRINTHEX(6, mtctx->jobs[u].prefix.start, 12); - POOL_add(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[u]); - - frameStartPos += jobSize; - dstBufferPos += dstBufferCapacity; - remainingSrcSize -= jobSize; - } } - - /* collect result */ - { size_t error = 0, dstPos = 0; - unsigned jobID; - for (jobID=0; jobID<nbJobs; jobID++) { - DEBUGLOG(5, "waiting for job %u ", jobID); - ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex); - while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) { - DEBUGLOG(5, "waiting for jobCompleted signal from job %u", jobID); - ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex); - } - ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex); - DEBUGLOG(5, "ready to write job %u ", jobID); - - { size_t const cSize = mtctx->jobs[jobID].cSize; - if (ZSTD_isError(cSize)) error = cSize; - if ((!error) && (dstPos + cSize > dstCapacity)) error = ERROR(dstSize_tooSmall); - if (jobID) { /* note : job 0 is written directly at dst, which is correct position */ - if (!error) - memmove((char*)dst + dstPos, mtctx->jobs[jobID].dstBuff.start, cSize); /* may overlap when job compressed within dst */ - if (jobID >= compressWithinDst) { /* job compressed into its own buffer, which must be released */ - DEBUGLOG(5, "releasing buffer %u>=%u", jobID, compressWithinDst); - ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff); - } } - mtctx->jobs[jobID].dstBuff = g_nullBuffer; - mtctx->jobs[jobID].cSize = 0; - dstPos += cSize ; - } - } /* for (jobID=0; jobID<nbJobs; jobID++) */ - - DEBUGLOG(4, "checksumFlag : %u ", params.fParams.checksumFlag); - if (params.fParams.checksumFlag) { - U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState); - if (dstPos + 4 > dstCapacity) { - error = ERROR(dstSize_tooSmall); - } else { - DEBUGLOG(4, "writing checksum : %08X \n", checksum); - MEM_writeLE32((char*)dst + dstPos, checksum); - dstPos += 4; - } } - - if (!error) DEBUGLOG(4, "compressed size : %u ", (U32)dstPos); - return error ? error : dstPos; - } -} - -size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - const ZSTD_CDict* cdict, - ZSTD_parameters params, - int overlapLog) -{ - ZSTD_CCtx_params cctxParams = mtctx->params; - cctxParams.cParams = params.cParams; - cctxParams.fParams = params.fParams; - assert(ZSTD_OVERLAPLOG_MIN <= overlapLog && overlapLog <= ZSTD_OVERLAPLOG_MAX); - cctxParams.overlapLog = overlapLog; - return ZSTDMT_compress_advanced_internal(mtctx, - dst, dstCapacity, - src, srcSize, - cdict, cctxParams); -} - - -size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - int compressionLevel) -{ - ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0); - int const overlapLog = ZSTDMT_overlapLog_default(params.cParams.strategy); - params.fParams.contentSizeFlag = 1; - return ZSTDMT_compress_advanced(mtctx, dst, dstCapacity, src, srcSize, NULL, params, overlapLog); -} - - /* ====================================== */ /* ======= Streaming API ======= */ /* ====================================== */ @@ -1432,16 +1234,6 @@ size_t ZSTDMT_initCStream_internal( if (params.jobSize != 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN; if (params.jobSize > (size_t)ZSTDMT_JOBSIZE_MAX) params.jobSize = (size_t)ZSTDMT_JOBSIZE_MAX; - mtctx->singleBlockingThread = (pledgedSrcSize <= ZSTDMT_JOBSIZE_MIN); /* do not trigger multi-threading when srcSize is too small */ - if (mtctx->singleBlockingThread) { - ZSTD_CCtx_params const singleThreadParams = ZSTDMT_initJobCCtxParams(¶ms); - DEBUGLOG(5, "ZSTDMT_initCStream_internal: switch to single blocking thread mode"); - assert(singleThreadParams.nbWorkers == 0); - return ZSTD_initCStream_internal(mtctx->cctxPool->cctx[0], - dict, dictSize, cdict, - &singleThreadParams, pledgedSrcSize); - } - DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers); if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */ @@ -1475,9 +1267,11 @@ size_t ZSTDMT_initCStream_internal( if (params.rsyncable) { /* Aim for the targetsectionSize as the average job size. */ - U32 const jobSizeMB = (U32)(mtctx->targetSectionSize >> 20); - U32 const rsyncBits = ZSTD_highbit32(jobSizeMB) + 20; - assert(jobSizeMB >= 1); + U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10); + U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10); + /* We refuse to create jobs < RSYNC_MIN_BLOCK_SIZE bytes, so make sure our + * expected job size is at least 4x larger. */ + assert(rsyncBits >= RSYNC_MIN_BLOCK_LOG + 2); DEBUGLOG(4, "rsyncLog = %u", rsyncBits); mtctx->rsync.hash = 0; mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1; @@ -1489,7 +1283,7 @@ size_t ZSTDMT_initCStream_internal( ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize)); { /* If ldm is enabled we need windowSize space. */ - size_t const windowSize = mtctx->params.ldmParams.enableLdm ? (1U << mtctx->params.cParams.windowLog) : 0; + size_t const windowSize = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable ? (1U << mtctx->params.cParams.windowLog) : 0; /* Two buffers of slack, plus extra space for the overlap * This is the minimum slack that LDM works with. One extra because * flush might waste up to targetSectionSize-1 bytes. Another extra @@ -1504,8 +1298,8 @@ size_t ZSTDMT_initCStream_internal( size_t const capacity = MAX(windowSize, sectionsSize) + slackSize; if (mtctx->roundBuff.capacity < capacity) { if (mtctx->roundBuff.buffer) - ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem); - mtctx->roundBuff.buffer = (BYTE*)ZSTD_malloc(capacity, mtctx->cMem); + ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem); + mtctx->roundBuff.buffer = (BYTE*)ZSTD_customMalloc(capacity, mtctx->cMem); if (mtctx->roundBuff.buffer == NULL) { mtctx->roundBuff.capacity = 0; return ERROR(memory_allocation); @@ -1530,53 +1324,6 @@ size_t ZSTDMT_initCStream_internal( return 0; } -size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx, - const void* dict, size_t dictSize, - ZSTD_parameters params, - unsigned long long pledgedSrcSize) -{ - ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */ - DEBUGLOG(4, "ZSTDMT_initCStream_advanced (pledgedSrcSize=%u)", (U32)pledgedSrcSize); - cctxParams.cParams = params.cParams; - cctxParams.fParams = params.fParams; - return ZSTDMT_initCStream_internal(mtctx, dict, dictSize, ZSTD_dct_auto, NULL, - cctxParams, pledgedSrcSize); -} - -size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx, - const ZSTD_CDict* cdict, - ZSTD_frameParameters fParams, - unsigned long long pledgedSrcSize) -{ - ZSTD_CCtx_params cctxParams = mtctx->params; - if (cdict==NULL) return ERROR(dictionary_wrong); /* method incompatible with NULL cdict */ - cctxParams.cParams = ZSTD_getCParamsFromCDict(cdict); - cctxParams.fParams = fParams; - return ZSTDMT_initCStream_internal(mtctx, NULL, 0 /*dictSize*/, ZSTD_dct_auto, cdict, - cctxParams, pledgedSrcSize); -} - - -/* ZSTDMT_resetCStream() : - * pledgedSrcSize can be zero == unknown (for the time being) - * prefer using ZSTD_CONTENTSIZE_UNKNOWN, - * as `0` might mean "empty" in the future */ -size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize) -{ - if (!pledgedSrcSize) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN; - return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, 0, mtctx->params, - pledgedSrcSize); -} - -size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel) { - ZSTD_parameters const params = ZSTD_getParams(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0); - ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */ - DEBUGLOG(4, "ZSTDMT_initCStream (cLevel=%i)", compressionLevel); - cctxParams.cParams = params.cParams; - cctxParams.fParams = params.fParams; - return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, NULL, cctxParams, ZSTD_CONTENTSIZE_UNKNOWN); -} - /* ZSTDMT_writeLastEmptyBlock() * Write a single empty block with an end-of-frame to finish a frame. @@ -1740,7 +1487,7 @@ static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, u assert(cSize >= mtctx->jobs[wJobID].dstFlushed); assert(mtctx->jobs[wJobID].dstBuff.start != NULL); if (toFlush > 0) { - memcpy((char*)output->dst + output->pos, + ZSTD_memcpy((char*)output->dst + output->pos, (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed, toFlush); } @@ -1811,17 +1558,21 @@ static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx) static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range) { BYTE const* const bufferStart = (BYTE const*)buffer.start; - BYTE const* const bufferEnd = bufferStart + buffer.capacity; BYTE const* const rangeStart = (BYTE const*)range.start; - BYTE const* const rangeEnd = range.size != 0 ? rangeStart + range.size : rangeStart; if (rangeStart == NULL || bufferStart == NULL) return 0; - /* Empty ranges cannot overlap */ - if (bufferStart == bufferEnd || rangeStart == rangeEnd) - return 0; - return bufferStart < rangeEnd && rangeStart < bufferEnd; + { + BYTE const* const bufferEnd = bufferStart + buffer.capacity; + BYTE const* const rangeEnd = rangeStart + range.size; + + /* Empty ranges cannot overlap */ + if (bufferStart == bufferEnd || rangeStart == rangeEnd) + return 0; + + return bufferStart < rangeEnd && rangeStart < bufferEnd; + } } static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window) @@ -1848,7 +1599,7 @@ static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window) static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer) { - if (mtctx->params.ldmParams.enableLdm) { + if (mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable) { ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex; DEBUGLOG(5, "ZSTDMT_waitForLdmComplete"); DEBUGLOG(5, "source [0x%zx, 0x%zx)", @@ -1894,7 +1645,7 @@ static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx) return 0; } ZSTDMT_waitForLdmComplete(mtctx, buffer); - memmove(start, mtctx->inBuff.prefix.start, prefixSize); + ZSTD_memmove(start, mtctx->inBuff.prefix.start, prefixSize); mtctx->inBuff.prefix.start = start; mtctx->roundBuff.pos = prefixSize; } @@ -1951,6 +1702,11 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input) if (!mtctx->params.rsyncable) /* Rsync is disabled. */ return syncPoint; + if (mtctx->inBuff.filled + input.size - input.pos < RSYNC_MIN_BLOCK_SIZE) + /* We don't emit synchronization points if it would produce too small blocks. + * We don't have enough input to find a synchronization point, so don't look. + */ + return syncPoint; if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH) /* Not enough to compute the hash. * We will miss any synchronization points in this RSYNC_LENGTH byte @@ -1961,23 +1717,41 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input) */ return syncPoint; /* Initialize the loop variables. */ - if (mtctx->inBuff.filled >= RSYNC_LENGTH) { - /* We have enough bytes buffered to initialize the hash. + if (mtctx->inBuff.filled < RSYNC_MIN_BLOCK_SIZE) { + /* We don't need to scan the first RSYNC_MIN_BLOCK_SIZE positions + * because they can't possibly be a sync point. So we can start + * part way through the input buffer. + */ + pos = RSYNC_MIN_BLOCK_SIZE - mtctx->inBuff.filled; + if (pos >= RSYNC_LENGTH) { + prev = istart + pos - RSYNC_LENGTH; + hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH); + } else { + assert(mtctx->inBuff.filled >= RSYNC_LENGTH); + prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH; + hash = ZSTD_rollingHash_compute(prev + pos, (RSYNC_LENGTH - pos)); + hash = ZSTD_rollingHash_append(hash, istart, pos); + } + } else { + /* We have enough bytes buffered to initialize the hash, + * and are have processed enough bytes to find a sync point. * Start scanning at the beginning of the input. */ + assert(mtctx->inBuff.filled >= RSYNC_MIN_BLOCK_SIZE); + assert(RSYNC_MIN_BLOCK_SIZE >= RSYNC_LENGTH); pos = 0; prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH; hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH); - } else { - /* We don't have enough bytes buffered to initialize the hash, but - * we know we have at least RSYNC_LENGTH bytes total. - * Start scanning after the first RSYNC_LENGTH bytes less the bytes - * already buffered. - */ - pos = RSYNC_LENGTH - mtctx->inBuff.filled; - prev = (BYTE const*)mtctx->inBuff.buffer.start - pos; - hash = ZSTD_rollingHash_compute(mtctx->inBuff.buffer.start, mtctx->inBuff.filled); - hash = ZSTD_rollingHash_append(hash, istart, pos); + if ((hash & hitMask) == hitMask) { + /* We're already at a sync point so don't load any more until + * we're able to flush this sync point. + * This likely happened because the job table was full so we + * couldn't add our job. + */ + syncPoint.toLoad = 0; + syncPoint.flush = 1; + return syncPoint; + } } /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll * through the input. If we hit a synchronization point, then cut the @@ -1989,8 +1763,9 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input) */ for (; pos < syncPoint.toLoad; ++pos) { BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH]; - /* if (pos >= RSYNC_LENGTH) assert(ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash); */ + assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash); hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower); + assert(mtctx->inBuff.filled + pos >= RSYNC_MIN_BLOCK_SIZE); if ((hash & hitMask) == hitMask) { syncPoint.toLoad = pos + 1; syncPoint.flush = 1; @@ -2022,34 +1797,11 @@ size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, assert(output->pos <= output->size); assert(input->pos <= input->size); - if (mtctx->singleBlockingThread) { /* delegate to single-thread (synchronous) */ - return ZSTD_compressStream2(mtctx->cctxPool->cctx[0], output, input, endOp); - } - if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) { /* current frame being ended. Only flush/end are allowed */ return ERROR(stage_wrong); } - /* single-pass shortcut (note : synchronous-mode) */ - if ( (!mtctx->params.rsyncable) /* rsyncable mode is disabled */ - && (mtctx->nextJobID == 0) /* just started */ - && (mtctx->inBuff.filled == 0) /* nothing buffered */ - && (!mtctx->jobReady) /* no job already created */ - && (endOp == ZSTD_e_end) /* end order */ - && (output->size - output->pos >= ZSTD_compressBound(input->size - input->pos)) ) { /* enough space in dst */ - size_t const cSize = ZSTDMT_compress_advanced_internal(mtctx, - (char*)output->dst + output->pos, output->size - output->pos, - (const char*)input->src + input->pos, input->size - input->pos, - mtctx->cdict, mtctx->params); - if (ZSTD_isError(cSize)) return cSize; - input->pos = input->size; - output->pos += cSize; - mtctx->allJobsCompleted = 1; - mtctx->frameEnded = 1; - return 0; - } - /* fill input buffer */ if ( (!mtctx->jobReady) && (input->size > input->pos) ) { /* support NULL input */ @@ -2072,13 +1824,21 @@ size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize); DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u", (U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize); - memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad); + ZSTD_memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad); input->pos += syncPoint.toLoad; mtctx->inBuff.filled += syncPoint.toLoad; forwardInputProgress = syncPoint.toLoad>0; } - if ((input->pos < input->size) && (endOp == ZSTD_e_end)) - endOp = ZSTD_e_flush; /* can't end now : not all input consumed */ + } + if ((input->pos < input->size) && (endOp == ZSTD_e_end)) { + /* Can't end yet because the input is not fully consumed. + * We are in one of these cases: + * - mtctx->inBuff is NULL & empty: we couldn't get an input buffer so don't create a new job. + * - We filled the input buffer: flush this job but don't end the frame. + * - We hit a synchronization point: flush this job but don't end the frame. + */ + assert(mtctx->inBuff.filled == 0 || mtctx->inBuff.filled == mtctx->targetSectionSize || mtctx->params.rsyncable); + endOp = ZSTD_e_flush; } if ( (mtctx->jobReady) @@ -2097,47 +1857,3 @@ size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, return remainingToFlush; } } - - -size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input) -{ - FORWARD_IF_ERROR( ZSTDMT_compressStream_generic(mtctx, output, input, ZSTD_e_continue) , ""); - - /* recommended next input size : fill current input buffer */ - return mtctx->targetSectionSize - mtctx->inBuff.filled; /* note : could be zero when input buffer is fully filled and no more availability to create new job */ -} - - -static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_EndDirective endFrame) -{ - size_t const srcSize = mtctx->inBuff.filled; - DEBUGLOG(5, "ZSTDMT_flushStream_internal"); - - if ( mtctx->jobReady /* one job ready for a worker to pick up */ - || (srcSize > 0) /* still some data within input buffer */ - || ((endFrame==ZSTD_e_end) && !mtctx->frameEnded)) { /* need a last 0-size block to end frame */ - DEBUGLOG(5, "ZSTDMT_flushStream_internal : create a new job (%u bytes, end:%u)", - (U32)srcSize, (U32)endFrame); - FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, srcSize, endFrame) , ""); - } - - /* check if there is any data available to flush */ - return ZSTDMT_flushProduced(mtctx, output, 1 /* blockToFlush */, endFrame); -} - - -size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output) -{ - DEBUGLOG(5, "ZSTDMT_flushStream"); - if (mtctx->singleBlockingThread) - return ZSTD_flushStream(mtctx->cctxPool->cctx[0], output); - return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_flush); -} - -size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output) -{ - DEBUGLOG(4, "ZSTDMT_endStream"); - if (mtctx->singleBlockingThread) - return ZSTD_endStream(mtctx->cctxPool->cctx[0], output); - return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_end); -} diff --git a/thirdparty/zstd/compress/zstdmt_compress.h b/thirdparty/zstd/compress/zstdmt_compress.h index 89914eb7f8..271eb1ac71 100644 --- a/thirdparty/zstd/compress/zstdmt_compress.h +++ b/thirdparty/zstd/compress/zstdmt_compress.h @@ -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 @@ -19,113 +19,60 @@ /* Note : This is an internal API. * These APIs used to be exposed with ZSTDLIB_API, * because it used to be the only way to invoke MT compression. - * Now, it's recommended to use ZSTD_compress2 and ZSTD_compressStream2() - * instead. - * - * If you depend on these APIs and can't switch, then define - * ZSTD_LEGACY_MULTITHREADED_API when making the dynamic library. - * However, we may completely remove these functions in a future - * release, so please switch soon. + * Now, you must use ZSTD_compress2 and ZSTD_compressStream2() instead. * * This API requires ZSTD_MULTITHREAD to be defined during compilation, * otherwise ZSTDMT_createCCtx*() will fail. */ -#ifdef ZSTD_LEGACY_MULTITHREADED_API -# define ZSTDMT_API ZSTDLIB_API -#else -# define ZSTDMT_API -#endif - /* === Dependencies === */ -#include <stddef.h> /* size_t */ +#include "../common/zstd_deps.h" /* size_t */ #define ZSTD_STATIC_LINKING_ONLY /* ZSTD_parameters */ #include "../zstd.h" /* ZSTD_inBuffer, ZSTD_outBuffer, ZSTDLIB_API */ /* === Constants === */ -#ifndef ZSTDMT_NBWORKERS_MAX -# define ZSTDMT_NBWORKERS_MAX 200 +#ifndef ZSTDMT_NBWORKERS_MAX /* a different value can be selected at compile time */ +# define ZSTDMT_NBWORKERS_MAX ((sizeof(void*)==4) /*32-bit*/ ? 64 : 256) #endif -#ifndef ZSTDMT_JOBSIZE_MIN -# define ZSTDMT_JOBSIZE_MIN (1 MB) +#ifndef ZSTDMT_JOBSIZE_MIN /* a different value can be selected at compile time */ +# define ZSTDMT_JOBSIZE_MIN (512 KB) #endif #define ZSTDMT_JOBLOG_MAX (MEM_32bits() ? 29 : 30) #define ZSTDMT_JOBSIZE_MAX (MEM_32bits() ? (512 MB) : (1024 MB)) +/* ======================================================== + * === Private interface, for use by ZSTD_compress.c === + * === Not exposed in libzstd. Never invoke directly === + * ======================================================== */ + /* === Memory management === */ typedef struct ZSTDMT_CCtx_s ZSTDMT_CCtx; /* Requires ZSTD_MULTITHREAD to be defined during compilation, otherwise it will return NULL. */ -ZSTDMT_API ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers); -/* Requires ZSTD_MULTITHREAD to be defined during compilation, otherwise it will return NULL. */ -ZSTDMT_API ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, - ZSTD_customMem cMem); -ZSTDMT_API size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx); - -ZSTDMT_API size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx); - - -/* === Simple one-pass compression function === */ - -ZSTDMT_API size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - int compressionLevel); - +ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, + ZSTD_customMem cMem, + ZSTD_threadPool *pool); +size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx); +size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx); /* === Streaming functions === */ -ZSTDMT_API size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel); -ZSTDMT_API size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize); /**< if srcSize is not known at reset time, use ZSTD_CONTENTSIZE_UNKNOWN. Note: for compatibility with older programs, 0 means the same as ZSTD_CONTENTSIZE_UNKNOWN, but it will change in the future to mean "empty" */ - -ZSTDMT_API size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx); -ZSTDMT_API size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input); - -ZSTDMT_API size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output); /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */ -ZSTDMT_API size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output); /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */ - - -/* === Advanced functions and parameters === */ - -ZSTDMT_API size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx, - void* dst, size_t dstCapacity, - const void* src, size_t srcSize, - const ZSTD_CDict* cdict, - ZSTD_parameters params, - int overlapLog); - -ZSTDMT_API size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx, - const void* dict, size_t dictSize, /* dict can be released after init, a local copy is preserved within zcs */ - ZSTD_parameters params, - unsigned long long pledgedSrcSize); /* pledgedSrcSize is optional and can be zero == unknown */ - -ZSTDMT_API size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx, - const ZSTD_CDict* cdict, - ZSTD_frameParameters fparams, - unsigned long long pledgedSrcSize); /* note : zero means empty */ - -/* ZSTDMT_parameter : - * List of parameters that can be set using ZSTDMT_setMTCtxParameter() */ -typedef enum { - ZSTDMT_p_jobSize, /* Each job is compressed in parallel. By default, this value is dynamically determined depending on compression parameters. Can be set explicitly here. */ - ZSTDMT_p_overlapLog, /* Each job may reload a part of previous job to enhance compression ratio; 0 == no overlap, 6(default) == use 1/8th of window, >=9 == use full window. This is a "sticky" parameter : its value will be re-used on next compression job */ - ZSTDMT_p_rsyncable /* Enables rsyncable mode. */ -} ZSTDMT_parameter; - -/* ZSTDMT_setMTCtxParameter() : - * allow setting individual parameters, one at a time, among a list of enums defined in ZSTDMT_parameter. - * The function must be called typically after ZSTD_createCCtx() but __before ZSTDMT_init*() !__ - * Parameters not explicitly reset by ZSTDMT_init*() remain the same in consecutive compression sessions. - * @return : 0, or an error code (which can be tested using ZSTD_isError()) */ -ZSTDMT_API size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int value); - -/* ZSTDMT_getMTCtxParameter() : - * Query the ZSTDMT_CCtx for a parameter value. - * @return : 0, or an error code (which can be tested using ZSTD_isError()) */ -ZSTDMT_API size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int* value); +size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx); +/*! ZSTDMT_initCStream_internal() : + * Private use only. Init streaming operation. + * expects params to be valid. + * must receive dict, or cdict, or none, but not both. + * mtctx can be freshly constructed or reused from a prior compression. + * If mtctx is reused, memory allocations from the prior compression may not be freed, + * even if they are not needed for the current compression. + * @return : 0, or an error code */ +size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* mtctx, + const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType, + const ZSTD_CDict* cdict, + ZSTD_CCtx_params params, unsigned long long pledgedSrcSize); /*! ZSTDMT_compressStream_generic() : * Combines ZSTDMT_compressStream() with optional ZSTDMT_flushStream() or ZSTDMT_endStream() @@ -134,16 +81,10 @@ ZSTDMT_API size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter * 0 if fully flushed * or an error code * note : needs to be init using any ZSTD_initCStream*() variant */ -ZSTDMT_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, - ZSTD_outBuffer* output, - ZSTD_inBuffer* input, - ZSTD_EndDirective endOp); - - -/* ======================================================== - * === Private interface, for use by ZSTD_compress.c === - * === Not exposed in libzstd. Never invoke directly === - * ======================================================== */ +size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, + ZSTD_outBuffer* output, + ZSTD_inBuffer* input, + ZSTD_EndDirective endOp); /*! ZSTDMT_toFlushNow() * Tell how many bytes are ready to be flushed immediately. @@ -153,15 +94,6 @@ ZSTDMT_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx, * therefore flushing is limited by speed of oldest job. */ size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx); -/*! ZSTDMT_CCtxParam_setMTCtxParameter() - * like ZSTDMT_setMTCtxParameter(), but into a ZSTD_CCtx_Params */ -size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params, ZSTDMT_parameter parameter, int value); - -/*! ZSTDMT_CCtxParam_setNbWorkers() - * Set nbWorkers, and clamp it. - * Also reset jobSize and overlapLog */ -size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers); - /*! ZSTDMT_updateCParams_whileCompressing() : * Updates only a selected set of compression parameters, to remain compatible with current frame. * New parameters will be applied to next compression job. */ @@ -174,17 +106,6 @@ void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_p ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx); -/*! ZSTDMT_initCStream_internal() : - * Private use only. Init streaming operation. - * expects params to be valid. - * must receive dict, or cdict, or none, but not both. - * @return : 0, or an error code */ -size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs, - const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType, - const ZSTD_CDict* cdict, - ZSTD_CCtx_params params, unsigned long long pledgedSrcSize); - - #if defined (__cplusplus) } #endif |