From: Elliot Gorokhovsky Date: Tue, 21 Jun 2022 21:27:19 +0000 (-0700) Subject: "Short cache" optimization for level 1-4 DMS (+5-30% compression speed) (#3152) X-Git-Tag: v1.5.4^2~195 X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=f6ef14329f396eb8b2c1290790e7547d070d9511;p=thirdparty%2Fzstd.git "Short cache" optimization for level 1-4 DMS (+5-30% compression speed) (#3152) * first attempt at fast DMS short cache * significant wins for some scenarios * fix all clang regressions * nits * fix 1.5% gcc11 regression on hot 110Kdict scenario * fix CI * nit * Add tags to doublefast hash table * use tags in doublefast DMS * Fix CI * Clean up some hardcoded logic / constants * Switch forCCtx to an enum * nit * add short cache to ip+1 long search * Move tag size into hashLog * Minor nits * Truncate dictionaries greater than 16MB in short cache mode * Helper function for tag comparison * Cap short cache hashLog at 24 to prevent overflow * size_t dictTagsMatch -> int dictTagsMatch * nit * Clean up and comment dictionary truncation * Move ZSTD_tableFillPurpose_e next to ZSTD_dictTableLoadMethod_e * Comment and expand helper functions * Asserts and documentation * nit --- diff --git a/lib/compress/zstd_compress.c b/lib/compress/zstd_compress.c index ec2a2fe27..46a9dbe66 100644 --- a/lib/compress/zstd_compress.c +++ b/lib/compress/zstd_compress.c @@ -275,6 +275,12 @@ static ZSTD_paramSwitch_e ZSTD_resolveEnableLdm(ZSTD_paramSwitch_e mode, return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable; } +/* Returns 1 if compression parameters are such that CDict hashtable and chaintable indices are tagged. + * If so, the tags need to be removed in ZSTD_resetCCtx_byCopyingCDict. */ +static int ZSTD_CDictIndicesAreTagged(const ZSTD_compressionParameters* const cParams) { + return cParams->strategy == ZSTD_fast || cParams->strategy == ZSTD_dfast; +} + static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams( ZSTD_compressionParameters cParams) { @@ -1367,6 +1373,13 @@ ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar, if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN) cPar.windowLog = ZSTD_WINDOWLOG_ABSOLUTEMIN; /* minimum wlog required for valid frame header */ + if (mode == ZSTD_cpm_createCDict && ZSTD_CDictIndicesAreTagged(&cPar)) { + U32 const maxShortCacheHashLog = 32 - ZSTD_SHORT_CACHE_TAG_BITS; + if (cPar.hashLog > maxShortCacheHashLog) { + cPar.hashLog = maxShortCacheHashLog; + } + } + return cPar; } @@ -2096,6 +2109,22 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx, return 0; } +static void ZSTD_copyCDictTableIntoCCtx(U32* dst, U32 const* src, size_t tableSize, + ZSTD_compressionParameters const* cParams) { + if (ZSTD_CDictIndicesAreTagged(cParams)){ + /* Remove tags from the CDict table if they are present. + * See docs on "short cache" in zstd_compress_internal.h for context. */ + size_t i; + for (i = 0; i < tableSize; i++) { + U32 const taggedIndex = src[i]; + U32 const index = taggedIndex >> ZSTD_SHORT_CACHE_TAG_BITS; + dst[i] = index; + } + } else { + ZSTD_memcpy(dst, src, tableSize * sizeof(U32)); + } +} + static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict, ZSTD_CCtx_params params, @@ -2131,14 +2160,15 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx, : 0; size_t const hSize = (size_t)1 << cdict_cParams->hashLog; - ZSTD_memcpy(cctx->blockState.matchState.hashTable, - cdict->matchState.hashTable, - hSize * sizeof(U32)); + ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.hashTable, + cdict->matchState.hashTable, + hSize, cdict_cParams); + /* 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)); + ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.chainTable, + cdict->matchState.chainTable, + chainSize, cdict_cParams); } /* copy tag table */ if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) { @@ -4205,7 +4235,8 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, ZSTD_cwksp* ws, ZSTD_CCtx_params const* params, const void* src, size_t srcSize, - ZSTD_dictTableLoadMethod_e dtlm) + ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp) { const BYTE* ip = (const BYTE*) src; const BYTE* const iend = ip + srcSize; @@ -4214,22 +4245,37 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, /* Assert that the ms params match the params we're being given */ ZSTD_assertEqualCParams(params->cParams, ms->cParams); - if (srcSize > ZSTD_CHUNKSIZE_MAX) { + { /* Ensure large dictionaries can't cause index overflow */ + /* 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 - ZSTD_WINDOW_START_INDEX; - /* 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)); + U32 maxDictSize = ZSTD_CURRENT_MAX - ZSTD_WINDOW_START_INDEX; + + int const CDictTaggedIndices = ZSTD_CDictIndicesAreTagged(¶ms->cParams); + if (CDictTaggedIndices && tfp == ZSTD_tfp_forCDict) { + /* Some dictionary matchfinders in zstd use "short cache", + * which treats the lower ZSTD_SHORT_CACHE_TAG_BITS of each + * CDict hashtable entry as a tag rather than as part of an index. + * When short cache is used, we need to truncate the dictionary + * so that its indices don't overlap with the tag. */ + U32 const shortCacheMaxDictSize = (1u << (32 - ZSTD_SHORT_CACHE_TAG_BITS)) - ZSTD_WINDOW_START_INDEX; + maxDictSize = MIN(maxDictSize, shortCacheMaxDictSize); + assert(!loadLdmDict); + } + /* If the dictionary is too large, only load the suffix of the dictionary. */ if (srcSize > maxDictSize) { ip = iend - maxDictSize; src = ip; srcSize = maxDictSize; - } + } } + + if (srcSize > ZSTD_CHUNKSIZE_MAX) { + /* 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)); } DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder); @@ -4252,10 +4298,10 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms, switch(params->cParams.strategy) { case ZSTD_fast: - ZSTD_fillHashTable(ms, iend, dtlm); + ZSTD_fillHashTable(ms, iend, dtlm, tfp); break; case ZSTD_dfast: - ZSTD_fillDoubleHashTable(ms, iend, dtlm); + ZSTD_fillDoubleHashTable(ms, iend, dtlm, tfp); break; case ZSTD_greedy: @@ -4421,6 +4467,7 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs, ZSTD_CCtx_params const* params, const void* dict, size_t dictSize, ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp, void* workspace) { const BYTE* dictPtr = (const BYTE*)dict; @@ -4439,7 +4486,7 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs, { size_t const dictContentSize = (size_t)(dictEnd - dictPtr); FORWARD_IF_ERROR(ZSTD_loadDictionaryContent( - ms, NULL, ws, params, dictPtr, dictContentSize, dtlm), ""); + ms, NULL, ws, params, dictPtr, dictContentSize, dtlm, tfp), ""); } return dictID; } @@ -4455,6 +4502,7 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs, const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType, ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp, void* workspace) { DEBUGLOG(4, "ZSTD_compress_insertDictionary (dictSize=%u)", (U32)dictSize); @@ -4467,13 +4515,13 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs, /* dict restricted modes */ if (dictContentType == ZSTD_dct_rawContent) - return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm); + return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm, tfp); if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) { if (dictContentType == ZSTD_dct_auto) { DEBUGLOG(4, "raw content dictionary detected"); return ZSTD_loadDictionaryContent( - ms, ls, ws, params, dict, dictSize, dtlm); + ms, ls, ws, params, dict, dictSize, dtlm, tfp); } RETURN_ERROR_IF(dictContentType == ZSTD_dct_fullDict, dictionary_wrong, ""); assert(0); /* impossible */ @@ -4481,7 +4529,7 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs, /* dict as full zstd dictionary */ return ZSTD_loadZstdDictionary( - bs, ms, ws, params, dict, dictSize, dtlm, workspace); + bs, ms, ws, params, dict, dictSize, dtlm, tfp, workspace); } #define ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF (128 KB) @@ -4524,11 +4572,11 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx, cctx->blockState.prevCBlock, &cctx->blockState.matchState, &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, cdict->dictContent, cdict->dictContentSize, cdict->dictContentType, dtlm, - cctx->entropyWorkspace) + ZSTD_tfp_forCCtx, cctx->entropyWorkspace) : ZSTD_compress_insertDictionary( cctx->blockState.prevCBlock, &cctx->blockState.matchState, &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, dict, dictSize, - dictContentType, dtlm, cctx->entropyWorkspace); + dictContentType, dtlm, ZSTD_tfp_forCCtx, cctx->entropyWorkspace); FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed"); assert(dictID <= UINT_MAX); cctx->dictID = (U32)dictID; @@ -4832,7 +4880,7 @@ static size_t ZSTD_initCDict_internal( { size_t const dictID = ZSTD_compress_insertDictionary( &cdict->cBlockState, &cdict->matchState, NULL, &cdict->workspace, ¶ms, cdict->dictContent, cdict->dictContentSize, - dictContentType, ZSTD_dtlm_full, cdict->entropyWorkspace); + dictContentType, ZSTD_dtlm_full, ZSTD_tfp_forCDict, cdict->entropyWorkspace); FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed"); assert(dictID <= (size_t)(U32)-1); cdict->dictID = (U32)dictID; diff --git a/lib/compress/zstd_compress_internal.h b/lib/compress/zstd_compress_internal.h index bbae53303..870bcc8be 100644 --- a/lib/compress/zstd_compress_internal.h +++ b/lib/compress/zstd_compress_internal.h @@ -434,6 +434,7 @@ struct ZSTD_CCtx_s { }; typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e; +typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e; typedef enum { ZSTD_noDict = 0, @@ -745,32 +746,36 @@ ZSTD_count_2segments(const BYTE* ip, const BYTE* match, * Hashes ***************************************/ static const U32 prime3bytes = 506832829U; -static U32 ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes) >> (32-h) ; } +static U32 ZSTD_hash3(U32 u, U32 h) { assert(h <= 32); return ((u << (32-24)) * prime3bytes) >> (32-h) ; } MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */ static const U32 prime4bytes = 2654435761U; -static U32 ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; } +static U32 ZSTD_hash4(U32 u, U32 h) { assert(h <= 32); return (u * prime4bytes) >> (32-h) ; } static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); } static const U64 prime5bytes = 889523592379ULL; -static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u << (64-40)) * prime5bytes) >> (64-h)) ; } +static size_t ZSTD_hash5(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u << (64-40)) * prime5bytes) >> (64-h)) ; } static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); } static const U64 prime6bytes = 227718039650203ULL; -static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u << (64-48)) * prime6bytes) >> (64-h)) ; } +static size_t ZSTD_hash6(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u << (64-48)) * prime6bytes) >> (64-h)) ; } static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); } static const U64 prime7bytes = 58295818150454627ULL; -static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u << (64-56)) * prime7bytes) >> (64-h)) ; } +static size_t ZSTD_hash7(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u << (64-56)) * prime7bytes) >> (64-h)) ; } static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); } 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_hash8(U64 u, U32 h) { assert(h <= 64); 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 FORCE_INLINE_ATTR size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls) { + /* Although some of these hashes do support hBits up to 64, some do not. + * To be on the safe side, always avoid hBits > 32. */ + assert(hBits <= 32); + switch(mls) { default: @@ -1264,6 +1269,42 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max) #endif +/* Short Cache */ + +/* Normally, zstd matchfinders follow this flow: + * 1. Compute hash at ip + * 2. Load index from hashTable[hash] + * 3. Check if *ip == *(base + index) + * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss. + * + * Short cache is an optimization which allows us to avoid step 3 most of the time + * when the data doesn't actually match. With short cache, the flow becomes: + * 1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip. + * 2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works. + * 3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue. + * + * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to + * dictMatchState matchfinders. + */ +#define ZSTD_SHORT_CACHE_TAG_BITS 8 +#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1) + +/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable. + * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */ +MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) { + size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS; + U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK); + assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0); + hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag; +} + +/* Helper function for short cache matchfinders. + * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */ +MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) { + U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK; + U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK; + return tag1 == tag2; +} #if defined (__cplusplus) } diff --git a/lib/compress/zstd_double_fast.c b/lib/compress/zstd_double_fast.c index d8412ef0b..6697ba0a9 100644 --- a/lib/compress/zstd_double_fast.c +++ b/lib/compress/zstd_double_fast.c @@ -11,8 +11,43 @@ #include "zstd_compress_internal.h" #include "zstd_double_fast.h" +static void ZSTD_fillDoubleHashTableForCDict(ZSTD_matchState_t* ms, + void const* end, ZSTD_dictTableLoadMethod_e dtlm) +{ + const ZSTD_compressionParameters* const cParams = &ms->cParams; + U32* const hashLarge = ms->hashTable; + U32 const hBitsL = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; + U32 const mls = cParams->minMatch; + U32* const hashSmall = ms->chainTable; + U32 const hBitsS = cParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS; + const BYTE* const base = ms->window.base; + const BYTE* ip = base + ms->nextToUpdate; + const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE; + const U32 fastHashFillStep = 3; -void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, + /* Always insert every fastHashFillStep position into the hash tables. + * Insert the other positions into the large hash table if their entry + * is empty. + */ + for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) { + U32 const curr = (U32)(ip - base); + U32 i; + for (i = 0; i < fastHashFillStep; ++i) { + size_t const smHashAndTag = ZSTD_hashPtr(ip + i, hBitsS, mls); + size_t const lgHashAndTag = ZSTD_hashPtr(ip + i, hBitsL, 8); + if (i == 0) { + ZSTD_writeTaggedIndex(hashSmall, smHashAndTag, curr + i); + } + if (i == 0 || hashLarge[lgHashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) { + ZSTD_writeTaggedIndex(hashLarge, lgHashAndTag, curr + i); + } + /* Only load extra positions for ZSTD_dtlm_full */ + if (dtlm == ZSTD_dtlm_fast) + break; + } } +} + +static void ZSTD_fillDoubleHashTableForCCtx(ZSTD_matchState_t* ms, void const* end, ZSTD_dictTableLoadMethod_e dtlm) { const ZSTD_compressionParameters* const cParams = &ms->cParams; @@ -43,7 +78,19 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, /* Only load extra positions for ZSTD_dtlm_full */ if (dtlm == ZSTD_dtlm_fast) break; - } } + } } +} + +void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, + const void* const end, + ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp) +{ + if (tfp == ZSTD_tfp_forCDict) { + ZSTD_fillDoubleHashTableForCDict(ms, end, dtlm); + } else { + ZSTD_fillDoubleHashTableForCCtx(ms, end, dtlm); + } } @@ -289,8 +336,8 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( 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 dictHBitsL = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; + const U32 dictHBitsS = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS; const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictStart)); DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic"); @@ -312,8 +359,12 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( U32 offset; size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8); 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); + size_t const dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8); + size_t const dictHashAndTagS = ZSTD_hashPtr(ip, dictHBitsS, mls); + U32 const dictMatchIndexAndTagL = dictHashLong[dictHashAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS]; + U32 const dictMatchIndexAndTagS = dictHashSmall[dictHashAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS]; + int const dictTagsMatchL = ZSTD_comparePackedTags(dictMatchIndexAndTagL, dictHashAndTagL); + int const dictTagsMatchS = ZSTD_comparePackedTags(dictMatchIndexAndTagS, dictHashAndTagS); U32 const curr = (U32)(ip-base); U32 const matchIndexL = hashLong[h2]; U32 matchIndexS = hashSmall[h]; @@ -343,9 +394,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */ goto _match_found; } - } else { + } else if (dictTagsMatchL) { /* check dictMatchState long match */ - U32 const dictMatchIndexL = dictHashLong[dictHL]; + U32 const dictMatchIndexL = dictMatchIndexAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS; const BYTE* dictMatchL = dictBase + dictMatchIndexL; assert(dictMatchL < dictEnd); @@ -361,9 +412,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( if (MEM_read32(match) == MEM_read32(ip)) { goto _search_next_long; } - } else { + } else if (dictTagsMatchS) { /* check dictMatchState short match */ - U32 const dictMatchIndexS = dictHashSmall[dictHS]; + U32 const dictMatchIndexS = dictMatchIndexAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS; match = dictBase + dictMatchIndexS; matchIndexS = dictMatchIndexS + dictIndexDelta; @@ -378,10 +429,11 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic( continue; _search_next_long: - { size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8); - size_t const dictHLNext = ZSTD_hashPtr(ip+1, dictHBitsL, 8); + size_t const dictHashAndTagL3 = ZSTD_hashPtr(ip+1, dictHBitsL, 8); U32 const matchIndexL3 = hashLong[hl3]; + U32 const dictMatchIndexAndTagL3 = dictHashLong[dictHashAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS]; + int const dictTagsMatchL3 = ZSTD_comparePackedTags(dictMatchIndexAndTagL3, dictHashAndTagL3); const BYTE* matchL3 = base + matchIndexL3; hashLong[hl3] = curr + 1; @@ -394,9 +446,9 @@ _search_next_long: while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */ goto _match_found; } - } else { + } else if (dictTagsMatchL3) { /* check dict long +1 match */ - U32 const dictMatchIndexL3 = dictHashLong[dictHLNext]; + U32 const dictMatchIndexL3 = dictMatchIndexAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS; const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3; assert(dictMatchL3 < dictEnd); if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) { diff --git a/lib/compress/zstd_double_fast.h b/lib/compress/zstd_double_fast.h index e16b7b03a..6d8ee8c65 100644 --- a/lib/compress/zstd_double_fast.h +++ b/lib/compress/zstd_double_fast.h @@ -19,7 +19,8 @@ extern "C" { #include "zstd_compress_internal.h" /* ZSTD_CCtx, size_t */ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms, - void const* end, ZSTD_dictTableLoadMethod_e dtlm); + void const* end, ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp); size_t ZSTD_compressBlock_doubleFast( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); diff --git a/lib/compress/zstd_fast.c b/lib/compress/zstd_fast.c index b0ec972c6..de7336907 100644 --- a/lib/compress/zstd_fast.c +++ b/lib/compress/zstd_fast.c @@ -11,8 +11,42 @@ #include "zstd_compress_internal.h" /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */ #include "zstd_fast.h" +static void ZSTD_fillHashTableForCDict(ZSTD_matchState_t* ms, + const void* const end, + ZSTD_dictTableLoadMethod_e dtlm) +{ + const ZSTD_compressionParameters* const cParams = &ms->cParams; + U32* const hashTable = ms->hashTable; + U32 const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; + U32 const mls = cParams->minMatch; + const BYTE* const base = ms->window.base; + const BYTE* ip = base + ms->nextToUpdate; + const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE; + const U32 fastHashFillStep = 3; -void ZSTD_fillHashTable(ZSTD_matchState_t* ms, + /* Currently, we always use ZSTD_dtlm_full for filling CDict tables. + * Feel free to remove this assert if there's a good reason! */ + assert(dtlm == ZSTD_dtlm_full); + + /* Always insert every fastHashFillStep position into the hash table. + * Insert the other positions if their hash entry is empty. + */ + for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) { + U32 const curr = (U32)(ip - base); + { size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls); + ZSTD_writeTaggedIndex(hashTable, hashAndTag, 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 hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls); + if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) { /* not yet filled */ + ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p); + } } } } +} + +static void ZSTD_fillHashTableForCCtx(ZSTD_matchState_t* ms, const void* const end, ZSTD_dictTableLoadMethod_e dtlm) { @@ -25,6 +59,10 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms, const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE; const U32 fastHashFillStep = 3; + /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables. + * Feel free to remove this assert if there's a good reason! */ + assert(dtlm == ZSTD_dtlm_fast); + /* Always insert every fastHashFillStep position into the hash table. * Insert the other positions if their hash entry is empty. */ @@ -42,6 +80,18 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms, } } } } } +void ZSTD_fillHashTable(ZSTD_matchState_t* ms, + const void* const end, + ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp) +{ + if (tfp == ZSTD_tfp_forCDict) { + ZSTD_fillHashTableForCDict(ms, end, dtlm); + } else { + ZSTD_fillHashTableForCCtx(ms, end, dtlm); + } +} + /** * If you squint hard enough (and ignore repcodes), the search operation at any @@ -435,7 +485,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( const BYTE* const dictEnd = dms->window.nextSrc; const U32 dictIndexDelta = prefixStartIndex - (U32)(dictEnd - dictBase); const U32 dictAndPrefixLength = (U32)(istart - prefixStart + dictEnd - dictStart); - const U32 dictHLog = dictCParams->hashLog; + const U32 dictHBits = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; /* if a dictionary is still attached, it necessarily means that * it is within window size. So we just check it. */ @@ -463,8 +513,11 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( while (ip1 <= ilimit) { /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */ size_t mLength; size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls); - const size_t dictHash0 = ZSTD_hashPtr(ip0, dictHLog, mls); - U32 dictMatchIndex = dictHashTable[dictHash0]; + + size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls); + U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS]; + int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0); + U32 matchIndex = hashTable[hash0]; U32 curr = (U32)(ip0 - base); size_t step = stepSize; @@ -479,7 +532,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( dictBase + (repIndex - dictIndexDelta) : base + repIndex; const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls); - const size_t dictHash1 = ZSTD_hashPtr(ip1, dictHLog, mls); + size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls); hashTable[hash0] = curr; /* update hash table */ if (((U32) ((prefixStartIndex - 1) - repIndex) >= @@ -490,26 +543,33 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( ip0++; ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength); break; - } else if (matchIndex <= prefixStartIndex) { - /* We only look for a dict match if the normal matchIndex is invalid */ + } + + if (dictTagsMatch) { + /* Found a possible dict match */ + const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS; const BYTE* dictMatch = dictBase + dictMatchIndex; if (dictMatchIndex > dictStartIndex && MEM_read32(dictMatch) == MEM_read32(ip0)) { - /* found a dict match */ - U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta); - mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4; - while (((ip0 > anchor) & (dictMatch > dictStart)) - && (ip0[-1] == dictMatch[-1])) { - ip0--; - dictMatch--; - mLength++; - } /* catch up */ - offset_2 = offset_1; - offset_1 = offset; - ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength); - break; + /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */ + if (matchIndex <= prefixStartIndex) { + U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta); + mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4; + while (((ip0 > anchor) & (dictMatch > dictStart)) + && (ip0[-1] == dictMatch[-1])) { + ip0--; + dictMatch--; + mLength++; + } /* catch up */ + offset_2 = offset_1; + offset_1 = offset; + ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength); + break; + } } - } else if (MEM_read32(match) == MEM_read32(ip0)) { + } + + if (matchIndex > prefixStartIndex && MEM_read32(match) == MEM_read32(ip0)) { /* found a regular match */ U32 const offset = (U32) (ip0 - match); mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4; @@ -526,7 +586,8 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic( } /* Prepare for next iteration */ - dictMatchIndex = dictHashTable[dictHash1]; + dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS]; + dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1); matchIndex = hashTable[hash1]; if (ip1 >= nextStep) { diff --git a/lib/compress/zstd_fast.h b/lib/compress/zstd_fast.h index 0d4a0c109..3bfeb2c5f 100644 --- a/lib/compress/zstd_fast.h +++ b/lib/compress/zstd_fast.h @@ -19,7 +19,8 @@ extern "C" { #include "zstd_compress_internal.h" void ZSTD_fillHashTable(ZSTD_matchState_t* ms, - void const* end, ZSTD_dictTableLoadMethod_e dtlm); + void const* end, ZSTD_dictTableLoadMethod_e dtlm, + ZSTD_tableFillPurpose_e tfp); size_t ZSTD_compressBlock_fast( ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], void const* src, size_t srcSize); diff --git a/lib/compress/zstd_ldm.c b/lib/compress/zstd_ldm.c index e1d2f7410..c14c62454 100644 --- a/lib/compress/zstd_ldm.c +++ b/lib/compress/zstd_ldm.c @@ -242,11 +242,11 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms, switch(ms->cParams.strategy) { case ZSTD_fast: - ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast); + ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx); break; case ZSTD_dfast: - ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast); + ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx); break; case ZSTD_greedy: