zstd/common/fse_decompress.c

0001 /* ******************************************************************
0002  * FSE : Finite State Entropy decoder
0003  * Copyright (c) Yann Collet, Facebook, Inc.
0004  *
0005  *  You can contact the author at :
0006  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
0007  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
0008  *
0009  * This source code is licensed under both the BSD-style license (found in the
0010  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
0011  * in the COPYING file in the root directory of this source tree).
0012  * You may select, at your option, one of the above-listed licenses.
0013 ****************************************************************** */
0014
0015
0016 /* **************************************************************
0017 *  Includes
0018 ****************************************************************/
0019 #include "debug.h"      /* assert */
0020 #include "bitstream.h"
0021 #include "compiler.h"
0022 #define FSE_STATIC_LINKING_ONLY
0023 #include "fse.h"
0024 #include "error_private.h"
0025 #define ZSTD_DEPS_NEED_MALLOC
0026 #include "zstd_deps.h"
0027
0028
0029 /* **************************************************************
0030 *  Error Management
0031 ****************************************************************/
0032 #define FSE_isError ERR_isError
0033 #define FSE_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */
0034
0035
0036 /* **************************************************************
0037 *  Templates
0038 ****************************************************************/
0039 /*
0040   designed to be included
0041   for type-specific functions (template emulation in C)
0042   Objective is to write these functions only once, for improved maintenance
0043 */
0044
0045 /* safety checks */
0046 #ifndef FSE_FUNCTION_EXTENSION
0047 #  error "FSE_FUNCTION_EXTENSION must be defined"
0048 #endif
0049 #ifndef FSE_FUNCTION_TYPE
0050 #  error "FSE_FUNCTION_TYPE must be defined"
0051 #endif
0052
0053 /* Function names */
0054 #define FSE_CAT(X,Y) X##Y
0055 #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
0056 #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
0057
0058
0059 /* Function templates */
0060 FSE_DTable* FSE_createDTable (unsigned tableLog)
0061 {
0062     if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
0063     return (FSE_DTable*)ZSTD_malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
0064 }
0065
0066 void FSE_freeDTable (FSE_DTable* dt)
0067 {
0068     ZSTD_free(dt);
0069 }
0070
0071 static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
0072 {
0073     void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
0074     FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
0075     U16* symbolNext = (U16*)workSpace;
0076     BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
0077
0078     U32 const maxSV1 = maxSymbolValue + 1;
0079     U32 const tableSize = 1 << tableLog;
0080     U32 highThreshold = tableSize-1;
0081
0082     /* Sanity Checks */
0083     if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
0084     if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
0085     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
0086
0087     /* Init, lay down lowprob symbols */
0088     {   FSE_DTableHeader DTableH;
0089         DTableH.tableLog = (U16)tableLog;
0090         DTableH.fastMode = 1;
0091         {   S16 const largeLimit= (S16)(1 << (tableLog-1));
0092             U32 s;
0093             for (s=0; s<maxSV1; s++) {
0094                 if (normalizedCounter[s]==-1) {
0095                     tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
0096                     symbolNext[s] = 1;
0097                 } else {
0098                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
0099                     symbolNext[s] = normalizedCounter[s];
0100         }   }   }
0101         ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
0102     }
0103
0104     /* Spread symbols */
0105     if (highThreshold == tableSize - 1) {
0106         size_t const tableMask = tableSize-1;
0107         size_t const step = FSE_TABLESTEP(tableSize);
0108         /* First lay down the symbols in order.
0109          * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
0110          * misses since small blocks generally have small table logs, so nearly
0111          * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
0112          * our buffer to handle the over-write.
0113          */
0114         {
0115             U64 const add = 0x0101010101010101ull;
0116             size_t pos = 0;
0117             U64 sv = 0;
0118             U32 s;
0119             for (s=0; s<maxSV1; ++s, sv += add) {
0120                 int i;
0121                 int const n = normalizedCounter[s];
0122                 MEM_write64(spread + pos, sv);
0123                 for (i = 8; i < n; i += 8) {
0124                     MEM_write64(spread + pos + i, sv);
0125                 }
0126                 pos += n;
0127             }
0128         }
0129         /* Now we spread those positions across the table.
0130          * The benefit of doing it in two stages is that we avoid the the
0131          * variable size inner loop, which caused lots of branch misses.
0132          * Now we can run through all the positions without any branch misses.
0133          * We unroll the loop twice, since that is what emperically worked best.
0134          */
0135         {
0136             size_t position = 0;
0137             size_t s;
0138             size_t const unroll = 2;
0139             assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
0140             for (s = 0; s < (size_t)tableSize; s += unroll) {
0141                 size_t u;
0142                 for (u = 0; u < unroll; ++u) {
0143                     size_t const uPosition = (position + (u * step)) & tableMask;
0144                     tableDecode[uPosition].symbol = spread[s + u];
0145                 }
0146                 position = (position + (unroll * step)) & tableMask;
0147             }
0148             assert(position == 0);
0149         }
0150     } else {
0151         U32 const tableMask = tableSize-1;
0152         U32 const step = FSE_TABLESTEP(tableSize);
0153         U32 s, position = 0;
0154         for (s=0; s<maxSV1; s++) {
0155             int i;
0156             for (i=0; i<normalizedCounter[s]; i++) {
0157                 tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
0158                 position = (position + step) & tableMask;
0159                 while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
0160         }   }
0161         if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
0162     }
0163
0164     /* Build Decoding table */
0165     {   U32 u;
0166         for (u=0; u<tableSize; u++) {
0167             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
0168             U32 const nextState = symbolNext[symbol]++;
0169             tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
0170             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
0171     }   }
0172
0173     return 0;
0174 }
0175
0176 size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
0177 {
0178     return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
0179 }
0180
0181
0182 #ifndef FSE_COMMONDEFS_ONLY
0183
0184 /*-*******************************************************
0185 *  Decompression (Byte symbols)
0186 *********************************************************/
0187 size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
0188 {
0189     void* ptr = dt;
0190     FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
0191     void* dPtr = dt + 1;
0192     FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
0193
0194     DTableH->tableLog = 0;
0195     DTableH->fastMode = 0;
0196
0197     cell->newState = 0;
0198     cell->symbol = symbolValue;
0199     cell->nbBits = 0;
0200
0201     return 0;
0202 }
0203
0204
0205 size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
0206 {
0207     void* ptr = dt;
0208     FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
0209     void* dPtr = dt + 1;
0210     FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
0211     const unsigned tableSize = 1 << nbBits;
0212     const unsigned tableMask = tableSize - 1;
0213     const unsigned maxSV1 = tableMask+1;
0214     unsigned s;
0215
0216     /* Sanity checks */
0217     if (nbBits < 1) return ERROR(GENERIC);         /* min size */
0218
0219     /* Build Decoding Table */
0220     DTableH->tableLog = (U16)nbBits;
0221     DTableH->fastMode = 1;
0222     for (s=0; s<maxSV1; s++) {
0223         dinfo[s].newState = 0;
0224         dinfo[s].symbol = (BYTE)s;
0225         dinfo[s].nbBits = (BYTE)nbBits;
0226     }
0227
0228     return 0;
0229 }
0230
0231 FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
0232           void* dst, size_t maxDstSize,
0233     const void* cSrc, size_t cSrcSize,
0234     const FSE_DTable* dt, const unsigned fast)
0235 {
0236     BYTE* const ostart = (BYTE*) dst;
0237     BYTE* op = ostart;
0238     BYTE* const omax = op + maxDstSize;
0239     BYTE* const olimit = omax-3;
0240
0241     BIT_DStream_t bitD;
0242     FSE_DState_t state1;
0243     FSE_DState_t state2;
0244
0245     /* Init */
0246     CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));
0247
0248     FSE_initDState(&state1, &bitD, dt);
0249     FSE_initDState(&state2, &bitD, dt);
0250
0251 #define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
0252
0253     /* 4 symbols per loop */
0254     for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
0255         op[0] = FSE_GETSYMBOL(&state1);
0256
0257         if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
0258             BIT_reloadDStream(&bitD);
0259
0260         op[1] = FSE_GETSYMBOL(&state2);
0261
0262         if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
0263             { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
0264
0265         op[2] = FSE_GETSYMBOL(&state1);
0266
0267         if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
0268             BIT_reloadDStream(&bitD);
0269
0270         op[3] = FSE_GETSYMBOL(&state2);
0271     }
0272
0273     /* tail */
0274     /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
0275     while (1) {
0276         if (op>(omax-2)) return ERROR(dstSize_tooSmall);
0277         *op++ = FSE_GETSYMBOL(&state1);
0278         if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
0279             *op++ = FSE_GETSYMBOL(&state2);
0280             break;
0281         }
0282
0283         if (op>(omax-2)) return ERROR(dstSize_tooSmall);
0284         *op++ = FSE_GETSYMBOL(&state2);
0285         if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
0286             *op++ = FSE_GETSYMBOL(&state1);
0287             break;
0288     }   }
0289
0290     return op-ostart;
0291 }
0292
0293
0294 size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
0295                             const void* cSrc, size_t cSrcSize,
0296                             const FSE_DTable* dt)
0297 {
0298     const void* ptr = dt;
0299     const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
0300     const U32 fastMode = DTableH->fastMode;
0301
0302     /* select fast mode (static) */
0303     if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
0304     return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
0305 }
0306
0307
0308 size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
0309 {
0310     return FSE_decompress_wksp_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, /* bmi2 */ 0);
0311 }
0312
0313 typedef struct {
0314     short ncount[FSE_MAX_SYMBOL_VALUE + 1];
0315     FSE_DTable dtable[1]; /* Dynamically sized */
0316 } FSE_DecompressWksp;
0317
0318
0319 FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
0320         void* dst, size_t dstCapacity,
0321         const void* cSrc, size_t cSrcSize,
0322         unsigned maxLog, void* workSpace, size_t wkspSize,
0323         int bmi2)
0324 {
0325     const BYTE* const istart = (const BYTE*)cSrc;
0326     const BYTE* ip = istart;
0327     unsigned tableLog;
0328     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
0329     FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
0330
0331     DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
0332     if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
0333
0334     /* normal FSE decoding mode */
0335     {
0336         size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
0337         if (FSE_isError(NCountLength)) return NCountLength;
0338         if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
0339         assert(NCountLength <= cSrcSize);
0340         ip += NCountLength;
0341         cSrcSize -= NCountLength;
0342     }
0343
0344     if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
0345     workSpace = wksp->dtable + FSE_DTABLE_SIZE_U32(tableLog);
0346     wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
0347
0348     CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
0349
0350     {
0351         const void* ptr = wksp->dtable;
0352         const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
0353         const U32 fastMode = DTableH->fastMode;
0354
0355         /* select fast mode (static) */
0356         if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
0357         return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
0358     }
0359 }
0360
0361 /* Avoids the FORCE_INLINE of the _body() function. */
0362 static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
0363 {
0364     return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
0365 }
0366
0367 #if DYNAMIC_BMI2
0368 TARGET_ATTRIBUTE("bmi2") static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
0369 {
0370     return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
0371 }
0372 #endif
0373
0374 size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
0375 {
0376 #if DYNAMIC_BMI2
0377     if (bmi2) {
0378         return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
0379     }
0380 #endif
0381     (void)bmi2;
0382     return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
0383 }
0384
0385
0386 typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
0387
0388
0389
0390 #endif   /* FSE_COMMONDEFS_ONLY */