module/zstd/lib/compress/huf_compress.c

   1 /* ******************************************************************
   2  * Huffman encoder, part of New Generation Entropy library
   3  * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
   4  *
   5  *  You can contact the author at :
   6  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
   7  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
   8  *
   9  * This source code is licensed under both the BSD-style license (found in the
  10  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
  11  * in the COPYING file in the root directory of this source tree).
  12  * You may select, at your option, one of the above-listed licenses.
  13 ****************************************************************** */
  14
  15 /* **************************************************************
  16 *  Compiler specifics
  17 ****************************************************************/
  18 #ifdef _MSC_VER    /* Visual Studio */
  19 #  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
  20 #endif
  21
  22
  23 /* **************************************************************
  24 *  Includes
  25 ****************************************************************/
  26 #include <string.h>     /* memcpy, memset */
  27 #include <stdio.h>      /* printf (debug) */
  28 #include "../common/compiler.h"
  29 #include "../common/bitstream.h"
  30 #include "hist.h"
  31 #define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
  32 #include "../common/fse.h"        /* header compression */
  33 #define HUF_STATIC_LINKING_ONLY
  34 #include "../common/huf.h"
  35 #include "../common/error_private.h"
  36
  37
  38 /* **************************************************************
  39 *  Error Management
  40 ****************************************************************/
  41 #define HUF_isError ERR_isError
  42 #define HUF_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */
  43
  44
  45 /* **************************************************************
  46 *  Utils
  47 ****************************************************************/
  48 unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
  49 {
  50     return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
  51 }
  52
  53
  54 /* *******************************************************
  55 *  HUF : Huffman block compression
  56 *********************************************************/
  57 /* HUF_compressWeights() :
  58  * Same as FSE_compress(), but dedicated to huff0's weights compression.
  59  * The use case needs much less stack memory.
  60  * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
  61  */
  62 #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
  63 static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weightTable, size_t wtSize)
  64 {
  65     BYTE* const ostart = (BYTE*) dst;
  66     BYTE* op = ostart;
  67     BYTE* const oend = ostart + dstSize;
  68
  69     unsigned maxSymbolValue = HUF_TABLELOG_MAX;
  70     U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
  71
  72     FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
  73     BYTE scratchBuffer[1<<MAX_FSE_TABLELOG_FOR_HUFF_HEADER];
  74
  75     unsigned count[HUF_TABLELOG_MAX+1];
  76     S16 norm[HUF_TABLELOG_MAX+1];
  77
  78     /* init conditions */
  79     if (wtSize <= 1) return 0;  /* Not compressible */
  80
  81     /* Scan input and build symbol stats */
  82     {   unsigned const maxCount = HIST_count_simple(count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
  83         if (maxCount == wtSize) return 1;   /* only a single symbol in src : rle */
  84         if (maxCount == 1) return 0;        /* each symbol present maximum once => not compressible */
  85     }
  86
  87     tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
  88     CHECK_F( FSE_normalizeCount(norm, tableLog, count, wtSize, maxSymbolValue) );
  89
  90     /* Write table description header */
  91     {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), norm, maxSymbolValue, tableLog) );
  92         op += hSize;
  93     }
  94
  95     /* Compress */
  96     CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, sizeof(scratchBuffer)) );
  97     {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, CTable) );
  98         if (cSize == 0) return 0;   /* not enough space for compressed data */
  99         op += cSize;
 100     }
 101
 102     return (size_t)(op-ostart);
 103 }
 104
 105
 106 struct HUF_CElt_s {
 107   U16  val;
 108   BYTE nbBits;
 109 };   /* typedef'd to HUF_CElt within "huf.h" */
 110
 111 /*! HUF_writeCTable() :
 112     `CTable` : Huffman tree to save, using huf representation.
 113     @return : size of saved CTable */
 114 size_t HUF_writeCTable (void* dst, size_t maxDstSize,
 115                         const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
 116 {
 117     BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
 118     BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
 119     BYTE* op = (BYTE*)dst;
 120     U32 n;
 121
 122      /* check conditions */
 123     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
 124
 125     /* convert to weight */
 126     bitsToWeight[0] = 0;
 127     for (n=1; n<huffLog+1; n++)
 128         bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
 129     for (n=0; n<maxSymbolValue; n++)
 130         huffWeight[n] = bitsToWeight[CTable[n].nbBits];
 131
 132     /* attempt weights compression by FSE */
 133     {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, huffWeight, maxSymbolValue) );
 134         if ((hSize>1) & (hSize < maxSymbolValue/2)) {   /* FSE compressed */
 135             op[0] = (BYTE)hSize;
 136             return hSize+1;
 137     }   }
 138
 139     /* write raw values as 4-bits (max : 15) */
 140     if (maxSymbolValue > (256-128)) return ERROR(GENERIC);   /* should not happen : likely means source cannot be compressed */
 141     if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall);   /* not enough space within dst buffer */
 142     op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
 143     huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
 144     for (n=0; n<maxSymbolValue; n+=2)
 145         op[(n/2)+1] = (BYTE)((huffWeight[n] << 4) + huffWeight[n+1]);
 146     return ((maxSymbolValue+1)/2) + 1;
 147 }
 148
 149
 150 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
 151 {
 152     BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];   /* init not required, even though some static analyzer may complain */
 153     U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
 154     U32 tableLog = 0;
 155     U32 nbSymbols = 0;
 156
 157     /* get symbol weights */
 158     CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
 159
 160     /* check result */
 161     if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
 162     if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
 163
 164     /* Prepare base value per rank */
 165     {   U32 n, nextRankStart = 0;
 166         for (n=1; n<=tableLog; n++) {
 167             U32 current = nextRankStart;
 168             nextRankStart += (rankVal[n] << (n-1));
 169             rankVal[n] = current;
 170     }   }
 171
 172     /* fill nbBits */
 173     *hasZeroWeights = 0;
 174     {   U32 n; for (n=0; n<nbSymbols; n++) {
 175             const U32 w = huffWeight[n];
 176             *hasZeroWeights |= (w == 0);
 177             CTable[n].nbBits = (BYTE)(tableLog + 1 - w) & -(w != 0);
 178     }   }
 179
 180     /* fill val */
 181     {   U16 nbPerRank[HUF_TABLELOG_MAX+2]  = {0};  /* support w=0=>n=tableLog+1 */
 182         U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
 183         { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; }
 184         /* determine stating value per rank */
 185         valPerRank[tableLog+1] = 0;   /* for w==0 */
 186         {   U16 min = 0;
 187             U32 n; for (n=tableLog; n>0; n--) {  /* start at n=tablelog <-> w=1 */
 188                 valPerRank[n] = min;     /* get starting value within each rank */
 189                 min += nbPerRank[n];
 190                 min >>= 1;
 191         }   }
 192         /* assign value within rank, symbol order */
 193         { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; }
 194     }
 195
 196     *maxSymbolValuePtr = nbSymbols - 1;
 197     return readSize;
 198 }
 199
 200 U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue)
 201 {
 202     const HUF_CElt* table = (const HUF_CElt*)symbolTable;
 203     assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
 204     return table[symbolValue].nbBits;
 205 }
 206
 207
 208 typedef struct nodeElt_s {
 209     U32 count;
 210     U16 parent;
 211     BYTE byte;
 212     BYTE nbBits;
 213 } nodeElt;
 214
 215 static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
 216 {
 217     const U32 largestBits = huffNode[lastNonNull].nbBits;
 218     if (largestBits <= maxNbBits) return largestBits;   /* early exit : no elt > maxNbBits */
 219
 220     /* there are several too large elements (at least >= 2) */
 221     {   int totalCost = 0;
 222         const U32 baseCost = 1 << (largestBits - maxNbBits);
 223         int n = (int)lastNonNull;
 224
 225         while (huffNode[n].nbBits > maxNbBits) {
 226             totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
 227             huffNode[n].nbBits = (BYTE)maxNbBits;
 228             n --;
 229         }  /* n stops at huffNode[n].nbBits <= maxNbBits */
 230         while (huffNode[n].nbBits == maxNbBits) n--;   /* n end at index of smallest symbol using < maxNbBits */
 231
 232         /* renorm totalCost */
 233         totalCost >>= (largestBits - maxNbBits);  /* note : totalCost is necessarily a multiple of baseCost */
 234
 235         /* repay normalized cost */
 236         {   U32 const noSymbol = 0xF0F0F0F0;
 237             U32 rankLast[HUF_TABLELOG_MAX+2];
 238
 239             /* Get pos of last (smallest) symbol per rank */
 240             memset(rankLast, 0xF0, sizeof(rankLast));
 241             {   U32 currentNbBits = maxNbBits;
 242                 int pos;
 243                 for (pos=n ; pos >= 0; pos--) {
 244                     if (huffNode[pos].nbBits >= currentNbBits) continue;
 245                     currentNbBits = huffNode[pos].nbBits;   /* < maxNbBits */
 246                     rankLast[maxNbBits-currentNbBits] = (U32)pos;
 247             }   }
 248
 249             while (totalCost > 0) {
 250                 U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1;
 251                 for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
 252                     U32 const highPos = rankLast[nBitsToDecrease];
 253                     U32 const lowPos = rankLast[nBitsToDecrease-1];
 254                     if (highPos == noSymbol) continue;
 255                     if (lowPos == noSymbol) break;
 256                     {   U32 const highTotal = huffNode[highPos].count;
 257                         U32 const lowTotal = 2 * huffNode[lowPos].count;
 258                         if (highTotal <= lowTotal) break;
 259                 }   }
 260                 /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
 261                 /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */
 262                 while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
 263                     nBitsToDecrease ++;
 264                 totalCost -= 1 << (nBitsToDecrease-1);
 265                 if (rankLast[nBitsToDecrease-1] == noSymbol)
 266                     rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];   /* this rank is no longer empty */
 267                 huffNode[rankLast[nBitsToDecrease]].nbBits ++;
 268                 if (rankLast[nBitsToDecrease] == 0)    /* special case, reached largest symbol */
 269                     rankLast[nBitsToDecrease] = noSymbol;
 270                 else {
 271                     rankLast[nBitsToDecrease]--;
 272                     if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
 273                         rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
 274             }   }   /* while (totalCost > 0) */
 275
 276             while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
 277                 if (rankLast[1] == noSymbol) {  /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
 278                     while (huffNode[n].nbBits == maxNbBits) n--;
 279                     huffNode[n+1].nbBits--;
 280                     assert(n >= 0);
 281                     rankLast[1] = (U32)(n+1);
 282                     totalCost++;
 283                     continue;
 284                 }
 285                 huffNode[ rankLast[1] + 1 ].nbBits--;
 286                 rankLast[1]++;
 287                 totalCost ++;
 288     }   }   }   /* there are several too large elements (at least >= 2) */
 289
 290     return maxNbBits;
 291 }
 292
 293 typedef struct {
 294     U32 base;
 295     U32 current;
 296 } rankPos;
 297
 298 typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
 299
 300 #define RANK_POSITION_TABLE_SIZE 32
 301
 302 typedef struct {
 303   huffNodeTable huffNodeTbl;
 304   rankPos rankPosition[RANK_POSITION_TABLE_SIZE];
 305 } HUF_buildCTable_wksp_tables;
 306
 307 static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValue, rankPos* rankPosition)
 308 {
 309     U32 n;
 310
 311     memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
 312     for (n=0; n<=maxSymbolValue; n++) {
 313         U32 r = BIT_highbit32(count[n] + 1);
 314         rankPosition[r].base ++;
 315     }
 316     for (n=30; n>0; n--) rankPosition[n-1].base += rankPosition[n].base;
 317     for (n=0; n<32; n++) rankPosition[n].current = rankPosition[n].base;
 318     for (n=0; n<=maxSymbolValue; n++) {
 319         U32 const c = count[n];
 320         U32 const r = BIT_highbit32(c+1) + 1;
 321         U32 pos = rankPosition[r].current++;
 322         while ((pos > rankPosition[r].base) && (c > huffNode[pos-1].count)) {
 323             huffNode[pos] = huffNode[pos-1];
 324             pos--;
 325         }
 326         huffNode[pos].count = c;
 327         huffNode[pos].byte  = (BYTE)n;
 328     }
 329 }
 330
 331
 332 /** HUF_buildCTable_wksp() :
 333  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
 334  *  `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as sizeof(HUF_buildCTable_wksp_tables).
 335  */
 336 #define STARTNODE (HUF_SYMBOLVALUE_MAX+1)
 337
 338 size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
 339 {
 340     HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)workSpace;
 341     nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
 342     nodeElt* const huffNode = huffNode0+1;
 343     int nonNullRank;
 344     int lowS, lowN;
 345     int nodeNb = STARTNODE;
 346     int n, nodeRoot;
 347
 348     /* safety checks */
 349     if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
 350     if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
 351       return ERROR(workSpace_tooSmall);
 352     if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
 353     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
 354       return ERROR(maxSymbolValue_tooLarge);
 355     memset(huffNode0, 0, sizeof(huffNodeTable));
 356
 357     /* sort, decreasing order */
 358     HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
 359
 360     /* init for parents */
 361     nonNullRank = (int)maxSymbolValue;
 362     while(huffNode[nonNullRank].count == 0) nonNullRank--;
 363     lowS = nonNullRank; nodeRoot = nodeNb + lowS - 1; lowN = nodeNb;
 364     huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count;
 365     huffNode[lowS].parent = huffNode[lowS-1].parent = (U16)nodeNb;
 366     nodeNb++; lowS-=2;
 367     for (n=nodeNb; n<=nodeRoot; n++) huffNode[n].count = (U32)(1U<<30);
 368     huffNode0[0].count = (U32)(1U<<31);  /* fake entry, strong barrier */
 369
 370     /* create parents */
 371     while (nodeNb <= nodeRoot) {
 372         int const n1 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
 373         int const n2 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
 374         huffNode[nodeNb].count = huffNode[n1].count + huffNode[n2].count;
 375         huffNode[n1].parent = huffNode[n2].parent = (U16)nodeNb;
 376         nodeNb++;
 377     }
 378
 379     /* distribute weights (unlimited tree height) */
 380     huffNode[nodeRoot].nbBits = 0;
 381     for (n=nodeRoot-1; n>=STARTNODE; n--)
 382         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
 383     for (n=0; n<=nonNullRank; n++)
 384         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
 385
 386     /* enforce maxTableLog */
 387     maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
 388
 389     /* fill result into tree (val, nbBits) */
 390     {   U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
 391         U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
 392         int const alphabetSize = (int)(maxSymbolValue + 1);
 393         if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
 394         for (n=0; n<=nonNullRank; n++)
 395             nbPerRank[huffNode[n].nbBits]++;
 396         /* determine stating value per rank */
 397         {   U16 min = 0;
 398             for (n=(int)maxNbBits; n>0; n--) {
 399                 valPerRank[n] = min;      /* get starting value within each rank */
 400                 min += nbPerRank[n];
 401                 min >>= 1;
 402         }   }
 403         for (n=0; n<alphabetSize; n++)
 404             tree[huffNode[n].byte].nbBits = huffNode[n].nbBits;   /* push nbBits per symbol, symbol order */
 405         for (n=0; n<alphabetSize; n++)
 406             tree[n].val = valPerRank[tree[n].nbBits]++;   /* assign value within rank, symbol order */
 407     }
 408
 409     return maxNbBits;
 410 }
 411
 412 /** HUF_buildCTable() :
 413  * @return : maxNbBits
 414  *  Note : count is used before tree is written, so they can safely overlap
 415  */
 416 size_t HUF_buildCTable (HUF_CElt* tree, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits)
 417 {
 418     HUF_buildCTable_wksp_tables workspace;
 419     return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, &workspace, sizeof(workspace));
 420 }
 421
 422 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
 423 {
 424     size_t nbBits = 0;
 425     int s;
 426     for (s = 0; s <= (int)maxSymbolValue; ++s) {
 427         nbBits += CTable[s].nbBits * count[s];
 428     }
 429     return nbBits >> 3;
 430 }
 431
 432 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
 433   int bad = 0;
 434   int s;
 435   for (s = 0; s <= (int)maxSymbolValue; ++s) {
 436     bad |= (count[s] != 0) & (CTable[s].nbBits == 0);
 437   }
 438   return !bad;
 439 }
 440
 441 size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
 442
 443 FORCE_INLINE_TEMPLATE void
 444 HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
 445 {
 446     BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
 447 }
 448
 449 #define HUF_FLUSHBITS(s)  BIT_flushBits(s)
 450
 451 #define HUF_FLUSHBITS_1(stream) \
 452     if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
 453
 454 #define HUF_FLUSHBITS_2(stream) \
 455     if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream)
 456
 457 FORCE_INLINE_TEMPLATE size_t
 458 HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
 459                                    const void* src, size_t srcSize,
 460                                    const HUF_CElt* CTable)
 461 {
 462     const BYTE* ip = (const BYTE*) src;
 463     BYTE* const ostart = (BYTE*)dst;
 464     BYTE* const oend = ostart + dstSize;
 465     BYTE* op = ostart;
 466     size_t n;
 467     BIT_CStream_t bitC;
 468
 469     /* init */
 470     if (dstSize < 8) return 0;   /* not enough space to compress */
 471     { size_t const initErr = BIT_initCStream(&bitC, op, (size_t)(oend-op));
 472       if (HUF_isError(initErr)) return 0; }
 473
 474     n = srcSize & ~3;  /* join to mod 4 */
 475     switch (srcSize & 3)
 476     {
 477         case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
 478                  HUF_FLUSHBITS_2(&bitC);
 479                  /* fall-through */
 480         case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
 481                  HUF_FLUSHBITS_1(&bitC);
 482                  /* fall-through */
 483         case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
 484                  HUF_FLUSHBITS(&bitC);
 485                  /* fall-through */
 486         case 0 : /* fall-through */
 487         default: break;
 488     }
 489
 490     for (; n>0; n-=4) {  /* note : n&3==0 at this stage */
 491         HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
 492         HUF_FLUSHBITS_1(&bitC);
 493         HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
 494         HUF_FLUSHBITS_2(&bitC);
 495         HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
 496         HUF_FLUSHBITS_1(&bitC);
 497         HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
 498         HUF_FLUSHBITS(&bitC);
 499     }
 500
 501     return BIT_closeCStream(&bitC);
 502 }
 503
 504 #if DYNAMIC_BMI2
 505
 506 static TARGET_ATTRIBUTE("bmi2") size_t
 507 HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
 508                                    const void* src, size_t srcSize,
 509                                    const HUF_CElt* CTable)
 510 {
 511     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 512 }
 513
 514 static size_t
 515 HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
 516                                       const void* src, size_t srcSize,
 517                                       const HUF_CElt* CTable)
 518 {
 519     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 520 }
 521
 522 static size_t
 523 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
 524                               const void* src, size_t srcSize,
 525                               const HUF_CElt* CTable, const int bmi2)
 526 {
 527     if (bmi2) {
 528         return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
 529     }
 530     return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
 531 }
 532
 533 #else
 534
 535 static size_t
 536 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
 537                               const void* src, size_t srcSize,
 538                               const HUF_CElt* CTable, const int bmi2)
 539 {
 540     (void)bmi2;
 541     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 542 }
 543
 544 #endif
 545
 546 size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
 547 {
 548     return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
 549 }
 550
 551
 552 static size_t
 553 HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
 554                               const void* src, size_t srcSize,
 555                               const HUF_CElt* CTable, int bmi2)
 556 {
 557     size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
 558     const BYTE* ip = (const BYTE*) src;
 559     const BYTE* const iend = ip + srcSize;
 560     BYTE* const ostart = (BYTE*) dst;
 561     BYTE* const oend = ostart + dstSize;
 562     BYTE* op = ostart;
 563
 564     if (dstSize < 6 + 1 + 1 + 1 + 8) return 0;   /* minimum space to compress successfully */
 565     if (srcSize < 12) return 0;   /* no saving possible : too small input */
 566     op += 6;   /* jumpTable */
 567
 568     assert(op <= oend);
 569     {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
 570         if (cSize==0) return 0;
 571         assert(cSize <= 65535);
 572         MEM_writeLE16(ostart, (U16)cSize);
 573         op += cSize;
 574     }
 575
 576     ip += segmentSize;
 577     assert(op <= oend);
 578     {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
 579         if (cSize==0) return 0;
 580         assert(cSize <= 65535);
 581         MEM_writeLE16(ostart+2, (U16)cSize);
 582         op += cSize;
 583     }
 584
 585     ip += segmentSize;
 586     assert(op <= oend);
 587     {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
 588         if (cSize==0) return 0;
 589         assert(cSize <= 65535);
 590         MEM_writeLE16(ostart+4, (U16)cSize);
 591         op += cSize;
 592     }
 593
 594     ip += segmentSize;
 595     assert(op <= oend);
 596     assert(ip <= iend);
 597     {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) );
 598         if (cSize==0) return 0;
 599         op += cSize;
 600     }
 601
 602     return (size_t)(op-ostart);
 603 }
 604
 605 size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
 606 {
 607     return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
 608 }
 609
 610 typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
 611
 612 static size_t HUF_compressCTable_internal(
 613                 BYTE* const ostart, BYTE* op, BYTE* const oend,
 614                 const void* src, size_t srcSize,
 615                 HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int bmi2)
 616 {
 617     size_t const cSize = (nbStreams==HUF_singleStream) ?
 618                          HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2) :
 619                          HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2);
 620     if (HUF_isError(cSize)) { return cSize; }
 621     if (cSize==0) { return 0; }   /* uncompressible */
 622     op += cSize;
 623     /* check compressibility */
 624     assert(op >= ostart);
 625     if ((size_t)(op-ostart) >= srcSize-1) { return 0; }
 626     return (size_t)(op-ostart);
 627 }
 628
 629 typedef struct {
 630     unsigned count[HUF_SYMBOLVALUE_MAX + 1];
 631     HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
 632     HUF_buildCTable_wksp_tables buildCTable_wksp;
 633 } HUF_compress_tables_t;
 634
 635 /* HUF_compress_internal() :
 636  * `workSpace` must a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
 637 static size_t
 638 HUF_compress_internal (void* dst, size_t dstSize,
 639                  const void* src, size_t srcSize,
 640                        unsigned maxSymbolValue, unsigned huffLog,
 641                        HUF_nbStreams_e nbStreams,
 642                        void* workSpace, size_t wkspSize,
 643                        HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
 644                  const int bmi2)
 645 {
 646     HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace;
 647     BYTE* const ostart = (BYTE*)dst;
 648     BYTE* const oend = ostart + dstSize;
 649     BYTE* op = ostart;
 650
 651     HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE);
 652
 653     /* checks & inits */
 654     if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
 655     if (wkspSize < HUF_WORKSPACE_SIZE) return ERROR(workSpace_tooSmall);
 656     if (!srcSize) return 0;  /* Uncompressed */
 657     if (!dstSize) return 0;  /* cannot fit anything within dst budget */
 658     if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);   /* current block size limit */
 659     if (huffLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
 660     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
 661     if (!maxSymbolValue) maxSymbolValue = HUF_SYMBOLVALUE_MAX;
 662     if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
 663
 664     /* Heuristic : If old table is valid, use it for small inputs */
 665     if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
 666         return HUF_compressCTable_internal(ostart, op, oend,
 667                                            src, srcSize,
 668                                            nbStreams, oldHufTable, bmi2);
 669     }
 670
 671     /* Scan input and build symbol stats */
 672     {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, workSpace, wkspSize) );
 673         if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
 674         if (largest <= (srcSize >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
 675     }
 676
 677     /* Check validity of previous table */
 678     if ( repeat
 679       && *repeat == HUF_repeat_check
 680       && !HUF_validateCTable(oldHufTable, table->count, maxSymbolValue)) {
 681         *repeat = HUF_repeat_none;
 682     }
 683     /* Heuristic : use existing table for small inputs */
 684     if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
 685         return HUF_compressCTable_internal(ostart, op, oend,
 686                                            src, srcSize,
 687                                            nbStreams, oldHufTable, bmi2);
 688     }
 689
 690     /* Build Huffman Tree */
 691     huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
 692     {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
 693                                             maxSymbolValue, huffLog,
 694                                             &table->buildCTable_wksp, sizeof(table->buildCTable_wksp));
 695         CHECK_F(maxBits);
 696         huffLog = (U32)maxBits;
 697         /* Zero unused symbols in CTable, so we can check it for validity */
 698         memset(table->CTable + (maxSymbolValue + 1), 0,
 699                sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt)));
 700     }
 701
 702     /* Write table description header */
 703     {   CHECK_V_F(hSize, HUF_writeCTable (op, dstSize, table->CTable, maxSymbolValue, huffLog) );
 704         /* Check if using previous huffman table is beneficial */
 705         if (repeat && *repeat != HUF_repeat_none) {
 706             size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
 707             size_t const newSize = HUF_estimateCompressedSize(table->CTable, table->count, maxSymbolValue);
 708             if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
 709                 return HUF_compressCTable_internal(ostart, op, oend,
 710                                                    src, srcSize,
 711                                                    nbStreams, oldHufTable, bmi2);
 712         }   }
 713
 714         /* Use the new huffman table */
 715         if (hSize + 12ul >= srcSize) { return 0; }
 716         op += hSize;
 717         if (repeat) { *repeat = HUF_repeat_none; }
 718         if (oldHufTable)
 719             memcpy(oldHufTable, table->CTable, sizeof(table->CTable));  /* Save new table */
 720     }
 721     return HUF_compressCTable_internal(ostart, op, oend,
 722                                        src, srcSize,
 723                                        nbStreams, table->CTable, bmi2);
 724 }
 725
 726
 727 size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
 728                       const void* src, size_t srcSize,
 729                       unsigned maxSymbolValue, unsigned huffLog,
 730                       void* workSpace, size_t wkspSize)
 731 {
 732     return HUF_compress_internal(dst, dstSize, src, srcSize,
 733                                  maxSymbolValue, huffLog, HUF_singleStream,
 734                                  workSpace, wkspSize,
 735                                  NULL, NULL, 0, 0 /*bmi2*/);
 736 }
 737
 738 size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
 739                       const void* src, size_t srcSize,
 740                       unsigned maxSymbolValue, unsigned huffLog,
 741                       void* workSpace, size_t wkspSize,
 742                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
 743 {
 744     return HUF_compress_internal(dst, dstSize, src, srcSize,
 745                                  maxSymbolValue, huffLog, HUF_singleStream,
 746                                  workSpace, wkspSize, hufTable,
 747                                  repeat, preferRepeat, bmi2);
 748 }
 749
 750 size_t HUF_compress1X (void* dst, size_t dstSize,
 751                  const void* src, size_t srcSize,
 752                  unsigned maxSymbolValue, unsigned huffLog)
 753 {
 754     unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
 755     return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
 756 }
 757
 758 /* HUF_compress4X_repeat():
 759  * compress input using 4 streams.
 760  * provide workspace to generate compression tables */
 761 size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
 762                       const void* src, size_t srcSize,
 763                       unsigned maxSymbolValue, unsigned huffLog,
 764                       void* workSpace, size_t wkspSize)
 765 {
 766     return HUF_compress_internal(dst, dstSize, src, srcSize,
 767                                  maxSymbolValue, huffLog, HUF_fourStreams,
 768                                  workSpace, wkspSize,
 769                                  NULL, NULL, 0, 0 /*bmi2*/);
 770 }
 771
 772 /* HUF_compress4X_repeat():
 773  * compress input using 4 streams.
 774  * re-use an existing huffman compression table */
 775 size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
 776                       const void* src, size_t srcSize,
 777                       unsigned maxSymbolValue, unsigned huffLog,
 778                       void* workSpace, size_t wkspSize,
 779                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
 780 {
 781     return HUF_compress_internal(dst, dstSize, src, srcSize,
 782                                  maxSymbolValue, huffLog, HUF_fourStreams,
 783                                  workSpace, wkspSize,
 784                                  hufTable, repeat, preferRepeat, bmi2);
 785 }
 786
 787 size_t HUF_compress2 (void* dst, size_t dstSize,
 788                 const void* src, size_t srcSize,
 789                 unsigned maxSymbolValue, unsigned huffLog)
 790 {
 791     unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
 792     return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
 793 }
 794
 795 size_t HUF_compress (void* dst, size_t maxDstSize, const void* src, size_t srcSize)
 796 {
 797     return HUF_compress2(dst, maxDstSize, src, srcSize, 255, HUF_TABLELOG_DEFAULT);
 798 }