1 // SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0-only 2 /* 3 * Copyright (c) 2016-2020, Przemyslaw Skibinski, Yann Collet, Facebook, Inc. 4 * All rights reserved. 5 * 6 * This source code is licensed under both the BSD-style license (found in the 7 * LICENSE file in the root directory of this source tree) and the GPLv2 (found 8 * in the COPYING file in the root directory of this source tree). 9 * You may select, at your option, one of the above-listed licenses. 10 */ 11 12 #include "zstd_compress_internal.h" 13 #include "hist.h" 14 #include "zstd_opt.h" 15 16 17 #define ZSTD_LITFREQ_ADD 2 /* scaling factor for litFreq, so that frequencies adapt faster to new stats */ 18 #define ZSTD_FREQ_DIV 4 /* log factor when using previous stats to init next stats */ 19 #define ZSTD_MAX_PRICE (1<<30) 20 21 #define ZSTD_PREDEF_THRESHOLD 1024 /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */ 22 23 24 /*-************************************* 25 * Price functions for optimal parser 26 ***************************************/ 27 28 #if 0 /* approximation at bit level */ 29 # define BITCOST_ACCURACY 0 30 # define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY) 31 # define WEIGHT(stat) ((void)opt, ZSTD_bitWeight(stat)) 32 #elif 0 /* fractional bit accuracy */ 33 # define BITCOST_ACCURACY 8 34 # define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY) 35 # define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat)) 36 #else /* opt==approx, ultra==accurate */ 37 # define BITCOST_ACCURACY 8 38 # define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY) 39 # define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat)) 40 #endif 41 42 MEM_STATIC U32 ZSTD_bitWeight(U32 stat) 43 { 44 return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER); 45 } 46 47 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat) 48 { 49 U32 const stat = rawStat + 1; 50 U32 const hb = ZSTD_highbit32(stat); 51 U32 const BWeight = hb * BITCOST_MULTIPLIER; 52 U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb; 53 U32 const weight = BWeight + FWeight; 54 assert(hb + BITCOST_ACCURACY < 31); 55 return weight; 56 } 57 58 #if (DEBUGLEVEL>=2) 59 /* debugging function, 60 * @return price in bytes as fractional value 61 * for debug messages only */ 62 MEM_STATIC double ZSTD_fCost(U32 price) 63 { 64 return (double)price / (BITCOST_MULTIPLIER*8); 65 } 66 #endif 67 68 static int ZSTD_compressedLiterals(optState_t const* const optPtr) 69 { 70 return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed; 71 } 72 73 static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel) 74 { 75 if (ZSTD_compressedLiterals(optPtr)) 76 optPtr->litSumBasePrice = WEIGHT(optPtr->litSum, optLevel); 77 optPtr->litLengthSumBasePrice = WEIGHT(optPtr->litLengthSum, optLevel); 78 optPtr->matchLengthSumBasePrice = WEIGHT(optPtr->matchLengthSum, optLevel); 79 optPtr->offCodeSumBasePrice = WEIGHT(optPtr->offCodeSum, optLevel); 80 } 81 82 83 /* ZSTD_downscaleStat() : 84 * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus) 85 * return the resulting sum of elements */ 86 static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus) 87 { 88 U32 s, sum=0; 89 DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1); 90 assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31); 91 for (s=0; s<lastEltIndex+1; s++) { 92 table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus)); 93 sum += table[s]; 94 } 95 return sum; 96 } 97 98 /* ZSTD_rescaleFreqs() : 99 * if first block (detected by optPtr->litLengthSum == 0) : init statistics 100 * take hints from dictionary if there is one 101 * or init from zero, using src for literals stats, or flat 1 for match symbols 102 * otherwise downscale existing stats, to be used as seed for next block. 103 */ 104 static void 105 ZSTD_rescaleFreqs(optState_t* const optPtr, 106 const BYTE* const src, size_t const srcSize, 107 int const optLevel) 108 { 109 int const compressedLiterals = ZSTD_compressedLiterals(optPtr); 110 DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize); 111 optPtr->priceType = zop_dynamic; 112 113 if (optPtr->litLengthSum == 0) { /* first block : init */ 114 if (srcSize <= ZSTD_PREDEF_THRESHOLD) { /* heuristic */ 115 DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef"); 116 optPtr->priceType = zop_predef; 117 } 118 119 assert(optPtr->symbolCosts != NULL); 120 if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) { 121 /* huffman table presumed generated by dictionary */ 122 optPtr->priceType = zop_dynamic; 123 124 if (compressedLiterals) { 125 unsigned lit; 126 assert(optPtr->litFreq != NULL); 127 optPtr->litSum = 0; 128 for (lit=0; lit<=MaxLit; lit++) { 129 U32 const scaleLog = 11; /* scale to 2K */ 130 U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit); 131 assert(bitCost <= scaleLog); 132 optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/; 133 optPtr->litSum += optPtr->litFreq[lit]; 134 } } 135 136 { unsigned ll; 137 FSE_CState_t llstate; 138 FSE_initCState(&llstate, optPtr->symbolCosts->fse.litlengthCTable); 139 optPtr->litLengthSum = 0; 140 for (ll=0; ll<=MaxLL; ll++) { 141 U32 const scaleLog = 10; /* scale to 1K */ 142 U32 const bitCost = FSE_getMaxNbBits(llstate.symbolTT, ll); 143 assert(bitCost < scaleLog); 144 optPtr->litLengthFreq[ll] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/; 145 optPtr->litLengthSum += optPtr->litLengthFreq[ll]; 146 } } 147 148 { unsigned ml; 149 FSE_CState_t mlstate; 150 FSE_initCState(&mlstate, optPtr->symbolCosts->fse.matchlengthCTable); 151 optPtr->matchLengthSum = 0; 152 for (ml=0; ml<=MaxML; ml++) { 153 U32 const scaleLog = 10; 154 U32 const bitCost = FSE_getMaxNbBits(mlstate.symbolTT, ml); 155 assert(bitCost < scaleLog); 156 optPtr->matchLengthFreq[ml] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/; 157 optPtr->matchLengthSum += optPtr->matchLengthFreq[ml]; 158 } } 159 160 { unsigned of; 161 FSE_CState_t ofstate; 162 FSE_initCState(&ofstate, optPtr->symbolCosts->fse.offcodeCTable); 163 optPtr->offCodeSum = 0; 164 for (of=0; of<=MaxOff; of++) { 165 U32 const scaleLog = 10; 166 U32 const bitCost = FSE_getMaxNbBits(ofstate.symbolTT, of); 167 assert(bitCost < scaleLog); 168 optPtr->offCodeFreq[of] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/; 169 optPtr->offCodeSum += optPtr->offCodeFreq[of]; 170 } } 171 172 } else { /* not a dictionary */ 173 174 assert(optPtr->litFreq != NULL); 175 if (compressedLiterals) { 176 unsigned lit = MaxLit; 177 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize); /* use raw first block to init statistics */ 178 optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1); 179 } 180 181 { unsigned ll; 182 for (ll=0; ll<=MaxLL; ll++) 183 optPtr->litLengthFreq[ll] = 1; 184 } 185 optPtr->litLengthSum = MaxLL+1; 186 187 { unsigned ml; 188 for (ml=0; ml<=MaxML; ml++) 189 optPtr->matchLengthFreq[ml] = 1; 190 } 191 optPtr->matchLengthSum = MaxML+1; 192 193 { unsigned of; 194 for (of=0; of<=MaxOff; of++) 195 optPtr->offCodeFreq[of] = 1; 196 } 197 optPtr->offCodeSum = MaxOff+1; 198 199 } 200 201 } else { /* new block : re-use previous statistics, scaled down */ 202 203 if (compressedLiterals) 204 optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1); 205 optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0); 206 optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0); 207 optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0); 208 } 209 210 ZSTD_setBasePrices(optPtr, optLevel); 211 } 212 213 /* ZSTD_rawLiteralsCost() : 214 * price of literals (only) in specified segment (which length can be 0). 215 * does not include price of literalLength symbol */ 216 static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength, 217 const optState_t* const optPtr, 218 int optLevel) 219 { 220 if (litLength == 0) return 0; 221 222 if (!ZSTD_compressedLiterals(optPtr)) 223 return (litLength << 3) * BITCOST_MULTIPLIER; /* Uncompressed - 8 bytes per literal. */ 224 225 if (optPtr->priceType == zop_predef) 226 return (litLength*6) * BITCOST_MULTIPLIER; /* 6 bit per literal - no statistic used */ 227 228 /* dynamic statistics */ 229 { U32 price = litLength * optPtr->litSumBasePrice; 230 U32 u; 231 for (u=0; u < litLength; u++) { 232 assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice); /* literal cost should never be negative */ 233 price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel); 234 } 235 return price; 236 } 237 } 238 239 /* ZSTD_litLengthPrice() : 240 * cost of literalLength symbol */ 241 static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel) 242 { 243 if (optPtr->priceType == zop_predef) return WEIGHT(litLength, optLevel); 244 245 /* dynamic statistics */ 246 { U32 const llCode = ZSTD_LLcode(litLength); 247 return (LL_bits[llCode] * BITCOST_MULTIPLIER) 248 + optPtr->litLengthSumBasePrice 249 - WEIGHT(optPtr->litLengthFreq[llCode], optLevel); 250 } 251 } 252 253 /* ZSTD_getMatchPrice() : 254 * Provides the cost of the match part (offset + matchLength) of a sequence 255 * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence. 256 * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */ 257 FORCE_INLINE_TEMPLATE U32 258 ZSTD_getMatchPrice(U32 const offset, 259 U32 const matchLength, 260 const optState_t* const optPtr, 261 int const optLevel) 262 { 263 U32 price; 264 U32 const offCode = ZSTD_highbit32(offset+1); 265 U32 const mlBase = matchLength - MINMATCH; 266 assert(matchLength >= MINMATCH); 267 268 if (optPtr->priceType == zop_predef) /* fixed scheme, do not use statistics */ 269 return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER); 270 271 /* dynamic statistics */ 272 price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel)); 273 if ((optLevel<2) /*static*/ && offCode >= 20) 274 price += (offCode-19)*2 * BITCOST_MULTIPLIER; /* handicap for long distance offsets, favor decompression speed */ 275 276 /* match Length */ 277 { U32 const mlCode = ZSTD_MLcode(mlBase); 278 price += (ML_bits[mlCode] * BITCOST_MULTIPLIER) + (optPtr->matchLengthSumBasePrice - WEIGHT(optPtr->matchLengthFreq[mlCode], optLevel)); 279 } 280 281 price += BITCOST_MULTIPLIER / 5; /* heuristic : make matches a bit more costly to favor less sequences -> faster decompression speed */ 282 283 DEBUGLOG(8, "ZSTD_getMatchPrice(ml:%u) = %u", matchLength, price); 284 return price; 285 } 286 287 /* ZSTD_updateStats() : 288 * assumption : literals + litLengtn <= iend */ 289 static void ZSTD_updateStats(optState_t* const optPtr, 290 U32 litLength, const BYTE* literals, 291 U32 offsetCode, U32 matchLength) 292 { 293 /* literals */ 294 if (ZSTD_compressedLiterals(optPtr)) { 295 U32 u; 296 for (u=0; u < litLength; u++) 297 optPtr->litFreq[literals[u]] += ZSTD_LITFREQ_ADD; 298 optPtr->litSum += litLength*ZSTD_LITFREQ_ADD; 299 } 300 301 /* literal Length */ 302 { U32 const llCode = ZSTD_LLcode(litLength); 303 optPtr->litLengthFreq[llCode]++; 304 optPtr->litLengthSum++; 305 } 306 307 /* match offset code (0-2=>repCode; 3+=>offset+2) */ 308 { U32 const offCode = ZSTD_highbit32(offsetCode+1); 309 assert(offCode <= MaxOff); 310 optPtr->offCodeFreq[offCode]++; 311 optPtr->offCodeSum++; 312 } 313 314 /* match Length */ 315 { U32 const mlBase = matchLength - MINMATCH; 316 U32 const mlCode = ZSTD_MLcode(mlBase); 317 optPtr->matchLengthFreq[mlCode]++; 318 optPtr->matchLengthSum++; 319 } 320 } 321 322 323 /* ZSTD_readMINMATCH() : 324 * function safe only for comparisons 325 * assumption : memPtr must be at least 4 bytes before end of buffer */ 326 MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length) 327 { 328 switch (length) 329 { 330 default : 331 case 4 : return MEM_read32(memPtr); 332 case 3 : if (MEM_isLittleEndian()) 333 return MEM_read32(memPtr)<<8; 334 else 335 return MEM_read32(memPtr)>>8; 336 } 337 } 338 339 340 /* Update hashTable3 up to ip (excluded) 341 Assumption : always within prefix (i.e. not within extDict) */ 342 static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms, 343 U32* nextToUpdate3, 344 const BYTE* const ip) 345 { 346 U32* const hashTable3 = ms->hashTable3; 347 U32 const hashLog3 = ms->hashLog3; 348 const BYTE* const base = ms->window.base; 349 U32 idx = *nextToUpdate3; 350 U32 const target = (U32)(ip - base); 351 size_t const hash3 = ZSTD_hash3Ptr(ip, hashLog3); 352 assert(hashLog3 > 0); 353 354 while(idx < target) { 355 hashTable3[ZSTD_hash3Ptr(base+idx, hashLog3)] = idx; 356 idx++; 357 } 358 359 *nextToUpdate3 = target; 360 return hashTable3[hash3]; 361 } 362 363 364 /*-************************************* 365 * Binary Tree search 366 ***************************************/ 367 /** ZSTD_insertBt1() : add one or multiple positions to tree. 368 * ip : assumed <= iend-8 . 369 * @return : nb of positions added */ 370 static U32 ZSTD_insertBt1( 371 ZSTD_matchState_t* ms, 372 const BYTE* const ip, const BYTE* const iend, 373 U32 const mls, const int extDict) 374 { 375 const ZSTD_compressionParameters* const cParams = &ms->cParams; 376 U32* const hashTable = ms->hashTable; 377 U32 const hashLog = cParams->hashLog; 378 size_t const h = ZSTD_hashPtr(ip, hashLog, mls); 379 U32* const bt = ms->chainTable; 380 U32 const btLog = cParams->chainLog - 1; 381 U32 const btMask = (1 << btLog) - 1; 382 U32 matchIndex = hashTable[h]; 383 size_t commonLengthSmaller=0, commonLengthLarger=0; 384 const BYTE* const base = ms->window.base; 385 const BYTE* const dictBase = ms->window.dictBase; 386 const U32 dictLimit = ms->window.dictLimit; 387 const BYTE* const dictEnd = dictBase + dictLimit; 388 const BYTE* const prefixStart = base + dictLimit; 389 const BYTE* match; 390 const U32 current = (U32)(ip-base); 391 const U32 btLow = btMask >= current ? 0 : current - btMask; 392 U32* smallerPtr = bt + 2*(current&btMask); 393 U32* largerPtr = smallerPtr + 1; 394 U32 dummy32; /* to be nullified at the end */ 395 U32 const windowLow = ms->window.lowLimit; 396 U32 matchEndIdx = current+8+1; 397 size_t bestLength = 8; 398 U32 nbCompares = 1U << cParams->searchLog; 399 #ifdef ZSTD_C_PREDICT 400 U32 predictedSmall = *(bt + 2*((current-1)&btMask) + 0); 401 U32 predictedLarge = *(bt + 2*((current-1)&btMask) + 1); 402 predictedSmall += (predictedSmall>0); 403 predictedLarge += (predictedLarge>0); 404 #endif /* ZSTD_C_PREDICT */ 405 406 DEBUGLOG(8, "ZSTD_insertBt1 (%u)", current); 407 408 assert(ip <= iend-8); /* required for h calculation */ 409 hashTable[h] = current; /* Update Hash Table */ 410 411 assert(windowLow > 0); 412 while (nbCompares-- && (matchIndex >= windowLow)) { 413 U32* const nextPtr = bt + 2*(matchIndex & btMask); 414 size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ 415 assert(matchIndex < current); 416 417 #ifdef ZSTD_C_PREDICT /* note : can create issues when hlog small <= 11 */ 418 const U32* predictPtr = bt + 2*((matchIndex-1) & btMask); /* written this way, as bt is a roll buffer */ 419 if (matchIndex == predictedSmall) { 420 /* no need to check length, result known */ 421 *smallerPtr = matchIndex; 422 if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop the search */ 423 smallerPtr = nextPtr+1; /* new "smaller" => larger of match */ 424 matchIndex = nextPtr[1]; /* new matchIndex larger than previous (closer to current) */ 425 predictedSmall = predictPtr[1] + (predictPtr[1]>0); 426 continue; 427 } 428 if (matchIndex == predictedLarge) { 429 *largerPtr = matchIndex; 430 if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop the search */ 431 largerPtr = nextPtr; 432 matchIndex = nextPtr[0]; 433 predictedLarge = predictPtr[0] + (predictPtr[0]>0); 434 continue; 435 } 436 #endif 437 438 if (!extDict || (matchIndex+matchLength >= dictLimit)) { 439 assert(matchIndex+matchLength >= dictLimit); /* might be wrong if actually extDict */ 440 match = base + matchIndex; 441 matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend); 442 } else { 443 match = dictBase + matchIndex; 444 matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart); 445 if (matchIndex+matchLength >= dictLimit) 446 match = base + matchIndex; /* to prepare for next usage of match[matchLength] */ 447 } 448 449 if (matchLength > bestLength) { 450 bestLength = matchLength; 451 if (matchLength > matchEndIdx - matchIndex) 452 matchEndIdx = matchIndex + (U32)matchLength; 453 } 454 455 if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */ 456 break; /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */ 457 } 458 459 if (match[matchLength] < ip[matchLength]) { /* necessarily within buffer */ 460 /* match is smaller than current */ 461 *smallerPtr = matchIndex; /* update smaller idx */ 462 commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */ 463 if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop searching */ 464 smallerPtr = nextPtr+1; /* new "candidate" => larger than match, which was smaller than target */ 465 matchIndex = nextPtr[1]; /* new matchIndex, larger than previous and closer to current */ 466 } else { 467 /* match is larger than current */ 468 *largerPtr = matchIndex; 469 commonLengthLarger = matchLength; 470 if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop searching */ 471 largerPtr = nextPtr; 472 matchIndex = nextPtr[0]; 473 } } 474 475 *smallerPtr = *largerPtr = 0; 476 { U32 positions = 0; 477 if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384)); /* speed optimization */ 478 assert(matchEndIdx > current + 8); 479 return MAX(positions, matchEndIdx - (current + 8)); 480 } 481 } 482 483 FORCE_INLINE_TEMPLATE 484 void ZSTD_updateTree_internal( 485 ZSTD_matchState_t* ms, 486 const BYTE* const ip, const BYTE* const iend, 487 const U32 mls, const ZSTD_dictMode_e dictMode) 488 { 489 const BYTE* const base = ms->window.base; 490 U32 const target = (U32)(ip - base); 491 U32 idx = ms->nextToUpdate; 492 DEBUGLOG(6, "ZSTD_updateTree_internal, from %u to %u (dictMode:%u)", 493 idx, target, dictMode); 494 495 while(idx < target) { 496 U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict); 497 assert(idx < (U32)(idx + forward)); 498 idx += forward; 499 } 500 assert((size_t)(ip - base) <= (size_t)(U32)(-1)); 501 assert((size_t)(iend - base) <= (size_t)(U32)(-1)); 502 ms->nextToUpdate = target; 503 } 504 505 void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) { 506 ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict); 507 } 508 509 FORCE_INLINE_TEMPLATE 510 U32 ZSTD_insertBtAndGetAllMatches ( 511 ZSTD_match_t* matches, /* store result (found matches) in this table (presumed large enough) */ 512 ZSTD_matchState_t* ms, 513 U32* nextToUpdate3, 514 const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode, 515 const U32 rep[ZSTD_REP_NUM], 516 U32 const ll0, /* tells if associated literal length is 0 or not. This value must be 0 or 1 */ 517 const U32 lengthToBeat, 518 U32 const mls /* template */) 519 { 520 const ZSTD_compressionParameters* const cParams = &ms->cParams; 521 U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1); 522 const BYTE* const base = ms->window.base; 523 U32 const current = (U32)(ip-base); 524 U32 const hashLog = cParams->hashLog; 525 U32 const minMatch = (mls==3) ? 3 : 4; 526 U32* const hashTable = ms->hashTable; 527 size_t const h = ZSTD_hashPtr(ip, hashLog, mls); 528 U32 matchIndex = hashTable[h]; 529 U32* const bt = ms->chainTable; 530 U32 const btLog = cParams->chainLog - 1; 531 U32 const btMask= (1U << btLog) - 1; 532 size_t commonLengthSmaller=0, commonLengthLarger=0; 533 const BYTE* const dictBase = ms->window.dictBase; 534 U32 const dictLimit = ms->window.dictLimit; 535 const BYTE* const dictEnd = dictBase + dictLimit; 536 const BYTE* const prefixStart = base + dictLimit; 537 U32 const btLow = (btMask >= current) ? 0 : current - btMask; 538 U32 const windowLow = ZSTD_getLowestMatchIndex(ms, current, cParams->windowLog); 539 U32 const matchLow = windowLow ? windowLow : 1; 540 U32* smallerPtr = bt + 2*(current&btMask); 541 U32* largerPtr = bt + 2*(current&btMask) + 1; 542 U32 matchEndIdx = current+8+1; /* farthest referenced position of any match => detects repetitive patterns */ 543 U32 dummy32; /* to be nullified at the end */ 544 U32 mnum = 0; 545 U32 nbCompares = 1U << cParams->searchLog; 546 547 const ZSTD_matchState_t* dms = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL; 548 const ZSTD_compressionParameters* const dmsCParams = 549 dictMode == ZSTD_dictMatchState ? &dms->cParams : NULL; 550 const BYTE* const dmsBase = dictMode == ZSTD_dictMatchState ? dms->window.base : NULL; 551 const BYTE* const dmsEnd = dictMode == ZSTD_dictMatchState ? dms->window.nextSrc : NULL; 552 U32 const dmsHighLimit = dictMode == ZSTD_dictMatchState ? (U32)(dmsEnd - dmsBase) : 0; 553 U32 const dmsLowLimit = dictMode == ZSTD_dictMatchState ? dms->window.lowLimit : 0; 554 U32 const dmsIndexDelta = dictMode == ZSTD_dictMatchState ? windowLow - dmsHighLimit : 0; 555 U32 const dmsHashLog = dictMode == ZSTD_dictMatchState ? dmsCParams->hashLog : hashLog; 556 U32 const dmsBtLog = dictMode == ZSTD_dictMatchState ? dmsCParams->chainLog - 1 : btLog; 557 U32 const dmsBtMask = dictMode == ZSTD_dictMatchState ? (1U << dmsBtLog) - 1 : 0; 558 U32 const dmsBtLow = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit; 559 560 size_t bestLength = lengthToBeat-1; 561 DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", current); 562 563 /* check repCode */ 564 assert(ll0 <= 1); /* necessarily 1 or 0 */ 565 { U32 const lastR = ZSTD_REP_NUM + ll0; 566 U32 repCode; 567 for (repCode = ll0; repCode < lastR; repCode++) { 568 U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode]; 569 U32 const repIndex = current - repOffset; 570 U32 repLen = 0; 571 assert(current >= dictLimit); 572 if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < current-dictLimit) { /* equivalent to `current > repIndex >= dictLimit` */ 573 /* We must validate the repcode offset because when we're using a dictionary the 574 * valid offset range shrinks when the dictionary goes out of bounds. 575 */ 576 if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) { 577 repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch; 578 } 579 } else { /* repIndex < dictLimit || repIndex >= current */ 580 const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ? 581 dmsBase + repIndex - dmsIndexDelta : 582 dictBase + repIndex; 583 assert(current >= windowLow); 584 if ( dictMode == ZSTD_extDict 585 && ( ((repOffset-1) /*intentional overflow*/ < current - windowLow) /* equivalent to `current > repIndex >= windowLow` */ 586 & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */) 587 && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) { 588 repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch; 589 } 590 if (dictMode == ZSTD_dictMatchState 591 && ( ((repOffset-1) /*intentional overflow*/ < current - (dmsLowLimit + dmsIndexDelta)) /* equivalent to `current > repIndex >= dmsLowLimit` */ 592 & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */ 593 && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) { 594 repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch; 595 } } 596 /* save longer solution */ 597 if (repLen > bestLength) { 598 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u", 599 repCode, ll0, repOffset, repLen); 600 bestLength = repLen; 601 matches[mnum].off = repCode - ll0; 602 matches[mnum].len = (U32)repLen; 603 mnum++; 604 if ( (repLen > sufficient_len) 605 | (ip+repLen == iLimit) ) { /* best possible */ 606 return mnum; 607 } } } } 608 609 /* HC3 match finder */ 610 if ((mls == 3) /*static*/ && (bestLength < mls)) { 611 U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip); 612 if ((matchIndex3 >= matchLow) 613 & (current - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) { 614 size_t mlen; 615 if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) { 616 const BYTE* const match = base + matchIndex3; 617 mlen = ZSTD_count(ip, match, iLimit); 618 } else { 619 const BYTE* const match = dictBase + matchIndex3; 620 mlen = ZSTD_count_2segments(ip, match, iLimit, dictEnd, prefixStart); 621 } 622 623 /* save best solution */ 624 if (mlen >= mls /* == 3 > bestLength */) { 625 DEBUGLOG(8, "found small match with hlog3, of length %u", 626 (U32)mlen); 627 bestLength = mlen; 628 assert(current > matchIndex3); 629 assert(mnum==0); /* no prior solution */ 630 matches[0].off = (current - matchIndex3) + ZSTD_REP_MOVE; 631 matches[0].len = (U32)mlen; 632 mnum = 1; 633 if ( (mlen > sufficient_len) | 634 (ip+mlen == iLimit) ) { /* best possible length */ 635 ms->nextToUpdate = current+1; /* skip insertion */ 636 return 1; 637 } } } 638 /* no dictMatchState lookup: dicts don't have a populated HC3 table */ 639 } 640 641 hashTable[h] = current; /* Update Hash Table */ 642 643 while (nbCompares-- && (matchIndex >= matchLow)) { 644 U32* const nextPtr = bt + 2*(matchIndex & btMask); 645 const BYTE* match; 646 size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ 647 assert(current > matchIndex); 648 649 if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) { 650 assert(matchIndex+matchLength >= dictLimit); /* ensure the condition is correct when !extDict */ 651 match = base + matchIndex; 652 if (matchIndex >= dictLimit) assert(memcmp(match, ip, matchLength) == 0); /* ensure early section of match is equal as expected */ 653 matchLength += ZSTD_count(ip+matchLength, match+matchLength, iLimit); 654 } else { 655 match = dictBase + matchIndex; 656 assert(memcmp(match, ip, matchLength) == 0); /* ensure early section of match is equal as expected */ 657 matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dictEnd, prefixStart); 658 if (matchIndex+matchLength >= dictLimit) 659 match = base + matchIndex; /* prepare for match[matchLength] read */ 660 } 661 662 if (matchLength > bestLength) { 663 DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)", 664 (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE); 665 assert(matchEndIdx > matchIndex); 666 if (matchLength > matchEndIdx - matchIndex) 667 matchEndIdx = matchIndex + (U32)matchLength; 668 bestLength = matchLength; 669 matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE; 670 matches[mnum].len = (U32)matchLength; 671 mnum++; 672 if ( (matchLength > ZSTD_OPT_NUM) 673 | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) { 674 if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */ 675 break; /* drop, to preserve bt consistency (miss a little bit of compression) */ 676 } 677 } 678 679 if (match[matchLength] < ip[matchLength]) { 680 /* match smaller than current */ 681 *smallerPtr = matchIndex; /* update smaller idx */ 682 commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */ 683 if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop the search */ 684 smallerPtr = nextPtr+1; /* new candidate => larger than match, which was smaller than current */ 685 matchIndex = nextPtr[1]; /* new matchIndex, larger than previous, closer to current */ 686 } else { 687 *largerPtr = matchIndex; 688 commonLengthLarger = matchLength; 689 if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop the search */ 690 largerPtr = nextPtr; 691 matchIndex = nextPtr[0]; 692 } } 693 694 *smallerPtr = *largerPtr = 0; 695 696 if (dictMode == ZSTD_dictMatchState && nbCompares) { 697 size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls); 698 U32 dictMatchIndex = dms->hashTable[dmsH]; 699 const U32* const dmsBt = dms->chainTable; 700 commonLengthSmaller = commonLengthLarger = 0; 701 while (nbCompares-- && (dictMatchIndex > dmsLowLimit)) { 702 const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask); 703 size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */ 704 const BYTE* match = dmsBase + dictMatchIndex; 705 matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dmsEnd, prefixStart); 706 if (dictMatchIndex+matchLength >= dmsHighLimit) 707 match = base + dictMatchIndex + dmsIndexDelta; /* to prepare for next usage of match[matchLength] */ 708 709 if (matchLength > bestLength) { 710 matchIndex = dictMatchIndex + dmsIndexDelta; 711 DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)", 712 (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE); 713 if (matchLength > matchEndIdx - matchIndex) 714 matchEndIdx = matchIndex + (U32)matchLength; 715 bestLength = matchLength; 716 matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE; 717 matches[mnum].len = (U32)matchLength; 718 mnum++; 719 if ( (matchLength > ZSTD_OPT_NUM) 720 | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) { 721 break; /* drop, to guarantee consistency (miss a little bit of compression) */ 722 } 723 } 724 725 if (dictMatchIndex <= dmsBtLow) { break; } /* beyond tree size, stop the search */ 726 if (match[matchLength] < ip[matchLength]) { 727 commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */ 728 dictMatchIndex = nextPtr[1]; /* new matchIndex larger than previous (closer to current) */ 729 } else { 730 /* match is larger than current */ 731 commonLengthLarger = matchLength; 732 dictMatchIndex = nextPtr[0]; 733 } 734 } 735 } 736 737 assert(matchEndIdx > current+8); 738 ms->nextToUpdate = matchEndIdx - 8; /* skip repetitive patterns */ 739 return mnum; 740 } 741 742 743 FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches ( 744 ZSTD_match_t* matches, /* store result (match found, increasing size) in this table */ 745 ZSTD_matchState_t* ms, 746 U32* nextToUpdate3, 747 const BYTE* ip, const BYTE* const iHighLimit, const ZSTD_dictMode_e dictMode, 748 const U32 rep[ZSTD_REP_NUM], 749 U32 const ll0, 750 U32 const lengthToBeat) 751 { 752 const ZSTD_compressionParameters* const cParams = &ms->cParams; 753 U32 const matchLengthSearch = cParams->minMatch; 754 DEBUGLOG(8, "ZSTD_BtGetAllMatches"); 755 if (ip < ms->window.base + ms->nextToUpdate) return 0; /* skipped area */ 756 ZSTD_updateTree_internal(ms, ip, iHighLimit, matchLengthSearch, dictMode); 757 switch(matchLengthSearch) 758 { 759 case 3 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 3); 760 default : 761 case 4 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 4); 762 case 5 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 5); 763 case 7 : 764 case 6 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 6); 765 } 766 } 767 768 769 /*-******************************* 770 * Optimal parser 771 *********************************/ 772 773 774 static U32 ZSTD_totalLen(ZSTD_optimal_t sol) 775 { 776 return sol.litlen + sol.mlen; 777 } 778 779 #if 0 /* debug */ 780 781 static void 782 listStats(const U32* table, int lastEltID) 783 { 784 int const nbElts = lastEltID + 1; 785 int enb; 786 for (enb=0; enb < nbElts; enb++) { 787 (void)table; 788 /* RAWLOG(2, "%3i:%3i, ", enb, table[enb]); */ 789 RAWLOG(2, "%4i,", table[enb]); 790 } 791 RAWLOG(2, " \n"); 792 } 793 794 #endif 795 796 FORCE_INLINE_TEMPLATE size_t 797 ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms, 798 seqStore_t* seqStore, 799 U32 rep[ZSTD_REP_NUM], 800 const void* src, size_t srcSize, 801 const int optLevel, 802 const ZSTD_dictMode_e dictMode) 803 { 804 optState_t* const optStatePtr = &ms->opt; 805 const BYTE* const istart = (const BYTE*)src; 806 const BYTE* ip = istart; 807 const BYTE* anchor = istart; 808 const BYTE* const iend = istart + srcSize; 809 const BYTE* const ilimit = iend - 8; 810 const BYTE* const base = ms->window.base; 811 const BYTE* const prefixStart = base + ms->window.dictLimit; 812 const ZSTD_compressionParameters* const cParams = &ms->cParams; 813 814 U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1); 815 U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4; 816 U32 nextToUpdate3 = ms->nextToUpdate; 817 818 ZSTD_optimal_t* const opt = optStatePtr->priceTable; 819 ZSTD_match_t* const matches = optStatePtr->matchTable; 820 ZSTD_optimal_t lastSequence; 821 822 /* init */ 823 DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u", 824 (U32)(ip - base), ms->window.dictLimit, ms->nextToUpdate); 825 assert(optLevel <= 2); 826 ZSTD_rescaleFreqs(optStatePtr, (const BYTE*)src, srcSize, optLevel); 827 ip += (ip==prefixStart); 828 829 /* Match Loop */ 830 while (ip < ilimit) { 831 U32 cur, last_pos = 0; 832 833 /* find first match */ 834 { U32 const litlen = (U32)(ip - anchor); 835 U32 const ll0 = !litlen; 836 U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch); 837 if (!nbMatches) { ip++; continue; } 838 839 /* initialize opt[0] */ 840 { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; } 841 opt[0].mlen = 0; /* means is_a_literal */ 842 opt[0].litlen = litlen; 843 /* We don't need to include the actual price of the literals because 844 * it is static for the duration of the forward pass, and is included 845 * in every price. We include the literal length to avoid negative 846 * prices when we subtract the previous literal length. 847 */ 848 opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel); 849 850 /* large match -> immediate encoding */ 851 { U32 const maxML = matches[nbMatches-1].len; 852 U32 const maxOffset = matches[nbMatches-1].off; 853 DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series", 854 nbMatches, maxML, maxOffset, (U32)(ip-prefixStart)); 855 856 if (maxML > sufficient_len) { 857 lastSequence.litlen = litlen; 858 lastSequence.mlen = maxML; 859 lastSequence.off = maxOffset; 860 DEBUGLOG(6, "large match (%u>%u), immediate encoding", 861 maxML, sufficient_len); 862 cur = 0; 863 last_pos = ZSTD_totalLen(lastSequence); 864 goto _shortestPath; 865 } } 866 867 /* set prices for first matches starting position == 0 */ 868 { U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel); 869 U32 pos; 870 U32 matchNb; 871 for (pos = 1; pos < minMatch; pos++) { 872 opt[pos].price = ZSTD_MAX_PRICE; /* mlen, litlen and price will be fixed during forward scanning */ 873 } 874 for (matchNb = 0; matchNb < nbMatches; matchNb++) { 875 U32 const offset = matches[matchNb].off; 876 U32 const end = matches[matchNb].len; 877 for ( ; pos <= end ; pos++ ) { 878 U32 const matchPrice = ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel); 879 U32 const sequencePrice = literalsPrice + matchPrice; 880 DEBUGLOG(7, "rPos:%u => set initial price : %.2f", 881 pos, ZSTD_fCost(sequencePrice)); 882 opt[pos].mlen = pos; 883 opt[pos].off = offset; 884 opt[pos].litlen = litlen; 885 opt[pos].price = sequencePrice; 886 } } 887 last_pos = pos-1; 888 } 889 } 890 891 /* check further positions */ 892 for (cur = 1; cur <= last_pos; cur++) { 893 const BYTE* const inr = ip + cur; 894 assert(cur < ZSTD_OPT_NUM); 895 DEBUGLOG(7, "cPos:%zi==rPos:%u", inr-istart, cur) 896 897 /* Fix current position with one literal if cheaper */ 898 { U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1; 899 int const price = opt[cur-1].price 900 + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel) 901 + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel) 902 - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel); 903 assert(price < 1000000000); /* overflow check */ 904 if (price <= opt[cur].price) { 905 DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)", 906 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen, 907 opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]); 908 opt[cur].mlen = 0; 909 opt[cur].off = 0; 910 opt[cur].litlen = litlen; 911 opt[cur].price = price; 912 } else { 913 DEBUGLOG(7, "cPos:%zi==rPos:%u : literal would cost more (%.2f>%.2f) (hist:%u,%u,%u)", 914 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), 915 opt[cur].rep[0], opt[cur].rep[1], opt[cur].rep[2]); 916 } 917 } 918 919 /* Set the repcodes of the current position. We must do it here 920 * because we rely on the repcodes of the 2nd to last sequence being 921 * correct to set the next chunks repcodes during the backward 922 * traversal. 923 */ 924 ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(repcodes_t)); 925 assert(cur >= opt[cur].mlen); 926 if (opt[cur].mlen != 0) { 927 U32 const prev = cur - opt[cur].mlen; 928 repcodes_t newReps = ZSTD_updateRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0); 929 memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t)); 930 } else { 931 memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t)); 932 } 933 934 /* last match must start at a minimum distance of 8 from oend */ 935 if (inr > ilimit) continue; 936 937 if (cur == last_pos) break; 938 939 if ( (optLevel==0) /*static_test*/ 940 && (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) { 941 DEBUGLOG(7, "move to next rPos:%u : price is <=", cur+1); 942 continue; /* skip unpromising positions; about ~+6% speed, -0.01 ratio */ 943 } 944 945 { U32 const ll0 = (opt[cur].mlen != 0); 946 U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0; 947 U32 const previousPrice = opt[cur].price; 948 U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel); 949 U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch); 950 U32 matchNb; 951 if (!nbMatches) { 952 DEBUGLOG(7, "rPos:%u : no match found", cur); 953 continue; 954 } 955 956 { U32 const maxML = matches[nbMatches-1].len; 957 DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of maxLength=%u", 958 inr-istart, cur, nbMatches, maxML); 959 960 if ( (maxML > sufficient_len) 961 || (cur + maxML >= ZSTD_OPT_NUM) ) { 962 lastSequence.mlen = maxML; 963 lastSequence.off = matches[nbMatches-1].off; 964 lastSequence.litlen = litlen; 965 cur -= (opt[cur].mlen==0) ? opt[cur].litlen : 0; /* last sequence is actually only literals, fix cur to last match - note : may underflow, in which case, it's first sequence, and it's okay */ 966 last_pos = cur + ZSTD_totalLen(lastSequence); 967 if (cur > ZSTD_OPT_NUM) cur = 0; /* underflow => first match */ 968 goto _shortestPath; 969 } } 970 971 /* set prices using matches found at position == cur */ 972 for (matchNb = 0; matchNb < nbMatches; matchNb++) { 973 U32 const offset = matches[matchNb].off; 974 U32 const lastML = matches[matchNb].len; 975 U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch; 976 U32 mlen; 977 978 DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u", 979 matchNb, matches[matchNb].off, lastML, litlen); 980 981 for (mlen = lastML; mlen >= startML; mlen--) { /* scan downward */ 982 U32 const pos = cur + mlen; 983 int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel); 984 985 if ((pos > last_pos) || (price < opt[pos].price)) { 986 DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)", 987 pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price)); 988 while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; } /* fill empty positions */ 989 opt[pos].mlen = mlen; 990 opt[pos].off = offset; 991 opt[pos].litlen = litlen; 992 opt[pos].price = price; 993 } else { 994 DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)", 995 pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price)); 996 if (optLevel==0) break; /* early update abort; gets ~+10% speed for about -0.01 ratio loss */ 997 } 998 } } } 999 } /* for (cur = 1; cur <= last_pos; cur++) */ 1000 1001 lastSequence = opt[last_pos]; 1002 cur = last_pos > ZSTD_totalLen(lastSequence) ? last_pos - ZSTD_totalLen(lastSequence) : 0; /* single sequence, and it starts before `ip` */ 1003 assert(cur < ZSTD_OPT_NUM); /* control overflow*/ 1004 1005 _shortestPath: /* cur, last_pos, best_mlen, best_off have to be set */ 1006 assert(opt[0].mlen == 0); 1007 1008 /* Set the next chunk's repcodes based on the repcodes of the beginning 1009 * of the last match, and the last sequence. This avoids us having to 1010 * update them while traversing the sequences. 1011 */ 1012 if (lastSequence.mlen != 0) { 1013 repcodes_t reps = ZSTD_updateRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0); 1014 memcpy(rep, &reps, sizeof(reps)); 1015 } else { 1016 memcpy(rep, opt[cur].rep, sizeof(repcodes_t)); 1017 } 1018 1019 { U32 const storeEnd = cur + 1; 1020 U32 storeStart = storeEnd; 1021 U32 seqPos = cur; 1022 1023 DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)", 1024 last_pos, cur); (void)last_pos; 1025 assert(storeEnd < ZSTD_OPT_NUM); 1026 DEBUGLOG(6, "last sequence copied into pos=%u (llen=%u,mlen=%u,ofc=%u)", 1027 storeEnd, lastSequence.litlen, lastSequence.mlen, lastSequence.off); 1028 opt[storeEnd] = lastSequence; 1029 while (seqPos > 0) { 1030 U32 const backDist = ZSTD_totalLen(opt[seqPos]); 1031 storeStart--; 1032 DEBUGLOG(6, "sequence from rPos=%u copied into pos=%u (llen=%u,mlen=%u,ofc=%u)", 1033 seqPos, storeStart, opt[seqPos].litlen, opt[seqPos].mlen, opt[seqPos].off); 1034 opt[storeStart] = opt[seqPos]; 1035 seqPos = (seqPos > backDist) ? seqPos - backDist : 0; 1036 } 1037 1038 /* save sequences */ 1039 DEBUGLOG(6, "sending selected sequences into seqStore") 1040 { U32 storePos; 1041 for (storePos=storeStart; storePos <= storeEnd; storePos++) { 1042 U32 const llen = opt[storePos].litlen; 1043 U32 const mlen = opt[storePos].mlen; 1044 U32 const offCode = opt[storePos].off; 1045 U32 const advance = llen + mlen; 1046 DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u", 1047 anchor - istart, (unsigned)llen, (unsigned)mlen); 1048 1049 if (mlen==0) { /* only literals => must be last "sequence", actually starting a new stream of sequences */ 1050 assert(storePos == storeEnd); /* must be last sequence */ 1051 ip = anchor + llen; /* last "sequence" is a bunch of literals => don't progress anchor */ 1052 continue; /* will finish */ 1053 } 1054 1055 assert(anchor + llen <= iend); 1056 ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen); 1057 ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen-MINMATCH); 1058 anchor += advance; 1059 ip = anchor; 1060 } } 1061 ZSTD_setBasePrices(optStatePtr, optLevel); 1062 } 1063 } /* while (ip < ilimit) */ 1064 1065 /* Return the last literals size */ 1066 return (size_t)(iend - anchor); 1067 } 1068 1069 1070 size_t ZSTD_compressBlock_btopt( 1071 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1072 const void* src, size_t srcSize) 1073 { 1074 DEBUGLOG(5, "ZSTD_compressBlock_btopt"); 1075 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict); 1076 } 1077 1078 1079 /* used in 2-pass strategy */ 1080 static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus) 1081 { 1082 U32 s, sum=0; 1083 assert(ZSTD_FREQ_DIV+bonus >= 0); 1084 for (s=0; s<lastEltIndex+1; s++) { 1085 table[s] <<= ZSTD_FREQ_DIV+bonus; 1086 table[s]--; 1087 sum += table[s]; 1088 } 1089 return sum; 1090 } 1091 1092 /* used in 2-pass strategy */ 1093 MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr) 1094 { 1095 if (ZSTD_compressedLiterals(optPtr)) 1096 optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0); 1097 optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0); 1098 optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0); 1099 optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0); 1100 } 1101 1102 /* ZSTD_initStats_ultra(): 1103 * make a first compression pass, just to seed stats with more accurate starting values. 1104 * only works on first block, with no dictionary and no ldm. 1105 * this function cannot error, hence its contract must be respected. 1106 */ 1107 static void 1108 ZSTD_initStats_ultra(ZSTD_matchState_t* ms, 1109 seqStore_t* seqStore, 1110 U32 rep[ZSTD_REP_NUM], 1111 const void* src, size_t srcSize) 1112 { 1113 U32 tmpRep[ZSTD_REP_NUM]; /* updated rep codes will sink here */ 1114 memcpy(tmpRep, rep, sizeof(tmpRep)); 1115 1116 DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize); 1117 assert(ms->opt.litLengthSum == 0); /* first block */ 1118 assert(seqStore->sequences == seqStore->sequencesStart); /* no ldm */ 1119 assert(ms->window.dictLimit == ms->window.lowLimit); /* no dictionary */ 1120 assert(ms->window.dictLimit - ms->nextToUpdate <= 1); /* no prefix (note: intentional overflow, defined as 2-complement) */ 1121 1122 ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); /* generate stats into ms->opt*/ 1123 1124 /* invalidate first scan from history */ 1125 ZSTD_resetSeqStore(seqStore); 1126 ms->window.base -= srcSize; 1127 ms->window.dictLimit += (U32)srcSize; 1128 ms->window.lowLimit = ms->window.dictLimit; 1129 ms->nextToUpdate = ms->window.dictLimit; 1130 1131 /* re-inforce weight of collected statistics */ 1132 ZSTD_upscaleStats(&ms->opt); 1133 } 1134 1135 size_t ZSTD_compressBlock_btultra( 1136 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1137 const void* src, size_t srcSize) 1138 { 1139 DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize); 1140 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); 1141 } 1142 1143 size_t ZSTD_compressBlock_btultra2( 1144 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1145 const void* src, size_t srcSize) 1146 { 1147 U32 const current = (U32)((const BYTE*)src - ms->window.base); 1148 DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize); 1149 1150 /* 2-pass strategy: 1151 * this strategy makes a first pass over first block to collect statistics 1152 * and seed next round's statistics with it. 1153 * After 1st pass, function forgets everything, and starts a new block. 1154 * Consequently, this can only work if no data has been previously loaded in tables, 1155 * aka, no dictionary, no prefix, no ldm preprocessing. 1156 * The compression ratio gain is generally small (~0.5% on first block), 1157 * the cost is 2x cpu time on first block. */ 1158 assert(srcSize <= ZSTD_BLOCKSIZE_MAX); 1159 if ( (ms->opt.litLengthSum==0) /* first block */ 1160 && (seqStore->sequences == seqStore->sequencesStart) /* no ldm */ 1161 && (ms->window.dictLimit == ms->window.lowLimit) /* no dictionary */ 1162 && (current == ms->window.dictLimit) /* start of frame, nothing already loaded nor skipped */ 1163 && (srcSize > ZSTD_PREDEF_THRESHOLD) 1164 ) { 1165 ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize); 1166 } 1167 1168 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict); 1169 } 1170 1171 size_t ZSTD_compressBlock_btopt_dictMatchState( 1172 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1173 const void* src, size_t srcSize) 1174 { 1175 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState); 1176 } 1177 1178 size_t ZSTD_compressBlock_btultra_dictMatchState( 1179 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1180 const void* src, size_t srcSize) 1181 { 1182 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState); 1183 } 1184 1185 size_t ZSTD_compressBlock_btopt_extDict( 1186 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1187 const void* src, size_t srcSize) 1188 { 1189 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict); 1190 } 1191 1192 size_t ZSTD_compressBlock_btultra_extDict( 1193 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 1194 const void* src, size_t srcSize) 1195 { 1196 return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict); 1197 } 1198 1199 /* note : no btultra2 variant for extDict nor dictMatchState, 1200 * because btultra2 is not meant to work with dictionaries 1201 * and is only specific for the first block (no prefix) */ 1202