xref: /openbmc/linux/lib/zstd/compress/zstd_opt.c (revision 6e10e219)
1 /*
2  * Copyright (c) Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
3  * All rights reserved.
4  *
5  * This source code is licensed under both the BSD-style license (found in the
6  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7  * in the COPYING file in the root directory of this source tree).
8  * You may select, at your option, one of the above-listed licenses.
9  */
10 
11 /*
12  * Disable inlining for the optimal parser for the kernel build.
13  * It is unlikely to be used in the kernel, and where it is used
14  * latency shouldn't matter because it is very slow to begin with.
15  * We prefer a ~180KB binary size win over faster optimal parsing.
16  *
17  * TODO(https://github.com/facebook/zstd/issues/2862):
18  * Improve the code size of the optimal parser in general, so we
19  * don't need this hack for the kernel build.
20  */
21 #define ZSTD_NO_INLINE 1
22 
23 #include "zstd_compress_internal.h"
24 #include "hist.h"
25 #include "zstd_opt.h"
26 
27 
28 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
29 #define ZSTD_FREQ_DIV       4   /* log factor when using previous stats to init next stats */
30 #define ZSTD_MAX_PRICE     (1<<30)
31 
32 #define ZSTD_PREDEF_THRESHOLD 1024   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
33 
34 
35 /*-*************************************
36 *  Price functions for optimal parser
37 ***************************************/
38 
39 #if 0    /* approximation at bit level */
40 #  define BITCOST_ACCURACY 0
41 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
42 #  define WEIGHT(stat)  ((void)opt, ZSTD_bitWeight(stat))
43 #elif 0  /* fractional bit accuracy */
44 #  define BITCOST_ACCURACY 8
45 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
46 #  define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat))
47 #else    /* opt==approx, ultra==accurate */
48 #  define BITCOST_ACCURACY 8
49 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
50 #  define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
51 #endif
52 
53 MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
54 {
55     return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
56 }
57 
58 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
59 {
60     U32 const stat = rawStat + 1;
61     U32 const hb = ZSTD_highbit32(stat);
62     U32 const BWeight = hb * BITCOST_MULTIPLIER;
63     U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
64     U32 const weight = BWeight + FWeight;
65     assert(hb + BITCOST_ACCURACY < 31);
66     return weight;
67 }
68 
69 #if (DEBUGLEVEL>=2)
70 /* debugging function,
71  * @return price in bytes as fractional value
72  * for debug messages only */
73 MEM_STATIC double ZSTD_fCost(U32 price)
74 {
75     return (double)price / (BITCOST_MULTIPLIER*8);
76 }
77 #endif
78 
79 static int ZSTD_compressedLiterals(optState_t const* const optPtr)
80 {
81     return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed;
82 }
83 
84 static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
85 {
86     if (ZSTD_compressedLiterals(optPtr))
87         optPtr->litSumBasePrice = WEIGHT(optPtr->litSum, optLevel);
88     optPtr->litLengthSumBasePrice = WEIGHT(optPtr->litLengthSum, optLevel);
89     optPtr->matchLengthSumBasePrice = WEIGHT(optPtr->matchLengthSum, optLevel);
90     optPtr->offCodeSumBasePrice = WEIGHT(optPtr->offCodeSum, optLevel);
91 }
92 
93 
94 /* ZSTD_downscaleStat() :
95  * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus)
96  * return the resulting sum of elements */
97 static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus)
98 {
99     U32 s, sum=0;
100     DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1);
101     assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31);
102     for (s=0; s<lastEltIndex+1; s++) {
103         table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus));
104         sum += table[s];
105     }
106     return sum;
107 }
108 
109 /* ZSTD_rescaleFreqs() :
110  * if first block (detected by optPtr->litLengthSum == 0) : init statistics
111  *    take hints from dictionary if there is one
112  *    or init from zero, using src for literals stats, or flat 1 for match symbols
113  * otherwise downscale existing stats, to be used as seed for next block.
114  */
115 static void
116 ZSTD_rescaleFreqs(optState_t* const optPtr,
117             const BYTE* const src, size_t const srcSize,
118                   int const optLevel)
119 {
120     int const compressedLiterals = ZSTD_compressedLiterals(optPtr);
121     DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
122     optPtr->priceType = zop_dynamic;
123 
124     if (optPtr->litLengthSum == 0) {  /* first block : init */
125         if (srcSize <= ZSTD_PREDEF_THRESHOLD) {  /* heuristic */
126             DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef");
127             optPtr->priceType = zop_predef;
128         }
129 
130         assert(optPtr->symbolCosts != NULL);
131         if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {
132             /* huffman table presumed generated by dictionary */
133             optPtr->priceType = zop_dynamic;
134 
135             if (compressedLiterals) {
136                 unsigned lit;
137                 assert(optPtr->litFreq != NULL);
138                 optPtr->litSum = 0;
139                 for (lit=0; lit<=MaxLit; lit++) {
140                     U32 const scaleLog = 11;   /* scale to 2K */
141                     U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit);
142                     assert(bitCost <= scaleLog);
143                     optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
144                     optPtr->litSum += optPtr->litFreq[lit];
145             }   }
146 
147             {   unsigned ll;
148                 FSE_CState_t llstate;
149                 FSE_initCState(&llstate, optPtr->symbolCosts->fse.litlengthCTable);
150                 optPtr->litLengthSum = 0;
151                 for (ll=0; ll<=MaxLL; ll++) {
152                     U32 const scaleLog = 10;   /* scale to 1K */
153                     U32 const bitCost = FSE_getMaxNbBits(llstate.symbolTT, ll);
154                     assert(bitCost < scaleLog);
155                     optPtr->litLengthFreq[ll] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
156                     optPtr->litLengthSum += optPtr->litLengthFreq[ll];
157             }   }
158 
159             {   unsigned ml;
160                 FSE_CState_t mlstate;
161                 FSE_initCState(&mlstate, optPtr->symbolCosts->fse.matchlengthCTable);
162                 optPtr->matchLengthSum = 0;
163                 for (ml=0; ml<=MaxML; ml++) {
164                     U32 const scaleLog = 10;
165                     U32 const bitCost = FSE_getMaxNbBits(mlstate.symbolTT, ml);
166                     assert(bitCost < scaleLog);
167                     optPtr->matchLengthFreq[ml] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
168                     optPtr->matchLengthSum += optPtr->matchLengthFreq[ml];
169             }   }
170 
171             {   unsigned of;
172                 FSE_CState_t ofstate;
173                 FSE_initCState(&ofstate, optPtr->symbolCosts->fse.offcodeCTable);
174                 optPtr->offCodeSum = 0;
175                 for (of=0; of<=MaxOff; of++) {
176                     U32 const scaleLog = 10;
177                     U32 const bitCost = FSE_getMaxNbBits(ofstate.symbolTT, of);
178                     assert(bitCost < scaleLog);
179                     optPtr->offCodeFreq[of] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
180                     optPtr->offCodeSum += optPtr->offCodeFreq[of];
181             }   }
182 
183         } else {  /* not a dictionary */
184 
185             assert(optPtr->litFreq != NULL);
186             if (compressedLiterals) {
187                 unsigned lit = MaxLit;
188                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
189                 optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
190             }
191 
192             {   unsigned ll;
193                 for (ll=0; ll<=MaxLL; ll++)
194                     optPtr->litLengthFreq[ll] = 1;
195             }
196             optPtr->litLengthSum = MaxLL+1;
197 
198             {   unsigned ml;
199                 for (ml=0; ml<=MaxML; ml++)
200                     optPtr->matchLengthFreq[ml] = 1;
201             }
202             optPtr->matchLengthSum = MaxML+1;
203 
204             {   unsigned of;
205                 for (of=0; of<=MaxOff; of++)
206                     optPtr->offCodeFreq[of] = 1;
207             }
208             optPtr->offCodeSum = MaxOff+1;
209 
210         }
211 
212     } else {   /* new block : re-use previous statistics, scaled down */
213 
214         if (compressedLiterals)
215             optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
216         optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0);
217         optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0);
218         optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0);
219     }
220 
221     ZSTD_setBasePrices(optPtr, optLevel);
222 }
223 
224 /* ZSTD_rawLiteralsCost() :
225  * price of literals (only) in specified segment (which length can be 0).
226  * does not include price of literalLength symbol */
227 static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
228                                 const optState_t* const optPtr,
229                                 int optLevel)
230 {
231     if (litLength == 0) return 0;
232 
233     if (!ZSTD_compressedLiterals(optPtr))
234         return (litLength << 3) * BITCOST_MULTIPLIER;  /* Uncompressed - 8 bytes per literal. */
235 
236     if (optPtr->priceType == zop_predef)
237         return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */
238 
239     /* dynamic statistics */
240     {   U32 price = litLength * optPtr->litSumBasePrice;
241         U32 u;
242         for (u=0; u < litLength; u++) {
243             assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice);   /* literal cost should never be negative */
244             price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel);
245         }
246         return price;
247     }
248 }
249 
250 /* ZSTD_litLengthPrice() :
251  * cost of literalLength symbol */
252 static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel)
253 {
254     if (optPtr->priceType == zop_predef) return WEIGHT(litLength, optLevel);
255 
256     /* dynamic statistics */
257     {   U32 const llCode = ZSTD_LLcode(litLength);
258         return (LL_bits[llCode] * BITCOST_MULTIPLIER)
259              + optPtr->litLengthSumBasePrice
260              - WEIGHT(optPtr->litLengthFreq[llCode], optLevel);
261     }
262 }
263 
264 /* ZSTD_getMatchPrice() :
265  * Provides the cost of the match part (offset + matchLength) of a sequence
266  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
267  * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */
268 FORCE_INLINE_TEMPLATE U32
269 ZSTD_getMatchPrice(U32 const offset,
270                    U32 const matchLength,
271              const optState_t* const optPtr,
272                    int const optLevel)
273 {
274     U32 price;
275     U32 const offCode = ZSTD_highbit32(offset+1);
276     U32 const mlBase = matchLength - MINMATCH;
277     assert(matchLength >= MINMATCH);
278 
279     if (optPtr->priceType == zop_predef)  /* fixed scheme, do not use statistics */
280         return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER);
281 
282     /* dynamic statistics */
283     price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
284     if ((optLevel<2) /*static*/ && offCode >= 20)
285         price += (offCode-19)*2 * BITCOST_MULTIPLIER; /* handicap for long distance offsets, favor decompression speed */
286 
287     /* match Length */
288     {   U32 const mlCode = ZSTD_MLcode(mlBase);
289         price += (ML_bits[mlCode] * BITCOST_MULTIPLIER) + (optPtr->matchLengthSumBasePrice - WEIGHT(optPtr->matchLengthFreq[mlCode], optLevel));
290     }
291 
292     price += BITCOST_MULTIPLIER / 5;   /* heuristic : make matches a bit more costly to favor less sequences -> faster decompression speed */
293 
294     DEBUGLOG(8, "ZSTD_getMatchPrice(ml:%u) = %u", matchLength, price);
295     return price;
296 }
297 
298 /* ZSTD_updateStats() :
299  * assumption : literals + litLengtn <= iend */
300 static void ZSTD_updateStats(optState_t* const optPtr,
301                              U32 litLength, const BYTE* literals,
302                              U32 offsetCode, U32 matchLength)
303 {
304     /* literals */
305     if (ZSTD_compressedLiterals(optPtr)) {
306         U32 u;
307         for (u=0; u < litLength; u++)
308             optPtr->litFreq[literals[u]] += ZSTD_LITFREQ_ADD;
309         optPtr->litSum += litLength*ZSTD_LITFREQ_ADD;
310     }
311 
312     /* literal Length */
313     {   U32 const llCode = ZSTD_LLcode(litLength);
314         optPtr->litLengthFreq[llCode]++;
315         optPtr->litLengthSum++;
316     }
317 
318     /* match offset code (0-2=>repCode; 3+=>offset+2) */
319     {   U32 const offCode = ZSTD_highbit32(offsetCode+1);
320         assert(offCode <= MaxOff);
321         optPtr->offCodeFreq[offCode]++;
322         optPtr->offCodeSum++;
323     }
324 
325     /* match Length */
326     {   U32 const mlBase = matchLength - MINMATCH;
327         U32 const mlCode = ZSTD_MLcode(mlBase);
328         optPtr->matchLengthFreq[mlCode]++;
329         optPtr->matchLengthSum++;
330     }
331 }
332 
333 
334 /* ZSTD_readMINMATCH() :
335  * function safe only for comparisons
336  * assumption : memPtr must be at least 4 bytes before end of buffer */
337 MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
338 {
339     switch (length)
340     {
341     default :
342     case 4 : return MEM_read32(memPtr);
343     case 3 : if (MEM_isLittleEndian())
344                 return MEM_read32(memPtr)<<8;
345              else
346                 return MEM_read32(memPtr)>>8;
347     }
348 }
349 
350 
351 /* Update hashTable3 up to ip (excluded)
352    Assumption : always within prefix (i.e. not within extDict) */
353 static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
354                                               U32* nextToUpdate3,
355                                               const BYTE* const ip)
356 {
357     U32* const hashTable3 = ms->hashTable3;
358     U32 const hashLog3 = ms->hashLog3;
359     const BYTE* const base = ms->window.base;
360     U32 idx = *nextToUpdate3;
361     U32 const target = (U32)(ip - base);
362     size_t const hash3 = ZSTD_hash3Ptr(ip, hashLog3);
363     assert(hashLog3 > 0);
364 
365     while(idx < target) {
366         hashTable3[ZSTD_hash3Ptr(base+idx, hashLog3)] = idx;
367         idx++;
368     }
369 
370     *nextToUpdate3 = target;
371     return hashTable3[hash3];
372 }
373 
374 
375 /*-*************************************
376 *  Binary Tree search
377 ***************************************/
378 /* ZSTD_insertBt1() : add one or multiple positions to tree.
379  *  ip : assumed <= iend-8 .
380  * @return : nb of positions added */
381 static U32 ZSTD_insertBt1(
382                 ZSTD_matchState_t* ms,
383                 const BYTE* const ip, const BYTE* const iend,
384                 U32 const mls, const int extDict)
385 {
386     const ZSTD_compressionParameters* const cParams = &ms->cParams;
387     U32*   const hashTable = ms->hashTable;
388     U32    const hashLog = cParams->hashLog;
389     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
390     U32*   const bt = ms->chainTable;
391     U32    const btLog  = cParams->chainLog - 1;
392     U32    const btMask = (1 << btLog) - 1;
393     U32 matchIndex = hashTable[h];
394     size_t commonLengthSmaller=0, commonLengthLarger=0;
395     const BYTE* const base = ms->window.base;
396     const BYTE* const dictBase = ms->window.dictBase;
397     const U32 dictLimit = ms->window.dictLimit;
398     const BYTE* const dictEnd = dictBase + dictLimit;
399     const BYTE* const prefixStart = base + dictLimit;
400     const BYTE* match;
401     const U32 curr = (U32)(ip-base);
402     const U32 btLow = btMask >= curr ? 0 : curr - btMask;
403     U32* smallerPtr = bt + 2*(curr&btMask);
404     U32* largerPtr  = smallerPtr + 1;
405     U32 dummy32;   /* to be nullified at the end */
406     U32 const windowLow = ms->window.lowLimit;
407     U32 matchEndIdx = curr+8+1;
408     size_t bestLength = 8;
409     U32 nbCompares = 1U << cParams->searchLog;
410 #ifdef ZSTD_C_PREDICT
411     U32 predictedSmall = *(bt + 2*((curr-1)&btMask) + 0);
412     U32 predictedLarge = *(bt + 2*((curr-1)&btMask) + 1);
413     predictedSmall += (predictedSmall>0);
414     predictedLarge += (predictedLarge>0);
415 #endif /* ZSTD_C_PREDICT */
416 
417     DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr);
418 
419     assert(ip <= iend-8);   /* required for h calculation */
420     hashTable[h] = curr;   /* Update Hash Table */
421 
422     assert(windowLow > 0);
423     for (; nbCompares && (matchIndex >= windowLow); --nbCompares) {
424         U32* const nextPtr = bt + 2*(matchIndex & btMask);
425         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
426         assert(matchIndex < curr);
427 
428 #ifdef ZSTD_C_PREDICT   /* note : can create issues when hlog small <= 11 */
429         const U32* predictPtr = bt + 2*((matchIndex-1) & btMask);   /* written this way, as bt is a roll buffer */
430         if (matchIndex == predictedSmall) {
431             /* no need to check length, result known */
432             *smallerPtr = matchIndex;
433             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
434             smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
435             matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
436             predictedSmall = predictPtr[1] + (predictPtr[1]>0);
437             continue;
438         }
439         if (matchIndex == predictedLarge) {
440             *largerPtr = matchIndex;
441             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
442             largerPtr = nextPtr;
443             matchIndex = nextPtr[0];
444             predictedLarge = predictPtr[0] + (predictPtr[0]>0);
445             continue;
446         }
447 #endif
448 
449         if (!extDict || (matchIndex+matchLength >= dictLimit)) {
450             assert(matchIndex+matchLength >= dictLimit);   /* might be wrong if actually extDict */
451             match = base + matchIndex;
452             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
453         } else {
454             match = dictBase + matchIndex;
455             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
456             if (matchIndex+matchLength >= dictLimit)
457                 match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
458         }
459 
460         if (matchLength > bestLength) {
461             bestLength = matchLength;
462             if (matchLength > matchEndIdx - matchIndex)
463                 matchEndIdx = matchIndex + (U32)matchLength;
464         }
465 
466         if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
467             break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
468         }
469 
470         if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
471             /* match is smaller than current */
472             *smallerPtr = matchIndex;             /* update smaller idx */
473             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
474             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
475             smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
476             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
477         } else {
478             /* match is larger than current */
479             *largerPtr = matchIndex;
480             commonLengthLarger = matchLength;
481             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
482             largerPtr = nextPtr;
483             matchIndex = nextPtr[0];
484     }   }
485 
486     *smallerPtr = *largerPtr = 0;
487     {   U32 positions = 0;
488         if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384));   /* speed optimization */
489         assert(matchEndIdx > curr + 8);
490         return MAX(positions, matchEndIdx - (curr + 8));
491     }
492 }
493 
494 FORCE_INLINE_TEMPLATE
495 void ZSTD_updateTree_internal(
496                 ZSTD_matchState_t* ms,
497                 const BYTE* const ip, const BYTE* const iend,
498                 const U32 mls, const ZSTD_dictMode_e dictMode)
499 {
500     const BYTE* const base = ms->window.base;
501     U32 const target = (U32)(ip - base);
502     U32 idx = ms->nextToUpdate;
503     DEBUGLOG(6, "ZSTD_updateTree_internal, from %u to %u  (dictMode:%u)",
504                 idx, target, dictMode);
505 
506     while(idx < target) {
507         U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict);
508         assert(idx < (U32)(idx + forward));
509         idx += forward;
510     }
511     assert((size_t)(ip - base) <= (size_t)(U32)(-1));
512     assert((size_t)(iend - base) <= (size_t)(U32)(-1));
513     ms->nextToUpdate = target;
514 }
515 
516 void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
517     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
518 }
519 
520 FORCE_INLINE_TEMPLATE
521 U32 ZSTD_insertBtAndGetAllMatches (
522                     ZSTD_match_t* matches,   /* store result (found matches) in this table (presumed large enough) */
523                     ZSTD_matchState_t* ms,
524                     U32* nextToUpdate3,
525                     const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode,
526                     const U32 rep[ZSTD_REP_NUM],
527                     U32 const ll0,   /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
528                     const U32 lengthToBeat,
529                     U32 const mls /* template */)
530 {
531     const ZSTD_compressionParameters* const cParams = &ms->cParams;
532     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
533     const BYTE* const base = ms->window.base;
534     U32 const curr = (U32)(ip-base);
535     U32 const hashLog = cParams->hashLog;
536     U32 const minMatch = (mls==3) ? 3 : 4;
537     U32* const hashTable = ms->hashTable;
538     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
539     U32 matchIndex  = hashTable[h];
540     U32* const bt   = ms->chainTable;
541     U32 const btLog = cParams->chainLog - 1;
542     U32 const btMask= (1U << btLog) - 1;
543     size_t commonLengthSmaller=0, commonLengthLarger=0;
544     const BYTE* const dictBase = ms->window.dictBase;
545     U32 const dictLimit = ms->window.dictLimit;
546     const BYTE* const dictEnd = dictBase + dictLimit;
547     const BYTE* const prefixStart = base + dictLimit;
548     U32 const btLow = (btMask >= curr) ? 0 : curr - btMask;
549     U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
550     U32 const matchLow = windowLow ? windowLow : 1;
551     U32* smallerPtr = bt + 2*(curr&btMask);
552     U32* largerPtr  = bt + 2*(curr&btMask) + 1;
553     U32 matchEndIdx = curr+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
554     U32 dummy32;   /* to be nullified at the end */
555     U32 mnum = 0;
556     U32 nbCompares = 1U << cParams->searchLog;
557 
558     const ZSTD_matchState_t* dms    = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL;
559     const ZSTD_compressionParameters* const dmsCParams =
560                                       dictMode == ZSTD_dictMatchState ? &dms->cParams : NULL;
561     const BYTE* const dmsBase       = dictMode == ZSTD_dictMatchState ? dms->window.base : NULL;
562     const BYTE* const dmsEnd        = dictMode == ZSTD_dictMatchState ? dms->window.nextSrc : NULL;
563     U32         const dmsHighLimit  = dictMode == ZSTD_dictMatchState ? (U32)(dmsEnd - dmsBase) : 0;
564     U32         const dmsLowLimit   = dictMode == ZSTD_dictMatchState ? dms->window.lowLimit : 0;
565     U32         const dmsIndexDelta = dictMode == ZSTD_dictMatchState ? windowLow - dmsHighLimit : 0;
566     U32         const dmsHashLog    = dictMode == ZSTD_dictMatchState ? dmsCParams->hashLog : hashLog;
567     U32         const dmsBtLog      = dictMode == ZSTD_dictMatchState ? dmsCParams->chainLog - 1 : btLog;
568     U32         const dmsBtMask     = dictMode == ZSTD_dictMatchState ? (1U << dmsBtLog) - 1 : 0;
569     U32         const dmsBtLow      = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit;
570 
571     size_t bestLength = lengthToBeat-1;
572     DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", curr);
573 
574     /* check repCode */
575     assert(ll0 <= 1);   /* necessarily 1 or 0 */
576     {   U32 const lastR = ZSTD_REP_NUM + ll0;
577         U32 repCode;
578         for (repCode = ll0; repCode < lastR; repCode++) {
579             U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
580             U32 const repIndex = curr - repOffset;
581             U32 repLen = 0;
582             assert(curr >= dictLimit);
583             if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < curr-dictLimit) {  /* equivalent to `curr > repIndex >= dictLimit` */
584                 /* We must validate the repcode offset because when we're using a dictionary the
585                  * valid offset range shrinks when the dictionary goes out of bounds.
586                  */
587                 if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) {
588                     repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch;
589                 }
590             } else {  /* repIndex < dictLimit || repIndex >= curr */
591                 const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ?
592                                              dmsBase + repIndex - dmsIndexDelta :
593                                              dictBase + repIndex;
594                 assert(curr >= windowLow);
595                 if ( dictMode == ZSTD_extDict
596                   && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow)  /* equivalent to `curr > repIndex >= windowLow` */
597                      & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */)
598                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
599                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
600                 }
601                 if (dictMode == ZSTD_dictMatchState
602                   && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `curr > repIndex >= dmsLowLimit` */
603                      & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */
604                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
605                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch;
606             }   }
607             /* save longer solution */
608             if (repLen > bestLength) {
609                 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
610                             repCode, ll0, repOffset, repLen);
611                 bestLength = repLen;
612                 matches[mnum].off = repCode - ll0;
613                 matches[mnum].len = (U32)repLen;
614                 mnum++;
615                 if ( (repLen > sufficient_len)
616                    | (ip+repLen == iLimit) ) {  /* best possible */
617                     return mnum;
618     }   }   }   }
619 
620     /* HC3 match finder */
621     if ((mls == 3) /*static*/ && (bestLength < mls)) {
622         U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip);
623         if ((matchIndex3 >= matchLow)
624           & (curr - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
625             size_t mlen;
626             if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) {
627                 const BYTE* const match = base + matchIndex3;
628                 mlen = ZSTD_count(ip, match, iLimit);
629             } else {
630                 const BYTE* const match = dictBase + matchIndex3;
631                 mlen = ZSTD_count_2segments(ip, match, iLimit, dictEnd, prefixStart);
632             }
633 
634             /* save best solution */
635             if (mlen >= mls /* == 3 > bestLength */) {
636                 DEBUGLOG(8, "found small match with hlog3, of length %u",
637                             (U32)mlen);
638                 bestLength = mlen;
639                 assert(curr > matchIndex3);
640                 assert(mnum==0);  /* no prior solution */
641                 matches[0].off = (curr - matchIndex3) + ZSTD_REP_MOVE;
642                 matches[0].len = (U32)mlen;
643                 mnum = 1;
644                 if ( (mlen > sufficient_len) |
645                      (ip+mlen == iLimit) ) {  /* best possible length */
646                     ms->nextToUpdate = curr+1;  /* skip insertion */
647                     return 1;
648         }   }   }
649         /* no dictMatchState lookup: dicts don't have a populated HC3 table */
650     }
651 
652     hashTable[h] = curr;   /* Update Hash Table */
653 
654     for (; nbCompares && (matchIndex >= matchLow); --nbCompares) {
655         U32* const nextPtr = bt + 2*(matchIndex & btMask);
656         const BYTE* match;
657         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
658         assert(curr > matchIndex);
659 
660         if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) {
661             assert(matchIndex+matchLength >= dictLimit);  /* ensure the condition is correct when !extDict */
662             match = base + matchIndex;
663             if (matchIndex >= dictLimit) assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
664             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iLimit);
665         } else {
666             match = dictBase + matchIndex;
667             assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
668             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dictEnd, prefixStart);
669             if (matchIndex+matchLength >= dictLimit)
670                 match = base + matchIndex;   /* prepare for match[matchLength] read */
671         }
672 
673         if (matchLength > bestLength) {
674             DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)",
675                     (U32)matchLength, curr - matchIndex, curr - matchIndex + ZSTD_REP_MOVE);
676             assert(matchEndIdx > matchIndex);
677             if (matchLength > matchEndIdx - matchIndex)
678                 matchEndIdx = matchIndex + (U32)matchLength;
679             bestLength = matchLength;
680             matches[mnum].off = (curr - matchIndex) + ZSTD_REP_MOVE;
681             matches[mnum].len = (U32)matchLength;
682             mnum++;
683             if ( (matchLength > ZSTD_OPT_NUM)
684                | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
685                 if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */
686                 break; /* drop, to preserve bt consistency (miss a little bit of compression) */
687             }
688         }
689 
690         if (match[matchLength] < ip[matchLength]) {
691             /* match smaller than current */
692             *smallerPtr = matchIndex;             /* update smaller idx */
693             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
694             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
695             smallerPtr = nextPtr+1;               /* new candidate => larger than match, which was smaller than current */
696             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous, closer to current */
697         } else {
698             *largerPtr = matchIndex;
699             commonLengthLarger = matchLength;
700             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
701             largerPtr = nextPtr;
702             matchIndex = nextPtr[0];
703     }   }
704 
705     *smallerPtr = *largerPtr = 0;
706 
707     assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
708     if (dictMode == ZSTD_dictMatchState && nbCompares) {
709         size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls);
710         U32 dictMatchIndex = dms->hashTable[dmsH];
711         const U32* const dmsBt = dms->chainTable;
712         commonLengthSmaller = commonLengthLarger = 0;
713         for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) {
714             const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask);
715             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
716             const BYTE* match = dmsBase + dictMatchIndex;
717             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dmsEnd, prefixStart);
718             if (dictMatchIndex+matchLength >= dmsHighLimit)
719                 match = base + dictMatchIndex + dmsIndexDelta;   /* to prepare for next usage of match[matchLength] */
720 
721             if (matchLength > bestLength) {
722                 matchIndex = dictMatchIndex + dmsIndexDelta;
723                 DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)",
724                         (U32)matchLength, curr - matchIndex, curr - matchIndex + ZSTD_REP_MOVE);
725                 if (matchLength > matchEndIdx - matchIndex)
726                     matchEndIdx = matchIndex + (U32)matchLength;
727                 bestLength = matchLength;
728                 matches[mnum].off = (curr - matchIndex) + ZSTD_REP_MOVE;
729                 matches[mnum].len = (U32)matchLength;
730                 mnum++;
731                 if ( (matchLength > ZSTD_OPT_NUM)
732                    | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
733                     break;   /* drop, to guarantee consistency (miss a little bit of compression) */
734                 }
735             }
736 
737             if (dictMatchIndex <= dmsBtLow) { break; }   /* beyond tree size, stop the search */
738             if (match[matchLength] < ip[matchLength]) {
739                 commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
740                 dictMatchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
741             } else {
742                 /* match is larger than current */
743                 commonLengthLarger = matchLength;
744                 dictMatchIndex = nextPtr[0];
745             }
746         }
747     }
748 
749     assert(matchEndIdx > curr+8);
750     ms->nextToUpdate = matchEndIdx - 8;  /* skip repetitive patterns */
751     return mnum;
752 }
753 
754 
755 FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches (
756                         ZSTD_match_t* matches,   /* store result (match found, increasing size) in this table */
757                         ZSTD_matchState_t* ms,
758                         U32* nextToUpdate3,
759                         const BYTE* ip, const BYTE* const iHighLimit, const ZSTD_dictMode_e dictMode,
760                         const U32 rep[ZSTD_REP_NUM],
761                         U32 const ll0,
762                         U32 const lengthToBeat)
763 {
764     const ZSTD_compressionParameters* const cParams = &ms->cParams;
765     U32 const matchLengthSearch = cParams->minMatch;
766     DEBUGLOG(8, "ZSTD_BtGetAllMatches");
767     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
768     ZSTD_updateTree_internal(ms, ip, iHighLimit, matchLengthSearch, dictMode);
769     switch(matchLengthSearch)
770     {
771     case 3 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 3);
772     default :
773     case 4 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 4);
774     case 5 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 5);
775     case 7 :
776     case 6 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 6);
777     }
778 }
779 
780 /* ***********************
781 *  LDM helper functions  *
782 *************************/
783 
784 /* Struct containing info needed to make decision about ldm inclusion */
785 typedef struct {
786     rawSeqStore_t seqStore;         /* External match candidates store for this block */
787     U32 startPosInBlock;            /* Start position of the current match candidate */
788     U32 endPosInBlock;              /* End position of the current match candidate */
789     U32 offset;                     /* Offset of the match candidate */
790 } ZSTD_optLdm_t;
791 
792 /* ZSTD_optLdm_skipRawSeqStoreBytes():
793  * Moves forward in rawSeqStore by nbBytes, which will update the fields 'pos' and 'posInSequence'.
794  */
795 static void ZSTD_optLdm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
796     U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
797     while (currPos && rawSeqStore->pos < rawSeqStore->size) {
798         rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
799         if (currPos >= currSeq.litLength + currSeq.matchLength) {
800             currPos -= currSeq.litLength + currSeq.matchLength;
801             rawSeqStore->pos++;
802         } else {
803             rawSeqStore->posInSequence = currPos;
804             break;
805         }
806     }
807     if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
808         rawSeqStore->posInSequence = 0;
809     }
810 }
811 
812 /* ZSTD_opt_getNextMatchAndUpdateSeqStore():
813  * Calculates the beginning and end of the next match in the current block.
814  * Updates 'pos' and 'posInSequence' of the ldmSeqStore.
815  */
816 static void ZSTD_opt_getNextMatchAndUpdateSeqStore(ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
817                                                    U32 blockBytesRemaining) {
818     rawSeq currSeq;
819     U32 currBlockEndPos;
820     U32 literalsBytesRemaining;
821     U32 matchBytesRemaining;
822 
823     /* Setting match end position to MAX to ensure we never use an LDM during this block */
824     if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
825         optLdm->startPosInBlock = UINT_MAX;
826         optLdm->endPosInBlock = UINT_MAX;
827         return;
828     }
829     /* Calculate appropriate bytes left in matchLength and litLength after adjusting
830        based on ldmSeqStore->posInSequence */
831     currSeq = optLdm->seqStore.seq[optLdm->seqStore.pos];
832     assert(optLdm->seqStore.posInSequence <= currSeq.litLength + currSeq.matchLength);
833     currBlockEndPos = currPosInBlock + blockBytesRemaining;
834     literalsBytesRemaining = (optLdm->seqStore.posInSequence < currSeq.litLength) ?
835             currSeq.litLength - (U32)optLdm->seqStore.posInSequence :
836             0;
837     matchBytesRemaining = (literalsBytesRemaining == 0) ?
838             currSeq.matchLength - ((U32)optLdm->seqStore.posInSequence - currSeq.litLength) :
839             currSeq.matchLength;
840 
841     /* If there are more literal bytes than bytes remaining in block, no ldm is possible */
842     if (literalsBytesRemaining >= blockBytesRemaining) {
843         optLdm->startPosInBlock = UINT_MAX;
844         optLdm->endPosInBlock = UINT_MAX;
845         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, blockBytesRemaining);
846         return;
847     }
848 
849     /* Matches may be < MINMATCH by this process. In that case, we will reject them
850        when we are deciding whether or not to add the ldm */
851     optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining;
852     optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining;
853     optLdm->offset = currSeq.offset;
854 
855     if (optLdm->endPosInBlock > currBlockEndPos) {
856         /* Match ends after the block ends, we can't use the whole match */
857         optLdm->endPosInBlock = currBlockEndPos;
858         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, currBlockEndPos - currPosInBlock);
859     } else {
860         /* Consume nb of bytes equal to size of sequence left */
861         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, literalsBytesRemaining + matchBytesRemaining);
862     }
863 }
864 
865 /* ZSTD_optLdm_maybeAddMatch():
866  * Adds a match if it's long enough, based on it's 'matchStartPosInBlock'
867  * and 'matchEndPosInBlock', into 'matches'. Maintains the correct ordering of 'matches'
868  */
869 static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
870                                       ZSTD_optLdm_t* optLdm, U32 currPosInBlock) {
871     U32 posDiff = currPosInBlock - optLdm->startPosInBlock;
872     /* Note: ZSTD_match_t actually contains offCode and matchLength (before subtracting MINMATCH) */
873     U32 candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;
874     U32 candidateOffCode = optLdm->offset + ZSTD_REP_MOVE;
875 
876     /* Ensure that current block position is not outside of the match */
877     if (currPosInBlock < optLdm->startPosInBlock
878       || currPosInBlock >= optLdm->endPosInBlock
879       || candidateMatchLength < MINMATCH) {
880         return;
881     }
882 
883     if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
884         DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offCode: %u matchLength %u) at block position=%u",
885                  candidateOffCode, candidateMatchLength, currPosInBlock);
886         matches[*nbMatches].len = candidateMatchLength;
887         matches[*nbMatches].off = candidateOffCode;
888         (*nbMatches)++;
889     }
890 }
891 
892 /* ZSTD_optLdm_processMatchCandidate():
893  * Wrapper function to update ldm seq store and call ldm functions as necessary.
894  */
895 static void ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm, ZSTD_match_t* matches, U32* nbMatches,
896                                               U32 currPosInBlock, U32 remainingBytes) {
897     if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
898         return;
899     }
900 
901     if (currPosInBlock >= optLdm->endPosInBlock) {
902         if (currPosInBlock > optLdm->endPosInBlock) {
903             /* The position at which ZSTD_optLdm_processMatchCandidate() is called is not necessarily
904              * at the end of a match from the ldm seq store, and will often be some bytes
905              * over beyond matchEndPosInBlock. As such, we need to correct for these "overshoots"
906              */
907             U32 posOvershoot = currPosInBlock - optLdm->endPosInBlock;
908             ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot);
909         }
910         ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
911     }
912     ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
913 }
914 
915 /*-*******************************
916 *  Optimal parser
917 *********************************/
918 
919 
920 static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
921 {
922     return sol.litlen + sol.mlen;
923 }
924 
925 #if 0 /* debug */
926 
927 static void
928 listStats(const U32* table, int lastEltID)
929 {
930     int const nbElts = lastEltID + 1;
931     int enb;
932     for (enb=0; enb < nbElts; enb++) {
933         (void)table;
934         /* RAWLOG(2, "%3i:%3i,  ", enb, table[enb]); */
935         RAWLOG(2, "%4i,", table[enb]);
936     }
937     RAWLOG(2, " \n");
938 }
939 
940 #endif
941 
942 FORCE_INLINE_TEMPLATE size_t
943 ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
944                                seqStore_t* seqStore,
945                                U32 rep[ZSTD_REP_NUM],
946                          const void* src, size_t srcSize,
947                          const int optLevel,
948                          const ZSTD_dictMode_e dictMode)
949 {
950     optState_t* const optStatePtr = &ms->opt;
951     const BYTE* const istart = (const BYTE*)src;
952     const BYTE* ip = istart;
953     const BYTE* anchor = istart;
954     const BYTE* const iend = istart + srcSize;
955     const BYTE* const ilimit = iend - 8;
956     const BYTE* const base = ms->window.base;
957     const BYTE* const prefixStart = base + ms->window.dictLimit;
958     const ZSTD_compressionParameters* const cParams = &ms->cParams;
959 
960     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
961     U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4;
962     U32 nextToUpdate3 = ms->nextToUpdate;
963 
964     ZSTD_optimal_t* const opt = optStatePtr->priceTable;
965     ZSTD_match_t* const matches = optStatePtr->matchTable;
966     ZSTD_optimal_t lastSequence;
967     ZSTD_optLdm_t optLdm;
968 
969     optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
970     optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
971     ZSTD_opt_getNextMatchAndUpdateSeqStore(&optLdm, (U32)(ip-istart), (U32)(iend-ip));
972 
973     /* init */
974     DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u",
975                 (U32)(ip - base), ms->window.dictLimit, ms->nextToUpdate);
976     assert(optLevel <= 2);
977     ZSTD_rescaleFreqs(optStatePtr, (const BYTE*)src, srcSize, optLevel);
978     ip += (ip==prefixStart);
979 
980     /* Match Loop */
981     while (ip < ilimit) {
982         U32 cur, last_pos = 0;
983 
984         /* find first match */
985         {   U32 const litlen = (U32)(ip - anchor);
986             U32 const ll0 = !litlen;
987             U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch);
988             ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
989                                               (U32)(ip-istart), (U32)(iend - ip));
990             if (!nbMatches) { ip++; continue; }
991 
992             /* initialize opt[0] */
993             { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; }
994             opt[0].mlen = 0;  /* means is_a_literal */
995             opt[0].litlen = litlen;
996             /* We don't need to include the actual price of the literals because
997              * it is static for the duration of the forward pass, and is included
998              * in every price. We include the literal length to avoid negative
999              * prices when we subtract the previous literal length.
1000              */
1001             opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
1002 
1003             /* large match -> immediate encoding */
1004             {   U32 const maxML = matches[nbMatches-1].len;
1005                 U32 const maxOffset = matches[nbMatches-1].off;
1006                 DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series",
1007                             nbMatches, maxML, maxOffset, (U32)(ip-prefixStart));
1008 
1009                 if (maxML > sufficient_len) {
1010                     lastSequence.litlen = litlen;
1011                     lastSequence.mlen = maxML;
1012                     lastSequence.off = maxOffset;
1013                     DEBUGLOG(6, "large match (%u>%u), immediate encoding",
1014                                 maxML, sufficient_len);
1015                     cur = 0;
1016                     last_pos = ZSTD_totalLen(lastSequence);
1017                     goto _shortestPath;
1018             }   }
1019 
1020             /* set prices for first matches starting position == 0 */
1021             {   U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
1022                 U32 pos;
1023                 U32 matchNb;
1024                 for (pos = 1; pos < minMatch; pos++) {
1025                     opt[pos].price = ZSTD_MAX_PRICE;   /* mlen, litlen and price will be fixed during forward scanning */
1026                 }
1027                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
1028                     U32 const offset = matches[matchNb].off;
1029                     U32 const end = matches[matchNb].len;
1030                     for ( ; pos <= end ; pos++ ) {
1031                         U32 const matchPrice = ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel);
1032                         U32 const sequencePrice = literalsPrice + matchPrice;
1033                         DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
1034                                     pos, ZSTD_fCost(sequencePrice));
1035                         opt[pos].mlen = pos;
1036                         opt[pos].off = offset;
1037                         opt[pos].litlen = litlen;
1038                         opt[pos].price = sequencePrice;
1039                 }   }
1040                 last_pos = pos-1;
1041             }
1042         }
1043 
1044         /* check further positions */
1045         for (cur = 1; cur <= last_pos; cur++) {
1046             const BYTE* const inr = ip + cur;
1047             assert(cur < ZSTD_OPT_NUM);
1048             DEBUGLOG(7, "cPos:%zi==rPos:%u", inr-istart, cur)
1049 
1050             /* Fix current position with one literal if cheaper */
1051             {   U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
1052                 int const price = opt[cur-1].price
1053                                 + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
1054                                 + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
1055                                 - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
1056                 assert(price < 1000000000); /* overflow check */
1057                 if (price <= opt[cur].price) {
1058                     DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
1059                                 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
1060                                 opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]);
1061                     opt[cur].mlen = 0;
1062                     opt[cur].off = 0;
1063                     opt[cur].litlen = litlen;
1064                     opt[cur].price = price;
1065                 } else {
1066                     DEBUGLOG(7, "cPos:%zi==rPos:%u : literal would cost more (%.2f>%.2f) (hist:%u,%u,%u)",
1067                                 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price),
1068                                 opt[cur].rep[0], opt[cur].rep[1], opt[cur].rep[2]);
1069                 }
1070             }
1071 
1072             /* Set the repcodes of the current position. We must do it here
1073              * because we rely on the repcodes of the 2nd to last sequence being
1074              * correct to set the next chunks repcodes during the backward
1075              * traversal.
1076              */
1077             ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(repcodes_t));
1078             assert(cur >= opt[cur].mlen);
1079             if (opt[cur].mlen != 0) {
1080                 U32 const prev = cur - opt[cur].mlen;
1081                 repcodes_t newReps = ZSTD_updateRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
1082                 ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
1083             } else {
1084                 ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
1085             }
1086 
1087             /* last match must start at a minimum distance of 8 from oend */
1088             if (inr > ilimit) continue;
1089 
1090             if (cur == last_pos) break;
1091 
1092             if ( (optLevel==0) /*static_test*/
1093               && (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) {
1094                 DEBUGLOG(7, "move to next rPos:%u : price is <=", cur+1);
1095                 continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
1096             }
1097 
1098             {   U32 const ll0 = (opt[cur].mlen != 0);
1099                 U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
1100                 U32 const previousPrice = opt[cur].price;
1101                 U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
1102                 U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch);
1103                 U32 matchNb;
1104 
1105                 ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
1106                                                   (U32)(inr-istart), (U32)(iend-inr));
1107 
1108                 if (!nbMatches) {
1109                     DEBUGLOG(7, "rPos:%u : no match found", cur);
1110                     continue;
1111                 }
1112 
1113                 {   U32 const maxML = matches[nbMatches-1].len;
1114                     DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of maxLength=%u",
1115                                 inr-istart, cur, nbMatches, maxML);
1116 
1117                     if ( (maxML > sufficient_len)
1118                       || (cur + maxML >= ZSTD_OPT_NUM) ) {
1119                         lastSequence.mlen = maxML;
1120                         lastSequence.off = matches[nbMatches-1].off;
1121                         lastSequence.litlen = litlen;
1122                         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 */
1123                         last_pos = cur + ZSTD_totalLen(lastSequence);
1124                         if (cur > ZSTD_OPT_NUM) cur = 0;   /* underflow => first match */
1125                         goto _shortestPath;
1126                 }   }
1127 
1128                 /* set prices using matches found at position == cur */
1129                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
1130                     U32 const offset = matches[matchNb].off;
1131                     U32 const lastML = matches[matchNb].len;
1132                     U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
1133                     U32 mlen;
1134 
1135                     DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u",
1136                                 matchNb, matches[matchNb].off, lastML, litlen);
1137 
1138                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
1139                         U32 const pos = cur + mlen;
1140                         int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
1141 
1142                         if ((pos > last_pos) || (price < opt[pos].price)) {
1143                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
1144                                         pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
1145                             while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; }   /* fill empty positions */
1146                             opt[pos].mlen = mlen;
1147                             opt[pos].off = offset;
1148                             opt[pos].litlen = litlen;
1149                             opt[pos].price = price;
1150                         } else {
1151                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)",
1152                                         pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
1153                             if (optLevel==0) break;  /* early update abort; gets ~+10% speed for about -0.01 ratio loss */
1154                         }
1155             }   }   }
1156         }  /* for (cur = 1; cur <= last_pos; cur++) */
1157 
1158         lastSequence = opt[last_pos];
1159         cur = last_pos > ZSTD_totalLen(lastSequence) ? last_pos - ZSTD_totalLen(lastSequence) : 0;  /* single sequence, and it starts before `ip` */
1160         assert(cur < ZSTD_OPT_NUM);  /* control overflow*/
1161 
1162 _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
1163         assert(opt[0].mlen == 0);
1164 
1165         /* Set the next chunk's repcodes based on the repcodes of the beginning
1166          * of the last match, and the last sequence. This avoids us having to
1167          * update them while traversing the sequences.
1168          */
1169         if (lastSequence.mlen != 0) {
1170             repcodes_t reps = ZSTD_updateRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
1171             ZSTD_memcpy(rep, &reps, sizeof(reps));
1172         } else {
1173             ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
1174         }
1175 
1176         {   U32 const storeEnd = cur + 1;
1177             U32 storeStart = storeEnd;
1178             U32 seqPos = cur;
1179 
1180             DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)",
1181                         last_pos, cur); (void)last_pos;
1182             assert(storeEnd < ZSTD_OPT_NUM);
1183             DEBUGLOG(6, "last sequence copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
1184                         storeEnd, lastSequence.litlen, lastSequence.mlen, lastSequence.off);
1185             opt[storeEnd] = lastSequence;
1186             while (seqPos > 0) {
1187                 U32 const backDist = ZSTD_totalLen(opt[seqPos]);
1188                 storeStart--;
1189                 DEBUGLOG(6, "sequence from rPos=%u copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
1190                             seqPos, storeStart, opt[seqPos].litlen, opt[seqPos].mlen, opt[seqPos].off);
1191                 opt[storeStart] = opt[seqPos];
1192                 seqPos = (seqPos > backDist) ? seqPos - backDist : 0;
1193             }
1194 
1195             /* save sequences */
1196             DEBUGLOG(6, "sending selected sequences into seqStore")
1197             {   U32 storePos;
1198                 for (storePos=storeStart; storePos <= storeEnd; storePos++) {
1199                     U32 const llen = opt[storePos].litlen;
1200                     U32 const mlen = opt[storePos].mlen;
1201                     U32 const offCode = opt[storePos].off;
1202                     U32 const advance = llen + mlen;
1203                     DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u",
1204                                 anchor - istart, (unsigned)llen, (unsigned)mlen);
1205 
1206                     if (mlen==0) {  /* only literals => must be last "sequence", actually starting a new stream of sequences */
1207                         assert(storePos == storeEnd);   /* must be last sequence */
1208                         ip = anchor + llen;     /* last "sequence" is a bunch of literals => don't progress anchor */
1209                         continue;   /* will finish */
1210                     }
1211 
1212                     assert(anchor + llen <= iend);
1213                     ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen);
1214                     ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen-MINMATCH);
1215                     anchor += advance;
1216                     ip = anchor;
1217             }   }
1218             ZSTD_setBasePrices(optStatePtr, optLevel);
1219         }
1220     }   /* while (ip < ilimit) */
1221 
1222     /* Return the last literals size */
1223     return (size_t)(iend - anchor);
1224 }
1225 
1226 
1227 size_t ZSTD_compressBlock_btopt(
1228         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1229         const void* src, size_t srcSize)
1230 {
1231     DEBUGLOG(5, "ZSTD_compressBlock_btopt");
1232     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict);
1233 }
1234 
1235 
1236 /* used in 2-pass strategy */
1237 static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus)
1238 {
1239     U32 s, sum=0;
1240     assert(ZSTD_FREQ_DIV+bonus >= 0);
1241     for (s=0; s<lastEltIndex+1; s++) {
1242         table[s] <<= ZSTD_FREQ_DIV+bonus;
1243         table[s]--;
1244         sum += table[s];
1245     }
1246     return sum;
1247 }
1248 
1249 /* used in 2-pass strategy */
1250 MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr)
1251 {
1252     if (ZSTD_compressedLiterals(optPtr))
1253         optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0);
1254     optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0);
1255     optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0);
1256     optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0);
1257 }
1258 
1259 /* ZSTD_initStats_ultra():
1260  * make a first compression pass, just to seed stats with more accurate starting values.
1261  * only works on first block, with no dictionary and no ldm.
1262  * this function cannot error, hence its contract must be respected.
1263  */
1264 static void
1265 ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
1266                      seqStore_t* seqStore,
1267                      U32 rep[ZSTD_REP_NUM],
1268                const void* src, size_t srcSize)
1269 {
1270     U32 tmpRep[ZSTD_REP_NUM];  /* updated rep codes will sink here */
1271     ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep));
1272 
1273     DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize);
1274     assert(ms->opt.litLengthSum == 0);    /* first block */
1275     assert(seqStore->sequences == seqStore->sequencesStart);   /* no ldm */
1276     assert(ms->window.dictLimit == ms->window.lowLimit);   /* no dictionary */
1277     assert(ms->window.dictLimit - ms->nextToUpdate <= 1);  /* no prefix (note: intentional overflow, defined as 2-complement) */
1278 
1279     ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);   /* generate stats into ms->opt*/
1280 
1281     /* invalidate first scan from history */
1282     ZSTD_resetSeqStore(seqStore);
1283     ms->window.base -= srcSize;
1284     ms->window.dictLimit += (U32)srcSize;
1285     ms->window.lowLimit = ms->window.dictLimit;
1286     ms->nextToUpdate = ms->window.dictLimit;
1287 
1288     /* re-inforce weight of collected statistics */
1289     ZSTD_upscaleStats(&ms->opt);
1290 }
1291 
1292 size_t ZSTD_compressBlock_btultra(
1293         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1294         const void* src, size_t srcSize)
1295 {
1296     DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
1297     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
1298 }
1299 
1300 size_t ZSTD_compressBlock_btultra2(
1301         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1302         const void* src, size_t srcSize)
1303 {
1304     U32 const curr = (U32)((const BYTE*)src - ms->window.base);
1305     DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);
1306 
1307     /* 2-pass strategy:
1308      * this strategy makes a first pass over first block to collect statistics
1309      * and seed next round's statistics with it.
1310      * After 1st pass, function forgets everything, and starts a new block.
1311      * Consequently, this can only work if no data has been previously loaded in tables,
1312      * aka, no dictionary, no prefix, no ldm preprocessing.
1313      * The compression ratio gain is generally small (~0.5% on first block),
1314      * the cost is 2x cpu time on first block. */
1315     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
1316     if ( (ms->opt.litLengthSum==0)   /* first block */
1317       && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
1318       && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
1319       && (curr == ms->window.dictLimit)   /* start of frame, nothing already loaded nor skipped */
1320       && (srcSize > ZSTD_PREDEF_THRESHOLD)
1321       ) {
1322         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
1323     }
1324 
1325     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
1326 }
1327 
1328 size_t ZSTD_compressBlock_btopt_dictMatchState(
1329         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1330         const void* src, size_t srcSize)
1331 {
1332     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState);
1333 }
1334 
1335 size_t ZSTD_compressBlock_btultra_dictMatchState(
1336         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1337         const void* src, size_t srcSize)
1338 {
1339     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState);
1340 }
1341 
1342 size_t ZSTD_compressBlock_btopt_extDict(
1343         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1344         const void* src, size_t srcSize)
1345 {
1346     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict);
1347 }
1348 
1349 size_t ZSTD_compressBlock_btultra_extDict(
1350         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1351         const void* src, size_t srcSize)
1352 {
1353     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict);
1354 }
1355 
1356 /* note : no btultra2 variant for extDict nor dictMatchState,
1357  * because btultra2 is not meant to work with dictionaries
1358  * and is only specific for the first block (no prefix) */
1359