xref: /openbmc/linux/lib/zstd/compress/zstd_opt.c (revision 63705da3)
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 #include "zstd_compress_internal.h"
12 #include "hist.h"
13 #include "zstd_opt.h"
14 
15 
16 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
17 #define ZSTD_FREQ_DIV       4   /* log factor when using previous stats to init next stats */
18 #define ZSTD_MAX_PRICE     (1<<30)
19 
20 #define ZSTD_PREDEF_THRESHOLD 1024   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
21 
22 
23 /*-*************************************
24 *  Price functions for optimal parser
25 ***************************************/
26 
27 #if 0    /* approximation at bit level */
28 #  define BITCOST_ACCURACY 0
29 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
30 #  define WEIGHT(stat)  ((void)opt, ZSTD_bitWeight(stat))
31 #elif 0  /* fractional bit accuracy */
32 #  define BITCOST_ACCURACY 8
33 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
34 #  define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat))
35 #else    /* opt==approx, ultra==accurate */
36 #  define BITCOST_ACCURACY 8
37 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
38 #  define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
39 #endif
40 
41 MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
42 {
43     return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
44 }
45 
46 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
47 {
48     U32 const stat = rawStat + 1;
49     U32 const hb = ZSTD_highbit32(stat);
50     U32 const BWeight = hb * BITCOST_MULTIPLIER;
51     U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
52     U32 const weight = BWeight + FWeight;
53     assert(hb + BITCOST_ACCURACY < 31);
54     return weight;
55 }
56 
57 #if (DEBUGLEVEL>=2)
58 /* debugging function,
59  * @return price in bytes as fractional value
60  * for debug messages only */
61 MEM_STATIC double ZSTD_fCost(U32 price)
62 {
63     return (double)price / (BITCOST_MULTIPLIER*8);
64 }
65 #endif
66 
67 static int ZSTD_compressedLiterals(optState_t const* const optPtr)
68 {
69     return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed;
70 }
71 
72 static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
73 {
74     if (ZSTD_compressedLiterals(optPtr))
75         optPtr->litSumBasePrice = WEIGHT(optPtr->litSum, optLevel);
76     optPtr->litLengthSumBasePrice = WEIGHT(optPtr->litLengthSum, optLevel);
77     optPtr->matchLengthSumBasePrice = WEIGHT(optPtr->matchLengthSum, optLevel);
78     optPtr->offCodeSumBasePrice = WEIGHT(optPtr->offCodeSum, optLevel);
79 }
80 
81 
82 /* ZSTD_downscaleStat() :
83  * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus)
84  * return the resulting sum of elements */
85 static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus)
86 {
87     U32 s, sum=0;
88     DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1);
89     assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31);
90     for (s=0; s<lastEltIndex+1; s++) {
91         table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus));
92         sum += table[s];
93     }
94     return sum;
95 }
96 
97 /* ZSTD_rescaleFreqs() :
98  * if first block (detected by optPtr->litLengthSum == 0) : init statistics
99  *    take hints from dictionary if there is one
100  *    or init from zero, using src for literals stats, or flat 1 for match symbols
101  * otherwise downscale existing stats, to be used as seed for next block.
102  */
103 static void
104 ZSTD_rescaleFreqs(optState_t* const optPtr,
105             const BYTE* const src, size_t const srcSize,
106                   int const optLevel)
107 {
108     int const compressedLiterals = ZSTD_compressedLiterals(optPtr);
109     DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
110     optPtr->priceType = zop_dynamic;
111 
112     if (optPtr->litLengthSum == 0) {  /* first block : init */
113         if (srcSize <= ZSTD_PREDEF_THRESHOLD) {  /* heuristic */
114             DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef");
115             optPtr->priceType = zop_predef;
116         }
117 
118         assert(optPtr->symbolCosts != NULL);
119         if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {
120             /* huffman table presumed generated by dictionary */
121             optPtr->priceType = zop_dynamic;
122 
123             if (compressedLiterals) {
124                 unsigned lit;
125                 assert(optPtr->litFreq != NULL);
126                 optPtr->litSum = 0;
127                 for (lit=0; lit<=MaxLit; lit++) {
128                     U32 const scaleLog = 11;   /* scale to 2K */
129                     U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit);
130                     assert(bitCost <= scaleLog);
131                     optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
132                     optPtr->litSum += optPtr->litFreq[lit];
133             }   }
134 
135             {   unsigned ll;
136                 FSE_CState_t llstate;
137                 FSE_initCState(&llstate, optPtr->symbolCosts->fse.litlengthCTable);
138                 optPtr->litLengthSum = 0;
139                 for (ll=0; ll<=MaxLL; ll++) {
140                     U32 const scaleLog = 10;   /* scale to 1K */
141                     U32 const bitCost = FSE_getMaxNbBits(llstate.symbolTT, ll);
142                     assert(bitCost < scaleLog);
143                     optPtr->litLengthFreq[ll] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
144                     optPtr->litLengthSum += optPtr->litLengthFreq[ll];
145             }   }
146 
147             {   unsigned ml;
148                 FSE_CState_t mlstate;
149                 FSE_initCState(&mlstate, optPtr->symbolCosts->fse.matchlengthCTable);
150                 optPtr->matchLengthSum = 0;
151                 for (ml=0; ml<=MaxML; ml++) {
152                     U32 const scaleLog = 10;
153                     U32 const bitCost = FSE_getMaxNbBits(mlstate.symbolTT, ml);
154                     assert(bitCost < scaleLog);
155                     optPtr->matchLengthFreq[ml] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
156                     optPtr->matchLengthSum += optPtr->matchLengthFreq[ml];
157             }   }
158 
159             {   unsigned of;
160                 FSE_CState_t ofstate;
161                 FSE_initCState(&ofstate, optPtr->symbolCosts->fse.offcodeCTable);
162                 optPtr->offCodeSum = 0;
163                 for (of=0; of<=MaxOff; of++) {
164                     U32 const scaleLog = 10;
165                     U32 const bitCost = FSE_getMaxNbBits(ofstate.symbolTT, of);
166                     assert(bitCost < scaleLog);
167                     optPtr->offCodeFreq[of] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
168                     optPtr->offCodeSum += optPtr->offCodeFreq[of];
169             }   }
170 
171         } else {  /* not a dictionary */
172 
173             assert(optPtr->litFreq != NULL);
174             if (compressedLiterals) {
175                 unsigned lit = MaxLit;
176                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
177                 optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
178             }
179 
180             {   unsigned ll;
181                 for (ll=0; ll<=MaxLL; ll++)
182                     optPtr->litLengthFreq[ll] = 1;
183             }
184             optPtr->litLengthSum = MaxLL+1;
185 
186             {   unsigned ml;
187                 for (ml=0; ml<=MaxML; ml++)
188                     optPtr->matchLengthFreq[ml] = 1;
189             }
190             optPtr->matchLengthSum = MaxML+1;
191 
192             {   unsigned of;
193                 for (of=0; of<=MaxOff; of++)
194                     optPtr->offCodeFreq[of] = 1;
195             }
196             optPtr->offCodeSum = MaxOff+1;
197 
198         }
199 
200     } else {   /* new block : re-use previous statistics, scaled down */
201 
202         if (compressedLiterals)
203             optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
204         optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0);
205         optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0);
206         optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0);
207     }
208 
209     ZSTD_setBasePrices(optPtr, optLevel);
210 }
211 
212 /* ZSTD_rawLiteralsCost() :
213  * price of literals (only) in specified segment (which length can be 0).
214  * does not include price of literalLength symbol */
215 static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
216                                 const optState_t* const optPtr,
217                                 int optLevel)
218 {
219     if (litLength == 0) return 0;
220 
221     if (!ZSTD_compressedLiterals(optPtr))
222         return (litLength << 3) * BITCOST_MULTIPLIER;  /* Uncompressed - 8 bytes per literal. */
223 
224     if (optPtr->priceType == zop_predef)
225         return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */
226 
227     /* dynamic statistics */
228     {   U32 price = litLength * optPtr->litSumBasePrice;
229         U32 u;
230         for (u=0; u < litLength; u++) {
231             assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice);   /* literal cost should never be negative */
232             price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel);
233         }
234         return price;
235     }
236 }
237 
238 /* ZSTD_litLengthPrice() :
239  * cost of literalLength symbol */
240 static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel)
241 {
242     if (optPtr->priceType == zop_predef) return WEIGHT(litLength, optLevel);
243 
244     /* dynamic statistics */
245     {   U32 const llCode = ZSTD_LLcode(litLength);
246         return (LL_bits[llCode] * BITCOST_MULTIPLIER)
247              + optPtr->litLengthSumBasePrice
248              - WEIGHT(optPtr->litLengthFreq[llCode], optLevel);
249     }
250 }
251 
252 /* ZSTD_getMatchPrice() :
253  * Provides the cost of the match part (offset + matchLength) of a sequence
254  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
255  * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */
256 FORCE_INLINE_TEMPLATE U32
257 ZSTD_getMatchPrice(U32 const offset,
258                    U32 const matchLength,
259              const optState_t* const optPtr,
260                    int const optLevel)
261 {
262     U32 price;
263     U32 const offCode = ZSTD_highbit32(offset+1);
264     U32 const mlBase = matchLength - MINMATCH;
265     assert(matchLength >= MINMATCH);
266 
267     if (optPtr->priceType == zop_predef)  /* fixed scheme, do not use statistics */
268         return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER);
269 
270     /* dynamic statistics */
271     price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
272     if ((optLevel<2) /*static*/ && offCode >= 20)
273         price += (offCode-19)*2 * BITCOST_MULTIPLIER; /* handicap for long distance offsets, favor decompression speed */
274 
275     /* match Length */
276     {   U32 const mlCode = ZSTD_MLcode(mlBase);
277         price += (ML_bits[mlCode] * BITCOST_MULTIPLIER) + (optPtr->matchLengthSumBasePrice - WEIGHT(optPtr->matchLengthFreq[mlCode], optLevel));
278     }
279 
280     price += BITCOST_MULTIPLIER / 5;   /* heuristic : make matches a bit more costly to favor less sequences -> faster decompression speed */
281 
282     DEBUGLOG(8, "ZSTD_getMatchPrice(ml:%u) = %u", matchLength, price);
283     return price;
284 }
285 
286 /* ZSTD_updateStats() :
287  * assumption : literals + litLengtn <= iend */
288 static void ZSTD_updateStats(optState_t* const optPtr,
289                              U32 litLength, const BYTE* literals,
290                              U32 offsetCode, U32 matchLength)
291 {
292     /* literals */
293     if (ZSTD_compressedLiterals(optPtr)) {
294         U32 u;
295         for (u=0; u < litLength; u++)
296             optPtr->litFreq[literals[u]] += ZSTD_LITFREQ_ADD;
297         optPtr->litSum += litLength*ZSTD_LITFREQ_ADD;
298     }
299 
300     /* literal Length */
301     {   U32 const llCode = ZSTD_LLcode(litLength);
302         optPtr->litLengthFreq[llCode]++;
303         optPtr->litLengthSum++;
304     }
305 
306     /* match offset code (0-2=>repCode; 3+=>offset+2) */
307     {   U32 const offCode = ZSTD_highbit32(offsetCode+1);
308         assert(offCode <= MaxOff);
309         optPtr->offCodeFreq[offCode]++;
310         optPtr->offCodeSum++;
311     }
312 
313     /* match Length */
314     {   U32 const mlBase = matchLength - MINMATCH;
315         U32 const mlCode = ZSTD_MLcode(mlBase);
316         optPtr->matchLengthFreq[mlCode]++;
317         optPtr->matchLengthSum++;
318     }
319 }
320 
321 
322 /* ZSTD_readMINMATCH() :
323  * function safe only for comparisons
324  * assumption : memPtr must be at least 4 bytes before end of buffer */
325 MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
326 {
327     switch (length)
328     {
329     default :
330     case 4 : return MEM_read32(memPtr);
331     case 3 : if (MEM_isLittleEndian())
332                 return MEM_read32(memPtr)<<8;
333              else
334                 return MEM_read32(memPtr)>>8;
335     }
336 }
337 
338 
339 /* Update hashTable3 up to ip (excluded)
340    Assumption : always within prefix (i.e. not within extDict) */
341 static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
342                                               U32* nextToUpdate3,
343                                               const BYTE* const ip)
344 {
345     U32* const hashTable3 = ms->hashTable3;
346     U32 const hashLog3 = ms->hashLog3;
347     const BYTE* const base = ms->window.base;
348     U32 idx = *nextToUpdate3;
349     U32 const target = (U32)(ip - base);
350     size_t const hash3 = ZSTD_hash3Ptr(ip, hashLog3);
351     assert(hashLog3 > 0);
352 
353     while(idx < target) {
354         hashTable3[ZSTD_hash3Ptr(base+idx, hashLog3)] = idx;
355         idx++;
356     }
357 
358     *nextToUpdate3 = target;
359     return hashTable3[hash3];
360 }
361 
362 
363 /*-*************************************
364 *  Binary Tree search
365 ***************************************/
366 /* ZSTD_insertBt1() : add one or multiple positions to tree.
367  *  ip : assumed <= iend-8 .
368  * @return : nb of positions added */
369 static U32 ZSTD_insertBt1(
370                 ZSTD_matchState_t* ms,
371                 const BYTE* const ip, const BYTE* const iend,
372                 U32 const mls, const int extDict)
373 {
374     const ZSTD_compressionParameters* const cParams = &ms->cParams;
375     U32*   const hashTable = ms->hashTable;
376     U32    const hashLog = cParams->hashLog;
377     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
378     U32*   const bt = ms->chainTable;
379     U32    const btLog  = cParams->chainLog - 1;
380     U32    const btMask = (1 << btLog) - 1;
381     U32 matchIndex = hashTable[h];
382     size_t commonLengthSmaller=0, commonLengthLarger=0;
383     const BYTE* const base = ms->window.base;
384     const BYTE* const dictBase = ms->window.dictBase;
385     const U32 dictLimit = ms->window.dictLimit;
386     const BYTE* const dictEnd = dictBase + dictLimit;
387     const BYTE* const prefixStart = base + dictLimit;
388     const BYTE* match;
389     const U32 curr = (U32)(ip-base);
390     const U32 btLow = btMask >= curr ? 0 : curr - btMask;
391     U32* smallerPtr = bt + 2*(curr&btMask);
392     U32* largerPtr  = smallerPtr + 1;
393     U32 dummy32;   /* to be nullified at the end */
394     U32 const windowLow = ms->window.lowLimit;
395     U32 matchEndIdx = curr+8+1;
396     size_t bestLength = 8;
397     U32 nbCompares = 1U << cParams->searchLog;
398 #ifdef ZSTD_C_PREDICT
399     U32 predictedSmall = *(bt + 2*((curr-1)&btMask) + 0);
400     U32 predictedLarge = *(bt + 2*((curr-1)&btMask) + 1);
401     predictedSmall += (predictedSmall>0);
402     predictedLarge += (predictedLarge>0);
403 #endif /* ZSTD_C_PREDICT */
404 
405     DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr);
406 
407     assert(ip <= iend-8);   /* required for h calculation */
408     hashTable[h] = curr;   /* Update Hash Table */
409 
410     assert(windowLow > 0);
411     for (; nbCompares && (matchIndex >= windowLow); --nbCompares) {
412         U32* const nextPtr = bt + 2*(matchIndex & btMask);
413         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
414         assert(matchIndex < curr);
415 
416 #ifdef ZSTD_C_PREDICT   /* note : can create issues when hlog small <= 11 */
417         const U32* predictPtr = bt + 2*((matchIndex-1) & btMask);   /* written this way, as bt is a roll buffer */
418         if (matchIndex == predictedSmall) {
419             /* no need to check length, result known */
420             *smallerPtr = matchIndex;
421             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
422             smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
423             matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
424             predictedSmall = predictPtr[1] + (predictPtr[1]>0);
425             continue;
426         }
427         if (matchIndex == predictedLarge) {
428             *largerPtr = matchIndex;
429             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
430             largerPtr = nextPtr;
431             matchIndex = nextPtr[0];
432             predictedLarge = predictPtr[0] + (predictPtr[0]>0);
433             continue;
434         }
435 #endif
436 
437         if (!extDict || (matchIndex+matchLength >= dictLimit)) {
438             assert(matchIndex+matchLength >= dictLimit);   /* might be wrong if actually extDict */
439             match = base + matchIndex;
440             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
441         } else {
442             match = dictBase + matchIndex;
443             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
444             if (matchIndex+matchLength >= dictLimit)
445                 match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
446         }
447 
448         if (matchLength > bestLength) {
449             bestLength = matchLength;
450             if (matchLength > matchEndIdx - matchIndex)
451                 matchEndIdx = matchIndex + (U32)matchLength;
452         }
453 
454         if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
455             break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
456         }
457 
458         if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
459             /* match is smaller than current */
460             *smallerPtr = matchIndex;             /* update smaller idx */
461             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
462             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
463             smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
464             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
465         } else {
466             /* match is larger than current */
467             *largerPtr = matchIndex;
468             commonLengthLarger = matchLength;
469             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
470             largerPtr = nextPtr;
471             matchIndex = nextPtr[0];
472     }   }
473 
474     *smallerPtr = *largerPtr = 0;
475     {   U32 positions = 0;
476         if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384));   /* speed optimization */
477         assert(matchEndIdx > curr + 8);
478         return MAX(positions, matchEndIdx - (curr + 8));
479     }
480 }
481 
482 FORCE_INLINE_TEMPLATE
483 void ZSTD_updateTree_internal(
484                 ZSTD_matchState_t* ms,
485                 const BYTE* const ip, const BYTE* const iend,
486                 const U32 mls, const ZSTD_dictMode_e dictMode)
487 {
488     const BYTE* const base = ms->window.base;
489     U32 const target = (U32)(ip - base);
490     U32 idx = ms->nextToUpdate;
491     DEBUGLOG(6, "ZSTD_updateTree_internal, from %u to %u  (dictMode:%u)",
492                 idx, target, dictMode);
493 
494     while(idx < target) {
495         U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict);
496         assert(idx < (U32)(idx + forward));
497         idx += forward;
498     }
499     assert((size_t)(ip - base) <= (size_t)(U32)(-1));
500     assert((size_t)(iend - base) <= (size_t)(U32)(-1));
501     ms->nextToUpdate = target;
502 }
503 
504 void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
505     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
506 }
507 
508 FORCE_INLINE_TEMPLATE
509 U32 ZSTD_insertBtAndGetAllMatches (
510                     ZSTD_match_t* matches,   /* store result (found matches) in this table (presumed large enough) */
511                     ZSTD_matchState_t* ms,
512                     U32* nextToUpdate3,
513                     const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode,
514                     const U32 rep[ZSTD_REP_NUM],
515                     U32 const ll0,   /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
516                     const U32 lengthToBeat,
517                     U32 const mls /* template */)
518 {
519     const ZSTD_compressionParameters* const cParams = &ms->cParams;
520     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
521     const BYTE* const base = ms->window.base;
522     U32 const curr = (U32)(ip-base);
523     U32 const hashLog = cParams->hashLog;
524     U32 const minMatch = (mls==3) ? 3 : 4;
525     U32* const hashTable = ms->hashTable;
526     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
527     U32 matchIndex  = hashTable[h];
528     U32* const bt   = ms->chainTable;
529     U32 const btLog = cParams->chainLog - 1;
530     U32 const btMask= (1U << btLog) - 1;
531     size_t commonLengthSmaller=0, commonLengthLarger=0;
532     const BYTE* const dictBase = ms->window.dictBase;
533     U32 const dictLimit = ms->window.dictLimit;
534     const BYTE* const dictEnd = dictBase + dictLimit;
535     const BYTE* const prefixStart = base + dictLimit;
536     U32 const btLow = (btMask >= curr) ? 0 : curr - btMask;
537     U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
538     U32 const matchLow = windowLow ? windowLow : 1;
539     U32* smallerPtr = bt + 2*(curr&btMask);
540     U32* largerPtr  = bt + 2*(curr&btMask) + 1;
541     U32 matchEndIdx = curr+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
542     U32 dummy32;   /* to be nullified at the end */
543     U32 mnum = 0;
544     U32 nbCompares = 1U << cParams->searchLog;
545 
546     const ZSTD_matchState_t* dms    = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL;
547     const ZSTD_compressionParameters* const dmsCParams =
548                                       dictMode == ZSTD_dictMatchState ? &dms->cParams : NULL;
549     const BYTE* const dmsBase       = dictMode == ZSTD_dictMatchState ? dms->window.base : NULL;
550     const BYTE* const dmsEnd        = dictMode == ZSTD_dictMatchState ? dms->window.nextSrc : NULL;
551     U32         const dmsHighLimit  = dictMode == ZSTD_dictMatchState ? (U32)(dmsEnd - dmsBase) : 0;
552     U32         const dmsLowLimit   = dictMode == ZSTD_dictMatchState ? dms->window.lowLimit : 0;
553     U32         const dmsIndexDelta = dictMode == ZSTD_dictMatchState ? windowLow - dmsHighLimit : 0;
554     U32         const dmsHashLog    = dictMode == ZSTD_dictMatchState ? dmsCParams->hashLog : hashLog;
555     U32         const dmsBtLog      = dictMode == ZSTD_dictMatchState ? dmsCParams->chainLog - 1 : btLog;
556     U32         const dmsBtMask     = dictMode == ZSTD_dictMatchState ? (1U << dmsBtLog) - 1 : 0;
557     U32         const dmsBtLow      = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit;
558 
559     size_t bestLength = lengthToBeat-1;
560     DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", curr);
561 
562     /* check repCode */
563     assert(ll0 <= 1);   /* necessarily 1 or 0 */
564     {   U32 const lastR = ZSTD_REP_NUM + ll0;
565         U32 repCode;
566         for (repCode = ll0; repCode < lastR; repCode++) {
567             U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
568             U32 const repIndex = curr - repOffset;
569             U32 repLen = 0;
570             assert(curr >= dictLimit);
571             if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < curr-dictLimit) {  /* equivalent to `curr > repIndex >= dictLimit` */
572                 /* We must validate the repcode offset because when we're using a dictionary the
573                  * valid offset range shrinks when the dictionary goes out of bounds.
574                  */
575                 if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) {
576                     repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch;
577                 }
578             } else {  /* repIndex < dictLimit || repIndex >= curr */
579                 const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ?
580                                              dmsBase + repIndex - dmsIndexDelta :
581                                              dictBase + repIndex;
582                 assert(curr >= windowLow);
583                 if ( dictMode == ZSTD_extDict
584                   && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow)  /* equivalent to `curr > repIndex >= windowLow` */
585                      & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */)
586                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
587                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
588                 }
589                 if (dictMode == ZSTD_dictMatchState
590                   && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `curr > repIndex >= dmsLowLimit` */
591                      & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */
592                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
593                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch;
594             }   }
595             /* save longer solution */
596             if (repLen > bestLength) {
597                 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
598                             repCode, ll0, repOffset, repLen);
599                 bestLength = repLen;
600                 matches[mnum].off = repCode - ll0;
601                 matches[mnum].len = (U32)repLen;
602                 mnum++;
603                 if ( (repLen > sufficient_len)
604                    | (ip+repLen == iLimit) ) {  /* best possible */
605                     return mnum;
606     }   }   }   }
607 
608     /* HC3 match finder */
609     if ((mls == 3) /*static*/ && (bestLength < mls)) {
610         U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip);
611         if ((matchIndex3 >= matchLow)
612           & (curr - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
613             size_t mlen;
614             if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) {
615                 const BYTE* const match = base + matchIndex3;
616                 mlen = ZSTD_count(ip, match, iLimit);
617             } else {
618                 const BYTE* const match = dictBase + matchIndex3;
619                 mlen = ZSTD_count_2segments(ip, match, iLimit, dictEnd, prefixStart);
620             }
621 
622             /* save best solution */
623             if (mlen >= mls /* == 3 > bestLength */) {
624                 DEBUGLOG(8, "found small match with hlog3, of length %u",
625                             (U32)mlen);
626                 bestLength = mlen;
627                 assert(curr > matchIndex3);
628                 assert(mnum==0);  /* no prior solution */
629                 matches[0].off = (curr - matchIndex3) + ZSTD_REP_MOVE;
630                 matches[0].len = (U32)mlen;
631                 mnum = 1;
632                 if ( (mlen > sufficient_len) |
633                      (ip+mlen == iLimit) ) {  /* best possible length */
634                     ms->nextToUpdate = curr+1;  /* skip insertion */
635                     return 1;
636         }   }   }
637         /* no dictMatchState lookup: dicts don't have a populated HC3 table */
638     }
639 
640     hashTable[h] = curr;   /* Update Hash Table */
641 
642     for (; nbCompares && (matchIndex >= matchLow); --nbCompares) {
643         U32* const nextPtr = bt + 2*(matchIndex & btMask);
644         const BYTE* match;
645         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
646         assert(curr > matchIndex);
647 
648         if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) {
649             assert(matchIndex+matchLength >= dictLimit);  /* ensure the condition is correct when !extDict */
650             match = base + matchIndex;
651             if (matchIndex >= dictLimit) assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
652             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iLimit);
653         } else {
654             match = dictBase + matchIndex;
655             assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
656             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dictEnd, prefixStart);
657             if (matchIndex+matchLength >= dictLimit)
658                 match = base + matchIndex;   /* prepare for match[matchLength] read */
659         }
660 
661         if (matchLength > bestLength) {
662             DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)",
663                     (U32)matchLength, curr - matchIndex, curr - matchIndex + ZSTD_REP_MOVE);
664             assert(matchEndIdx > matchIndex);
665             if (matchLength > matchEndIdx - matchIndex)
666                 matchEndIdx = matchIndex + (U32)matchLength;
667             bestLength = matchLength;
668             matches[mnum].off = (curr - matchIndex) + ZSTD_REP_MOVE;
669             matches[mnum].len = (U32)matchLength;
670             mnum++;
671             if ( (matchLength > ZSTD_OPT_NUM)
672                | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
673                 if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */
674                 break; /* drop, to preserve bt consistency (miss a little bit of compression) */
675             }
676         }
677 
678         if (match[matchLength] < ip[matchLength]) {
679             /* match smaller than current */
680             *smallerPtr = matchIndex;             /* update smaller idx */
681             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
682             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
683             smallerPtr = nextPtr+1;               /* new candidate => larger than match, which was smaller than current */
684             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous, closer to current */
685         } else {
686             *largerPtr = matchIndex;
687             commonLengthLarger = matchLength;
688             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
689             largerPtr = nextPtr;
690             matchIndex = nextPtr[0];
691     }   }
692 
693     *smallerPtr = *largerPtr = 0;
694 
695     assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
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         for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) {
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, curr - matchIndex, curr - matchIndex + ZSTD_REP_MOVE);
713                 if (matchLength > matchEndIdx - matchIndex)
714                     matchEndIdx = matchIndex + (U32)matchLength;
715                 bestLength = matchLength;
716                 matches[mnum].off = (curr - 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 > curr+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 *  LDM helper functions  *
770 *************************/
771 
772 /* Struct containing info needed to make decision about ldm inclusion */
773 typedef struct {
774     rawSeqStore_t seqStore;         /* External match candidates store for this block */
775     U32 startPosInBlock;            /* Start position of the current match candidate */
776     U32 endPosInBlock;              /* End position of the current match candidate */
777     U32 offset;                     /* Offset of the match candidate */
778 } ZSTD_optLdm_t;
779 
780 /* ZSTD_optLdm_skipRawSeqStoreBytes():
781  * Moves forward in rawSeqStore by nbBytes, which will update the fields 'pos' and 'posInSequence'.
782  */
783 static void ZSTD_optLdm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
784     U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
785     while (currPos && rawSeqStore->pos < rawSeqStore->size) {
786         rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
787         if (currPos >= currSeq.litLength + currSeq.matchLength) {
788             currPos -= currSeq.litLength + currSeq.matchLength;
789             rawSeqStore->pos++;
790         } else {
791             rawSeqStore->posInSequence = currPos;
792             break;
793         }
794     }
795     if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
796         rawSeqStore->posInSequence = 0;
797     }
798 }
799 
800 /* ZSTD_opt_getNextMatchAndUpdateSeqStore():
801  * Calculates the beginning and end of the next match in the current block.
802  * Updates 'pos' and 'posInSequence' of the ldmSeqStore.
803  */
804 static void ZSTD_opt_getNextMatchAndUpdateSeqStore(ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
805                                                    U32 blockBytesRemaining) {
806     rawSeq currSeq;
807     U32 currBlockEndPos;
808     U32 literalsBytesRemaining;
809     U32 matchBytesRemaining;
810 
811     /* Setting match end position to MAX to ensure we never use an LDM during this block */
812     if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
813         optLdm->startPosInBlock = UINT_MAX;
814         optLdm->endPosInBlock = UINT_MAX;
815         return;
816     }
817     /* Calculate appropriate bytes left in matchLength and litLength after adjusting
818        based on ldmSeqStore->posInSequence */
819     currSeq = optLdm->seqStore.seq[optLdm->seqStore.pos];
820     assert(optLdm->seqStore.posInSequence <= currSeq.litLength + currSeq.matchLength);
821     currBlockEndPos = currPosInBlock + blockBytesRemaining;
822     literalsBytesRemaining = (optLdm->seqStore.posInSequence < currSeq.litLength) ?
823             currSeq.litLength - (U32)optLdm->seqStore.posInSequence :
824             0;
825     matchBytesRemaining = (literalsBytesRemaining == 0) ?
826             currSeq.matchLength - ((U32)optLdm->seqStore.posInSequence - currSeq.litLength) :
827             currSeq.matchLength;
828 
829     /* If there are more literal bytes than bytes remaining in block, no ldm is possible */
830     if (literalsBytesRemaining >= blockBytesRemaining) {
831         optLdm->startPosInBlock = UINT_MAX;
832         optLdm->endPosInBlock = UINT_MAX;
833         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, blockBytesRemaining);
834         return;
835     }
836 
837     /* Matches may be < MINMATCH by this process. In that case, we will reject them
838        when we are deciding whether or not to add the ldm */
839     optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining;
840     optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining;
841     optLdm->offset = currSeq.offset;
842 
843     if (optLdm->endPosInBlock > currBlockEndPos) {
844         /* Match ends after the block ends, we can't use the whole match */
845         optLdm->endPosInBlock = currBlockEndPos;
846         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, currBlockEndPos - currPosInBlock);
847     } else {
848         /* Consume nb of bytes equal to size of sequence left */
849         ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, literalsBytesRemaining + matchBytesRemaining);
850     }
851 }
852 
853 /* ZSTD_optLdm_maybeAddMatch():
854  * Adds a match if it's long enough, based on it's 'matchStartPosInBlock'
855  * and 'matchEndPosInBlock', into 'matches'. Maintains the correct ordering of 'matches'
856  */
857 static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
858                                       ZSTD_optLdm_t* optLdm, U32 currPosInBlock) {
859     U32 posDiff = currPosInBlock - optLdm->startPosInBlock;
860     /* Note: ZSTD_match_t actually contains offCode and matchLength (before subtracting MINMATCH) */
861     U32 candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;
862     U32 candidateOffCode = optLdm->offset + ZSTD_REP_MOVE;
863 
864     /* Ensure that current block position is not outside of the match */
865     if (currPosInBlock < optLdm->startPosInBlock
866       || currPosInBlock >= optLdm->endPosInBlock
867       || candidateMatchLength < MINMATCH) {
868         return;
869     }
870 
871     if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
872         DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offCode: %u matchLength %u) at block position=%u",
873                  candidateOffCode, candidateMatchLength, currPosInBlock);
874         matches[*nbMatches].len = candidateMatchLength;
875         matches[*nbMatches].off = candidateOffCode;
876         (*nbMatches)++;
877     }
878 }
879 
880 /* ZSTD_optLdm_processMatchCandidate():
881  * Wrapper function to update ldm seq store and call ldm functions as necessary.
882  */
883 static void ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm, ZSTD_match_t* matches, U32* nbMatches,
884                                               U32 currPosInBlock, U32 remainingBytes) {
885     if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
886         return;
887     }
888 
889     if (currPosInBlock >= optLdm->endPosInBlock) {
890         if (currPosInBlock > optLdm->endPosInBlock) {
891             /* The position at which ZSTD_optLdm_processMatchCandidate() is called is not necessarily
892              * at the end of a match from the ldm seq store, and will often be some bytes
893              * over beyond matchEndPosInBlock. As such, we need to correct for these "overshoots"
894              */
895             U32 posOvershoot = currPosInBlock - optLdm->endPosInBlock;
896             ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot);
897         }
898         ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
899     }
900     ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
901 }
902 
903 /*-*******************************
904 *  Optimal parser
905 *********************************/
906 
907 
908 static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
909 {
910     return sol.litlen + sol.mlen;
911 }
912 
913 #if 0 /* debug */
914 
915 static void
916 listStats(const U32* table, int lastEltID)
917 {
918     int const nbElts = lastEltID + 1;
919     int enb;
920     for (enb=0; enb < nbElts; enb++) {
921         (void)table;
922         /* RAWLOG(2, "%3i:%3i,  ", enb, table[enb]); */
923         RAWLOG(2, "%4i,", table[enb]);
924     }
925     RAWLOG(2, " \n");
926 }
927 
928 #endif
929 
930 FORCE_INLINE_TEMPLATE size_t
931 ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
932                                seqStore_t* seqStore,
933                                U32 rep[ZSTD_REP_NUM],
934                          const void* src, size_t srcSize,
935                          const int optLevel,
936                          const ZSTD_dictMode_e dictMode)
937 {
938     optState_t* const optStatePtr = &ms->opt;
939     const BYTE* const istart = (const BYTE*)src;
940     const BYTE* ip = istart;
941     const BYTE* anchor = istart;
942     const BYTE* const iend = istart + srcSize;
943     const BYTE* const ilimit = iend - 8;
944     const BYTE* const base = ms->window.base;
945     const BYTE* const prefixStart = base + ms->window.dictLimit;
946     const ZSTD_compressionParameters* const cParams = &ms->cParams;
947 
948     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
949     U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4;
950     U32 nextToUpdate3 = ms->nextToUpdate;
951 
952     ZSTD_optimal_t* const opt = optStatePtr->priceTable;
953     ZSTD_match_t* const matches = optStatePtr->matchTable;
954     ZSTD_optimal_t lastSequence;
955     ZSTD_optLdm_t optLdm;
956 
957     optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
958     optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
959     ZSTD_opt_getNextMatchAndUpdateSeqStore(&optLdm, (U32)(ip-istart), (U32)(iend-ip));
960 
961     /* init */
962     DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u",
963                 (U32)(ip - base), ms->window.dictLimit, ms->nextToUpdate);
964     assert(optLevel <= 2);
965     ZSTD_rescaleFreqs(optStatePtr, (const BYTE*)src, srcSize, optLevel);
966     ip += (ip==prefixStart);
967 
968     /* Match Loop */
969     while (ip < ilimit) {
970         U32 cur, last_pos = 0;
971 
972         /* find first match */
973         {   U32 const litlen = (U32)(ip - anchor);
974             U32 const ll0 = !litlen;
975             U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch);
976             ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
977                                               (U32)(ip-istart), (U32)(iend - ip));
978             if (!nbMatches) { ip++; continue; }
979 
980             /* initialize opt[0] */
981             { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; }
982             opt[0].mlen = 0;  /* means is_a_literal */
983             opt[0].litlen = litlen;
984             /* We don't need to include the actual price of the literals because
985              * it is static for the duration of the forward pass, and is included
986              * in every price. We include the literal length to avoid negative
987              * prices when we subtract the previous literal length.
988              */
989             opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
990 
991             /* large match -> immediate encoding */
992             {   U32 const maxML = matches[nbMatches-1].len;
993                 U32 const maxOffset = matches[nbMatches-1].off;
994                 DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series",
995                             nbMatches, maxML, maxOffset, (U32)(ip-prefixStart));
996 
997                 if (maxML > sufficient_len) {
998                     lastSequence.litlen = litlen;
999                     lastSequence.mlen = maxML;
1000                     lastSequence.off = maxOffset;
1001                     DEBUGLOG(6, "large match (%u>%u), immediate encoding",
1002                                 maxML, sufficient_len);
1003                     cur = 0;
1004                     last_pos = ZSTD_totalLen(lastSequence);
1005                     goto _shortestPath;
1006             }   }
1007 
1008             /* set prices for first matches starting position == 0 */
1009             {   U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
1010                 U32 pos;
1011                 U32 matchNb;
1012                 for (pos = 1; pos < minMatch; pos++) {
1013                     opt[pos].price = ZSTD_MAX_PRICE;   /* mlen, litlen and price will be fixed during forward scanning */
1014                 }
1015                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
1016                     U32 const offset = matches[matchNb].off;
1017                     U32 const end = matches[matchNb].len;
1018                     for ( ; pos <= end ; pos++ ) {
1019                         U32 const matchPrice = ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel);
1020                         U32 const sequencePrice = literalsPrice + matchPrice;
1021                         DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
1022                                     pos, ZSTD_fCost(sequencePrice));
1023                         opt[pos].mlen = pos;
1024                         opt[pos].off = offset;
1025                         opt[pos].litlen = litlen;
1026                         opt[pos].price = sequencePrice;
1027                 }   }
1028                 last_pos = pos-1;
1029             }
1030         }
1031 
1032         /* check further positions */
1033         for (cur = 1; cur <= last_pos; cur++) {
1034             const BYTE* const inr = ip + cur;
1035             assert(cur < ZSTD_OPT_NUM);
1036             DEBUGLOG(7, "cPos:%zi==rPos:%u", inr-istart, cur)
1037 
1038             /* Fix current position with one literal if cheaper */
1039             {   U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
1040                 int const price = opt[cur-1].price
1041                                 + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
1042                                 + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
1043                                 - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
1044                 assert(price < 1000000000); /* overflow check */
1045                 if (price <= opt[cur].price) {
1046                     DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
1047                                 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
1048                                 opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]);
1049                     opt[cur].mlen = 0;
1050                     opt[cur].off = 0;
1051                     opt[cur].litlen = litlen;
1052                     opt[cur].price = price;
1053                 } else {
1054                     DEBUGLOG(7, "cPos:%zi==rPos:%u : literal would cost more (%.2f>%.2f) (hist:%u,%u,%u)",
1055                                 inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price),
1056                                 opt[cur].rep[0], opt[cur].rep[1], opt[cur].rep[2]);
1057                 }
1058             }
1059 
1060             /* Set the repcodes of the current position. We must do it here
1061              * because we rely on the repcodes of the 2nd to last sequence being
1062              * correct to set the next chunks repcodes during the backward
1063              * traversal.
1064              */
1065             ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(repcodes_t));
1066             assert(cur >= opt[cur].mlen);
1067             if (opt[cur].mlen != 0) {
1068                 U32 const prev = cur - opt[cur].mlen;
1069                 repcodes_t newReps = ZSTD_updateRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
1070                 ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
1071             } else {
1072                 ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
1073             }
1074 
1075             /* last match must start at a minimum distance of 8 from oend */
1076             if (inr > ilimit) continue;
1077 
1078             if (cur == last_pos) break;
1079 
1080             if ( (optLevel==0) /*static_test*/
1081               && (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) {
1082                 DEBUGLOG(7, "move to next rPos:%u : price is <=", cur+1);
1083                 continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
1084             }
1085 
1086             {   U32 const ll0 = (opt[cur].mlen != 0);
1087                 U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
1088                 U32 const previousPrice = opt[cur].price;
1089                 U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
1090                 U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch);
1091                 U32 matchNb;
1092 
1093                 ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
1094                                                   (U32)(inr-istart), (U32)(iend-inr));
1095 
1096                 if (!nbMatches) {
1097                     DEBUGLOG(7, "rPos:%u : no match found", cur);
1098                     continue;
1099                 }
1100 
1101                 {   U32 const maxML = matches[nbMatches-1].len;
1102                     DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of maxLength=%u",
1103                                 inr-istart, cur, nbMatches, maxML);
1104 
1105                     if ( (maxML > sufficient_len)
1106                       || (cur + maxML >= ZSTD_OPT_NUM) ) {
1107                         lastSequence.mlen = maxML;
1108                         lastSequence.off = matches[nbMatches-1].off;
1109                         lastSequence.litlen = litlen;
1110                         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 */
1111                         last_pos = cur + ZSTD_totalLen(lastSequence);
1112                         if (cur > ZSTD_OPT_NUM) cur = 0;   /* underflow => first match */
1113                         goto _shortestPath;
1114                 }   }
1115 
1116                 /* set prices using matches found at position == cur */
1117                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
1118                     U32 const offset = matches[matchNb].off;
1119                     U32 const lastML = matches[matchNb].len;
1120                     U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
1121                     U32 mlen;
1122 
1123                     DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u",
1124                                 matchNb, matches[matchNb].off, lastML, litlen);
1125 
1126                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
1127                         U32 const pos = cur + mlen;
1128                         int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
1129 
1130                         if ((pos > last_pos) || (price < opt[pos].price)) {
1131                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
1132                                         pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
1133                             while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; }   /* fill empty positions */
1134                             opt[pos].mlen = mlen;
1135                             opt[pos].off = offset;
1136                             opt[pos].litlen = litlen;
1137                             opt[pos].price = price;
1138                         } else {
1139                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)",
1140                                         pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
1141                             if (optLevel==0) break;  /* early update abort; gets ~+10% speed for about -0.01 ratio loss */
1142                         }
1143             }   }   }
1144         }  /* for (cur = 1; cur <= last_pos; cur++) */
1145 
1146         lastSequence = opt[last_pos];
1147         cur = last_pos > ZSTD_totalLen(lastSequence) ? last_pos - ZSTD_totalLen(lastSequence) : 0;  /* single sequence, and it starts before `ip` */
1148         assert(cur < ZSTD_OPT_NUM);  /* control overflow*/
1149 
1150 _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
1151         assert(opt[0].mlen == 0);
1152 
1153         /* Set the next chunk's repcodes based on the repcodes of the beginning
1154          * of the last match, and the last sequence. This avoids us having to
1155          * update them while traversing the sequences.
1156          */
1157         if (lastSequence.mlen != 0) {
1158             repcodes_t reps = ZSTD_updateRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
1159             ZSTD_memcpy(rep, &reps, sizeof(reps));
1160         } else {
1161             ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
1162         }
1163 
1164         {   U32 const storeEnd = cur + 1;
1165             U32 storeStart = storeEnd;
1166             U32 seqPos = cur;
1167 
1168             DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)",
1169                         last_pos, cur); (void)last_pos;
1170             assert(storeEnd < ZSTD_OPT_NUM);
1171             DEBUGLOG(6, "last sequence copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
1172                         storeEnd, lastSequence.litlen, lastSequence.mlen, lastSequence.off);
1173             opt[storeEnd] = lastSequence;
1174             while (seqPos > 0) {
1175                 U32 const backDist = ZSTD_totalLen(opt[seqPos]);
1176                 storeStart--;
1177                 DEBUGLOG(6, "sequence from rPos=%u copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
1178                             seqPos, storeStart, opt[seqPos].litlen, opt[seqPos].mlen, opt[seqPos].off);
1179                 opt[storeStart] = opt[seqPos];
1180                 seqPos = (seqPos > backDist) ? seqPos - backDist : 0;
1181             }
1182 
1183             /* save sequences */
1184             DEBUGLOG(6, "sending selected sequences into seqStore")
1185             {   U32 storePos;
1186                 for (storePos=storeStart; storePos <= storeEnd; storePos++) {
1187                     U32 const llen = opt[storePos].litlen;
1188                     U32 const mlen = opt[storePos].mlen;
1189                     U32 const offCode = opt[storePos].off;
1190                     U32 const advance = llen + mlen;
1191                     DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u",
1192                                 anchor - istart, (unsigned)llen, (unsigned)mlen);
1193 
1194                     if (mlen==0) {  /* only literals => must be last "sequence", actually starting a new stream of sequences */
1195                         assert(storePos == storeEnd);   /* must be last sequence */
1196                         ip = anchor + llen;     /* last "sequence" is a bunch of literals => don't progress anchor */
1197                         continue;   /* will finish */
1198                     }
1199 
1200                     assert(anchor + llen <= iend);
1201                     ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen);
1202                     ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen-MINMATCH);
1203                     anchor += advance;
1204                     ip = anchor;
1205             }   }
1206             ZSTD_setBasePrices(optStatePtr, optLevel);
1207         }
1208     }   /* while (ip < ilimit) */
1209 
1210     /* Return the last literals size */
1211     return (size_t)(iend - anchor);
1212 }
1213 
1214 
1215 size_t ZSTD_compressBlock_btopt(
1216         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1217         const void* src, size_t srcSize)
1218 {
1219     DEBUGLOG(5, "ZSTD_compressBlock_btopt");
1220     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict);
1221 }
1222 
1223 
1224 /* used in 2-pass strategy */
1225 static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus)
1226 {
1227     U32 s, sum=0;
1228     assert(ZSTD_FREQ_DIV+bonus >= 0);
1229     for (s=0; s<lastEltIndex+1; s++) {
1230         table[s] <<= ZSTD_FREQ_DIV+bonus;
1231         table[s]--;
1232         sum += table[s];
1233     }
1234     return sum;
1235 }
1236 
1237 /* used in 2-pass strategy */
1238 MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr)
1239 {
1240     if (ZSTD_compressedLiterals(optPtr))
1241         optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0);
1242     optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0);
1243     optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0);
1244     optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0);
1245 }
1246 
1247 /* ZSTD_initStats_ultra():
1248  * make a first compression pass, just to seed stats with more accurate starting values.
1249  * only works on first block, with no dictionary and no ldm.
1250  * this function cannot error, hence its contract must be respected.
1251  */
1252 static void
1253 ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
1254                      seqStore_t* seqStore,
1255                      U32 rep[ZSTD_REP_NUM],
1256                const void* src, size_t srcSize)
1257 {
1258     U32 tmpRep[ZSTD_REP_NUM];  /* updated rep codes will sink here */
1259     ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep));
1260 
1261     DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize);
1262     assert(ms->opt.litLengthSum == 0);    /* first block */
1263     assert(seqStore->sequences == seqStore->sequencesStart);   /* no ldm */
1264     assert(ms->window.dictLimit == ms->window.lowLimit);   /* no dictionary */
1265     assert(ms->window.dictLimit - ms->nextToUpdate <= 1);  /* no prefix (note: intentional overflow, defined as 2-complement) */
1266 
1267     ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);   /* generate stats into ms->opt*/
1268 
1269     /* invalidate first scan from history */
1270     ZSTD_resetSeqStore(seqStore);
1271     ms->window.base -= srcSize;
1272     ms->window.dictLimit += (U32)srcSize;
1273     ms->window.lowLimit = ms->window.dictLimit;
1274     ms->nextToUpdate = ms->window.dictLimit;
1275 
1276     /* re-inforce weight of collected statistics */
1277     ZSTD_upscaleStats(&ms->opt);
1278 }
1279 
1280 size_t ZSTD_compressBlock_btultra(
1281         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1282         const void* src, size_t srcSize)
1283 {
1284     DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
1285     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
1286 }
1287 
1288 size_t ZSTD_compressBlock_btultra2(
1289         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1290         const void* src, size_t srcSize)
1291 {
1292     U32 const curr = (U32)((const BYTE*)src - ms->window.base);
1293     DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);
1294 
1295     /* 2-pass strategy:
1296      * this strategy makes a first pass over first block to collect statistics
1297      * and seed next round's statistics with it.
1298      * After 1st pass, function forgets everything, and starts a new block.
1299      * Consequently, this can only work if no data has been previously loaded in tables,
1300      * aka, no dictionary, no prefix, no ldm preprocessing.
1301      * The compression ratio gain is generally small (~0.5% on first block),
1302      * the cost is 2x cpu time on first block. */
1303     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
1304     if ( (ms->opt.litLengthSum==0)   /* first block */
1305       && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
1306       && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
1307       && (curr == ms->window.dictLimit)   /* start of frame, nothing already loaded nor skipped */
1308       && (srcSize > ZSTD_PREDEF_THRESHOLD)
1309       ) {
1310         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
1311     }
1312 
1313     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
1314 }
1315 
1316 size_t ZSTD_compressBlock_btopt_dictMatchState(
1317         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1318         const void* src, size_t srcSize)
1319 {
1320     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState);
1321 }
1322 
1323 size_t ZSTD_compressBlock_btultra_dictMatchState(
1324         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1325         const void* src, size_t srcSize)
1326 {
1327     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState);
1328 }
1329 
1330 size_t ZSTD_compressBlock_btopt_extDict(
1331         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1332         const void* src, size_t srcSize)
1333 {
1334     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict);
1335 }
1336 
1337 size_t ZSTD_compressBlock_btultra_extDict(
1338         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1339         const void* src, size_t srcSize)
1340 {
1341     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict);
1342 }
1343 
1344 /* note : no btultra2 variant for extDict nor dictMatchState,
1345  * because btultra2 is not meant to work with dictionaries
1346  * and is only specific for the first block (no prefix) */
1347