xref: /linux/lib/zstd/compress/zstd_lazy.c (revision acbf6de674ef7b1b5870b25e7b3c695bf84273d0)
1 /*
2  * Copyright (c) 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 "zstd_lazy.h"
13 
14 
15 /*-*************************************
16 *  Binary Tree search
17 ***************************************/
18 
19 static void
20 ZSTD_updateDUBT(ZSTD_matchState_t* ms,
21                 const BYTE* ip, const BYTE* iend,
22                 U32 mls)
23 {
24     const ZSTD_compressionParameters* const cParams = &ms->cParams;
25     U32* const hashTable = ms->hashTable;
26     U32  const hashLog = cParams->hashLog;
27 
28     U32* const bt = ms->chainTable;
29     U32  const btLog  = cParams->chainLog - 1;
30     U32  const btMask = (1 << btLog) - 1;
31 
32     const BYTE* const base = ms->window.base;
33     U32 const target = (U32)(ip - base);
34     U32 idx = ms->nextToUpdate;
35 
36     if (idx != target)
37         DEBUGLOG(7, "ZSTD_updateDUBT, from %u to %u (dictLimit:%u)",
38                     idx, target, ms->window.dictLimit);
39     assert(ip + 8 <= iend);   /* condition for ZSTD_hashPtr */
40     (void)iend;
41 
42     assert(idx >= ms->window.dictLimit);   /* condition for valid base+idx */
43     for ( ; idx < target ; idx++) {
44         size_t const h  = ZSTD_hashPtr(base + idx, hashLog, mls);   /* assumption : ip + 8 <= iend */
45         U32    const matchIndex = hashTable[h];
46 
47         U32*   const nextCandidatePtr = bt + 2*(idx&btMask);
48         U32*   const sortMarkPtr  = nextCandidatePtr + 1;
49 
50         DEBUGLOG(8, "ZSTD_updateDUBT: insert %u", idx);
51         hashTable[h] = idx;   /* Update Hash Table */
52         *nextCandidatePtr = matchIndex;   /* update BT like a chain */
53         *sortMarkPtr = ZSTD_DUBT_UNSORTED_MARK;
54     }
55     ms->nextToUpdate = target;
56 }
57 
58 
59 /* ZSTD_insertDUBT1() :
60  *  sort one already inserted but unsorted position
61  *  assumption : curr >= btlow == (curr - btmask)
62  *  doesn't fail */
63 static void
64 ZSTD_insertDUBT1(const ZSTD_matchState_t* ms,
65                  U32 curr, const BYTE* inputEnd,
66                  U32 nbCompares, U32 btLow,
67                  const ZSTD_dictMode_e dictMode)
68 {
69     const ZSTD_compressionParameters* const cParams = &ms->cParams;
70     U32* const bt = ms->chainTable;
71     U32  const btLog  = cParams->chainLog - 1;
72     U32  const btMask = (1 << btLog) - 1;
73     size_t commonLengthSmaller=0, commonLengthLarger=0;
74     const BYTE* const base = ms->window.base;
75     const BYTE* const dictBase = ms->window.dictBase;
76     const U32 dictLimit = ms->window.dictLimit;
77     const BYTE* const ip = (curr>=dictLimit) ? base + curr : dictBase + curr;
78     const BYTE* const iend = (curr>=dictLimit) ? inputEnd : dictBase + dictLimit;
79     const BYTE* const dictEnd = dictBase + dictLimit;
80     const BYTE* const prefixStart = base + dictLimit;
81     const BYTE* match;
82     U32* smallerPtr = bt + 2*(curr&btMask);
83     U32* largerPtr  = smallerPtr + 1;
84     U32 matchIndex = *smallerPtr;   /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */
85     U32 dummy32;   /* to be nullified at the end */
86     U32 const windowValid = ms->window.lowLimit;
87     U32 const maxDistance = 1U << cParams->windowLog;
88     U32 const windowLow = (curr - windowValid > maxDistance) ? curr - maxDistance : windowValid;
89 
90 
91     DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
92                 curr, dictLimit, windowLow);
93     assert(curr >= btLow);
94     assert(ip < iend);   /* condition for ZSTD_count */
95 
96     for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
97         U32* const nextPtr = bt + 2*(matchIndex & btMask);
98         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
99         assert(matchIndex < curr);
100         /* note : all candidates are now supposed sorted,
101          * but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK
102          * when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */
103 
104         if ( (dictMode != ZSTD_extDict)
105           || (matchIndex+matchLength >= dictLimit)  /* both in current segment*/
106           || (curr < dictLimit) /* both in extDict */) {
107             const BYTE* const mBase = ( (dictMode != ZSTD_extDict)
108                                      || (matchIndex+matchLength >= dictLimit)) ?
109                                         base : dictBase;
110             assert( (matchIndex+matchLength >= dictLimit)   /* might be wrong if extDict is incorrectly set to 0 */
111                  || (curr < dictLimit) );
112             match = mBase + matchIndex;
113             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
114         } else {
115             match = dictBase + matchIndex;
116             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
117             if (matchIndex+matchLength >= dictLimit)
118                 match = base + matchIndex;   /* preparation for next read of match[matchLength] */
119         }
120 
121         DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
122                     curr, matchIndex, (U32)matchLength);
123 
124         if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
125             break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
126         }
127 
128         if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
129             /* match is smaller than current */
130             *smallerPtr = matchIndex;             /* update smaller idx */
131             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
132             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
133             DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is smaller : next => %u",
134                         matchIndex, btLow, nextPtr[1]);
135             smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
136             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
137         } else {
138             /* match is larger than current */
139             *largerPtr = matchIndex;
140             commonLengthLarger = matchLength;
141             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
142             DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is larger => %u",
143                         matchIndex, btLow, nextPtr[0]);
144             largerPtr = nextPtr;
145             matchIndex = nextPtr[0];
146     }   }
147 
148     *smallerPtr = *largerPtr = 0;
149 }
150 
151 
152 static size_t
153 ZSTD_DUBT_findBetterDictMatch (
154         const ZSTD_matchState_t* ms,
155         const BYTE* const ip, const BYTE* const iend,
156         size_t* offsetPtr,
157         size_t bestLength,
158         U32 nbCompares,
159         U32 const mls,
160         const ZSTD_dictMode_e dictMode)
161 {
162     const ZSTD_matchState_t * const dms = ms->dictMatchState;
163     const ZSTD_compressionParameters* const dmsCParams = &dms->cParams;
164     const U32 * const dictHashTable = dms->hashTable;
165     U32         const hashLog = dmsCParams->hashLog;
166     size_t      const h  = ZSTD_hashPtr(ip, hashLog, mls);
167     U32               dictMatchIndex = dictHashTable[h];
168 
169     const BYTE* const base = ms->window.base;
170     const BYTE* const prefixStart = base + ms->window.dictLimit;
171     U32         const curr = (U32)(ip-base);
172     const BYTE* const dictBase = dms->window.base;
173     const BYTE* const dictEnd = dms->window.nextSrc;
174     U32         const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base);
175     U32         const dictLowLimit = dms->window.lowLimit;
176     U32         const dictIndexDelta = ms->window.lowLimit - dictHighLimit;
177 
178     U32*        const dictBt = dms->chainTable;
179     U32         const btLog  = dmsCParams->chainLog - 1;
180     U32         const btMask = (1 << btLog) - 1;
181     U32         const btLow = (btMask >= dictHighLimit - dictLowLimit) ? dictLowLimit : dictHighLimit - btMask;
182 
183     size_t commonLengthSmaller=0, commonLengthLarger=0;
184 
185     (void)dictMode;
186     assert(dictMode == ZSTD_dictMatchState);
187 
188     for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) {
189         U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
190         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
191         const BYTE* match = dictBase + dictMatchIndex;
192         matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
193         if (dictMatchIndex+matchLength >= dictHighLimit)
194             match = base + dictMatchIndex + dictIndexDelta;   /* to prepare for next usage of match[matchLength] */
195 
196         if (matchLength > bestLength) {
197             U32 matchIndex = dictMatchIndex + dictIndexDelta;
198             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
199                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
200                     curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, STORE_OFFSET(curr - matchIndex), dictMatchIndex, matchIndex);
201                 bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
202             }
203             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
204                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
205             }
206         }
207 
208         if (match[matchLength] < ip[matchLength]) {
209             if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
210             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
211             dictMatchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
212         } else {
213             /* match is larger than current */
214             if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
215             commonLengthLarger = matchLength;
216             dictMatchIndex = nextPtr[0];
217         }
218     }
219 
220     if (bestLength >= MINMATCH) {
221         U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
222         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
223                     curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
224     }
225     return bestLength;
226 
227 }
228 
229 
230 static size_t
231 ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
232                         const BYTE* const ip, const BYTE* const iend,
233                         size_t* offsetPtr,
234                         U32 const mls,
235                         const ZSTD_dictMode_e dictMode)
236 {
237     const ZSTD_compressionParameters* const cParams = &ms->cParams;
238     U32*   const hashTable = ms->hashTable;
239     U32    const hashLog = cParams->hashLog;
240     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
241     U32          matchIndex  = hashTable[h];
242 
243     const BYTE* const base = ms->window.base;
244     U32    const curr = (U32)(ip-base);
245     U32    const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
246 
247     U32*   const bt = ms->chainTable;
248     U32    const btLog  = cParams->chainLog - 1;
249     U32    const btMask = (1 << btLog) - 1;
250     U32    const btLow = (btMask >= curr) ? 0 : curr - btMask;
251     U32    const unsortLimit = MAX(btLow, windowLow);
252 
253     U32*         nextCandidate = bt + 2*(matchIndex&btMask);
254     U32*         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
255     U32          nbCompares = 1U << cParams->searchLog;
256     U32          nbCandidates = nbCompares;
257     U32          previousCandidate = 0;
258 
259     DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", curr);
260     assert(ip <= iend-8);   /* required for h calculation */
261     assert(dictMode != ZSTD_dedicatedDictSearch);
262 
263     /* reach end of unsorted candidates list */
264     while ( (matchIndex > unsortLimit)
265          && (*unsortedMark == ZSTD_DUBT_UNSORTED_MARK)
266          && (nbCandidates > 1) ) {
267         DEBUGLOG(8, "ZSTD_DUBT_findBestMatch: candidate %u is unsorted",
268                     matchIndex);
269         *unsortedMark = previousCandidate;  /* the unsortedMark becomes a reversed chain, to move up back to original position */
270         previousCandidate = matchIndex;
271         matchIndex = *nextCandidate;
272         nextCandidate = bt + 2*(matchIndex&btMask);
273         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
274         nbCandidates --;
275     }
276 
277     /* nullify last candidate if it's still unsorted
278      * simplification, detrimental to compression ratio, beneficial for speed */
279     if ( (matchIndex > unsortLimit)
280       && (*unsortedMark==ZSTD_DUBT_UNSORTED_MARK) ) {
281         DEBUGLOG(7, "ZSTD_DUBT_findBestMatch: nullify last unsorted candidate %u",
282                     matchIndex);
283         *nextCandidate = *unsortedMark = 0;
284     }
285 
286     /* batch sort stacked candidates */
287     matchIndex = previousCandidate;
288     while (matchIndex) {  /* will end on matchIndex == 0 */
289         U32* const nextCandidateIdxPtr = bt + 2*(matchIndex&btMask) + 1;
290         U32 const nextCandidateIdx = *nextCandidateIdxPtr;
291         ZSTD_insertDUBT1(ms, matchIndex, iend,
292                          nbCandidates, unsortLimit, dictMode);
293         matchIndex = nextCandidateIdx;
294         nbCandidates++;
295     }
296 
297     /* find longest match */
298     {   size_t commonLengthSmaller = 0, commonLengthLarger = 0;
299         const BYTE* const dictBase = ms->window.dictBase;
300         const U32 dictLimit = ms->window.dictLimit;
301         const BYTE* const dictEnd = dictBase + dictLimit;
302         const BYTE* const prefixStart = base + dictLimit;
303         U32* smallerPtr = bt + 2*(curr&btMask);
304         U32* largerPtr  = bt + 2*(curr&btMask) + 1;
305         U32 matchEndIdx = curr + 8 + 1;
306         U32 dummy32;   /* to be nullified at the end */
307         size_t bestLength = 0;
308 
309         matchIndex  = hashTable[h];
310         hashTable[h] = curr;   /* Update Hash Table */
311 
312         for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
313             U32* const nextPtr = bt + 2*(matchIndex & btMask);
314             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
315             const BYTE* match;
316 
317             if ((dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit)) {
318                 match = base + matchIndex;
319                 matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
320             } else {
321                 match = dictBase + matchIndex;
322                 matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
323                 if (matchIndex+matchLength >= dictLimit)
324                     match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
325             }
326 
327             if (matchLength > bestLength) {
328                 if (matchLength > matchEndIdx - matchIndex)
329                     matchEndIdx = matchIndex + (U32)matchLength;
330                 if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
331                     bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
332                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
333                     if (dictMode == ZSTD_dictMatchState) {
334                         nbCompares = 0; /* in addition to avoiding checking any
335                                          * further in this loop, make sure we
336                                          * skip checking in the dictionary. */
337                     }
338                     break;   /* drop, to guarantee consistency (miss a little bit of compression) */
339                 }
340             }
341 
342             if (match[matchLength] < ip[matchLength]) {
343                 /* match is smaller than current */
344                 *smallerPtr = matchIndex;             /* update smaller idx */
345                 commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
346                 if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
347                 smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
348                 matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
349             } else {
350                 /* match is larger than current */
351                 *largerPtr = matchIndex;
352                 commonLengthLarger = matchLength;
353                 if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
354                 largerPtr = nextPtr;
355                 matchIndex = nextPtr[0];
356         }   }
357 
358         *smallerPtr = *largerPtr = 0;
359 
360         assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
361         if (dictMode == ZSTD_dictMatchState && nbCompares) {
362             bestLength = ZSTD_DUBT_findBetterDictMatch(
363                     ms, ip, iend,
364                     offsetPtr, bestLength, nbCompares,
365                     mls, dictMode);
366         }
367 
368         assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
369         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
370         if (bestLength >= MINMATCH) {
371             U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
372             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
373                         curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
374         }
375         return bestLength;
376     }
377 }
378 
379 
380 /* ZSTD_BtFindBestMatch() : Tree updater, providing best match */
381 FORCE_INLINE_TEMPLATE size_t
382 ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
383                 const BYTE* const ip, const BYTE* const iLimit,
384                       size_t* offsetPtr,
385                 const U32 mls /* template */,
386                 const ZSTD_dictMode_e dictMode)
387 {
388     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
389     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
390     ZSTD_updateDUBT(ms, ip, iLimit, mls);
391     return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
392 }
393 
394 /* *********************************
395 * Dedicated dict search
396 ***********************************/
397 
398 void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
399 {
400     const BYTE* const base = ms->window.base;
401     U32 const target = (U32)(ip - base);
402     U32* const hashTable = ms->hashTable;
403     U32* const chainTable = ms->chainTable;
404     U32 const chainSize = 1 << ms->cParams.chainLog;
405     U32 idx = ms->nextToUpdate;
406     U32 const minChain = chainSize < target - idx ? target - chainSize : idx;
407     U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
408     U32 const cacheSize = bucketSize - 1;
409     U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
410     U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts;
411 
412     /* We know the hashtable is oversized by a factor of `bucketSize`.
413      * We are going to temporarily pretend `bucketSize == 1`, keeping only a
414      * single entry. We will use the rest of the space to construct a temporary
415      * chaintable.
416      */
417     U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
418     U32* const tmpHashTable = hashTable;
419     U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
420     U32 const tmpChainSize = (U32)((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
421     U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
422     U32 hashIdx;
423 
424     assert(ms->cParams.chainLog <= 24);
425     assert(ms->cParams.hashLog > ms->cParams.chainLog);
426     assert(idx != 0);
427     assert(tmpMinChain <= minChain);
428 
429     /* fill conventional hash table and conventional chain table */
430     for ( ; idx < target; idx++) {
431         U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch);
432         if (idx >= tmpMinChain) {
433             tmpChainTable[idx - tmpMinChain] = hashTable[h];
434         }
435         tmpHashTable[h] = idx;
436     }
437 
438     /* sort chains into ddss chain table */
439     {
440         U32 chainPos = 0;
441         for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) {
442             U32 count;
443             U32 countBeyondMinChain = 0;
444             U32 i = tmpHashTable[hashIdx];
445             for (count = 0; i >= tmpMinChain && count < cacheSize; count++) {
446                 /* skip through the chain to the first position that won't be
447                  * in the hash cache bucket */
448                 if (i < minChain) {
449                     countBeyondMinChain++;
450                 }
451                 i = tmpChainTable[i - tmpMinChain];
452             }
453             if (count == cacheSize) {
454                 for (count = 0; count < chainLimit;) {
455                     if (i < minChain) {
456                         if (!i || ++countBeyondMinChain > cacheSize) {
457                             /* only allow pulling `cacheSize` number of entries
458                              * into the cache or chainTable beyond `minChain`,
459                              * to replace the entries pulled out of the
460                              * chainTable into the cache. This lets us reach
461                              * back further without increasing the total number
462                              * of entries in the chainTable, guaranteeing the
463                              * DDSS chain table will fit into the space
464                              * allocated for the regular one. */
465                             break;
466                         }
467                     }
468                     chainTable[chainPos++] = i;
469                     count++;
470                     if (i < tmpMinChain) {
471                         break;
472                     }
473                     i = tmpChainTable[i - tmpMinChain];
474                 }
475             } else {
476                 count = 0;
477             }
478             if (count) {
479                 tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count;
480             } else {
481                 tmpHashTable[hashIdx] = 0;
482             }
483         }
484         assert(chainPos <= chainSize); /* I believe this is guaranteed... */
485     }
486 
487     /* move chain pointers into the last entry of each hash bucket */
488     for (hashIdx = (1 << hashLog); hashIdx; ) {
489         U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG;
490         U32 const chainPackedPointer = tmpHashTable[hashIdx];
491         U32 i;
492         for (i = 0; i < cacheSize; i++) {
493             hashTable[bucketIdx + i] = 0;
494         }
495         hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer;
496     }
497 
498     /* fill the buckets of the hash table */
499     for (idx = ms->nextToUpdate; idx < target; idx++) {
500         U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch)
501                    << ZSTD_LAZY_DDSS_BUCKET_LOG;
502         U32 i;
503         /* Shift hash cache down 1. */
504         for (i = cacheSize - 1; i; i--)
505             hashTable[h + i] = hashTable[h + i - 1];
506         hashTable[h] = idx;
507     }
508 
509     ms->nextToUpdate = target;
510 }
511 
512 /* Returns the longest match length found in the dedicated dict search structure.
513  * If none are longer than the argument ml, then ml will be returned.
514  */
515 FORCE_INLINE_TEMPLATE
516 size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
517                                             const ZSTD_matchState_t* const dms,
518                                             const BYTE* const ip, const BYTE* const iLimit,
519                                             const BYTE* const prefixStart, const U32 curr,
520                                             const U32 dictLimit, const size_t ddsIdx) {
521     const U32 ddsLowestIndex  = dms->window.dictLimit;
522     const BYTE* const ddsBase = dms->window.base;
523     const BYTE* const ddsEnd  = dms->window.nextSrc;
524     const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
525     const U32 ddsIndexDelta   = dictLimit - ddsSize;
526     const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
527     const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
528     U32 ddsAttempt;
529     U32 matchIndex;
530 
531     for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
532         PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
533     }
534 
535     {
536         U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
537         U32 const chainIndex = chainPackedPointer >> 8;
538 
539         PREFETCH_L1(&dms->chainTable[chainIndex]);
540     }
541 
542     for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
543         size_t currentMl=0;
544         const BYTE* match;
545         matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
546         match = ddsBase + matchIndex;
547 
548         if (!matchIndex) {
549             return ml;
550         }
551 
552         /* guaranteed by table construction */
553         (void)ddsLowestIndex;
554         assert(matchIndex >= ddsLowestIndex);
555         assert(match+4 <= ddsEnd);
556         if (MEM_read32(match) == MEM_read32(ip)) {
557             /* assumption : matchIndex <= dictLimit-4 (by table construction) */
558             currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
559         }
560 
561         /* save best solution */
562         if (currentMl > ml) {
563             ml = currentMl;
564             *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
565             if (ip+currentMl == iLimit) {
566                 /* best possible, avoids read overflow on next attempt */
567                 return ml;
568             }
569         }
570     }
571 
572     {
573         U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
574         U32 chainIndex = chainPackedPointer >> 8;
575         U32 const chainLength = chainPackedPointer & 0xFF;
576         U32 const chainAttempts = nbAttempts - ddsAttempt;
577         U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
578         U32 chainAttempt;
579 
580         for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
581             PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
582         }
583 
584         for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
585             size_t currentMl=0;
586             const BYTE* match;
587             matchIndex = dms->chainTable[chainIndex];
588             match = ddsBase + matchIndex;
589 
590             /* guaranteed by table construction */
591             assert(matchIndex >= ddsLowestIndex);
592             assert(match+4 <= ddsEnd);
593             if (MEM_read32(match) == MEM_read32(ip)) {
594                 /* assumption : matchIndex <= dictLimit-4 (by table construction) */
595                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
596             }
597 
598             /* save best solution */
599             if (currentMl > ml) {
600                 ml = currentMl;
601                 *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
602                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
603             }
604         }
605     }
606     return ml;
607 }
608 
609 
610 /* *********************************
611 *  Hash Chain
612 ***********************************/
613 #define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & (mask)]
614 
615 /* Update chains up to ip (excluded)
616    Assumption : always within prefix (i.e. not within extDict) */
617 FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
618                         ZSTD_matchState_t* ms,
619                         const ZSTD_compressionParameters* const cParams,
620                         const BYTE* ip, U32 const mls)
621 {
622     U32* const hashTable  = ms->hashTable;
623     const U32 hashLog = cParams->hashLog;
624     U32* const chainTable = ms->chainTable;
625     const U32 chainMask = (1 << cParams->chainLog) - 1;
626     const BYTE* const base = ms->window.base;
627     const U32 target = (U32)(ip - base);
628     U32 idx = ms->nextToUpdate;
629 
630     while(idx < target) { /* catch up */
631         size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
632         NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
633         hashTable[h] = idx;
634         idx++;
635     }
636 
637     ms->nextToUpdate = target;
638     return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
639 }
640 
641 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
642     const ZSTD_compressionParameters* const cParams = &ms->cParams;
643     return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
644 }
645 
646 /* inlining is important to hardwire a hot branch (template emulation) */
647 FORCE_INLINE_TEMPLATE
648 size_t ZSTD_HcFindBestMatch(
649                         ZSTD_matchState_t* ms,
650                         const BYTE* const ip, const BYTE* const iLimit,
651                         size_t* offsetPtr,
652                         const U32 mls, const ZSTD_dictMode_e dictMode)
653 {
654     const ZSTD_compressionParameters* const cParams = &ms->cParams;
655     U32* const chainTable = ms->chainTable;
656     const U32 chainSize = (1 << cParams->chainLog);
657     const U32 chainMask = chainSize-1;
658     const BYTE* const base = ms->window.base;
659     const BYTE* const dictBase = ms->window.dictBase;
660     const U32 dictLimit = ms->window.dictLimit;
661     const BYTE* const prefixStart = base + dictLimit;
662     const BYTE* const dictEnd = dictBase + dictLimit;
663     const U32 curr = (U32)(ip-base);
664     const U32 maxDistance = 1U << cParams->windowLog;
665     const U32 lowestValid = ms->window.lowLimit;
666     const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
667     const U32 isDictionary = (ms->loadedDictEnd != 0);
668     const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
669     const U32 minChain = curr > chainSize ? curr - chainSize : 0;
670     U32 nbAttempts = 1U << cParams->searchLog;
671     size_t ml=4-1;
672 
673     const ZSTD_matchState_t* const dms = ms->dictMatchState;
674     const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
675                          ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
676     const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
677                         ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
678 
679     U32 matchIndex;
680 
681     if (dictMode == ZSTD_dedicatedDictSearch) {
682         const U32* entry = &dms->hashTable[ddsIdx];
683         PREFETCH_L1(entry);
684     }
685 
686     /* HC4 match finder */
687     matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
688 
689     for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
690         size_t currentMl=0;
691         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
692             const BYTE* const match = base + matchIndex;
693             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
694             if (match[ml] == ip[ml])   /* potentially better */
695                 currentMl = ZSTD_count(ip, match, iLimit);
696         } else {
697             const BYTE* const match = dictBase + matchIndex;
698             assert(match+4 <= dictEnd);
699             if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
700                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
701         }
702 
703         /* save best solution */
704         if (currentMl > ml) {
705             ml = currentMl;
706             *offsetPtr = STORE_OFFSET(curr - matchIndex);
707             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
708         }
709 
710         if (matchIndex <= minChain) break;
711         matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
712     }
713 
714     assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
715     if (dictMode == ZSTD_dedicatedDictSearch) {
716         ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms,
717                                                   ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
718     } else if (dictMode == ZSTD_dictMatchState) {
719         const U32* const dmsChainTable = dms->chainTable;
720         const U32 dmsChainSize         = (1 << dms->cParams.chainLog);
721         const U32 dmsChainMask         = dmsChainSize - 1;
722         const U32 dmsLowestIndex       = dms->window.dictLimit;
723         const BYTE* const dmsBase      = dms->window.base;
724         const BYTE* const dmsEnd       = dms->window.nextSrc;
725         const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
726         const U32 dmsIndexDelta        = dictLimit - dmsSize;
727         const U32 dmsMinChain = dmsSize > dmsChainSize ? dmsSize - dmsChainSize : 0;
728 
729         matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];
730 
731         for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
732             size_t currentMl=0;
733             const BYTE* const match = dmsBase + matchIndex;
734             assert(match+4 <= dmsEnd);
735             if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
736                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
737 
738             /* save best solution */
739             if (currentMl > ml) {
740                 ml = currentMl;
741                 assert(curr > matchIndex + dmsIndexDelta);
742                 *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
743                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
744             }
745 
746             if (matchIndex <= dmsMinChain) break;
747 
748             matchIndex = dmsChainTable[matchIndex & dmsChainMask];
749         }
750     }
751 
752     return ml;
753 }
754 
755 /* *********************************
756 * (SIMD) Row-based matchfinder
757 ***********************************/
758 /* Constants for row-based hash */
759 #define ZSTD_ROW_HASH_TAG_OFFSET 16     /* byte offset of hashes in the match state's tagTable from the beginning of a row */
760 #define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
761 #define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
762 #define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */
763 
764 #define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1)
765 
766 typedef U64 ZSTD_VecMask;   /* Clarifies when we are interacting with a U64 representing a mask of matches */
767 
768 /* ZSTD_VecMask_next():
769  * Starting from the LSB, returns the idx of the next non-zero bit.
770  * Basically counting the nb of trailing zeroes.
771  */
772 static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
773     assert(val != 0);
774 #   if (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))
775     if (sizeof(size_t) == 4) {
776         U32 mostSignificantWord = (U32)(val >> 32);
777         U32 leastSignificantWord = (U32)val;
778         if (leastSignificantWord == 0) {
779             return 32 + (U32)__builtin_ctz(mostSignificantWord);
780         } else {
781             return (U32)__builtin_ctz(leastSignificantWord);
782         }
783     } else {
784         return (U32)__builtin_ctzll(val);
785     }
786 #   else
787     /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count
788      * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer
789      */
790     val = ~val & (val - 1ULL); /* Lowest set bit mask */
791     val = val - ((val >> 1) & 0x5555555555555555);
792     val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL);
793     return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
794 #   endif
795 }
796 
797 /* ZSTD_rotateRight_*():
798  * Rotates a bitfield to the right by "count" bits.
799  * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
800  */
801 FORCE_INLINE_TEMPLATE
802 U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
803     assert(count < 64);
804     count &= 0x3F; /* for fickle pattern recognition */
805     return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
806 }
807 
808 FORCE_INLINE_TEMPLATE
809 U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
810     assert(count < 32);
811     count &= 0x1F; /* for fickle pattern recognition */
812     return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
813 }
814 
815 FORCE_INLINE_TEMPLATE
816 U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
817     assert(count < 16);
818     count &= 0x0F; /* for fickle pattern recognition */
819     return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
820 }
821 
822 /* ZSTD_row_nextIndex():
823  * Returns the next index to insert at within a tagTable row, and updates the "head"
824  * value to reflect the update. Essentially cycles backwards from [0, {entries per row})
825  */
826 FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
827   U32 const next = (*tagRow - 1) & rowMask;
828   *tagRow = (BYTE)next;
829   return next;
830 }
831 
832 /* ZSTD_isAligned():
833  * Checks that a pointer is aligned to "align" bytes which must be a power of 2.
834  */
835 MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) {
836     assert((align & (align - 1)) == 0);
837     return (((size_t)ptr) & (align - 1)) == 0;
838 }
839 
840 /* ZSTD_row_prefetch():
841  * Performs prefetching for the hashTable and tagTable at a given row.
842  */
843 FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) {
844     PREFETCH_L1(hashTable + relRow);
845     if (rowLog >= 5) {
846         PREFETCH_L1(hashTable + relRow + 16);
847         /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */
848     }
849     PREFETCH_L1(tagTable + relRow);
850     if (rowLog == 6) {
851         PREFETCH_L1(tagTable + relRow + 32);
852     }
853     assert(rowLog == 4 || rowLog == 5 || rowLog == 6);
854     assert(ZSTD_isAligned(hashTable + relRow, 64));                 /* prefetched hash row always 64-byte aligned */
855     assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */
856 }
857 
858 /* ZSTD_row_fillHashCache():
859  * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
860  * but not beyond iLimit.
861  */
862 FORCE_INLINE_TEMPLATE void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
863                                    U32 const rowLog, U32 const mls,
864                                    U32 idx, const BYTE* const iLimit)
865 {
866     U32 const* const hashTable = ms->hashTable;
867     U16 const* const tagTable = ms->tagTable;
868     U32 const hashLog = ms->rowHashLog;
869     U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
870     U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);
871 
872     for (; idx < lim; ++idx) {
873         U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
874         U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
875         ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
876         ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
877     }
878 
879     DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1],
880                                                      ms->hashCache[2], ms->hashCache[3], ms->hashCache[4],
881                                                      ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]);
882 }
883 
884 /* ZSTD_row_nextCachedHash():
885  * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
886  * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
887  */
888 FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
889                                                   U16 const* tagTable, BYTE const* base,
890                                                   U32 idx, U32 const hashLog,
891                                                   U32 const rowLog, U32 const mls)
892 {
893     U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
894     U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
895     ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
896     {   U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
897         cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash;
898         return hash;
899     }
900 }
901 
902 /* ZSTD_row_update_internalImpl():
903  * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
904  */
905 FORCE_INLINE_TEMPLATE void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms,
906                                                         U32 updateStartIdx, U32 const updateEndIdx,
907                                                         U32 const mls, U32 const rowLog,
908                                                         U32 const rowMask, U32 const useCache)
909 {
910     U32* const hashTable = ms->hashTable;
911     U16* const tagTable = ms->tagTable;
912     U32 const hashLog = ms->rowHashLog;
913     const BYTE* const base = ms->window.base;
914 
915     DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
916     for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
917         U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls)
918                                   : (U32)ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
919         U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
920         U32* const row = hashTable + relRow;
921         BYTE* tagRow = (BYTE*)(tagTable + relRow);  /* Though tagTable is laid out as a table of U16, each tag is only 1 byte.
922                                                        Explicit cast allows us to get exact desired position within each row */
923         U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
924 
925         assert(hash == ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls));
926         ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK;
927         row[pos] = updateStartIdx;
928     }
929 }
930 
931 /* ZSTD_row_update_internal():
932  * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
933  * Skips sections of long matches as is necessary.
934  */
935 FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip,
936                                                     U32 const mls, U32 const rowLog,
937                                                     U32 const rowMask, U32 const useCache)
938 {
939     U32 idx = ms->nextToUpdate;
940     const BYTE* const base = ms->window.base;
941     const U32 target = (U32)(ip - base);
942     const U32 kSkipThreshold = 384;
943     const U32 kMaxMatchStartPositionsToUpdate = 96;
944     const U32 kMaxMatchEndPositionsToUpdate = 32;
945 
946     if (useCache) {
947         /* Only skip positions when using hash cache, i.e.
948          * if we are loading a dict, don't skip anything.
949          * If we decide to skip, then we only update a set number
950          * of positions at the beginning and end of the match.
951          */
952         if (UNLIKELY(target - idx > kSkipThreshold)) {
953             U32 const bound = idx + kMaxMatchStartPositionsToUpdate;
954             ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache);
955             idx = target - kMaxMatchEndPositionsToUpdate;
956             ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1);
957         }
958     }
959     assert(target >= idx);
960     ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache);
961     ms->nextToUpdate = target;
962 }
963 
964 /* ZSTD_row_update():
965  * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
966  * processing.
967  */
968 void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
969     const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
970     const U32 rowMask = (1u << rowLog) - 1;
971     const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
972 
973     DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
974     ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */);
975 }
976 
977 #if defined(ZSTD_ARCH_X86_SSE2)
978 FORCE_INLINE_TEMPLATE ZSTD_VecMask
979 ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
980 {
981     const __m128i comparisonMask = _mm_set1_epi8((char)tag);
982     int matches[4] = {0};
983     int i;
984     assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
985     for (i=0; i<nbChunks; i++) {
986         const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i));
987         const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask);
988         matches[i] = _mm_movemask_epi8(equalMask);
989     }
990     if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head);
991     if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
992     assert(nbChunks == 4);
993     return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
994 }
995 #endif
996 
997 /* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
998  * the hash at the nth position in a row of the tagTable.
999  * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
1000  * to match up with the actual layout of the entries within the hashTable */
1001 FORCE_INLINE_TEMPLATE ZSTD_VecMask
1002 ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries)
1003 {
1004     const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
1005     assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
1006     assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
1007 
1008 #if defined(ZSTD_ARCH_X86_SSE2)
1009 
1010     return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, head);
1011 
1012 #else /* SW or NEON-LE */
1013 
1014 # if defined(ZSTD_ARCH_ARM_NEON)
1015   /* This NEON path only works for little endian - otherwise use SWAR below */
1016     if (MEM_isLittleEndian()) {
1017         if (rowEntries == 16) {
1018             const uint8x16_t chunk = vld1q_u8(src);
1019             const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
1020             const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
1021             const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
1022             const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
1023             const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
1024             const U16 hi = (U16)vgetq_lane_u8(t3, 8);
1025             const U16 lo = (U16)vgetq_lane_u8(t3, 0);
1026             return ZSTD_rotateRight_U16((hi << 8) | lo, head);
1027         } else if (rowEntries == 32) {
1028             const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
1029             const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
1030             const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
1031             const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
1032             const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
1033             const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
1034             const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
1035             const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
1036             const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
1037             const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
1038             const uint8x8x2_t t3 = vuzp_u8(t2, t0);
1039             const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
1040             const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
1041             return ZSTD_rotateRight_U32(matches, head);
1042         } else { /* rowEntries == 64 */
1043             const uint8x16x4_t chunk = vld4q_u8(src);
1044             const uint8x16_t dup = vdupq_n_u8(tag);
1045             const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
1046             const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
1047             const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
1048             const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
1049 
1050             const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
1051             const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
1052             const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
1053             const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
1054             const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
1055             const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
1056             return ZSTD_rotateRight_U64(matches, head);
1057         }
1058     }
1059 # endif /* ZSTD_ARCH_ARM_NEON */
1060     /* SWAR */
1061     {   const size_t chunkSize = sizeof(size_t);
1062         const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
1063         const size_t xFF = ~((size_t)0);
1064         const size_t x01 = xFF / 0xFF;
1065         const size_t x80 = x01 << 7;
1066         const size_t splatChar = tag * x01;
1067         ZSTD_VecMask matches = 0;
1068         int i = rowEntries - chunkSize;
1069         assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
1070         if (MEM_isLittleEndian()) { /* runtime check so have two loops */
1071             const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
1072             do {
1073                 size_t chunk = MEM_readST(&src[i]);
1074                 chunk ^= splatChar;
1075                 chunk = (((chunk | x80) - x01) | chunk) & x80;
1076                 matches <<= chunkSize;
1077                 matches |= (chunk * extractMagic) >> shiftAmount;
1078                 i -= chunkSize;
1079             } while (i >= 0);
1080         } else { /* big endian: reverse bits during extraction */
1081             const size_t msb = xFF ^ (xFF >> 1);
1082             const size_t extractMagic = (msb / 0x1FF) | msb;
1083             do {
1084                 size_t chunk = MEM_readST(&src[i]);
1085                 chunk ^= splatChar;
1086                 chunk = (((chunk | x80) - x01) | chunk) & x80;
1087                 matches <<= chunkSize;
1088                 matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
1089                 i -= chunkSize;
1090             } while (i >= 0);
1091         }
1092         matches = ~matches;
1093         if (rowEntries == 16) {
1094             return ZSTD_rotateRight_U16((U16)matches, head);
1095         } else if (rowEntries == 32) {
1096             return ZSTD_rotateRight_U32((U32)matches, head);
1097         } else {
1098             return ZSTD_rotateRight_U64((U64)matches, head);
1099         }
1100     }
1101 #endif
1102 }
1103 
1104 /* The high-level approach of the SIMD row based match finder is as follows:
1105  * - Figure out where to insert the new entry:
1106  *      - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
1107  *      - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
1108  *        which row to insert into.
1109  *      - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
1110  *        be considered as a circular buffer with a "head" index that resides in the tagTable.
1111  *      - Also insert the "tag" into the equivalent row and position in the tagTable.
1112  *          - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
1113  *                  The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
1114  *                  for alignment/performance reasons, leaving some bytes unused.
1115  * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
1116  *   generate a bitfield that we can cycle through to check the collisions in the hash table.
1117  * - Pick the longest match.
1118  */
1119 FORCE_INLINE_TEMPLATE
1120 size_t ZSTD_RowFindBestMatch(
1121                         ZSTD_matchState_t* ms,
1122                         const BYTE* const ip, const BYTE* const iLimit,
1123                         size_t* offsetPtr,
1124                         const U32 mls, const ZSTD_dictMode_e dictMode,
1125                         const U32 rowLog)
1126 {
1127     U32* const hashTable = ms->hashTable;
1128     U16* const tagTable = ms->tagTable;
1129     U32* const hashCache = ms->hashCache;
1130     const U32 hashLog = ms->rowHashLog;
1131     const ZSTD_compressionParameters* const cParams = &ms->cParams;
1132     const BYTE* const base = ms->window.base;
1133     const BYTE* const dictBase = ms->window.dictBase;
1134     const U32 dictLimit = ms->window.dictLimit;
1135     const BYTE* const prefixStart = base + dictLimit;
1136     const BYTE* const dictEnd = dictBase + dictLimit;
1137     const U32 curr = (U32)(ip-base);
1138     const U32 maxDistance = 1U << cParams->windowLog;
1139     const U32 lowestValid = ms->window.lowLimit;
1140     const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
1141     const U32 isDictionary = (ms->loadedDictEnd != 0);
1142     const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
1143     const U32 rowEntries = (1U << rowLog);
1144     const U32 rowMask = rowEntries - 1;
1145     const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
1146     U32 nbAttempts = 1U << cappedSearchLog;
1147     size_t ml=4-1;
1148 
1149     /* DMS/DDS variables that may be referenced laster */
1150     const ZSTD_matchState_t* const dms = ms->dictMatchState;
1151 
1152     /* Initialize the following variables to satisfy static analyzer */
1153     size_t ddsIdx = 0;
1154     U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */
1155     U32 dmsTag = 0;
1156     U32* dmsRow = NULL;
1157     BYTE* dmsTagRow = NULL;
1158 
1159     if (dictMode == ZSTD_dedicatedDictSearch) {
1160         const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
1161         {   /* Prefetch DDS hashtable entry */
1162             ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG;
1163             PREFETCH_L1(&dms->hashTable[ddsIdx]);
1164         }
1165         ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0;
1166     }
1167 
1168     if (dictMode == ZSTD_dictMatchState) {
1169         /* Prefetch DMS rows */
1170         U32* const dmsHashTable = dms->hashTable;
1171         U16* const dmsTagTable = dms->tagTable;
1172         U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
1173         U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
1174         dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
1175         dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow);
1176         dmsRow = dmsHashTable + dmsRelRow;
1177         ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog);
1178     }
1179 
1180     /* Update the hashTable and tagTable up to (but not including) ip */
1181     ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
1182     {   /* Get the hash for ip, compute the appropriate row */
1183         U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls);
1184         U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
1185         U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
1186         U32* const row = hashTable + relRow;
1187         BYTE* tagRow = (BYTE*)(tagTable + relRow);
1188         U32 const head = *tagRow & rowMask;
1189         U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
1190         size_t numMatches = 0;
1191         size_t currMatch = 0;
1192         ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries);
1193 
1194         /* Cycle through the matches and prefetch */
1195         for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
1196             U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
1197             U32 const matchIndex = row[matchPos];
1198             assert(numMatches < rowEntries);
1199             if (matchIndex < lowLimit)
1200                 break;
1201             if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
1202                 PREFETCH_L1(base + matchIndex);
1203             } else {
1204                 PREFETCH_L1(dictBase + matchIndex);
1205             }
1206             matchBuffer[numMatches++] = matchIndex;
1207         }
1208 
1209         /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
1210            in ZSTD_row_update_internal() at the next search. */
1211         {
1212             U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
1213             tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag;
1214             row[pos] = ms->nextToUpdate++;
1215         }
1216 
1217         /* Return the longest match */
1218         for (; currMatch < numMatches; ++currMatch) {
1219             U32 const matchIndex = matchBuffer[currMatch];
1220             size_t currentMl=0;
1221             assert(matchIndex < curr);
1222             assert(matchIndex >= lowLimit);
1223 
1224             if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
1225                 const BYTE* const match = base + matchIndex;
1226                 assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
1227                 if (match[ml] == ip[ml])   /* potentially better */
1228                     currentMl = ZSTD_count(ip, match, iLimit);
1229             } else {
1230                 const BYTE* const match = dictBase + matchIndex;
1231                 assert(match+4 <= dictEnd);
1232                 if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
1233                     currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
1234             }
1235 
1236             /* Save best solution */
1237             if (currentMl > ml) {
1238                 ml = currentMl;
1239                 *offsetPtr = STORE_OFFSET(curr - matchIndex);
1240                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
1241             }
1242         }
1243     }
1244 
1245     assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
1246     if (dictMode == ZSTD_dedicatedDictSearch) {
1247         ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms,
1248                                                   ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
1249     } else if (dictMode == ZSTD_dictMatchState) {
1250         /* TODO: Measure and potentially add prefetching to DMS */
1251         const U32 dmsLowestIndex       = dms->window.dictLimit;
1252         const BYTE* const dmsBase      = dms->window.base;
1253         const BYTE* const dmsEnd       = dms->window.nextSrc;
1254         const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
1255         const U32 dmsIndexDelta        = dictLimit - dmsSize;
1256 
1257         {   U32 const head = *dmsTagRow & rowMask;
1258             U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
1259             size_t numMatches = 0;
1260             size_t currMatch = 0;
1261             ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries);
1262 
1263             for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
1264                 U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
1265                 U32 const matchIndex = dmsRow[matchPos];
1266                 if (matchIndex < dmsLowestIndex)
1267                     break;
1268                 PREFETCH_L1(dmsBase + matchIndex);
1269                 matchBuffer[numMatches++] = matchIndex;
1270             }
1271 
1272             /* Return the longest match */
1273             for (; currMatch < numMatches; ++currMatch) {
1274                 U32 const matchIndex = matchBuffer[currMatch];
1275                 size_t currentMl=0;
1276                 assert(matchIndex >= dmsLowestIndex);
1277                 assert(matchIndex < curr);
1278 
1279                 {   const BYTE* const match = dmsBase + matchIndex;
1280                     assert(match+4 <= dmsEnd);
1281                     if (MEM_read32(match) == MEM_read32(ip))
1282                         currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
1283                 }
1284 
1285                 if (currentMl > ml) {
1286                     ml = currentMl;
1287                     assert(curr > matchIndex + dmsIndexDelta);
1288                     *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
1289                     if (ip+currentMl == iLimit) break;
1290                 }
1291             }
1292         }
1293     }
1294     return ml;
1295 }
1296 
1297 
1298 /*
1299  * Generate search functions templated on (dictMode, mls, rowLog).
1300  * These functions are outlined for code size & compilation time.
1301  * ZSTD_searchMax() dispatches to the correct implementation function.
1302  *
1303  * TODO: The start of the search function involves loading and calculating a
1304  * bunch of constants from the ZSTD_matchState_t. These computations could be
1305  * done in an initialization function, and saved somewhere in the match state.
1306  * Then we could pass a pointer to the saved state instead of the match state,
1307  * and avoid duplicate computations.
1308  *
1309  * TODO: Move the match re-winding into searchMax. This improves compression
1310  * ratio, and unlocks further simplifications with the next TODO.
1311  *
1312  * TODO: Try moving the repcode search into searchMax. After the re-winding
1313  * and repcode search are in searchMax, there is no more logic in the match
1314  * finder loop that requires knowledge about the dictMode. So we should be
1315  * able to avoid force inlining it, and we can join the extDict loop with
1316  * the single segment loop. It should go in searchMax instead of its own
1317  * function to avoid having multiple virtual function calls per search.
1318  */
1319 
1320 #define ZSTD_BT_SEARCH_FN(dictMode, mls) ZSTD_BtFindBestMatch_##dictMode##_##mls
1321 #define ZSTD_HC_SEARCH_FN(dictMode, mls) ZSTD_HcFindBestMatch_##dictMode##_##mls
1322 #define ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog) ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog
1323 
1324 #define ZSTD_SEARCH_FN_ATTRS FORCE_NOINLINE
1325 
1326 #define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
1327     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
1328             ZSTD_matchState_t* ms,                                                     \
1329             const BYTE* ip, const BYTE* const iLimit,                                  \
1330             size_t* offBasePtr)                                                        \
1331     {                                                                                  \
1332         assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                           \
1333         return ZSTD_BtFindBestMatch(ms, ip, iLimit, offBasePtr, mls, ZSTD_##dictMode); \
1334     }                                                                                  \
1335 
1336 #define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
1337     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
1338             ZSTD_matchState_t* ms,                                                    \
1339             const BYTE* ip, const BYTE* const iLimit,                                 \
1340             size_t* offsetPtr)                                                        \
1341     {                                                                                 \
1342         assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                          \
1343         return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
1344     }                                                                                 \
1345 
1346 #define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
1347     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
1348             ZSTD_matchState_t* ms,                                                             \
1349             const BYTE* ip, const BYTE* const iLimit,                                          \
1350             size_t* offsetPtr)                                                                 \
1351     {                                                                                          \
1352         assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                                   \
1353         assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog);                               \
1354         return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
1355     }                                                                                          \
1356 
1357 #define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
1358     X(dictMode, mls, 4)                        \
1359     X(dictMode, mls, 5)                        \
1360     X(dictMode, mls, 6)
1361 
1362 #define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \
1363     ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4)      \
1364     ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5)      \
1365     ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6)
1366 
1367 #define ZSTD_FOR_EACH_MLS(X, dictMode) \
1368     X(dictMode, 4)                     \
1369     X(dictMode, 5)                     \
1370     X(dictMode, 6)
1371 
1372 #define ZSTD_FOR_EACH_DICT_MODE(X, ...) \
1373     X(__VA_ARGS__, noDict)              \
1374     X(__VA_ARGS__, extDict)             \
1375     X(__VA_ARGS__, dictMatchState)      \
1376     X(__VA_ARGS__, dedicatedDictSearch)
1377 
1378 /* Generate row search fns for each combination of (dictMode, mls, rowLog) */
1379 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_SEARCH_FN)
1380 /* Generate binary Tree search fns for each combination of (dictMode, mls) */
1381 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_SEARCH_FN)
1382 /* Generate hash chain search fns for each combination of (dictMode, mls) */
1383 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_SEARCH_FN)
1384 
1385 typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
1386 
1387 #define GEN_ZSTD_CALL_BT_SEARCH_FN(dictMode, mls)                         \
1388     case mls:                                                             \
1389         return ZSTD_BT_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
1390 #define GEN_ZSTD_CALL_HC_SEARCH_FN(dictMode, mls)                         \
1391     case mls:                                                             \
1392         return ZSTD_HC_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
1393 #define GEN_ZSTD_CALL_ROW_SEARCH_FN(dictMode, mls, rowLog)                         \
1394     case rowLog:                                                                   \
1395         return ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(ms, ip, iend, offsetPtr);
1396 
1397 #define ZSTD_SWITCH_MLS(X, dictMode)   \
1398     switch (mls) {                     \
1399         ZSTD_FOR_EACH_MLS(X, dictMode) \
1400     }
1401 
1402 #define ZSTD_SWITCH_ROWLOG(dictMode, mls)                                    \
1403     case mls:                                                                \
1404         switch (rowLog) {                                                    \
1405             ZSTD_FOR_EACH_ROWLOG(GEN_ZSTD_CALL_ROW_SEARCH_FN, dictMode, mls) \
1406         }                                                                    \
1407         ZSTD_UNREACHABLE;                                                    \
1408         break;
1409 
1410 #define ZSTD_SWITCH_SEARCH_METHOD(dictMode)                       \
1411     switch (searchMethod) {                                       \
1412         case search_hashChain:                                    \
1413             ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_HC_SEARCH_FN, dictMode) \
1414             break;                                                \
1415         case search_binaryTree:                                   \
1416             ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_BT_SEARCH_FN, dictMode) \
1417             break;                                                \
1418         case search_rowHash:                                      \
1419             ZSTD_SWITCH_MLS(ZSTD_SWITCH_ROWLOG, dictMode)         \
1420             break;                                                \
1421     }                                                             \
1422     ZSTD_UNREACHABLE;
1423 
1424 /*
1425  * Searches for the longest match at @p ip.
1426  * Dispatches to the correct implementation function based on the
1427  * (searchMethod, dictMode, mls, rowLog). We use switch statements
1428  * here instead of using an indirect function call through a function
1429  * pointer because after Spectre and Meltdown mitigations, indirect
1430  * function calls can be very costly, especially in the kernel.
1431  *
1432  * NOTE: dictMode and searchMethod should be templated, so those switch
1433  * statements should be optimized out. Only the mls & rowLog switches
1434  * should be left.
1435  *
1436  * @param ms The match state.
1437  * @param ip The position to search at.
1438  * @param iend The end of the input data.
1439  * @param[out] offsetPtr Stores the match offset into this pointer.
1440  * @param mls The minimum search length, in the range [4, 6].
1441  * @param rowLog The row log (if applicable), in the range [4, 6].
1442  * @param searchMethod The search method to use (templated).
1443  * @param dictMode The dictMode (templated).
1444  *
1445  * @returns The length of the longest match found, or < mls if no match is found.
1446  * If a match is found its offset is stored in @p offsetPtr.
1447  */
1448 FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
1449     ZSTD_matchState_t* ms,
1450     const BYTE* ip,
1451     const BYTE* iend,
1452     size_t* offsetPtr,
1453     U32 const mls,
1454     U32 const rowLog,
1455     searchMethod_e const searchMethod,
1456     ZSTD_dictMode_e const dictMode)
1457 {
1458     if (dictMode == ZSTD_noDict) {
1459         ZSTD_SWITCH_SEARCH_METHOD(noDict)
1460     } else if (dictMode == ZSTD_extDict) {
1461         ZSTD_SWITCH_SEARCH_METHOD(extDict)
1462     } else if (dictMode == ZSTD_dictMatchState) {
1463         ZSTD_SWITCH_SEARCH_METHOD(dictMatchState)
1464     } else if (dictMode == ZSTD_dedicatedDictSearch) {
1465         ZSTD_SWITCH_SEARCH_METHOD(dedicatedDictSearch)
1466     }
1467     ZSTD_UNREACHABLE;
1468     return 0;
1469 }
1470 
1471 /* *******************************
1472 *  Common parser - lazy strategy
1473 *********************************/
1474 
1475 FORCE_INLINE_TEMPLATE size_t
1476 ZSTD_compressBlock_lazy_generic(
1477                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
1478                         U32 rep[ZSTD_REP_NUM],
1479                         const void* src, size_t srcSize,
1480                         const searchMethod_e searchMethod, const U32 depth,
1481                         ZSTD_dictMode_e const dictMode)
1482 {
1483     const BYTE* const istart = (const BYTE*)src;
1484     const BYTE* ip = istart;
1485     const BYTE* anchor = istart;
1486     const BYTE* const iend = istart + srcSize;
1487     const BYTE* const ilimit = (searchMethod == search_rowHash) ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
1488     const BYTE* const base = ms->window.base;
1489     const U32 prefixLowestIndex = ms->window.dictLimit;
1490     const BYTE* const prefixLowest = base + prefixLowestIndex;
1491     const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
1492     const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
1493 
1494     U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
1495 
1496     const int isDMS = dictMode == ZSTD_dictMatchState;
1497     const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
1498     const int isDxS = isDMS || isDDS;
1499     const ZSTD_matchState_t* const dms = ms->dictMatchState;
1500     const U32 dictLowestIndex      = isDxS ? dms->window.dictLimit : 0;
1501     const BYTE* const dictBase     = isDxS ? dms->window.base : NULL;
1502     const BYTE* const dictLowest   = isDxS ? dictBase + dictLowestIndex : NULL;
1503     const BYTE* const dictEnd      = isDxS ? dms->window.nextSrc : NULL;
1504     const U32 dictIndexDelta       = isDxS ?
1505                                      prefixLowestIndex - (U32)(dictEnd - dictBase) :
1506                                      0;
1507     const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
1508 
1509     DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
1510     ip += (dictAndPrefixLength == 0);
1511     if (dictMode == ZSTD_noDict) {
1512         U32 const curr = (U32)(ip - base);
1513         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
1514         U32 const maxRep = curr - windowLow;
1515         if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
1516         if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
1517     }
1518     if (isDxS) {
1519         /* dictMatchState repCode checks don't currently handle repCode == 0
1520          * disabling. */
1521         assert(offset_1 <= dictAndPrefixLength);
1522         assert(offset_2 <= dictAndPrefixLength);
1523     }
1524 
1525     if (searchMethod == search_rowHash) {
1526         ZSTD_row_fillHashCache(ms, base, rowLog,
1527                             MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
1528                             ms->nextToUpdate, ilimit);
1529     }
1530 
1531     /* Match Loop */
1532 #if defined(__x86_64__)
1533     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
1534      * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
1535      */
1536     __asm__(".p2align 5");
1537 #endif
1538     while (ip < ilimit) {
1539         size_t matchLength=0;
1540         size_t offcode=STORE_REPCODE_1;
1541         const BYTE* start=ip+1;
1542         DEBUGLOG(7, "search baseline (depth 0)");
1543 
1544         /* check repCode */
1545         if (isDxS) {
1546             const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
1547             const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch)
1548                                 && repIndex < prefixLowestIndex) ?
1549                                    dictBase + (repIndex - dictIndexDelta) :
1550                                    base + repIndex;
1551             if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1552                 && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
1553                 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1554                 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1555                 if (depth==0) goto _storeSequence;
1556             }
1557         }
1558         if ( dictMode == ZSTD_noDict
1559           && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
1560             matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
1561             if (depth==0) goto _storeSequence;
1562         }
1563 
1564         /* first search (depth 0) */
1565         {   size_t offsetFound = 999999999;
1566             size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offsetFound, mls, rowLog, searchMethod, dictMode);
1567             if (ml2 > matchLength)
1568                 matchLength = ml2, start = ip, offcode=offsetFound;
1569         }
1570 
1571         if (matchLength < 4) {
1572             ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
1573             continue;
1574         }
1575 
1576         /* let's try to find a better solution */
1577         if (depth>=1)
1578         while (ip<ilimit) {
1579             DEBUGLOG(7, "search depth 1");
1580             ip ++;
1581             if ( (dictMode == ZSTD_noDict)
1582               && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
1583                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
1584                 int const gain2 = (int)(mlRep * 3);
1585                 int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1586                 if ((mlRep >= 4) && (gain2 > gain1))
1587                     matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
1588             }
1589             if (isDxS) {
1590                 const U32 repIndex = (U32)(ip - base) - offset_1;
1591                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1592                                dictBase + (repIndex - dictIndexDelta) :
1593                                base + repIndex;
1594                 if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1595                     && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1596                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1597                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1598                     int const gain2 = (int)(mlRep * 3);
1599                     int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1600                     if ((mlRep >= 4) && (gain2 > gain1))
1601                         matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
1602                 }
1603             }
1604             {   size_t offset2=999999999;
1605                 size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, dictMode);
1606                 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
1607                 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
1608                 if ((ml2 >= 4) && (gain2 > gain1)) {
1609                     matchLength = ml2, offcode = offset2, start = ip;
1610                     continue;   /* search a better one */
1611             }   }
1612 
1613             /* let's find an even better one */
1614             if ((depth==2) && (ip<ilimit)) {
1615                 DEBUGLOG(7, "search depth 2");
1616                 ip ++;
1617                 if ( (dictMode == ZSTD_noDict)
1618                   && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
1619                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
1620                     int const gain2 = (int)(mlRep * 4);
1621                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1622                     if ((mlRep >= 4) && (gain2 > gain1))
1623                         matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
1624                 }
1625                 if (isDxS) {
1626                     const U32 repIndex = (U32)(ip - base) - offset_1;
1627                     const BYTE* repMatch = repIndex < prefixLowestIndex ?
1628                                    dictBase + (repIndex - dictIndexDelta) :
1629                                    base + repIndex;
1630                     if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1631                         && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1632                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1633                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1634                         int const gain2 = (int)(mlRep * 4);
1635                         int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1636                         if ((mlRep >= 4) && (gain2 > gain1))
1637                             matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
1638                     }
1639                 }
1640                 {   size_t offset2=999999999;
1641                     size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, dictMode);
1642                     int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
1643                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
1644                     if ((ml2 >= 4) && (gain2 > gain1)) {
1645                         matchLength = ml2, offcode = offset2, start = ip;
1646                         continue;
1647             }   }   }
1648             break;  /* nothing found : store previous solution */
1649         }
1650 
1651         /* NOTE:
1652          * Pay attention that `start[-value]` can lead to strange undefined behavior
1653          * notably if `value` is unsigned, resulting in a large positive `-value`.
1654          */
1655         /* catch up */
1656         if (STORED_IS_OFFSET(offcode)) {
1657             if (dictMode == ZSTD_noDict) {
1658                 while ( ((start > anchor) & (start - STORED_OFFSET(offcode) > prefixLowest))
1659                      && (start[-1] == (start-STORED_OFFSET(offcode))[-1]) )  /* only search for offset within prefix */
1660                     { start--; matchLength++; }
1661             }
1662             if (isDxS) {
1663                 U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
1664                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
1665                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
1666                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
1667             }
1668             offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
1669         }
1670         /* store sequence */
1671 _storeSequence:
1672         {   size_t const litLength = (size_t)(start - anchor);
1673             ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
1674             anchor = ip = start + matchLength;
1675         }
1676 
1677         /* check immediate repcode */
1678         if (isDxS) {
1679             while (ip <= ilimit) {
1680                 U32 const current2 = (U32)(ip-base);
1681                 U32 const repIndex = current2 - offset_2;
1682                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1683                         dictBase - dictIndexDelta + repIndex :
1684                         base + repIndex;
1685                 if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
1686                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1687                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
1688                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
1689                     offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset_2 <=> offset_1 */
1690                     ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
1691                     ip += matchLength;
1692                     anchor = ip;
1693                     continue;
1694                 }
1695                 break;
1696             }
1697         }
1698 
1699         if (dictMode == ZSTD_noDict) {
1700             while ( ((ip <= ilimit) & (offset_2>0))
1701                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
1702                 /* store sequence */
1703                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
1704                 offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap repcodes */
1705                 ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
1706                 ip += matchLength;
1707                 anchor = ip;
1708                 continue;   /* faster when present ... (?) */
1709     }   }   }
1710 
1711     /* Save reps for next block */
1712     rep[0] = offset_1 ? offset_1 : savedOffset;
1713     rep[1] = offset_2 ? offset_2 : savedOffset;
1714 
1715     /* Return the last literals size */
1716     return (size_t)(iend - anchor);
1717 }
1718 
1719 
1720 size_t ZSTD_compressBlock_btlazy2(
1721         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1722         void const* src, size_t srcSize)
1723 {
1724     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
1725 }
1726 
1727 size_t ZSTD_compressBlock_lazy2(
1728         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1729         void const* src, size_t srcSize)
1730 {
1731     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
1732 }
1733 
1734 size_t ZSTD_compressBlock_lazy(
1735         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1736         void const* src, size_t srcSize)
1737 {
1738     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
1739 }
1740 
1741 size_t ZSTD_compressBlock_greedy(
1742         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1743         void const* src, size_t srcSize)
1744 {
1745     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
1746 }
1747 
1748 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
1749         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1750         void const* src, size_t srcSize)
1751 {
1752     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
1753 }
1754 
1755 size_t ZSTD_compressBlock_lazy2_dictMatchState(
1756         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1757         void const* src, size_t srcSize)
1758 {
1759     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
1760 }
1761 
1762 size_t ZSTD_compressBlock_lazy_dictMatchState(
1763         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1764         void const* src, size_t srcSize)
1765 {
1766     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
1767 }
1768 
1769 size_t ZSTD_compressBlock_greedy_dictMatchState(
1770         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1771         void const* src, size_t srcSize)
1772 {
1773     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
1774 }
1775 
1776 
1777 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
1778         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1779         void const* src, size_t srcSize)
1780 {
1781     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
1782 }
1783 
1784 size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
1785         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1786         void const* src, size_t srcSize)
1787 {
1788     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
1789 }
1790 
1791 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
1792         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1793         void const* src, size_t srcSize)
1794 {
1795     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
1796 }
1797 
1798 /* Row-based matchfinder */
1799 size_t ZSTD_compressBlock_lazy2_row(
1800         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1801         void const* src, size_t srcSize)
1802 {
1803     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
1804 }
1805 
1806 size_t ZSTD_compressBlock_lazy_row(
1807         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1808         void const* src, size_t srcSize)
1809 {
1810     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
1811 }
1812 
1813 size_t ZSTD_compressBlock_greedy_row(
1814         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1815         void const* src, size_t srcSize)
1816 {
1817     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
1818 }
1819 
1820 size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
1821         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1822         void const* src, size_t srcSize)
1823 {
1824     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
1825 }
1826 
1827 size_t ZSTD_compressBlock_lazy_dictMatchState_row(
1828         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1829         void const* src, size_t srcSize)
1830 {
1831     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
1832 }
1833 
1834 size_t ZSTD_compressBlock_greedy_dictMatchState_row(
1835         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1836         void const* src, size_t srcSize)
1837 {
1838     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
1839 }
1840 
1841 
1842 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
1843         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1844         void const* src, size_t srcSize)
1845 {
1846     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
1847 }
1848 
1849 size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
1850         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1851         void const* src, size_t srcSize)
1852 {
1853     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
1854 }
1855 
1856 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
1857         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1858         void const* src, size_t srcSize)
1859 {
1860     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
1861 }
1862 
1863 FORCE_INLINE_TEMPLATE
1864 size_t ZSTD_compressBlock_lazy_extDict_generic(
1865                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
1866                         U32 rep[ZSTD_REP_NUM],
1867                         const void* src, size_t srcSize,
1868                         const searchMethod_e searchMethod, const U32 depth)
1869 {
1870     const BYTE* const istart = (const BYTE*)src;
1871     const BYTE* ip = istart;
1872     const BYTE* anchor = istart;
1873     const BYTE* const iend = istart + srcSize;
1874     const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
1875     const BYTE* const base = ms->window.base;
1876     const U32 dictLimit = ms->window.dictLimit;
1877     const BYTE* const prefixStart = base + dictLimit;
1878     const BYTE* const dictBase = ms->window.dictBase;
1879     const BYTE* const dictEnd  = dictBase + dictLimit;
1880     const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
1881     const U32 windowLog = ms->cParams.windowLog;
1882     const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
1883     const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
1884 
1885     U32 offset_1 = rep[0], offset_2 = rep[1];
1886 
1887     DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
1888 
1889     /* init */
1890     ip += (ip == prefixStart);
1891     if (searchMethod == search_rowHash) {
1892         ZSTD_row_fillHashCache(ms, base, rowLog,
1893                                MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
1894                                ms->nextToUpdate, ilimit);
1895     }
1896 
1897     /* Match Loop */
1898 #if defined(__x86_64__)
1899     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
1900      * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
1901      */
1902     __asm__(".p2align 5");
1903 #endif
1904     while (ip < ilimit) {
1905         size_t matchLength=0;
1906         size_t offcode=STORE_REPCODE_1;
1907         const BYTE* start=ip+1;
1908         U32 curr = (U32)(ip-base);
1909 
1910         /* check repCode */
1911         {   const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog);
1912             const U32 repIndex = (U32)(curr+1 - offset_1);
1913             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1914             const BYTE* const repMatch = repBase + repIndex;
1915             if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */
1916                & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
1917             if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
1918                 /* repcode detected we should take it */
1919                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1920                 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1921                 if (depth==0) goto _storeSequence;
1922         }   }
1923 
1924         /* first search (depth 0) */
1925         {   size_t offsetFound = 999999999;
1926             size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offsetFound, mls, rowLog, searchMethod, ZSTD_extDict);
1927             if (ml2 > matchLength)
1928                 matchLength = ml2, start = ip, offcode=offsetFound;
1929         }
1930 
1931         if (matchLength < 4) {
1932             ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
1933             continue;
1934         }
1935 
1936         /* let's try to find a better solution */
1937         if (depth>=1)
1938         while (ip<ilimit) {
1939             ip ++;
1940             curr++;
1941             /* check repCode */
1942             if (offcode) {
1943                 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1944                 const U32 repIndex = (U32)(curr - offset_1);
1945                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1946                 const BYTE* const repMatch = repBase + repIndex;
1947                 if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
1948                    & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
1949                 if (MEM_read32(ip) == MEM_read32(repMatch)) {
1950                     /* repcode detected */
1951                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1952                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1953                     int const gain2 = (int)(repLength * 3);
1954                     int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1955                     if ((repLength >= 4) && (gain2 > gain1))
1956                         matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
1957             }   }
1958 
1959             /* search match, depth 1 */
1960             {   size_t offset2=999999999;
1961                 size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, ZSTD_extDict);
1962                 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
1963                 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
1964                 if ((ml2 >= 4) && (gain2 > gain1)) {
1965                     matchLength = ml2, offcode = offset2, start = ip;
1966                     continue;   /* search a better one */
1967             }   }
1968 
1969             /* let's find an even better one */
1970             if ((depth==2) && (ip<ilimit)) {
1971                 ip ++;
1972                 curr++;
1973                 /* check repCode */
1974                 if (offcode) {
1975                     const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1976                     const U32 repIndex = (U32)(curr - offset_1);
1977                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1978                     const BYTE* const repMatch = repBase + repIndex;
1979                     if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
1980                        & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
1981                     if (MEM_read32(ip) == MEM_read32(repMatch)) {
1982                         /* repcode detected */
1983                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1984                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1985                         int const gain2 = (int)(repLength * 4);
1986                         int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
1987                         if ((repLength >= 4) && (gain2 > gain1))
1988                             matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
1989                 }   }
1990 
1991                 /* search match, depth 2 */
1992                 {   size_t offset2=999999999;
1993                     size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, ZSTD_extDict);
1994                     int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
1995                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
1996                     if ((ml2 >= 4) && (gain2 > gain1)) {
1997                         matchLength = ml2, offcode = offset2, start = ip;
1998                         continue;
1999             }   }   }
2000             break;  /* nothing found : store previous solution */
2001         }
2002 
2003         /* catch up */
2004         if (STORED_IS_OFFSET(offcode)) {
2005             U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
2006             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
2007             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
2008             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
2009             offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
2010         }
2011 
2012         /* store sequence */
2013 _storeSequence:
2014         {   size_t const litLength = (size_t)(start - anchor);
2015             ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
2016             anchor = ip = start + matchLength;
2017         }
2018 
2019         /* check immediate repcode */
2020         while (ip <= ilimit) {
2021             const U32 repCurrent = (U32)(ip-base);
2022             const U32 windowLow = ZSTD_getLowestMatchIndex(ms, repCurrent, windowLog);
2023             const U32 repIndex = repCurrent - offset_2;
2024             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
2025             const BYTE* const repMatch = repBase + repIndex;
2026             if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
2027                & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
2028             if (MEM_read32(ip) == MEM_read32(repMatch)) {
2029                 /* repcode detected we should take it */
2030                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
2031                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
2032                 offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset history */
2033                 ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
2034                 ip += matchLength;
2035                 anchor = ip;
2036                 continue;   /* faster when present ... (?) */
2037             }
2038             break;
2039     }   }
2040 
2041     /* Save reps for next block */
2042     rep[0] = offset_1;
2043     rep[1] = offset_2;
2044 
2045     /* Return the last literals size */
2046     return (size_t)(iend - anchor);
2047 }
2048 
2049 
2050 size_t ZSTD_compressBlock_greedy_extDict(
2051         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2052         void const* src, size_t srcSize)
2053 {
2054     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
2055 }
2056 
2057 size_t ZSTD_compressBlock_lazy_extDict(
2058         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2059         void const* src, size_t srcSize)
2060 
2061 {
2062     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
2063 }
2064 
2065 size_t ZSTD_compressBlock_lazy2_extDict(
2066         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2067         void const* src, size_t srcSize)
2068 
2069 {
2070     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
2071 }
2072 
2073 size_t ZSTD_compressBlock_btlazy2_extDict(
2074         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2075         void const* src, size_t srcSize)
2076 
2077 {
2078     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
2079 }
2080 
2081 size_t ZSTD_compressBlock_greedy_extDict_row(
2082         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2083         void const* src, size_t srcSize)
2084 {
2085     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
2086 }
2087 
2088 size_t ZSTD_compressBlock_lazy_extDict_row(
2089         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2090         void const* src, size_t srcSize)
2091 
2092 {
2093     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
2094 }
2095 
2096 size_t ZSTD_compressBlock_lazy2_extDict_row(
2097         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2098         void const* src, size_t srcSize)
2099 
2100 {
2101     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
2102 }
2103