xref: /linux/lib/zstd/compress/zstd_ldm.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
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_ldm.h"
12 
13 #include "../common/debug.h"
14 #include <linux/xxhash.h>
15 #include "zstd_fast.h"          /* ZSTD_fillHashTable() */
16 #include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
17 #include "zstd_ldm_geartab.h"
18 
19 #define LDM_BUCKET_SIZE_LOG 3
20 #define LDM_MIN_MATCH_LENGTH 64
21 #define LDM_HASH_RLOG 7
22 
23 typedef struct {
24     U64 rolling;
25     U64 stopMask;
26 } ldmRollingHashState_t;
27 
28 /* ZSTD_ldm_gear_init():
29  *
30  * Initializes the rolling hash state such that it will honor the
31  * settings in params. */
32 static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params)
33 {
34     unsigned maxBitsInMask = MIN(params->minMatchLength, 64);
35     unsigned hashRateLog = params->hashRateLog;
36 
37     state->rolling = ~(U32)0;
38 
39     /* The choice of the splitting criterion is subject to two conditions:
40      *   1. it has to trigger on average every 2^(hashRateLog) bytes;
41      *   2. ideally, it has to depend on a window of minMatchLength bytes.
42      *
43      * In the gear hash algorithm, bit n depends on the last n bytes;
44      * so in order to obtain a good quality splitting criterion it is
45      * preferable to use bits with high weight.
46      *
47      * To match condition 1 we use a mask with hashRateLog bits set
48      * and, because of the previous remark, we make sure these bits
49      * have the highest possible weight while still respecting
50      * condition 2.
51      */
52     if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) {
53         state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog);
54     } else {
55         /* In this degenerate case we simply honor the hash rate. */
56         state->stopMask = ((U64)1 << hashRateLog) - 1;
57     }
58 }
59 
60 /* ZSTD_ldm_gear_reset()
61  * Feeds [data, data + minMatchLength) into the hash without registering any
62  * splits. This effectively resets the hash state. This is used when skipping
63  * over data, either at the beginning of a block, or skipping sections.
64  */
65 static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state,
66                                 BYTE const* data, size_t minMatchLength)
67 {
68     U64 hash = state->rolling;
69     size_t n = 0;
70 
71 #define GEAR_ITER_ONCE() do {                                  \
72         hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
73         n += 1;                                                \
74     } while (0)
75     while (n + 3 < minMatchLength) {
76         GEAR_ITER_ONCE();
77         GEAR_ITER_ONCE();
78         GEAR_ITER_ONCE();
79         GEAR_ITER_ONCE();
80     }
81     while (n < minMatchLength) {
82         GEAR_ITER_ONCE();
83     }
84 #undef GEAR_ITER_ONCE
85 }
86 
87 /* ZSTD_ldm_gear_feed():
88  *
89  * Registers in the splits array all the split points found in the first
90  * size bytes following the data pointer. This function terminates when
91  * either all the data has been processed or LDM_BATCH_SIZE splits are
92  * present in the splits array.
93  *
94  * Precondition: The splits array must not be full.
95  * Returns: The number of bytes processed. */
96 static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state,
97                                  BYTE const* data, size_t size,
98                                  size_t* splits, unsigned* numSplits)
99 {
100     size_t n;
101     U64 hash, mask;
102 
103     hash = state->rolling;
104     mask = state->stopMask;
105     n = 0;
106 
107 #define GEAR_ITER_ONCE() do { \
108         hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
109         n += 1; \
110         if (UNLIKELY((hash & mask) == 0)) { \
111             splits[*numSplits] = n; \
112             *numSplits += 1; \
113             if (*numSplits == LDM_BATCH_SIZE) \
114                 goto done; \
115         } \
116     } while (0)
117 
118     while (n + 3 < size) {
119         GEAR_ITER_ONCE();
120         GEAR_ITER_ONCE();
121         GEAR_ITER_ONCE();
122         GEAR_ITER_ONCE();
123     }
124     while (n < size) {
125         GEAR_ITER_ONCE();
126     }
127 
128 #undef GEAR_ITER_ONCE
129 
130 done:
131     state->rolling = hash;
132     return n;
133 }
134 
135 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
136                                ZSTD_compressionParameters const* cParams)
137 {
138     params->windowLog = cParams->windowLog;
139     ZSTD_STATIC_ASSERT(LDM_BUCKET_SIZE_LOG <= ZSTD_LDM_BUCKETSIZELOG_MAX);
140     DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
141     if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG;
142     if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH;
143     if (params->hashLog == 0) {
144         params->hashLog = MAX(ZSTD_HASHLOG_MIN, params->windowLog - LDM_HASH_RLOG);
145         assert(params->hashLog <= ZSTD_HASHLOG_MAX);
146     }
147     if (params->hashRateLog == 0) {
148         params->hashRateLog = params->windowLog < params->hashLog
149                                    ? 0
150                                    : params->windowLog - params->hashLog;
151     }
152     params->bucketSizeLog = MIN(params->bucketSizeLog, params->hashLog);
153 }
154 
155 size_t ZSTD_ldm_getTableSize(ldmParams_t params)
156 {
157     size_t const ldmHSize = ((size_t)1) << params.hashLog;
158     size_t const ldmBucketSizeLog = MIN(params.bucketSizeLog, params.hashLog);
159     size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
160     size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize)
161                            + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t));
162     return params.enableLdm == ZSTD_ps_enable ? totalSize : 0;
163 }
164 
165 size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
166 {
167     return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0;
168 }
169 
170 /* ZSTD_ldm_getBucket() :
171  *  Returns a pointer to the start of the bucket associated with hash. */
172 static ldmEntry_t* ZSTD_ldm_getBucket(
173         ldmState_t* ldmState, size_t hash, ldmParams_t const ldmParams)
174 {
175     return ldmState->hashTable + (hash << ldmParams.bucketSizeLog);
176 }
177 
178 /* ZSTD_ldm_insertEntry() :
179  *  Insert the entry with corresponding hash into the hash table */
180 static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
181                                  size_t const hash, const ldmEntry_t entry,
182                                  ldmParams_t const ldmParams)
183 {
184     BYTE* const pOffset = ldmState->bucketOffsets + hash;
185     unsigned const offset = *pOffset;
186 
187     *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry;
188     *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1));
189 
190 }
191 
192 /* ZSTD_ldm_countBackwardsMatch() :
193  *  Returns the number of bytes that match backwards before pIn and pMatch.
194  *
195  *  We count only bytes where pMatch >= pBase and pIn >= pAnchor. */
196 static size_t ZSTD_ldm_countBackwardsMatch(
197             const BYTE* pIn, const BYTE* pAnchor,
198             const BYTE* pMatch, const BYTE* pMatchBase)
199 {
200     size_t matchLength = 0;
201     while (pIn > pAnchor && pMatch > pMatchBase && pIn[-1] == pMatch[-1]) {
202         pIn--;
203         pMatch--;
204         matchLength++;
205     }
206     return matchLength;
207 }
208 
209 /* ZSTD_ldm_countBackwardsMatch_2segments() :
210  *  Returns the number of bytes that match backwards from pMatch,
211  *  even with the backwards match spanning 2 different segments.
212  *
213  *  On reaching `pMatchBase`, start counting from mEnd */
214 static size_t ZSTD_ldm_countBackwardsMatch_2segments(
215                     const BYTE* pIn, const BYTE* pAnchor,
216                     const BYTE* pMatch, const BYTE* pMatchBase,
217                     const BYTE* pExtDictStart, const BYTE* pExtDictEnd)
218 {
219     size_t matchLength = ZSTD_ldm_countBackwardsMatch(pIn, pAnchor, pMatch, pMatchBase);
220     if (pMatch - matchLength != pMatchBase || pMatchBase == pExtDictStart) {
221         /* If backwards match is entirely in the extDict or prefix, immediately return */
222         return matchLength;
223     }
224     DEBUGLOG(7, "ZSTD_ldm_countBackwardsMatch_2segments: found 2-parts backwards match (length in prefix==%zu)", matchLength);
225     matchLength += ZSTD_ldm_countBackwardsMatch(pIn - matchLength, pAnchor, pExtDictEnd, pExtDictStart);
226     DEBUGLOG(7, "final backwards match length = %zu", matchLength);
227     return matchLength;
228 }
229 
230 /* ZSTD_ldm_fillFastTables() :
231  *
232  *  Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies.
233  *  This is similar to ZSTD_loadDictionaryContent.
234  *
235  *  The tables for the other strategies are filled within their
236  *  block compressors. */
237 static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
238                                       void const* end)
239 {
240     const BYTE* const iend = (const BYTE*)end;
241 
242     switch(ms->cParams.strategy)
243     {
244     case ZSTD_fast:
245         ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast);
246         break;
247 
248     case ZSTD_dfast:
249         ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast);
250         break;
251 
252     case ZSTD_greedy:
253     case ZSTD_lazy:
254     case ZSTD_lazy2:
255     case ZSTD_btlazy2:
256     case ZSTD_btopt:
257     case ZSTD_btultra:
258     case ZSTD_btultra2:
259         break;
260     default:
261         assert(0);  /* not possible : not a valid strategy id */
262     }
263 
264     return 0;
265 }
266 
267 void ZSTD_ldm_fillHashTable(
268             ldmState_t* ldmState, const BYTE* ip,
269             const BYTE* iend, ldmParams_t const* params)
270 {
271     U32 const minMatchLength = params->minMatchLength;
272     U32 const hBits = params->hashLog - params->bucketSizeLog;
273     BYTE const* const base = ldmState->window.base;
274     BYTE const* const istart = ip;
275     ldmRollingHashState_t hashState;
276     size_t* const splits = ldmState->splitIndices;
277     unsigned numSplits;
278 
279     DEBUGLOG(5, "ZSTD_ldm_fillHashTable");
280 
281     ZSTD_ldm_gear_init(&hashState, params);
282     while (ip < iend) {
283         size_t hashed;
284         unsigned n;
285 
286         numSplits = 0;
287         hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits);
288 
289         for (n = 0; n < numSplits; n++) {
290             if (ip + splits[n] >= istart + minMatchLength) {
291                 BYTE const* const split = ip + splits[n] - minMatchLength;
292                 U64 const xxhash = xxh64(split, minMatchLength, 0);
293                 U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
294                 ldmEntry_t entry;
295 
296                 entry.offset = (U32)(split - base);
297                 entry.checksum = (U32)(xxhash >> 32);
298                 ZSTD_ldm_insertEntry(ldmState, hash, entry, *params);
299             }
300         }
301 
302         ip += hashed;
303     }
304 }
305 
306 
307 /* ZSTD_ldm_limitTableUpdate() :
308  *
309  *  Sets cctx->nextToUpdate to a position corresponding closer to anchor
310  *  if it is far way
311  *  (after a long match, only update tables a limited amount). */
312 static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor)
313 {
314     U32 const curr = (U32)(anchor - ms->window.base);
315     if (curr > ms->nextToUpdate + 1024) {
316         ms->nextToUpdate =
317             curr - MIN(512, curr - ms->nextToUpdate - 1024);
318     }
319 }
320 
321 static size_t ZSTD_ldm_generateSequences_internal(
322         ldmState_t* ldmState, rawSeqStore_t* rawSeqStore,
323         ldmParams_t const* params, void const* src, size_t srcSize)
324 {
325     /* LDM parameters */
326     int const extDict = ZSTD_window_hasExtDict(ldmState->window);
327     U32 const minMatchLength = params->minMatchLength;
328     U32 const entsPerBucket = 1U << params->bucketSizeLog;
329     U32 const hBits = params->hashLog - params->bucketSizeLog;
330     /* Prefix and extDict parameters */
331     U32 const dictLimit = ldmState->window.dictLimit;
332     U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
333     BYTE const* const base = ldmState->window.base;
334     BYTE const* const dictBase = extDict ? ldmState->window.dictBase : NULL;
335     BYTE const* const dictStart = extDict ? dictBase + lowestIndex : NULL;
336     BYTE const* const dictEnd = extDict ? dictBase + dictLimit : NULL;
337     BYTE const* const lowPrefixPtr = base + dictLimit;
338     /* Input bounds */
339     BYTE const* const istart = (BYTE const*)src;
340     BYTE const* const iend = istart + srcSize;
341     BYTE const* const ilimit = iend - HASH_READ_SIZE;
342     /* Input positions */
343     BYTE const* anchor = istart;
344     BYTE const* ip = istart;
345     /* Rolling hash state */
346     ldmRollingHashState_t hashState;
347     /* Arrays for staged-processing */
348     size_t* const splits = ldmState->splitIndices;
349     ldmMatchCandidate_t* const candidates = ldmState->matchCandidates;
350     unsigned numSplits;
351 
352     if (srcSize < minMatchLength)
353         return iend - anchor;
354 
355     /* Initialize the rolling hash state with the first minMatchLength bytes */
356     ZSTD_ldm_gear_init(&hashState, params);
357     ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength);
358     ip += minMatchLength;
359 
360     while (ip < ilimit) {
361         size_t hashed;
362         unsigned n;
363 
364         numSplits = 0;
365         hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip,
366                                     splits, &numSplits);
367 
368         for (n = 0; n < numSplits; n++) {
369             BYTE const* const split = ip + splits[n] - minMatchLength;
370             U64 const xxhash = xxh64(split, minMatchLength, 0);
371             U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
372 
373             candidates[n].split = split;
374             candidates[n].hash = hash;
375             candidates[n].checksum = (U32)(xxhash >> 32);
376             candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params);
377             PREFETCH_L1(candidates[n].bucket);
378         }
379 
380         for (n = 0; n < numSplits; n++) {
381             size_t forwardMatchLength = 0, backwardMatchLength = 0,
382                    bestMatchLength = 0, mLength;
383             U32 offset;
384             BYTE const* const split = candidates[n].split;
385             U32 const checksum = candidates[n].checksum;
386             U32 const hash = candidates[n].hash;
387             ldmEntry_t* const bucket = candidates[n].bucket;
388             ldmEntry_t const* cur;
389             ldmEntry_t const* bestEntry = NULL;
390             ldmEntry_t newEntry;
391 
392             newEntry.offset = (U32)(split - base);
393             newEntry.checksum = checksum;
394 
395             /* If a split point would generate a sequence overlapping with
396              * the previous one, we merely register it in the hash table and
397              * move on */
398             if (split < anchor) {
399                 ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
400                 continue;
401             }
402 
403             for (cur = bucket; cur < bucket + entsPerBucket; cur++) {
404                 size_t curForwardMatchLength, curBackwardMatchLength,
405                        curTotalMatchLength;
406                 if (cur->checksum != checksum || cur->offset <= lowestIndex) {
407                     continue;
408                 }
409                 if (extDict) {
410                     BYTE const* const curMatchBase =
411                         cur->offset < dictLimit ? dictBase : base;
412                     BYTE const* const pMatch = curMatchBase + cur->offset;
413                     BYTE const* const matchEnd =
414                         cur->offset < dictLimit ? dictEnd : iend;
415                     BYTE const* const lowMatchPtr =
416                         cur->offset < dictLimit ? dictStart : lowPrefixPtr;
417                     curForwardMatchLength =
418                         ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr);
419                     if (curForwardMatchLength < minMatchLength) {
420                         continue;
421                     }
422                     curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments(
423                             split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd);
424                 } else { /* !extDict */
425                     BYTE const* const pMatch = base + cur->offset;
426                     curForwardMatchLength = ZSTD_count(split, pMatch, iend);
427                     if (curForwardMatchLength < minMatchLength) {
428                         continue;
429                     }
430                     curBackwardMatchLength =
431                         ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr);
432                 }
433                 curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength;
434 
435                 if (curTotalMatchLength > bestMatchLength) {
436                     bestMatchLength = curTotalMatchLength;
437                     forwardMatchLength = curForwardMatchLength;
438                     backwardMatchLength = curBackwardMatchLength;
439                     bestEntry = cur;
440                 }
441             }
442 
443             /* No match found -- insert an entry into the hash table
444              * and process the next candidate match */
445             if (bestEntry == NULL) {
446                 ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
447                 continue;
448             }
449 
450             /* Match found */
451             offset = (U32)(split - base) - bestEntry->offset;
452             mLength = forwardMatchLength + backwardMatchLength;
453             {
454                 rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
455 
456                 /* Out of sequence storage */
457                 if (rawSeqStore->size == rawSeqStore->capacity)
458                     return ERROR(dstSize_tooSmall);
459                 seq->litLength = (U32)(split - backwardMatchLength - anchor);
460                 seq->matchLength = (U32)mLength;
461                 seq->offset = offset;
462                 rawSeqStore->size++;
463             }
464 
465             /* Insert the current entry into the hash table --- it must be
466              * done after the previous block to avoid clobbering bestEntry */
467             ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
468 
469             anchor = split + forwardMatchLength;
470 
471             /* If we find a match that ends after the data that we've hashed
472              * then we have a repeating, overlapping, pattern. E.g. all zeros.
473              * If one repetition of the pattern matches our `stopMask` then all
474              * repetitions will. We don't need to insert them all into out table,
475              * only the first one. So skip over overlapping matches.
476              * This is a major speed boost (20x) for compressing a single byte
477              * repeated, when that byte ends up in the table.
478              */
479             if (anchor > ip + hashed) {
480                 ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
481                 /* Continue the outer loop at anchor (ip + hashed == anchor). */
482                 ip = anchor - hashed;
483                 break;
484             }
485         }
486 
487         ip += hashed;
488     }
489 
490     return iend - anchor;
491 }
492 
493 /*! ZSTD_ldm_reduceTable() :
494  *  reduce table indexes by `reducerValue` */
495 static void ZSTD_ldm_reduceTable(ldmEntry_t* const table, U32 const size,
496                                  U32 const reducerValue)
497 {
498     U32 u;
499     for (u = 0; u < size; u++) {
500         if (table[u].offset < reducerValue) table[u].offset = 0;
501         else table[u].offset -= reducerValue;
502     }
503 }
504 
505 size_t ZSTD_ldm_generateSequences(
506         ldmState_t* ldmState, rawSeqStore_t* sequences,
507         ldmParams_t const* params, void const* src, size_t srcSize)
508 {
509     U32 const maxDist = 1U << params->windowLog;
510     BYTE const* const istart = (BYTE const*)src;
511     BYTE const* const iend = istart + srcSize;
512     size_t const kMaxChunkSize = 1 << 20;
513     size_t const nbChunks = (srcSize / kMaxChunkSize) + ((srcSize % kMaxChunkSize) != 0);
514     size_t chunk;
515     size_t leftoverSize = 0;
516 
517     assert(ZSTD_CHUNKSIZE_MAX >= kMaxChunkSize);
518     /* Check that ZSTD_window_update() has been called for this chunk prior
519      * to passing it to this function.
520      */
521     assert(ldmState->window.nextSrc >= (BYTE const*)src + srcSize);
522     /* The input could be very large (in zstdmt), so it must be broken up into
523      * chunks to enforce the maximum distance and handle overflow correction.
524      */
525     assert(sequences->pos <= sequences->size);
526     assert(sequences->size <= sequences->capacity);
527     for (chunk = 0; chunk < nbChunks && sequences->size < sequences->capacity; ++chunk) {
528         BYTE const* const chunkStart = istart + chunk * kMaxChunkSize;
529         size_t const remaining = (size_t)(iend - chunkStart);
530         BYTE const *const chunkEnd =
531             (remaining < kMaxChunkSize) ? iend : chunkStart + kMaxChunkSize;
532         size_t const chunkSize = chunkEnd - chunkStart;
533         size_t newLeftoverSize;
534         size_t const prevSize = sequences->size;
535 
536         assert(chunkStart < iend);
537         /* 1. Perform overflow correction if necessary. */
538         if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) {
539             U32 const ldmHSize = 1U << params->hashLog;
540             U32 const correction = ZSTD_window_correctOverflow(
541                 &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart);
542             ZSTD_ldm_reduceTable(ldmState->hashTable, ldmHSize, correction);
543             /* invalidate dictionaries on overflow correction */
544             ldmState->loadedDictEnd = 0;
545         }
546         /* 2. We enforce the maximum offset allowed.
547          *
548          * kMaxChunkSize should be small enough that we don't lose too much of
549          * the window through early invalidation.
550          * TODO: * Test the chunk size.
551          *       * Try invalidation after the sequence generation and test the
552          *         the offset against maxDist directly.
553          *
554          * NOTE: Because of dictionaries + sequence splitting we MUST make sure
555          * that any offset used is valid at the END of the sequence, since it may
556          * be split into two sequences. This condition holds when using
557          * ZSTD_window_enforceMaxDist(), but if we move to checking offsets
558          * against maxDist directly, we'll have to carefully handle that case.
559          */
560         ZSTD_window_enforceMaxDist(&ldmState->window, chunkEnd, maxDist, &ldmState->loadedDictEnd, NULL);
561         /* 3. Generate the sequences for the chunk, and get newLeftoverSize. */
562         newLeftoverSize = ZSTD_ldm_generateSequences_internal(
563             ldmState, sequences, params, chunkStart, chunkSize);
564         if (ZSTD_isError(newLeftoverSize))
565             return newLeftoverSize;
566         /* 4. We add the leftover literals from previous iterations to the first
567          *    newly generated sequence, or add the `newLeftoverSize` if none are
568          *    generated.
569          */
570         /* Prepend the leftover literals from the last call */
571         if (prevSize < sequences->size) {
572             sequences->seq[prevSize].litLength += (U32)leftoverSize;
573             leftoverSize = newLeftoverSize;
574         } else {
575             assert(newLeftoverSize == chunkSize);
576             leftoverSize += chunkSize;
577         }
578     }
579     return 0;
580 }
581 
582 void
583 ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
584 {
585     while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
586         rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
587         if (srcSize <= seq->litLength) {
588             /* Skip past srcSize literals */
589             seq->litLength -= (U32)srcSize;
590             return;
591         }
592         srcSize -= seq->litLength;
593         seq->litLength = 0;
594         if (srcSize < seq->matchLength) {
595             /* Skip past the first srcSize of the match */
596             seq->matchLength -= (U32)srcSize;
597             if (seq->matchLength < minMatch) {
598                 /* The match is too short, omit it */
599                 if (rawSeqStore->pos + 1 < rawSeqStore->size) {
600                     seq[1].litLength += seq[0].matchLength;
601                 }
602                 rawSeqStore->pos++;
603             }
604             return;
605         }
606         srcSize -= seq->matchLength;
607         seq->matchLength = 0;
608         rawSeqStore->pos++;
609     }
610 }
611 
612 /*
613  * If the sequence length is longer than remaining then the sequence is split
614  * between this block and the next.
615  *
616  * Returns the current sequence to handle, or if the rest of the block should
617  * be literals, it returns a sequence with offset == 0.
618  */
619 static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore,
620                                  U32 const remaining, U32 const minMatch)
621 {
622     rawSeq sequence = rawSeqStore->seq[rawSeqStore->pos];
623     assert(sequence.offset > 0);
624     /* Likely: No partial sequence */
625     if (remaining >= sequence.litLength + sequence.matchLength) {
626         rawSeqStore->pos++;
627         return sequence;
628     }
629     /* Cut the sequence short (offset == 0 ==> rest is literals). */
630     if (remaining <= sequence.litLength) {
631         sequence.offset = 0;
632     } else if (remaining < sequence.litLength + sequence.matchLength) {
633         sequence.matchLength = remaining - sequence.litLength;
634         if (sequence.matchLength < minMatch) {
635             sequence.offset = 0;
636         }
637     }
638     /* Skip past `remaining` bytes for the future sequences. */
639     ZSTD_ldm_skipSequences(rawSeqStore, remaining, minMatch);
640     return sequence;
641 }
642 
643 void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
644     U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
645     while (currPos && rawSeqStore->pos < rawSeqStore->size) {
646         rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
647         if (currPos >= currSeq.litLength + currSeq.matchLength) {
648             currPos -= currSeq.litLength + currSeq.matchLength;
649             rawSeqStore->pos++;
650         } else {
651             rawSeqStore->posInSequence = currPos;
652             break;
653         }
654     }
655     if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
656         rawSeqStore->posInSequence = 0;
657     }
658 }
659 
660 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
661     ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
662     ZSTD_paramSwitch_e useRowMatchFinder,
663     void const* src, size_t srcSize)
664 {
665     const ZSTD_compressionParameters* const cParams = &ms->cParams;
666     unsigned const minMatch = cParams->minMatch;
667     ZSTD_blockCompressor const blockCompressor =
668         ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
669     /* Input bounds */
670     BYTE const* const istart = (BYTE const*)src;
671     BYTE const* const iend = istart + srcSize;
672     /* Input positions */
673     BYTE const* ip = istart;
674 
675     DEBUGLOG(5, "ZSTD_ldm_blockCompress: srcSize=%zu", srcSize);
676     /* If using opt parser, use LDMs only as candidates rather than always accepting them */
677     if (cParams->strategy >= ZSTD_btopt) {
678         size_t lastLLSize;
679         ms->ldmSeqStore = rawSeqStore;
680         lastLLSize = blockCompressor(ms, seqStore, rep, src, srcSize);
681         ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore, srcSize);
682         return lastLLSize;
683     }
684 
685     assert(rawSeqStore->pos <= rawSeqStore->size);
686     assert(rawSeqStore->size <= rawSeqStore->capacity);
687     /* Loop through each sequence and apply the block compressor to the literals */
688     while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
689         /* maybeSplitSequence updates rawSeqStore->pos */
690         rawSeq const sequence = maybeSplitSequence(rawSeqStore,
691                                                    (U32)(iend - ip), minMatch);
692         int i;
693         /* End signal */
694         if (sequence.offset == 0)
695             break;
696 
697         assert(ip + sequence.litLength + sequence.matchLength <= iend);
698 
699         /* Fill tables for block compressor */
700         ZSTD_ldm_limitTableUpdate(ms, ip);
701         ZSTD_ldm_fillFastTables(ms, ip);
702         /* Run the block compressor */
703         DEBUGLOG(5, "pos %u : calling block compressor on segment of size %u", (unsigned)(ip-istart), sequence.litLength);
704         {
705             size_t const newLitLength =
706                 blockCompressor(ms, seqStore, rep, ip, sequence.litLength);
707             ip += sequence.litLength;
708             /* Update the repcodes */
709             for (i = ZSTD_REP_NUM - 1; i > 0; i--)
710                 rep[i] = rep[i-1];
711             rep[0] = sequence.offset;
712             /* Store the sequence */
713             ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
714                           STORE_OFFSET(sequence.offset),
715                           sequence.matchLength);
716             ip += sequence.matchLength;
717         }
718     }
719     /* Fill the tables for the block compressor */
720     ZSTD_ldm_limitTableUpdate(ms, ip);
721     ZSTD_ldm_fillFastTables(ms, ip);
722     /* Compress the last literals */
723     return blockCompressor(ms, seqStore, rep, ip, iend - ip);
724 }
725