4 * A match-finder for Lempel-Ziv compression based on bottom-up construction and
5 * traversal of the Longest Common Prefix (LCP) interval tree.
10 * The author dedicates this file to the public domain.
11 * You can do whatever you want with this file.
20 #include "wimlib/divsufsort.h"
21 #include "wimlib/lcpit_matchfinder.h"
22 #include "wimlib/util.h"
25 #define LCP_MAX (((u32)1 << LCP_BITS) - 1)
26 #define POS_MASK (((u32)1 << (32 - LCP_BITS)) - 1)
27 #define LCP_SHIFT (32 - LCP_BITS)
28 #define LCP_MASK (LCP_MAX << LCP_SHIFT)
29 #define MAX_NORMAL_BUFSIZE (POS_MASK + 1)
31 #define HUGE_LCP_BITS 7
32 #define HUGE_LCP_MAX (((u32)1 << HUGE_LCP_BITS) - 1)
33 #define HUGE_LCP_SHIFT (64 - HUGE_LCP_BITS)
34 #define HUGE_POS_MASK 0xFFFFFFFF
35 #define MAX_HUGE_BUFSIZE ((u64)HUGE_POS_MASK + 1)
36 #define HUGE_UNVISITED_TAG 0x100000000
38 #define PREFETCH_SAFETY 5
41 * Build the LCP (Longest Common Prefix) array in linear time.
43 * LCP[r] will be the length of the longest common prefix between the suffixes
44 * with positions SA[r - 1] and SA[r]. LCP[0] will be undefined.
46 * Algorithm taken from Kasai et al. (2001), but modified slightly:
48 * - With bytes there is no realistic way to reserve a unique symbol for
49 * end-of-buffer, so use explicit checks for end-of-buffer.
51 * - For decreased memory usage and improved memory locality, pack the two
52 * logically distinct SA and LCP arrays into a single array SA_and_LCP.
54 * - Since SA_and_LCP is accessed randomly, improve the cache behavior by
55 * reading several entries ahead in ISA and prefetching the upcoming
58 * - If an LCP value is less than the minimum match length, then store 0. This
59 * avoids having to do comparisons against the minimum match length later.
61 * - If an LCP value is greater than the "nice match length", then store the
62 * "nice match length". This caps the number of bits needed to store each
63 * LCP value, and this caps the depth of the LCP-interval tree, without
64 * usually hurting the compression ratio too much.
68 * Kasai et al. 2001. Linear-Time Longest-Common-Prefix Computation in
69 * Suffix Arrays and Its Applications. CPM '01 Proceedings of the 12th
70 * Annual Symposium on Combinatorial Pattern Matching pp. 181-192.
73 build_LCP(u32 SA_and_LCP[restrict], const u32 ISA[restrict],
74 const u8 T[restrict], const u32 n,
75 const u32 min_lcp, const u32 max_lcp)
78 for (u32 i = 0; i < n; i++) {
80 prefetch(&SA_and_LCP[ISA[i + PREFETCH_SAFETY]]);
82 const u32 j = SA_and_LCP[r - 1] & POS_MASK;
83 const u32 lim = min(n - i, n - j);
84 while (h < lim && T[i + h] == T[j + h])
87 if (stored_lcp < min_lcp)
89 else if (stored_lcp > max_lcp)
91 SA_and_LCP[r] |= stored_lcp << LCP_SHIFT;
99 * Use the suffix array accompanied with the longest-common-prefix array ---
100 * which in combination can be called the "enhanced suffix array" --- to
101 * simulate a bottom-up traversal of the corresponding suffix tree, or
102 * equivalently the lcp-interval tree. Do so in suffix rank order, but save the
103 * superinterval references needed for later bottom-up traversal of the tree in
104 * suffix position order.
106 * To enumerate the lcp-intervals, this algorithm scans the suffix array and its
107 * corresponding LCP array linearly. While doing so, it maintains a stack of
108 * lcp-intervals that are currently open, meaning that their left boundaries
109 * have been seen but their right boundaries have not. The bottom of the stack
110 * is the interval which covers the entire suffix array (this has lcp=0), and
111 * the top of the stack is the deepest interval that is currently open (this has
112 * the greatest lcp of any interval on the stack). When this algorithm opens an
113 * lcp-interval, it assigns it a unique index in intervals[] and pushes it onto
114 * the stack. When this algorithm closes an interval, it pops it from the stack
115 * and sets the intervals[] entry of that interval to the index and lcp of that
116 * interval's superinterval, which is the new top of the stack.
118 * This algorithm also set pos_data[pos] for each suffix position 'pos' to the
119 * index and lcp of the deepest lcp-interval containing it. Alternatively, we
120 * can interpret each suffix as being associated with a singleton lcp-interval,
121 * or leaf of the suffix tree. With this interpretation, an entry in pos_data[]
122 * is the superinterval reference for one of these singleton lcp-intervals and
123 * therefore is not fundamentally different from an entry in intervals[].
125 * To reduce memory usage, this algorithm re-uses the suffix array's memory to
126 * store the generated intervals[] array. This is possible because SA and LCP
127 * are accessed linearly, and no more than one interval is generated per suffix.
129 * The techniques used in this algorithm are described in various published
130 * papers. The generation of lcp-intervals from the suffix array (SA) and the
131 * longest-common-prefix array (LCP) is given as Algorithm BottomUpTraverse in
132 * Kasai et al. (2001) and Algorithm 4.1 ("Computation of lcp-intervals") in
133 * Abouelhoda et al. (2004). Both these papers note the equivalence between
134 * lcp-intervals (including the singleton lcp-interval for each suffix) and
135 * nodes of the suffix tree. Abouelhoda et al. (2004) furthermore applies
136 * bottom-up traversal of the lcp-interval tree to Lempel-Ziv factorization, as
137 * does Chen at al. (2008). Algorithm CPS1b of Chen et al. (2008) dynamically
138 * re-uses the suffix array during bottom-up traversal of the lcp-interval tree.
142 * Kasai et al. Linear-Time Longest-Common-Prefix Computation in Suffix
143 * Arrays and Its Applications. 2001. CPM '01 Proceedings of the 12th
144 * Annual Symposium on Combinatorial Pattern Matching pp. 181-192.
146 * M.I. Abouelhoda, S. Kurtz, E. Ohlebusch. 2004. Replacing Suffix Trees
147 * With Enhanced Suffix Arrays. Journal of Discrete Algorithms Volume 2
148 * Issue 1, March 2004, pp. 53-86.
150 * G. Chen, S.J. Puglisi, W.F. Smyth. 2008. Lempel-Ziv Factorization
151 * Using Less Time & Space. Mathematics in Computer Science June 2008,
152 * Volume 1, Issue 4, pp. 605-623.
155 build_LCPIT(u32 intervals[restrict], u32 pos_data[restrict], const u32 n)
157 u32 * const SA_and_LCP = intervals;
158 u32 next_interval_idx;
159 u32 open_intervals[LCP_MAX + 1];
160 u32 *top = open_intervals;
161 u32 prev_pos = SA_and_LCP[0] & POS_MASK;
165 next_interval_idx = 1;
167 for (u32 r = 1; r < n; r++) {
168 const u32 next_pos = SA_and_LCP[r] & POS_MASK;
169 const u32 next_lcp = SA_and_LCP[r] >> LCP_SHIFT;
170 const u32 top_lcp = *top >> LCP_SHIFT;
172 if (next_lcp == top_lcp) {
173 /* Continuing the deepest open interval */
174 pos_data[prev_pos] = *top;
175 } else if (next_lcp > top_lcp) {
176 /* Opening a new interval */
177 *++top = (next_lcp << LCP_SHIFT) | next_interval_idx++;
178 pos_data[prev_pos] = *top;
180 /* Closing the deepest open interval */
181 pos_data[prev_pos] = *top;
183 const u32 closed_interval_idx = *top-- & POS_MASK;
184 const u32 superinterval_lcp = *top >> LCP_SHIFT;
186 if (next_lcp == superinterval_lcp) {
187 /* Continuing the superinterval */
188 intervals[closed_interval_idx] = *top;
190 } else if (next_lcp > superinterval_lcp) {
191 /* Creating a new interval that is a
192 * superinterval of the one being
193 * closed, but still a subinterval of
194 * its superinterval */
195 *++top = (next_lcp << LCP_SHIFT) | next_interval_idx++;
196 intervals[closed_interval_idx] = *top;
199 /* Also closing the superinterval */
200 intervals[closed_interval_idx] = *top;
207 /* Close any still-open intervals. */
208 pos_data[prev_pos] = *top;
209 for (; top > open_intervals; top--)
210 intervals[*top & POS_MASK] = *(top - 1);
214 * Advance the LCP-interval tree matchfinder by one byte.
216 * If @record_matches is true, then matches are written to the @matches array
217 * sorted by strictly decreasing length and strictly decreasing offset, and the
218 * return value is the number of matches found. Otherwise, @matches is ignored
219 * and the return value is always 0.
221 * How this algorithm works:
223 * 'cur_pos' is the position of the current suffix, which is the suffix being
224 * matched against. 'cur_pos' starts at 0 and is incremented each time this
225 * function is called. This function finds each suffix with position less than
226 * 'cur_pos' that shares a prefix with the current suffix, but for each distinct
227 * prefix length it finds only the suffix with the greatest position (i.e. the
228 * most recently seen in the linear traversal by position). This function
229 * accomplishes this using the lcp-interval tree data structure that was built
230 * by build_LCPIT() and is updated dynamically by this function.
232 * The first step is to read 'pos_data[cur_pos]', which gives the index and lcp
233 * value of the deepest lcp-interval containing the current suffix --- or,
234 * equivalently, the parent of the conceptual singleton lcp-interval that
235 * contains the current suffix.
237 * The second step is to ascend the lcp-interval tree until reaching an interval
238 * that has not yet been visited, and link the intervals to the current suffix
239 * along the way. An lcp-interval has been visited if and only if it has been
240 * linked to a suffix. Initially, no lcp-intervals are linked to suffixes.
242 * The third step is to continue ascending the lcp-interval tree, but indirectly
243 * through suffix links rather than through the original superinterval
244 * references, and continuing to form links with the current suffix along the
245 * way. Each suffix visited during this step, except in a special case to
246 * handle outdated suffixes, is a match which can be written to matches[]. Each
247 * intervals[] entry contains the position of the next suffix to visit, which we
248 * shall call 'match_pos'; this is the most recently seen suffix that belongs to
249 * that lcp-interval. 'pos_data[match_pos]' then contains the lcp and interval
250 * index of the next lcp-interval that should be visited.
252 * We can view these arrays as describing a new set of links that gets overlaid
253 * on top of the original superinterval references of the lcp-interval tree.
254 * Each such link can connect a node of the lcp-interval tree to an ancestor
255 * more than one generation removed.
257 * For each one-byte advance, the current position becomes the most recently
258 * seen suffix for a continuous sequence of lcp-intervals from a leaf interval
259 * to the root interval. Conceptually, this algorithm needs to update all these
260 * nodes to link to 'cur_pos', and then update 'pos_data[cur_pos]' to a "null"
261 * link. But actually, if some of these nodes have the same most recently seen
262 * suffix, then this algorithm just visits the pos_data[] entry for that suffix
263 * and skips over all these nodes in one step. Updating the extra nodes is
264 * accomplished by creating a redirection from the previous suffix to the
267 * Using this shortcutting scheme, it's possible for a suffix to become out of
268 * date, which means that it is no longer the most recently seen suffix for the
269 * lcp-interval under consideration. This case is detected by noticing when the
270 * next lcp-interval link actually points deeper in the tree, and it is worked
271 * around by just continuing until we get to a link that actually takes us
272 * higher in the tree. This can be described as a lazy-update scheme.
275 lcpit_advance_one_byte(const u32 cur_pos,
276 u32 pos_data[restrict],
277 u32 intervals[restrict],
278 struct lz_match matches[restrict],
279 const bool record_matches)
284 struct lz_match *matchptr;
286 /* Get the deepest lcp-interval containing the current suffix. */
287 lcp = pos_data[cur_pos] >> LCP_SHIFT;
288 interval_idx = pos_data[cur_pos] & POS_MASK;
289 prefetch(&intervals[pos_data[cur_pos + 1] & POS_MASK]);
290 pos_data[cur_pos] = 0;
292 /* Ascend until we reach a visited interval, linking the unvisited
293 * intervals to the current suffix as we go. */
294 while (intervals[interval_idx] & LCP_MASK) {
295 const u32 superinterval_lcp = intervals[interval_idx] >> LCP_SHIFT;
296 const u32 superinterval_idx = intervals[interval_idx] & POS_MASK;
297 intervals[interval_idx] = cur_pos;
298 lcp = superinterval_lcp;
299 interval_idx = superinterval_idx;
302 match_pos = intervals[interval_idx] & POS_MASK;
303 if (match_pos == 0) {
304 /* Ambiguous case; just don't allow matches with position 0. */
305 if (interval_idx != 0)
306 intervals[interval_idx] = cur_pos;
310 /* Ascend indirectly via pos_data[] links. */
313 u32 next_interval_idx;
316 next_lcp = pos_data[match_pos] >> LCP_SHIFT;
317 next_interval_idx = pos_data[match_pos] & POS_MASK;
320 /* Suffix was out of date. */
321 match_pos = intervals[next_interval_idx];
323 intervals[interval_idx] = cur_pos;
324 pos_data[match_pos] = (lcp << LCP_SHIFT) | interval_idx;
325 if (record_matches) {
326 matchptr->length = lcp;
327 matchptr->offset = cur_pos - match_pos;
330 if (next_interval_idx == 0)
332 match_pos = intervals[next_interval_idx];
333 interval_idx = next_interval_idx;
336 return matchptr - matches;
339 /* Expand SA from 32 bits to 64 bits. */
341 expand_SA(void *p, u32 n)
343 typedef u32 _may_alias_attribute aliased_u32_t;
344 typedef u64 _may_alias_attribute aliased_u64_t;
346 aliased_u32_t *SA = p;
347 aliased_u64_t *SA64 = p;
355 /* Like build_LCP(), but for buffers larger than MAX_NORMAL_BUFSIZE. */
357 build_LCP_huge(u64 SA_and_LCP64[restrict], const u32 ISA[restrict],
358 const u8 T[restrict], const u32 n,
359 const u32 min_lcp, const u32 max_lcp)
362 for (u32 i = 0; i < n; i++) {
363 const u32 r = ISA[i];
364 prefetch(&SA_and_LCP64[ISA[i + PREFETCH_SAFETY]]);
366 const u32 j = SA_and_LCP64[r - 1] & HUGE_POS_MASK;
367 const u32 lim = min(n - i, n - j);
368 while (h < lim && T[i + h] == T[j + h])
371 if (stored_lcp < min_lcp)
373 else if (stored_lcp > max_lcp)
374 stored_lcp = max_lcp;
375 SA_and_LCP64[r] |= (u64)stored_lcp << HUGE_LCP_SHIFT;
383 * Like build_LCPIT(), but for buffers larger than MAX_NORMAL_BUFSIZE.
385 * This "huge" version is also slightly different in that the lcp value stored
386 * in each intervals[] entry is the lcp value for that interval, not its
387 * superinterval. This lcp value stays put in intervals[] and doesn't get moved
388 * to pos_data[] during lcpit_advance_one_byte_huge(). One consequence of this
389 * is that we have to use a special flag to distinguish visited from unvisited
390 * intervals. But overall, this scheme keeps the memory usage at 12n instead of
391 * 16n. (The non-huge version is 8n.)
394 build_LCPIT_huge(u64 intervals64[restrict], u32 pos_data[restrict], const u32 n)
396 u64 * const SA_and_LCP64 = intervals64;
397 u32 next_interval_idx;
398 u32 open_intervals[HUGE_LCP_MAX + 1];
399 u32 *top = open_intervals;
400 u32 prev_pos = SA_and_LCP64[0] & HUGE_POS_MASK;
404 next_interval_idx = 1;
406 for (u32 r = 1; r < n; r++) {
407 const u32 next_pos = SA_and_LCP64[r] & HUGE_POS_MASK;
408 const u32 next_lcp = SA_and_LCP64[r] >> HUGE_LCP_SHIFT;
409 const u32 top_lcp = intervals64[*top] >> HUGE_LCP_SHIFT;
411 if (next_lcp == top_lcp) {
412 /* Continuing the deepest open interval */
413 pos_data[prev_pos] = *top;
414 } else if (next_lcp > top_lcp) {
415 /* Opening a new interval */
416 intervals64[next_interval_idx] = (u64)next_lcp << HUGE_LCP_SHIFT;
417 pos_data[prev_pos] = next_interval_idx;
418 *++top = next_interval_idx++;
420 /* Closing the deepest open interval */
421 pos_data[prev_pos] = *top;
423 const u32 closed_interval_idx = *top--;
424 const u32 superinterval_lcp = intervals64[*top] >> HUGE_LCP_SHIFT;
426 if (next_lcp == superinterval_lcp) {
427 /* Continuing the superinterval */
428 intervals64[closed_interval_idx] |=
429 HUGE_UNVISITED_TAG | *top;
431 } else if (next_lcp > superinterval_lcp) {
432 /* Creating a new interval that is a
433 * superinterval of the one being
434 * closed, but still a subinterval of
435 * its superinterval */
436 intervals64[next_interval_idx] =
437 (u64)next_lcp << HUGE_LCP_SHIFT;
438 intervals64[closed_interval_idx] |=
439 HUGE_UNVISITED_TAG | next_interval_idx;
440 *++top = next_interval_idx++;
443 /* Also closing the superinterval */
444 intervals64[closed_interval_idx] |=
445 HUGE_UNVISITED_TAG | *top;
452 /* Close any still-open intervals. */
453 pos_data[prev_pos] = *top;
454 for (; top > open_intervals; top--)
455 intervals64[*top] |= HUGE_UNVISITED_TAG | *(top - 1);
458 /* Like lcpit_advance_one_byte(), but for buffers larger than
459 * MAX_NORMAL_BUFSIZE. */
461 lcpit_advance_one_byte_huge(const u32 cur_pos,
462 u32 pos_data[restrict],
463 u64 intervals64[restrict],
464 struct lz_match matches[restrict],
465 const bool record_matches)
470 struct lz_match *matchptr;
472 interval_idx = pos_data[cur_pos];
473 prefetch(&intervals64[pos_data[cur_pos + 1]]);
474 pos_data[cur_pos] = 0;
476 while (intervals64[interval_idx] & HUGE_UNVISITED_TAG) {
477 lcp = intervals64[interval_idx] >> HUGE_LCP_SHIFT;
480 const u32 superinterval_idx = intervals64[interval_idx] & HUGE_POS_MASK;
481 intervals64[interval_idx] = ((u64)lcp << HUGE_LCP_SHIFT) | cur_pos;
482 interval_idx = superinterval_idx;
485 match_pos = intervals64[interval_idx] & HUGE_POS_MASK;
486 lcp = intervals64[interval_idx] >> HUGE_LCP_SHIFT;
489 u32 next_interval_idx;
493 next_interval_idx = pos_data[match_pos];
494 next_lcp = intervals64[next_interval_idx] >> HUGE_LCP_SHIFT;
497 match_pos = intervals64[next_interval_idx] & HUGE_POS_MASK;
499 intervals64[interval_idx] = ((u64)lcp << HUGE_LCP_SHIFT) | cur_pos;
500 pos_data[match_pos] = interval_idx;
501 if (record_matches) {
502 matchptr->length = lcp;
503 matchptr->offset = cur_pos - match_pos;
506 match_pos = intervals64[next_interval_idx] & HUGE_POS_MASK;
507 interval_idx = next_interval_idx;
510 return matchptr - matches;
514 get_pos_data_size(size_t max_bufsize)
516 return (u64)max((u64)max_bufsize + PREFETCH_SAFETY,
517 DIVSUFSORT_TMP_LEN) * sizeof(u32);
521 get_intervals_size(size_t max_bufsize)
523 return (u64)max_bufsize * (max_bufsize <= MAX_NORMAL_BUFSIZE ?
524 sizeof(u32) : sizeof(u64));
528 * Calculate the number of bytes of memory needed for the LCP-interval tree
531 * @max_bufsize - maximum buffer size to support
533 * Returns the number of bytes required.
536 lcpit_matchfinder_get_needed_memory(size_t max_bufsize)
538 return get_pos_data_size(max_bufsize) + get_intervals_size(max_bufsize);
542 * Initialize the LCP-interval tree matchfinder.
544 * @mf - the matchfinder structure to initialize
545 * @max_bufsize - maximum buffer size to support
546 * @min_match_len - minimum match length in bytes
547 * @nice_match_len - only consider this many bytes of each match
549 * Returns true if successfully initialized; false if out of memory.
552 lcpit_matchfinder_init(struct lcpit_matchfinder *mf, size_t max_bufsize,
553 u32 min_match_len, u32 nice_match_len)
555 if (lcpit_matchfinder_get_needed_memory(max_bufsize) > SIZE_MAX)
557 if (max_bufsize > MAX_HUGE_BUFSIZE - PREFETCH_SAFETY)
560 mf->pos_data = MALLOC(get_pos_data_size(max_bufsize));
561 mf->intervals = MALLOC(get_intervals_size(max_bufsize));
562 if (!mf->pos_data || !mf->intervals) {
563 lcpit_matchfinder_destroy(mf);
567 mf->min_match_len = min_match_len;
568 mf->nice_match_len = min(nice_match_len,
569 (max_bufsize <= MAX_NORMAL_BUFSIZE) ?
570 LCP_MAX : HUGE_LCP_MAX);
571 for (u32 i = 0; i < PREFETCH_SAFETY; i++)
572 mf->pos_data[max_bufsize + i] = 0;
577 * Build the suffix array SA for the specified byte array T of length n.
579 * The suffix array is a sorted array of the byte array's suffixes, represented
580 * by indices into the byte array. It can equivalently be viewed as a mapping
581 * from suffix rank to suffix position.
583 * To build the suffix array, we use libdivsufsort, which uses an
584 * induced-sorting-based algorithm. In practice, this seems to be the fastest
585 * suffix array construction algorithm currently available.
589 * Y. Mori. libdivsufsort, a lightweight suffix-sorting library.
590 * https://code.google.com/p/libdivsufsort/.
592 * G. Nong, S. Zhang, and W.H. Chan. 2009. Linear Suffix Array
593 * Construction by Almost Pure Induced-Sorting. Data Compression
594 * Conference, 2009. DCC '09. pp. 193 - 202.
596 * S.J. Puglisi, W.F. Smyth, and A. Turpin. 2007. A Taxonomy of Suffix
597 * Array Construction Algorithms. ACM Computing Surveys (CSUR) Volume 39
598 * Issue 2, 2007 Article No. 4.
601 build_SA(u32 SA[], const u8 T[], u32 n, u32 *tmp)
603 /* Note: divsufsort() requires a fixed amount of temporary space. The
604 * implementation of divsufsort() has been modified from the original to
605 * use the provided temporary space instead of allocating its own, since
606 * we don't want to have to deal with malloc() failures here. */
607 divsufsort(T, SA, n, tmp);
611 * Build the inverse suffix array ISA from the suffix array SA.
613 * Whereas the suffix array is a mapping from suffix rank to suffix position,
614 * the inverse suffix array is a mapping from suffix position to suffix rank.
617 build_ISA(u32 ISA[restrict], const u32 SA[restrict], u32 n)
619 for (u32 r = 0; r < n; r++)
624 * Prepare the LCP-interval tree matchfinder for a new input buffer.
626 * @mf - the initialized matchfinder structure
627 * @T - the input buffer
628 * @n - size of the input buffer in bytes. This must be nonzero and can be at
629 * most the max_bufsize with which lcpit_matchfinder_init() was called.
632 lcpit_matchfinder_load_buffer(struct lcpit_matchfinder *mf, const u8 *T, u32 n)
634 /* intervals[] temporarily stores SA and LCP packed together.
635 * pos_data[] temporarily stores ISA.
636 * pos_data[] is also used as the temporary space for divsufsort(). */
638 build_SA(mf->intervals, T, n, mf->pos_data);
639 build_ISA(mf->pos_data, mf->intervals, n);
640 if (n <= MAX_NORMAL_BUFSIZE) {
641 build_LCP(mf->intervals, mf->pos_data, T, n,
642 mf->min_match_len, mf->nice_match_len);
643 build_LCPIT(mf->intervals, mf->pos_data, n);
644 mf->huge_mode = false;
646 expand_SA(mf->intervals, n);
647 build_LCP_huge(mf->intervals64, mf->pos_data, T, n,
648 mf->min_match_len, mf->nice_match_len);
649 build_LCPIT_huge(mf->intervals64, mf->pos_data, n);
650 mf->huge_mode = true;
652 mf->cur_pos = 0; /* starting at beginning of input buffer */
656 * Retrieve a list of matches with the next position.
658 * The matches will be recorded in the @matches array, ordered by strictly
659 * decreasing length and strictly decreasing offset.
661 * The return value is the number of matches found and written to @matches.
662 * This can be any value in [0, nice_match_len - min_match_len + 1].
665 lcpit_matchfinder_get_matches(struct lcpit_matchfinder *mf,
666 struct lz_match *matches)
669 return lcpit_advance_one_byte_huge(mf->cur_pos++, mf->pos_data,
670 mf->intervals64, matches, true);
672 return lcpit_advance_one_byte(mf->cur_pos++, mf->pos_data,
673 mf->intervals, matches, true);
677 * Skip the next @count bytes (don't search for matches at them). @count is
681 lcpit_matchfinder_skip_bytes(struct lcpit_matchfinder *mf, u32 count)
685 lcpit_advance_one_byte_huge(mf->cur_pos++, mf->pos_data,
686 mf->intervals64, NULL, false);
690 lcpit_advance_one_byte(mf->cur_pos++, mf->pos_data,
691 mf->intervals, NULL, false);
697 * Destroy an LCP-interval tree matchfinder that was previously initialized with
698 * lcpit_matchfinder_init().
701 lcpit_matchfinder_destroy(struct lcpit_matchfinder *mf)