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 LCP_SHIFT (32 - LCP_BITS)
27 #define LCP_MASK (LCP_MAX << LCP_SHIFT)
28 #define POS_MASK (((u32)1 << (32 - LCP_BITS)) - 1)
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_LCP_MASK ((u64)HUGE_LCP_MAX << HUGE_LCP_SHIFT)
35 #define HUGE_POS_MASK 0xFFFFFFFF
36 #define MAX_HUGE_BUFSIZE ((u64)HUGE_POS_MASK + 1)
37 #define HUGE_UNVISITED_TAG 0x100000000
39 #define PREFETCH_SAFETY 5
42 * Build the LCP (Longest Common Prefix) array in linear time.
44 * LCP[r] will be the length of the longest common prefix between the suffixes
45 * with positions SA[r - 1] and SA[r]. LCP[0] will be undefined.
47 * Algorithm taken from Kasai et al. (2001), but modified slightly:
49 * - With bytes there is no realistic way to reserve a unique symbol for
50 * end-of-buffer, so use explicit checks for end-of-buffer.
52 * - For decreased memory usage and improved memory locality, pack the two
53 * logically distinct SA and LCP arrays into a single array SA_and_LCP.
55 * - Since SA_and_LCP is accessed randomly, improve the cache behavior by
56 * reading several entries ahead in ISA and prefetching the upcoming
59 * - If an LCP value is less than the minimum match length, then store 0. This
60 * avoids having to do comparisons against the minimum match length later.
62 * - If an LCP value is greater than the "nice match length", then store the
63 * "nice match length". This caps the number of bits needed to store each
64 * LCP value, and this caps the depth of the LCP-interval tree, without
65 * usually hurting the compression ratio too much.
69 * Kasai et al. 2001. Linear-Time Longest-Common-Prefix Computation in
70 * Suffix Arrays and Its Applications. CPM '01 Proceedings of the 12th
71 * Annual Symposium on Combinatorial Pattern Matching pp. 181-192.
74 build_LCP(u32 SA_and_LCP[restrict], const u32 ISA[restrict],
75 const u8 T[restrict], const u32 n,
76 const u32 min_lcp, const u32 max_lcp)
79 for (u32 i = 0; i < n; i++) {
81 prefetch(&SA_and_LCP[ISA[i + PREFETCH_SAFETY]]);
83 const u32 j = SA_and_LCP[r - 1] & POS_MASK;
84 const u32 lim = min(n - i, n - j);
85 while (h < lim && T[i + h] == T[j + h])
88 if (stored_lcp < min_lcp)
90 else if (stored_lcp > max_lcp)
92 SA_and_LCP[r] |= stored_lcp << LCP_SHIFT;
100 * Use the suffix array accompanied with the longest-common-prefix array ---
101 * which in combination can be called the "enhanced suffix array" --- to
102 * simulate a bottom-up traversal of the corresponding suffix tree, or
103 * equivalently the lcp-interval tree. Do so in suffix rank order, but save the
104 * superinterval references needed for later bottom-up traversal of the tree in
105 * suffix position order.
107 * To enumerate the lcp-intervals, this algorithm scans the suffix array and its
108 * corresponding LCP array linearly. While doing so, it maintains a stack of
109 * lcp-intervals that are currently open, meaning that their left boundaries
110 * have been seen but their right boundaries have not. The bottom of the stack
111 * is the interval which covers the entire suffix array (this has lcp=0), and
112 * the top of the stack is the deepest interval that is currently open (this has
113 * the greatest lcp of any interval on the stack). When this algorithm opens an
114 * lcp-interval, it assigns it a unique index in intervals[] and pushes it onto
115 * the stack. When this algorithm closes an interval, it pops it from the stack
116 * and sets the intervals[] entry of that interval to the index and lcp of that
117 * interval's superinterval, which is the new top of the stack.
119 * This algorithm also set pos_data[pos] for each suffix position 'pos' to the
120 * index and lcp of the deepest lcp-interval containing it. Alternatively, we
121 * can interpret each suffix as being associated with a singleton lcp-interval,
122 * or leaf of the suffix tree. With this interpretation, an entry in pos_data[]
123 * is the superinterval reference for one of these singleton lcp-intervals and
124 * therefore is not fundamentally different from an entry in intervals[].
126 * To reduce memory usage, this algorithm re-uses the suffix array's memory to
127 * store the generated intervals[] array. This is possible because SA and LCP
128 * are accessed linearly, and no more than one interval is generated per suffix.
130 * The techniques used in this algorithm are described in various published
131 * papers. The generation of lcp-intervals from the suffix array (SA) and the
132 * longest-common-prefix array (LCP) is given as Algorithm BottomUpTraverse in
133 * Kasai et al. (2001) and Algorithm 4.1 ("Computation of lcp-intervals") in
134 * Abouelhoda et al. (2004). Both these papers note the equivalence between
135 * lcp-intervals (including the singleton lcp-interval for each suffix) and
136 * nodes of the suffix tree. Abouelhoda et al. (2004) furthermore applies
137 * bottom-up traversal of the lcp-interval tree to Lempel-Ziv factorization, as
138 * does Chen at al. (2008). Algorithm CPS1b of Chen et al. (2008) dynamically
139 * re-uses the suffix array during bottom-up traversal of the lcp-interval tree.
143 * Kasai et al. Linear-Time Longest-Common-Prefix Computation in Suffix
144 * Arrays and Its Applications. 2001. CPM '01 Proceedings of the 12th
145 * Annual Symposium on Combinatorial Pattern Matching pp. 181-192.
147 * M.I. Abouelhoda, S. Kurtz, E. Ohlebusch. 2004. Replacing Suffix Trees
148 * With Enhanced Suffix Arrays. Journal of Discrete Algorithms Volume 2
149 * Issue 1, March 2004, pp. 53-86.
151 * G. Chen, S.J. Puglisi, W.F. Smyth. 2008. Lempel-Ziv Factorization
152 * Using Less Time & Space. Mathematics in Computer Science June 2008,
153 * Volume 1, Issue 4, pp. 605-623.
156 build_LCPIT(u32 intervals[restrict], u32 pos_data[restrict], const u32 n)
158 u32 * const SA_and_LCP = intervals;
159 u32 next_interval_idx;
160 u32 open_intervals[LCP_MAX + 1];
161 u32 *top = open_intervals;
162 u32 prev_pos = SA_and_LCP[0] & POS_MASK;
166 next_interval_idx = 1;
168 for (u32 r = 1; r < n; r++) {
169 const u32 next_pos = SA_and_LCP[r] & POS_MASK;
170 const u32 next_lcp = SA_and_LCP[r] & LCP_MASK;
171 const u32 top_lcp = *top & LCP_MASK;
173 prefetch(&pos_data[SA_and_LCP[r + PREFETCH_SAFETY] & POS_MASK]);
175 if (next_lcp == top_lcp) {
176 /* Continuing the deepest open interval */
177 pos_data[prev_pos] = *top;
178 } else if (next_lcp > top_lcp) {
179 /* Opening a new interval */
180 *++top = next_lcp | next_interval_idx++;
181 pos_data[prev_pos] = *top;
183 /* Closing the deepest open interval */
184 pos_data[prev_pos] = *top;
186 const u32 closed_interval_idx = *top-- & POS_MASK;
187 const u32 superinterval_lcp = *top & LCP_MASK;
189 if (next_lcp == superinterval_lcp) {
190 /* Continuing the superinterval */
191 intervals[closed_interval_idx] = *top;
193 } else if (next_lcp > superinterval_lcp) {
194 /* Creating a new interval that is a
195 * superinterval of the one being
196 * closed, but still a subinterval of
197 * its superinterval */
198 *++top = next_lcp | next_interval_idx++;
199 intervals[closed_interval_idx] = *top;
202 /* Also closing the superinterval */
203 intervals[closed_interval_idx] = *top;
210 /* Close any still-open intervals. */
211 pos_data[prev_pos] = *top;
212 for (; top > open_intervals; top--)
213 intervals[*top & POS_MASK] = *(top - 1);
217 * Advance the LCP-interval tree matchfinder by one byte.
219 * If @record_matches is true, then matches are written to the @matches array
220 * sorted by strictly decreasing length and strictly decreasing offset, and the
221 * return value is the number of matches found. Otherwise, @matches is ignored
222 * and the return value is always 0.
224 * How this algorithm works:
226 * 'cur_pos' is the position of the current suffix, which is the suffix being
227 * matched against. 'cur_pos' starts at 0 and is incremented each time this
228 * function is called. This function finds each suffix with position less than
229 * 'cur_pos' that shares a prefix with the current suffix, but for each distinct
230 * prefix length it finds only the suffix with the greatest position (i.e. the
231 * most recently seen in the linear traversal by position). This function
232 * accomplishes this using the lcp-interval tree data structure that was built
233 * by build_LCPIT() and is updated dynamically by this function.
235 * The first step is to read 'pos_data[cur_pos]', which gives the index and lcp
236 * value of the deepest lcp-interval containing the current suffix --- or,
237 * equivalently, the parent of the conceptual singleton lcp-interval that
238 * contains the current suffix.
240 * The second step is to ascend the lcp-interval tree until reaching an interval
241 * that has not yet been visited, and link the intervals to the current suffix
242 * along the way. An lcp-interval has been visited if and only if it has been
243 * linked to a suffix. Initially, no lcp-intervals are linked to suffixes.
245 * The third step is to continue ascending the lcp-interval tree, but indirectly
246 * through suffix links rather than through the original superinterval
247 * references, and continuing to form links with the current suffix along the
248 * way. Each suffix visited during this step, except in a special case to
249 * handle outdated suffixes, is a match which can be written to matches[]. Each
250 * intervals[] entry contains the position of the next suffix to visit, which we
251 * shall call 'match_pos'; this is the most recently seen suffix that belongs to
252 * that lcp-interval. 'pos_data[match_pos]' then contains the lcp and interval
253 * index of the next lcp-interval that should be visited.
255 * We can view these arrays as describing a new set of links that gets overlaid
256 * on top of the original superinterval references of the lcp-interval tree.
257 * Each such link can connect a node of the lcp-interval tree to an ancestor
258 * more than one generation removed.
260 * For each one-byte advance, the current position becomes the most recently
261 * seen suffix for a continuous sequence of lcp-intervals from a leaf interval
262 * to the root interval. Conceptually, this algorithm needs to update all these
263 * nodes to link to 'cur_pos', and then update 'pos_data[cur_pos]' to a "null"
264 * link. But actually, if some of these nodes have the same most recently seen
265 * suffix, then this algorithm just visits the pos_data[] entry for that suffix
266 * and skips over all these nodes in one step. Updating the extra nodes is
267 * accomplished by creating a redirection from the previous suffix to the
270 * Using this shortcutting scheme, it's possible for a suffix to become out of
271 * date, which means that it is no longer the most recently seen suffix for the
272 * lcp-interval under consideration. This case is detected by noticing when the
273 * next lcp-interval link actually points deeper in the tree, and it is worked
274 * around by just continuing until we get to a link that actually takes us
275 * higher in the tree. This can be described as a lazy-update scheme.
278 lcpit_advance_one_byte(const u32 cur_pos,
279 u32 pos_data[restrict],
280 u32 intervals[restrict],
281 struct lz_match matches[restrict],
282 const bool record_matches)
287 struct lz_match *matchptr;
289 /* Get the deepest lcp-interval containing the current suffix. */
290 ref = pos_data[cur_pos];
292 /* Prefetch the deepest lcp-interval containing the *next* suffix. */
293 prefetch(&intervals[pos_data[cur_pos + 1] & POS_MASK]);
295 /* There is no "next suffix" after the current one. */
296 pos_data[cur_pos] = 0;
298 /* Ascend until we reach a visited interval, the root, or a child of the
299 * root. Link unvisited intervals to the current suffix as we go. */
300 while ((super_ref = intervals[ref & POS_MASK]) & LCP_MASK) {
301 intervals[ref & POS_MASK] = cur_pos;
305 if (super_ref == 0) {
306 /* In this case, the current interval may be any of:
308 * (2) an unvisited child of the root;
309 * (3) an interval last visited by suffix 0
311 * We could avoid the ambiguity with (3) by using an lcp
312 * placeholder value other than 0 to represent "visited", but
313 * it's fastest to use 0. So we just don't allow matches with
316 if (ref != 0) /* Not the root? */
317 intervals[ref & POS_MASK] = cur_pos;
321 /* Ascend indirectly via pos_data[] links. */
322 match_pos = super_ref;
325 while ((super_ref = pos_data[match_pos]) > ref)
326 match_pos = intervals[super_ref & POS_MASK];
327 intervals[ref & POS_MASK] = cur_pos;
328 pos_data[match_pos] = ref;
329 if (record_matches) {
330 matchptr->length = ref >> LCP_SHIFT;
331 matchptr->offset = cur_pos - match_pos;
337 match_pos = intervals[ref & POS_MASK];
339 return matchptr - matches;
342 /* Expand SA from 32 bits to 64 bits. */
344 expand_SA(void *p, u32 n)
346 typedef u32 _may_alias_attribute aliased_u32_t;
347 typedef u64 _may_alias_attribute aliased_u64_t;
349 aliased_u32_t *SA = p;
350 aliased_u64_t *SA64 = p;
358 /* Like build_LCP(), but for buffers larger than MAX_NORMAL_BUFSIZE. */
360 build_LCP_huge(u64 SA_and_LCP64[restrict], const u32 ISA[restrict],
361 const u8 T[restrict], const u32 n,
362 const u32 min_lcp, const u32 max_lcp)
365 for (u32 i = 0; i < n; i++) {
366 const u32 r = ISA[i];
367 prefetch(&SA_and_LCP64[ISA[i + PREFETCH_SAFETY]]);
369 const u32 j = SA_and_LCP64[r - 1] & HUGE_POS_MASK;
370 const u32 lim = min(n - i, n - j);
371 while (h < lim && T[i + h] == T[j + h])
374 if (stored_lcp < min_lcp)
376 else if (stored_lcp > max_lcp)
377 stored_lcp = max_lcp;
378 SA_and_LCP64[r] |= (u64)stored_lcp << HUGE_LCP_SHIFT;
386 * Like build_LCPIT(), but for buffers larger than MAX_NORMAL_BUFSIZE.
388 * This "huge" version is also slightly different in that the lcp value stored
389 * in each intervals[] entry is the lcp value for that interval, not its
390 * superinterval. This lcp value stays put in intervals[] and doesn't get moved
391 * to pos_data[] during lcpit_advance_one_byte_huge(). One consequence of this
392 * is that we have to use a special flag to distinguish visited from unvisited
393 * intervals. But overall, this scheme keeps the memory usage at 12n instead of
394 * 16n. (The non-huge version is 8n.)
397 build_LCPIT_huge(u64 intervals64[restrict], u32 pos_data[restrict], const u32 n)
399 u64 * const SA_and_LCP64 = intervals64;
400 u32 next_interval_idx;
401 u32 open_intervals[HUGE_LCP_MAX + 1];
402 u32 *top = open_intervals;
403 u32 prev_pos = SA_and_LCP64[0] & HUGE_POS_MASK;
407 next_interval_idx = 1;
409 for (u32 r = 1; r < n; r++) {
410 const u32 next_pos = SA_and_LCP64[r] & HUGE_POS_MASK;
411 const u64 next_lcp = SA_and_LCP64[r] & HUGE_LCP_MASK;
412 const u64 top_lcp = intervals64[*top];
414 prefetch(&pos_data[SA_and_LCP64[r + PREFETCH_SAFETY] & HUGE_POS_MASK]);
416 if (next_lcp == top_lcp) {
417 /* Continuing the deepest open interval */
418 pos_data[prev_pos] = *top;
419 } else if (next_lcp > top_lcp) {
420 /* Opening a new interval */
421 intervals64[next_interval_idx] = next_lcp;
422 pos_data[prev_pos] = next_interval_idx;
423 *++top = next_interval_idx++;
425 /* Closing the deepest open interval */
426 pos_data[prev_pos] = *top;
428 const u32 closed_interval_idx = *top--;
429 const u64 superinterval_lcp = intervals64[*top];
431 if (next_lcp == superinterval_lcp) {
432 /* Continuing the superinterval */
433 intervals64[closed_interval_idx] |=
434 HUGE_UNVISITED_TAG | *top;
436 } else if (next_lcp > superinterval_lcp) {
437 /* Creating a new interval that is a
438 * superinterval of the one being
439 * closed, but still a subinterval of
440 * its superinterval */
441 intervals64[next_interval_idx] = next_lcp;
442 intervals64[closed_interval_idx] |=
443 HUGE_UNVISITED_TAG | next_interval_idx;
444 *++top = next_interval_idx++;
447 /* Also closing the superinterval */
448 intervals64[closed_interval_idx] |=
449 HUGE_UNVISITED_TAG | *top;
456 /* Close any still-open intervals. */
457 pos_data[prev_pos] = *top;
458 for (; top > open_intervals; top--)
459 intervals64[*top] |= HUGE_UNVISITED_TAG | *(top - 1);
462 /* Like lcpit_advance_one_byte(), but for buffers larger than
463 * MAX_NORMAL_BUFSIZE. */
465 lcpit_advance_one_byte_huge(const u32 cur_pos,
466 u32 pos_data[restrict],
467 u64 intervals64[restrict],
468 struct lz_match matches[restrict],
469 const bool record_matches)
472 u32 next_interval_idx;
476 struct lz_match *matchptr;
478 interval_idx = pos_data[cur_pos];
479 prefetch(&pos_data[intervals64[pos_data[cur_pos + 1]] & HUGE_POS_MASK]);
480 prefetch(&intervals64[pos_data[cur_pos + 2]]);
481 pos_data[cur_pos] = 0;
483 while ((next = intervals64[interval_idx]) & HUGE_UNVISITED_TAG) {
484 intervals64[interval_idx] = (next & HUGE_LCP_MASK) | cur_pos;
485 interval_idx = next & HUGE_POS_MASK;
489 while (next & HUGE_LCP_MASK) {
492 match_pos = next & HUGE_POS_MASK;
493 next_interval_idx = pos_data[match_pos];
494 next = intervals64[next_interval_idx];
495 } while (next > cur);
496 intervals64[interval_idx] = (cur & HUGE_LCP_MASK) | cur_pos;
497 pos_data[match_pos] = interval_idx;
498 if (record_matches) {
499 matchptr->length = cur >> HUGE_LCP_SHIFT;
500 matchptr->offset = cur_pos - match_pos;
503 interval_idx = next_interval_idx;
505 return matchptr - matches;
509 get_pos_data_size(size_t max_bufsize)
511 return (u64)max((u64)max_bufsize + PREFETCH_SAFETY,
512 DIVSUFSORT_TMP_LEN) * sizeof(u32);
516 get_intervals_size(size_t max_bufsize)
518 return ((u64)max_bufsize + PREFETCH_SAFETY) *
519 (max_bufsize <= MAX_NORMAL_BUFSIZE ? sizeof(u32) : sizeof(u64));
523 * Calculate the number of bytes of memory needed for the LCP-interval tree
526 * @max_bufsize - maximum buffer size to support
528 * Returns the number of bytes required.
531 lcpit_matchfinder_get_needed_memory(size_t max_bufsize)
533 return get_pos_data_size(max_bufsize) + get_intervals_size(max_bufsize);
537 * Initialize the LCP-interval tree matchfinder.
539 * @mf - the matchfinder structure to initialize
540 * @max_bufsize - maximum buffer size to support
541 * @min_match_len - minimum match length in bytes
542 * @nice_match_len - only consider this many bytes of each match
544 * Returns true if successfully initialized; false if out of memory.
547 lcpit_matchfinder_init(struct lcpit_matchfinder *mf, size_t max_bufsize,
548 u32 min_match_len, u32 nice_match_len)
550 if (lcpit_matchfinder_get_needed_memory(max_bufsize) > SIZE_MAX)
552 if (max_bufsize > MAX_HUGE_BUFSIZE - PREFETCH_SAFETY)
555 mf->pos_data = MALLOC(get_pos_data_size(max_bufsize));
556 mf->intervals = MALLOC(get_intervals_size(max_bufsize));
557 if (!mf->pos_data || !mf->intervals) {
558 lcpit_matchfinder_destroy(mf);
562 mf->min_match_len = min_match_len;
563 mf->nice_match_len = min(nice_match_len,
564 (max_bufsize <= MAX_NORMAL_BUFSIZE) ?
565 LCP_MAX : HUGE_LCP_MAX);
570 * Build the suffix array SA for the specified byte array T of length n.
572 * The suffix array is a sorted array of the byte array's suffixes, represented
573 * by indices into the byte array. It can equivalently be viewed as a mapping
574 * from suffix rank to suffix position.
576 * To build the suffix array, we use libdivsufsort, which uses an
577 * induced-sorting-based algorithm. In practice, this seems to be the fastest
578 * suffix array construction algorithm currently available.
582 * Y. Mori. libdivsufsort, a lightweight suffix-sorting library.
583 * https://code.google.com/p/libdivsufsort/.
585 * G. Nong, S. Zhang, and W.H. Chan. 2009. Linear Suffix Array
586 * Construction by Almost Pure Induced-Sorting. Data Compression
587 * Conference, 2009. DCC '09. pp. 193 - 202.
589 * S.J. Puglisi, W.F. Smyth, and A. Turpin. 2007. A Taxonomy of Suffix
590 * Array Construction Algorithms. ACM Computing Surveys (CSUR) Volume 39
591 * Issue 2, 2007 Article No. 4.
594 build_SA(u32 SA[], const u8 T[], u32 n, u32 *tmp)
596 /* Note: divsufsort() requires a fixed amount of temporary space. The
597 * implementation of divsufsort() has been modified from the original to
598 * use the provided temporary space instead of allocating its own, since
599 * we don't want to have to deal with malloc() failures here. */
600 divsufsort(T, SA, n, tmp);
604 * Build the inverse suffix array ISA from the suffix array SA.
606 * Whereas the suffix array is a mapping from suffix rank to suffix position,
607 * the inverse suffix array is a mapping from suffix position to suffix rank.
610 build_ISA(u32 ISA[restrict], const u32 SA[restrict], u32 n)
612 for (u32 r = 0; r < n; r++)
617 * Prepare the LCP-interval tree matchfinder for a new input buffer.
619 * @mf - the initialized matchfinder structure
620 * @T - the input buffer
621 * @n - size of the input buffer in bytes. This must be nonzero and can be at
622 * most the max_bufsize with which lcpit_matchfinder_init() was called.
625 lcpit_matchfinder_load_buffer(struct lcpit_matchfinder *mf, const u8 *T, u32 n)
627 /* intervals[] temporarily stores SA and LCP packed together.
628 * pos_data[] temporarily stores ISA.
629 * pos_data[] is also used as the temporary space for divsufsort(). */
631 build_SA(mf->intervals, T, n, mf->pos_data);
632 build_ISA(mf->pos_data, mf->intervals, n);
633 if (n <= MAX_NORMAL_BUFSIZE) {
634 for (u32 i = 0; i < PREFETCH_SAFETY; i++) {
635 mf->intervals[n + i] = 0;
636 mf->pos_data[n + i] = 0;
638 build_LCP(mf->intervals, mf->pos_data, T, n,
639 mf->min_match_len, mf->nice_match_len);
640 build_LCPIT(mf->intervals, mf->pos_data, n);
641 mf->huge_mode = false;
643 for (u32 i = 0; i < PREFETCH_SAFETY; i++) {
644 mf->intervals64[n + i] = 0;
645 mf->pos_data[n + i] = 0;
647 expand_SA(mf->intervals, n);
648 build_LCP_huge(mf->intervals64, mf->pos_data, T, n,
649 mf->min_match_len, mf->nice_match_len);
650 build_LCPIT_huge(mf->intervals64, mf->pos_data, n);
651 mf->huge_mode = true;
653 mf->cur_pos = 0; /* starting at beginning of input buffer */
657 * Retrieve a list of matches with the next position.
659 * The matches will be recorded in the @matches array, ordered by strictly
660 * decreasing length and strictly decreasing offset.
662 * The return value is the number of matches found and written to @matches.
663 * This can be any value in [0, nice_match_len - min_match_len + 1].
666 lcpit_matchfinder_get_matches(struct lcpit_matchfinder *mf,
667 struct lz_match *matches)
670 return lcpit_advance_one_byte_huge(mf->cur_pos++, mf->pos_data,
671 mf->intervals64, matches, true);
673 return lcpit_advance_one_byte(mf->cur_pos++, mf->pos_data,
674 mf->intervals, matches, true);
678 * Skip the next @count bytes (don't search for matches at them). @count is
682 lcpit_matchfinder_skip_bytes(struct lcpit_matchfinder *mf, u32 count)
686 lcpit_advance_one_byte_huge(mf->cur_pos++, mf->pos_data,
687 mf->intervals64, NULL, false);
691 lcpit_advance_one_byte(mf->cur_pos++, mf->pos_data,
692 mf->intervals, NULL, false);
698 * Destroy an LCP-interval tree matchfinder that was previously initialized with
699 * lcpit_matchfinder_init().
702 lcpit_matchfinder_destroy(struct lcpit_matchfinder *mf)