Line data Source code
1 : /*
2 : * DFA routines
3 : * This file is #included by regexec.c.
4 : *
5 : * Copyright (c) 1998, 1999 Henry Spencer. All rights reserved.
6 : *
7 : * Development of this software was funded, in part, by Cray Research Inc.,
8 : * UUNET Communications Services Inc., Sun Microsystems Inc., and Scriptics
9 : * Corporation, none of whom are responsible for the results. The author
10 : * thanks all of them.
11 : *
12 : * Redistribution and use in source and binary forms -- with or without
13 : * modification -- are permitted for any purpose, provided that
14 : * redistributions in source form retain this entire copyright notice and
15 : * indicate the origin and nature of any modifications.
16 : *
17 : * I'd appreciate being given credit for this package in the documentation
18 : * of software which uses it, but that is not a requirement.
19 : *
20 : * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES,
21 : * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY
22 : * AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL
23 : * HENRY SPENCER BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24 : * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25 : * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26 : * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27 : * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28 : * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
29 : * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 : *
31 : * src/backend/regex/rege_dfa.c
32 : *
33 : */
34 :
35 : /*
36 : * longest - longest-preferred matching engine
37 : *
38 : * On success, returns match endpoint address. Returns NULL on no match.
39 : * Internal errors also return NULL, with v->err set.
40 : */
41 : static chr *
42 937362 : longest(struct vars *v,
43 : struct dfa *d,
44 : chr *start, /* where the match should start */
45 : chr *stop, /* match must end at or before here */
46 : int *hitstopp) /* record whether hit v->stop, if non-NULL */
47 : {
48 : chr *cp;
49 937362 : chr *realstop = (stop == v->stop) ? stop : stop + 1;
50 : color co;
51 : struct sset *css;
52 : struct sset *ss;
53 : chr *post;
54 : int i;
55 937362 : struct colormap *cm = d->cm;
56 :
57 : /* prevent "uninitialized variable" warnings */
58 937362 : if (hitstopp != NULL)
59 904276 : *hitstopp = 0;
60 :
61 : /* if this is a backref to a known string, just match against that */
62 937362 : if (d->backno >= 0)
63 : {
64 : assert((size_t) d->backno < v->nmatch);
65 1586 : if (v->pmatch[d->backno].rm_so >= 0)
66 : {
67 1232 : cp = dfa_backref(v, d, start, start, stop, false);
68 1232 : if (cp == v->stop && stop == v->stop && hitstopp != NULL)
69 0 : *hitstopp = 1;
70 1232 : return cp;
71 : }
72 : }
73 :
74 : /* fast path for matchall NFAs */
75 936130 : if (d->cnfa->flags & MATCHALL)
76 : {
77 4576 : size_t nchr = stop - start;
78 4576 : size_t maxmatchall = d->cnfa->maxmatchall;
79 :
80 4576 : if (nchr < d->cnfa->minmatchall)
81 330 : return NULL;
82 4246 : if (maxmatchall == DUPINF)
83 : {
84 2562 : if (stop == v->stop && hitstopp != NULL)
85 10 : *hitstopp = 1;
86 : }
87 : else
88 : {
89 1684 : if (stop == v->stop && nchr <= maxmatchall + 1 && hitstopp != NULL)
90 168 : *hitstopp = 1;
91 1684 : if (nchr > maxmatchall)
92 1144 : return start + maxmatchall;
93 : }
94 3102 : return stop;
95 : }
96 :
97 : /* initialize */
98 931554 : css = initialize(v, d, start);
99 931554 : if (css == NULL)
100 0 : return NULL;
101 931554 : cp = start;
102 :
103 : /* startup */
104 : FDEBUG(("+++ startup +++\n"));
105 931554 : if (cp == v->start)
106 : {
107 3282 : co = d->cnfa->bos[(v->eflags & REG_NOTBOL) ? 0 : 1];
108 : FDEBUG(("color %ld\n", (long) co));
109 : }
110 : else
111 : {
112 928272 : co = GETCOLOR(cm, *(cp - 1));
113 : FDEBUG(("char %c, color %ld\n", (char) *(cp - 1), (long) co));
114 : }
115 931554 : css = miss(v, d, css, co, cp, start);
116 931554 : if (css == NULL)
117 432 : return NULL;
118 931122 : css->lastseen = cp;
119 :
120 : /*
121 : * This is the main text-scanning loop. It seems worth having two copies
122 : * to avoid the overhead of REG_FTRACE tests here, even in REG_DEBUG
123 : * builds, when you're not actively tracing.
124 : */
125 : #ifdef REG_DEBUG
126 : if (v->eflags & REG_FTRACE)
127 : {
128 : while (cp < realstop)
129 : {
130 : FDEBUG(("+++ at c%d +++\n", (int) (css - d->ssets)));
131 : co = GETCOLOR(cm, *cp);
132 : FDEBUG(("char %c, color %ld\n", (char) *cp, (long) co));
133 : ss = css->outs[co];
134 : if (ss == NULL)
135 : {
136 : ss = miss(v, d, css, co, cp + 1, start);
137 : if (ss == NULL)
138 : break; /* NOTE BREAK OUT */
139 : }
140 : cp++;
141 : ss->lastseen = cp;
142 : css = ss;
143 : }
144 : }
145 : else
146 : #endif
147 : {
148 11850698 : while (cp < realstop)
149 : {
150 11825660 : co = GETCOLOR(cm, *cp);
151 11825660 : ss = css->outs[co];
152 11825660 : if (ss == NULL)
153 : {
154 2970300 : ss = miss(v, d, css, co, cp + 1, start);
155 2970300 : if (ss == NULL)
156 906084 : break; /* NOTE BREAK OUT */
157 : }
158 10919576 : cp++;
159 10919576 : ss->lastseen = cp;
160 10919576 : css = ss;
161 : }
162 : }
163 :
164 931122 : if (ISERR())
165 0 : return NULL;
166 :
167 : /* shutdown */
168 : FDEBUG(("+++ shutdown at c%d +++\n", (int) (css - d->ssets)));
169 931122 : if (cp == v->stop && stop == v->stop)
170 : {
171 13134 : if (hitstopp != NULL)
172 6290 : *hitstopp = 1;
173 13134 : co = d->cnfa->eos[(v->eflags & REG_NOTEOL) ? 0 : 1];
174 : FDEBUG(("color %ld\n", (long) co));
175 13134 : ss = miss(v, d, css, co, cp, start);
176 13134 : if (ISERR())
177 0 : return NULL;
178 : /* special case: match ended at eol? */
179 13134 : if (ss != NULL && (ss->flags & POSTSTATE))
180 6980 : return cp;
181 6154 : else if (ss != NULL)
182 0 : ss->lastseen = cp; /* to be tidy */
183 : }
184 :
185 : /* find last match, if any */
186 924142 : post = d->lastpost;
187 4789640 : for (ss = d->ssets, i = d->nssused; i > 0; ss++, i--)
188 3865498 : if ((ss->flags & POSTSTATE) && post != ss->lastseen &&
189 13908 : (post == NULL || post < ss->lastseen))
190 933630 : post = ss->lastseen;
191 924142 : if (post != NULL) /* found one */
192 920486 : return post - 1;
193 :
194 3656 : return NULL;
195 : }
196 :
197 : /*
198 : * shortest - shortest-preferred matching engine
199 : *
200 : * On success, returns match endpoint address. Returns NULL on no match.
201 : * Internal errors also return NULL, with v->err set.
202 : */
203 : static chr *
204 1874098 : shortest(struct vars *v,
205 : struct dfa *d,
206 : chr *start, /* where the match should start */
207 : chr *min, /* match must end at or after here */
208 : chr *max, /* match must end at or before here */
209 : chr **coldp, /* store coldstart pointer here, if non-NULL */
210 : int *hitstopp) /* record whether hit v->stop, if non-NULL */
211 : {
212 : chr *cp;
213 1874098 : chr *realmin = (min == v->stop) ? min : min + 1;
214 1874098 : chr *realmax = (max == v->stop) ? max : max + 1;
215 : color co;
216 : struct sset *css;
217 : struct sset *ss;
218 1874098 : struct colormap *cm = d->cm;
219 :
220 : /* prevent "uninitialized variable" warnings */
221 1874098 : if (coldp != NULL)
222 1872438 : *coldp = NULL;
223 1874098 : if (hitstopp != NULL)
224 318 : *hitstopp = 0;
225 :
226 : /* if this is a backref to a known string, just match against that */
227 1874098 : if (d->backno >= 0)
228 : {
229 : assert((size_t) d->backno < v->nmatch);
230 0 : if (v->pmatch[d->backno].rm_so >= 0)
231 : {
232 0 : cp = dfa_backref(v, d, start, min, max, true);
233 0 : if (cp != NULL && coldp != NULL)
234 0 : *coldp = start;
235 : /* there is no case where we should set *hitstopp */
236 0 : return cp;
237 : }
238 : }
239 :
240 : /* fast path for matchall NFAs */
241 1874098 : if (d->cnfa->flags & MATCHALL)
242 : {
243 1972 : size_t nchr = min - start;
244 :
245 1972 : if (d->cnfa->maxmatchall != DUPINF &&
246 24 : nchr > d->cnfa->maxmatchall)
247 0 : return NULL;
248 1972 : if ((max - start) < d->cnfa->minmatchall)
249 18 : return NULL;
250 1954 : if (nchr < d->cnfa->minmatchall)
251 130 : min = start + d->cnfa->minmatchall;
252 1954 : if (coldp != NULL)
253 894 : *coldp = start;
254 : /* there is no case where we should set *hitstopp */
255 1954 : return min;
256 : }
257 :
258 : /* initialize */
259 1872126 : css = initialize(v, d, start);
260 1872126 : if (css == NULL)
261 0 : return NULL;
262 1872126 : cp = start;
263 :
264 : /* startup */
265 : FDEBUG(("--- startup ---\n"));
266 1872126 : if (cp == v->start)
267 : {
268 975824 : co = d->cnfa->bos[(v->eflags & REG_NOTBOL) ? 0 : 1];
269 : FDEBUG(("color %ld\n", (long) co));
270 : }
271 : else
272 : {
273 896302 : co = GETCOLOR(cm, *(cp - 1));
274 : FDEBUG(("char %c, color %ld\n", (char) *(cp - 1), (long) co));
275 : }
276 1872126 : css = miss(v, d, css, co, cp, start);
277 1872126 : if (css == NULL)
278 12 : return NULL;
279 1872114 : css->lastseen = cp;
280 1872114 : ss = css;
281 :
282 : /*
283 : * This is the main text-scanning loop. It seems worth having two copies
284 : * to avoid the overhead of REG_FTRACE tests here, even in REG_DEBUG
285 : * builds, when you're not actively tracing.
286 : */
287 : #ifdef REG_DEBUG
288 : if (v->eflags & REG_FTRACE)
289 : {
290 : while (cp < realmax)
291 : {
292 : FDEBUG(("--- at c%d ---\n", (int) (css - d->ssets)));
293 : co = GETCOLOR(cm, *cp);
294 : FDEBUG(("char %c, color %ld\n", (char) *cp, (long) co));
295 : ss = css->outs[co];
296 : if (ss == NULL)
297 : {
298 : ss = miss(v, d, css, co, cp + 1, start);
299 : if (ss == NULL)
300 : break; /* NOTE BREAK OUT */
301 : }
302 : cp++;
303 : ss->lastseen = cp;
304 : css = ss;
305 : if ((ss->flags & POSTSTATE) && cp >= realmin)
306 : break; /* NOTE BREAK OUT */
307 : }
308 : }
309 : else
310 : #endif
311 : {
312 36030010 : while (cp < realmax)
313 : {
314 35552972 : co = GETCOLOR(cm, *cp);
315 35552972 : ss = css->outs[co];
316 35552972 : if (ss == NULL)
317 : {
318 6332690 : ss = miss(v, d, css, co, cp + 1, start);
319 6332690 : if (ss == NULL)
320 268322 : break; /* NOTE BREAK OUT */
321 : }
322 35284650 : cp++;
323 35284650 : ss->lastseen = cp;
324 35284650 : css = ss;
325 35284650 : if ((ss->flags & POSTSTATE) && cp >= realmin)
326 1126754 : break; /* NOTE BREAK OUT */
327 : }
328 : }
329 :
330 1872114 : if (ss == NULL)
331 268322 : return NULL;
332 :
333 1603792 : if (coldp != NULL) /* report last no-progress state set, if any */
334 1603256 : *coldp = lastcold(v, d);
335 :
336 1603792 : if ((ss->flags & POSTSTATE) && cp > min)
337 : {
338 : assert(cp >= realmin);
339 1126722 : cp--;
340 : }
341 477070 : else if (cp == v->stop && max == v->stop)
342 : {
343 477070 : co = d->cnfa->eos[(v->eflags & REG_NOTEOL) ? 0 : 1];
344 : FDEBUG(("color %ld\n", (long) co));
345 477070 : ss = miss(v, d, css, co, cp, start);
346 : /* match might have ended at eol */
347 477070 : if ((ss == NULL || !(ss->flags & POSTSTATE)) && hitstopp != NULL)
348 12 : *hitstopp = 1;
349 : }
350 :
351 1603792 : if (ss == NULL || !(ss->flags & POSTSTATE))
352 458412 : return NULL;
353 :
354 1145380 : return cp;
355 : }
356 :
357 : /*
358 : * matchuntil - incremental matching engine
359 : *
360 : * This is meant for use with a search-style NFA (that is, the pattern is
361 : * known to act as though it had a leading .*). We determine whether a
362 : * match exists starting at v->start and ending at probe. Multiple calls
363 : * require only O(N) time not O(N^2) so long as the probe values are
364 : * nondecreasing. *lastcss and *lastcp must be initialized to NULL before
365 : * starting a series of calls.
366 : *
367 : * Returns 1 if a match exists, 0 if not.
368 : * Internal errors also return 0, with v->err set.
369 : */
370 : static int
371 120 : matchuntil(struct vars *v,
372 : struct dfa *d,
373 : chr *probe, /* we want to know if a match ends here */
374 : struct sset **lastcss, /* state storage across calls */
375 : chr **lastcp) /* state storage across calls */
376 : {
377 120 : chr *cp = *lastcp;
378 : color co;
379 120 : struct sset *css = *lastcss;
380 : struct sset *ss;
381 120 : struct colormap *cm = d->cm;
382 :
383 : /* fast path for matchall NFAs */
384 120 : if (d->cnfa->flags & MATCHALL)
385 : {
386 36 : size_t nchr = probe - v->start;
387 :
388 36 : if (nchr < d->cnfa->minmatchall)
389 18 : return 0;
390 : /* maxmatchall will always be infinity, cf. makesearch() */
391 : assert(d->cnfa->maxmatchall == DUPINF);
392 18 : return 1;
393 : }
394 :
395 : /* initialize and startup, or restart, if necessary */
396 84 : if (cp == NULL || cp > probe)
397 : {
398 24 : cp = v->start;
399 24 : css = initialize(v, d, cp);
400 24 : if (css == NULL)
401 0 : return 0;
402 :
403 : FDEBUG((">>> startup >>>\n"));
404 24 : co = d->cnfa->bos[(v->eflags & REG_NOTBOL) ? 0 : 1];
405 : FDEBUG(("color %ld\n", (long) co));
406 :
407 24 : css = miss(v, d, css, co, cp, v->start);
408 24 : if (css == NULL)
409 0 : return 0;
410 24 : css->lastseen = cp;
411 : }
412 60 : else if (css == NULL)
413 : {
414 : /* we previously found that no match is possible beyond *lastcp */
415 0 : return 0;
416 : }
417 84 : ss = css;
418 :
419 : /*
420 : * This is the main text-scanning loop. It seems worth having two copies
421 : * to avoid the overhead of REG_FTRACE tests here, even in REG_DEBUG
422 : * builds, when you're not actively tracing.
423 : */
424 : #ifdef REG_DEBUG
425 : if (v->eflags & REG_FTRACE)
426 : {
427 : while (cp < probe)
428 : {
429 : FDEBUG((">>> at c%d >>>\n", (int) (css - d->ssets)));
430 : co = GETCOLOR(cm, *cp);
431 : FDEBUG(("char %c, color %ld\n", (char) *cp, (long) co));
432 : ss = css->outs[co];
433 : if (ss == NULL)
434 : {
435 : ss = miss(v, d, css, co, cp + 1, v->start);
436 : if (ss == NULL)
437 : break; /* NOTE BREAK OUT */
438 : }
439 : cp++;
440 : ss->lastseen = cp;
441 : css = ss;
442 : }
443 : }
444 : else
445 : #endif
446 : {
447 180 : while (cp < probe)
448 : {
449 96 : co = GETCOLOR(cm, *cp);
450 96 : ss = css->outs[co];
451 96 : if (ss == NULL)
452 : {
453 12 : ss = miss(v, d, css, co, cp + 1, v->start);
454 12 : if (ss == NULL)
455 0 : break; /* NOTE BREAK OUT */
456 : }
457 96 : cp++;
458 96 : ss->lastseen = cp;
459 96 : css = ss;
460 : }
461 : }
462 :
463 84 : *lastcss = ss;
464 84 : *lastcp = cp;
465 :
466 84 : if (ss == NULL)
467 0 : return 0; /* impossible match, or internal error */
468 :
469 : /* We need to process one more chr, or the EOS symbol, to check match */
470 84 : if (cp < v->stop)
471 : {
472 : FDEBUG((">>> at c%d >>>\n", (int) (css - d->ssets)));
473 84 : co = GETCOLOR(cm, *cp);
474 : FDEBUG(("char %c, color %ld\n", (char) *cp, (long) co));
475 84 : ss = css->outs[co];
476 84 : if (ss == NULL)
477 54 : ss = miss(v, d, css, co, cp + 1, v->start);
478 : }
479 : else
480 : {
481 : assert(cp == v->stop);
482 0 : co = d->cnfa->eos[(v->eflags & REG_NOTEOL) ? 0 : 1];
483 : FDEBUG(("color %ld\n", (long) co));
484 0 : ss = miss(v, d, css, co, cp, v->start);
485 : }
486 :
487 84 : if (ss == NULL || !(ss->flags & POSTSTATE))
488 60 : return 0;
489 :
490 24 : return 1;
491 : }
492 :
493 : /*
494 : * dfa_backref - find best match length for a known backref string
495 : *
496 : * When the backref's referent is already available, we can deliver an exact
497 : * answer with considerably less work than running the backref node's NFA.
498 : *
499 : * Return match endpoint for longest or shortest valid repeated match,
500 : * or NULL if there is no valid match.
501 : *
502 : * Should be in sync with cbrdissect(), although that has the different task
503 : * of checking a match to a predetermined section of the string.
504 : */
505 : static chr *
506 1232 : dfa_backref(struct vars *v,
507 : struct dfa *d,
508 : chr *start, /* where the match should start */
509 : chr *min, /* match must end at or after here */
510 : chr *max, /* match must end at or before here */
511 : bool shortest)
512 : {
513 1232 : int n = d->backno;
514 1232 : int backmin = d->backmin;
515 1232 : int backmax = d->backmax;
516 : size_t numreps;
517 : size_t minreps;
518 : size_t maxreps;
519 : size_t brlen;
520 : chr *brstring;
521 : chr *p;
522 :
523 : /* get the backreferenced string (caller should have checked this) */
524 1232 : if (v->pmatch[n].rm_so == -1)
525 0 : return NULL;
526 1232 : brstring = v->start + v->pmatch[n].rm_so;
527 1232 : brlen = v->pmatch[n].rm_eo - v->pmatch[n].rm_so;
528 :
529 : /* special-case zero-length backreference to avoid divide by zero */
530 1232 : if (brlen == 0)
531 : {
532 : /*
533 : * matches only a zero-length string, but any number of repetitions
534 : * can be considered to be present
535 : */
536 2 : if (min == start && backmin <= backmax)
537 2 : return start;
538 0 : return NULL;
539 : }
540 :
541 : /*
542 : * convert min and max into numbers of possible repetitions of the backref
543 : * string, rounding appropriately
544 : */
545 1230 : if (min <= start)
546 1230 : minreps = 0;
547 : else
548 0 : minreps = (min - start - 1) / brlen + 1;
549 1230 : maxreps = (max - start) / brlen;
550 :
551 : /* apply bounds, then see if there is any allowed match length */
552 1230 : if (minreps < backmin)
553 1194 : minreps = backmin;
554 1230 : if (backmax != DUPINF && maxreps > backmax)
555 594 : maxreps = backmax;
556 1230 : if (maxreps < minreps)
557 268 : return NULL;
558 :
559 : /* quick exit if zero-repetitions match is valid and preferred */
560 962 : if (shortest && minreps == 0)
561 0 : return start;
562 :
563 : /* okay, compare the actual string contents */
564 962 : p = start;
565 962 : numreps = 0;
566 1140 : while (numreps < maxreps)
567 : {
568 984 : if ((*v->g->compare) (brstring, p, brlen) != 0)
569 806 : break;
570 178 : p += brlen;
571 178 : numreps++;
572 178 : if (shortest && numreps >= minreps)
573 0 : break;
574 : }
575 :
576 962 : if (numreps >= minreps)
577 164 : return p;
578 798 : return NULL;
579 : }
580 :
581 : /*
582 : * lastcold - determine last point at which no progress had been made
583 : */
584 : static chr * /* endpoint, or NULL */
585 1603256 : lastcold(struct vars *v,
586 : struct dfa *d)
587 : {
588 : struct sset *ss;
589 : chr *nopr;
590 : int i;
591 :
592 1603256 : nopr = d->lastnopr;
593 1603256 : if (nopr == NULL)
594 1603252 : nopr = v->start;
595 8587324 : for (ss = d->ssets, i = d->nssused; i > 0; ss++, i--)
596 6984068 : if ((ss->flags & NOPROGRESS) && nopr < ss->lastseen)
597 2323616 : nopr = ss->lastseen;
598 1603256 : return nopr;
599 : }
600 :
601 : /*
602 : * newdfa - set up a fresh DFA
603 : *
604 : * Returns NULL (and sets v->err) on failure.
605 : */
606 : static struct dfa *
607 2800806 : newdfa(struct vars *v,
608 : struct cnfa *cnfa,
609 : struct colormap *cm,
610 : struct smalldfa *sml) /* preallocated space, may be NULL */
611 : {
612 : struct dfa *d;
613 2800806 : size_t nss = cnfa->nstates * 2;
614 2800806 : int wordsper = (cnfa->nstates + UBITS - 1) / UBITS;
615 2800806 : bool ismalloced = false;
616 :
617 : assert(cnfa != NULL && cnfa->nstates != 0);
618 :
619 2800806 : if (nss <= FEWSTATES && cnfa->ncolors <= FEWCOLORS)
620 : {
621 : assert(wordsper == 1);
622 2541208 : if (sml == NULL)
623 : {
624 23246 : sml = (struct smalldfa *) MALLOC(sizeof(struct smalldfa));
625 23246 : if (sml == NULL)
626 : {
627 0 : ERR(REG_ESPACE);
628 0 : return NULL;
629 : }
630 23246 : ismalloced = true;
631 : }
632 2541208 : d = &sml->dfa;
633 2541208 : d->ssets = sml->ssets;
634 2541208 : d->statesarea = sml->statesarea;
635 2541208 : d->work = &d->statesarea[nss];
636 2541208 : d->outsarea = sml->outsarea;
637 2541208 : d->incarea = sml->incarea;
638 2541208 : d->ismalloced = ismalloced;
639 2541208 : d->arraysmalloced = false; /* not separately allocated, anyway */
640 : }
641 : else
642 : {
643 259598 : d = (struct dfa *) MALLOC(sizeof(struct dfa));
644 259598 : if (d == NULL)
645 : {
646 0 : ERR(REG_ESPACE);
647 0 : return NULL;
648 : }
649 259598 : d->ssets = (struct sset *) MALLOC(nss * sizeof(struct sset));
650 259598 : d->statesarea = (unsigned *) MALLOC((nss + WORK) * wordsper *
651 : sizeof(unsigned));
652 259598 : d->work = &d->statesarea[nss * wordsper];
653 259598 : d->outsarea = (struct sset **) MALLOC(nss * cnfa->ncolors *
654 : sizeof(struct sset *));
655 259598 : d->incarea = (struct arcp *) MALLOC(nss * cnfa->ncolors *
656 : sizeof(struct arcp));
657 259598 : d->ismalloced = true;
658 259598 : d->arraysmalloced = true;
659 : /* now freedfa() will behave sanely */
660 259598 : if (d->ssets == NULL || d->statesarea == NULL ||
661 259598 : d->outsarea == NULL || d->incarea == NULL)
662 : {
663 0 : freedfa(d);
664 0 : ERR(REG_ESPACE);
665 0 : return NULL;
666 : }
667 : }
668 :
669 2800806 : d->nssets = (v->eflags & REG_SMALL) ? 7 : nss;
670 2800806 : d->nssused = 0;
671 2800806 : d->nstates = cnfa->nstates;
672 2800806 : d->ncolors = cnfa->ncolors;
673 2800806 : d->wordsper = wordsper;
674 2800806 : d->cnfa = cnfa;
675 2800806 : d->cm = cm;
676 2800806 : d->lastpost = NULL;
677 2800806 : d->lastnopr = NULL;
678 2800806 : d->search = d->ssets;
679 2800806 : d->backno = -1; /* may be set by caller */
680 2800806 : d->backmin = d->backmax = 0;
681 :
682 : /* initialization of sset fields is done as needed */
683 :
684 2800806 : return d;
685 : }
686 :
687 : /*
688 : * freedfa - free a DFA
689 : */
690 : static void
691 2800806 : freedfa(struct dfa *d)
692 : {
693 2800806 : if (d->arraysmalloced)
694 : {
695 259598 : if (d->ssets != NULL)
696 259598 : FREE(d->ssets);
697 259598 : if (d->statesarea != NULL)
698 259598 : FREE(d->statesarea);
699 259598 : if (d->outsarea != NULL)
700 259598 : FREE(d->outsarea);
701 259598 : if (d->incarea != NULL)
702 259598 : FREE(d->incarea);
703 : }
704 :
705 2800806 : if (d->ismalloced)
706 282844 : FREE(d);
707 2800806 : }
708 :
709 : /*
710 : * hash - construct a hash code for a bitvector
711 : *
712 : * There are probably better ways, but they're more expensive.
713 : */
714 : static unsigned
715 30742 : hash(unsigned *uv,
716 : int n)
717 : {
718 : int i;
719 : unsigned h;
720 :
721 30742 : h = 0;
722 148146 : for (i = 0; i < n; i++)
723 117404 : h ^= uv[i];
724 30742 : return h;
725 : }
726 :
727 : /*
728 : * initialize - hand-craft a cache entry for startup, otherwise get ready
729 : */
730 : static struct sset *
731 2803704 : initialize(struct vars *v,
732 : struct dfa *d,
733 : chr *start)
734 : {
735 : struct sset *ss;
736 : int i;
737 :
738 : /* is previous one still there? */
739 2803704 : if (d->nssused > 0 && (d->ssets[0].flags & STARTER))
740 5488 : ss = &d->ssets[0];
741 : else
742 : { /* no, must (re)build it */
743 2798216 : ss = getvacant(v, d, start, start);
744 2798216 : if (ss == NULL)
745 0 : return NULL;
746 5598732 : for (i = 0; i < d->wordsper; i++)
747 2800516 : ss->states[i] = 0;
748 2798216 : BSET(ss->states, d->cnfa->pre);
749 2798216 : ss->hash = HASH(ss->states, d->wordsper);
750 : assert(d->cnfa->pre != d->cnfa->post);
751 2798216 : ss->flags = STARTER | LOCKED | NOPROGRESS;
752 : /* lastseen dealt with below */
753 : }
754 :
755 5682368 : for (i = 0; i < d->nssused; i++)
756 2878664 : d->ssets[i].lastseen = NULL;
757 2803704 : ss->lastseen = start; /* maybe untrue, but harmless */
758 2803704 : d->lastpost = NULL;
759 2803704 : d->lastnopr = NULL;
760 2803704 : return ss;
761 : }
762 :
763 : /*
764 : * miss - handle a stateset cache miss
765 : *
766 : * css is the current stateset, co is the color of the current input character,
767 : * cp points to the character after that (which is where we may need to test
768 : * LACONs). start does not affect matching behavior but is needed for pickss'
769 : * heuristics about which stateset cache entry to replace.
770 : *
771 : * Ordinarily, returns the address of the next stateset (the one that is
772 : * valid after consuming the input character). Returns NULL if no valid
773 : * NFA states remain, ie we have a certain match failure.
774 : * Internal errors also return NULL, with v->err set.
775 : */
776 : static struct sset *
777 12596964 : miss(struct vars *v,
778 : struct dfa *d,
779 : struct sset *css,
780 : color co,
781 : chr *cp, /* next chr */
782 : chr *start) /* where the attempt got started */
783 : {
784 12596964 : struct cnfa *cnfa = d->cnfa;
785 : int i;
786 : unsigned h;
787 : struct carc *ca;
788 : struct sset *p;
789 : int ispseudocolor;
790 : int ispost;
791 : int noprogress;
792 : int gotstate;
793 : int dolacons;
794 : int sawlacons;
795 :
796 : /* for convenience, we can be called even if it might not be a miss */
797 12596964 : if (css->outs[co] != NULL)
798 : {
799 : FDEBUG(("hit\n"));
800 4492 : return css->outs[co];
801 : }
802 : FDEBUG(("miss\n"));
803 :
804 : /*
805 : * Checking for operation cancel in the inner text search loop seems
806 : * unduly expensive. As a compromise, check during cache misses.
807 : */
808 12592472 : INTERRUPT(v->re);
809 :
810 : /*
811 : * What set of states would we end up in after consuming the co character?
812 : * We first consider PLAIN arcs that consume the character, and then look
813 : * to see what LACON arcs could be traversed after consuming it.
814 : */
815 25274634 : for (i = 0; i < d->wordsper; i++)
816 12682162 : d->work[i] = 0; /* build new stateset bitmap in d->work */
817 12592472 : ispseudocolor = d->cm->cd[co].flags & PSEUDO;
818 12592472 : ispost = 0;
819 12592472 : noprogress = 1;
820 12592472 : gotstate = 0;
821 97867778 : for (i = 0; i < d->nstates; i++)
822 85275306 : if (ISBSET(css->states, i))
823 87284672 : for (ca = cnfa->states[i]; ca->co != COLORLESS; ca++)
824 65512584 : if (ca->co == co ||
825 59768928 : (ca->co == RAINBOW && !ispseudocolor))
826 : {
827 22732324 : BSET(d->work, ca->to);
828 22732324 : gotstate = 1;
829 22732324 : if (ca->to == cnfa->post)
830 2099310 : ispost = 1;
831 22732324 : if (!(cnfa->stflags[ca->to] & CNFA_NOPROGRESS))
832 7630160 : noprogress = 0;
833 : FDEBUG(("%d -> %d\n", i, ca->to));
834 : }
835 12592472 : if (!gotstate)
836 1639416 : return NULL; /* character cannot reach any new state */
837 10953056 : dolacons = (cnfa->flags & HASLACONS);
838 10953056 : sawlacons = 0;
839 : /* outer loop handles transitive closure of reachable-by-LACON states */
840 10954660 : while (dolacons)
841 : {
842 1604 : dolacons = 0;
843 14692 : for (i = 0; i < d->nstates; i++)
844 13088 : if (ISBSET(d->work, i))
845 5200 : for (ca = cnfa->states[i]; ca->co != COLORLESS; ca++)
846 : {
847 3120 : if (ca->co < cnfa->ncolors)
848 2748 : continue; /* not a LACON arc */
849 372 : if (ISBSET(d->work, ca->to))
850 132 : continue; /* arc would be a no-op anyway */
851 240 : sawlacons = 1; /* this LACON affects our result */
852 240 : if (!lacon(v, cnfa, cp, ca->co))
853 : {
854 126 : if (ISERR())
855 0 : return NULL;
856 126 : continue; /* LACON arc cannot be traversed */
857 : }
858 114 : if (ISERR())
859 0 : return NULL;
860 114 : BSET(d->work, ca->to);
861 114 : dolacons = 1;
862 114 : if (ca->to == cnfa->post)
863 0 : ispost = 1;
864 114 : if (!(cnfa->stflags[ca->to] & CNFA_NOPROGRESS))
865 114 : noprogress = 0;
866 : FDEBUG(("%d :> %d\n", i, ca->to));
867 : }
868 : }
869 10953056 : h = HASH(d->work, d->wordsper);
870 :
871 : /* Is this stateset already in the cache? */
872 38328488 : for (p = d->ssets, i = d->nssused; i > 0; p++, i--)
873 29682010 : if (HIT(h, d->work, p, d->wordsper))
874 : {
875 : FDEBUG(("cached c%d\n", (int) (p - d->ssets)));
876 : break; /* NOTE BREAK OUT */
877 : }
878 10953056 : if (i == 0)
879 : { /* nope, need a new cache entry */
880 8646478 : p = getvacant(v, d, cp, start);
881 8646478 : if (p == NULL)
882 0 : return NULL;
883 : assert(p != css);
884 17368564 : for (i = 0; i < d->wordsper; i++)
885 8722086 : p->states[i] = d->work[i];
886 8646478 : p->hash = h;
887 8646478 : p->flags = (ispost) ? POSTSTATE : 0;
888 8646478 : if (noprogress)
889 2801294 : p->flags |= NOPROGRESS;
890 : /* lastseen to be dealt with by caller */
891 : }
892 :
893 : /*
894 : * Link new stateset to old, unless a LACON affected the result, in which
895 : * case we don't create the link. That forces future transitions across
896 : * this same arc (same prior stateset and character color) to come through
897 : * miss() again, so that we can recheck the LACON(s), which might or might
898 : * not pass since context will be different.
899 : */
900 10953056 : if (!sawlacons)
901 : {
902 : FDEBUG(("c%d[%d]->c%d\n",
903 : (int) (css - d->ssets), co, (int) (p - d->ssets)));
904 10952880 : css->outs[co] = p;
905 10952880 : css->inchain[co] = p->ins;
906 10952880 : p->ins.ss = css;
907 10952880 : p->ins.co = co;
908 : }
909 10953056 : return p;
910 : }
911 :
912 : /*
913 : * lacon - lookaround-constraint checker for miss()
914 : */
915 : static int /* predicate: constraint satisfied? */
916 240 : lacon(struct vars *v,
917 : struct cnfa *pcnfa, /* parent cnfa */
918 : chr *cp,
919 : color co) /* "color" of the lookaround constraint */
920 : {
921 : int n;
922 : struct subre *sub;
923 : struct dfa *d;
924 : chr *end;
925 : int satisfied;
926 :
927 : /* Since this is recursive, it could be driven to stack overflow */
928 240 : if (STACK_TOO_DEEP(v->re))
929 : {
930 0 : ERR(REG_ETOOBIG);
931 0 : return 0;
932 : }
933 :
934 240 : n = co - pcnfa->ncolors;
935 : assert(n > 0 && n < v->g->nlacons && v->g->lacons != NULL);
936 : FDEBUG(("=== testing lacon %d\n", n));
937 240 : sub = &v->g->lacons[n];
938 240 : d = getladfa(v, n);
939 240 : if (d == NULL)
940 0 : return 0;
941 240 : if (LATYPE_IS_AHEAD(sub->latype))
942 : {
943 : /* used to use longest() here, but shortest() could be much cheaper */
944 120 : end = shortest(v, d, cp, cp, v->stop,
945 : (chr **) NULL, (int *) NULL);
946 120 : satisfied = LATYPE_IS_POS(sub->latype) ? (end != NULL) : (end == NULL);
947 : }
948 : else
949 : {
950 : /*
951 : * To avoid doing O(N^2) work when repeatedly testing a lookbehind
952 : * constraint in an N-character string, we use matchuntil() which can
953 : * cache the DFA state across calls. We only need to restart if the
954 : * probe point decreases, which is not common. The NFA we're using is
955 : * a search NFA, so it doesn't mind scanning over stuff before the
956 : * nominal match.
957 : */
958 120 : satisfied = matchuntil(v, d, cp, &v->lblastcss[n], &v->lblastcp[n]);
959 120 : if (!LATYPE_IS_POS(sub->latype))
960 0 : satisfied = !satisfied;
961 : }
962 : FDEBUG(("=== lacon %d satisfied %d\n", n, satisfied));
963 240 : return satisfied;
964 : }
965 :
966 : /*
967 : * getvacant - get a vacant state set
968 : *
969 : * This routine clears out the inarcs and outarcs, but does not otherwise
970 : * clear the innards of the state set -- that's up to the caller.
971 : */
972 : static struct sset *
973 11444694 : getvacant(struct vars *v,
974 : struct dfa *d,
975 : chr *cp,
976 : chr *start)
977 : {
978 : int i;
979 : struct sset *ss;
980 : struct sset *p;
981 : struct arcp ap;
982 : color co;
983 :
984 11444694 : ss = pickss(v, d, cp, start);
985 11444694 : if (ss == NULL)
986 0 : return NULL;
987 : assert(!(ss->flags & LOCKED));
988 :
989 : /* clear out its inarcs, including self-referential ones */
990 11444694 : ap = ss->ins;
991 11444718 : while ((p = ap.ss) != NULL)
992 : {
993 24 : co = ap.co;
994 : FDEBUG(("zapping c%d's %ld outarc\n", (int) (p - d->ssets), (long) co));
995 24 : p->outs[co] = NULL;
996 24 : ap = p->inchain[co];
997 24 : p->inchain[co].ss = NULL; /* paranoia */
998 : }
999 11444694 : ss->ins.ss = NULL;
1000 :
1001 : /* take it off the inarc chains of the ssets reached by its outarcs */
1002 99926822 : for (i = 0; i < d->ncolors; i++)
1003 : {
1004 88482128 : p = ss->outs[i];
1005 : assert(p != ss); /* not self-referential */
1006 88482128 : if (p == NULL)
1007 88481996 : continue; /* NOTE CONTINUE */
1008 : FDEBUG(("del outarc %d from c%d's in chn\n", i, (int) (p - d->ssets)));
1009 132 : if (p->ins.ss == ss && p->ins.co == i)
1010 120 : p->ins = ss->inchain[i];
1011 : else
1012 : {
1013 12 : struct arcp lastap = {NULL, 0};
1014 :
1015 : assert(p->ins.ss != NULL);
1016 24 : for (ap = p->ins; ap.ss != NULL &&
1017 24 : !(ap.ss == ss && ap.co == i);
1018 12 : ap = ap.ss->inchain[ap.co])
1019 12 : lastap = ap;
1020 : assert(ap.ss != NULL);
1021 12 : lastap.ss->inchain[lastap.co] = ss->inchain[i];
1022 : }
1023 132 : ss->outs[i] = NULL;
1024 132 : ss->inchain[i].ss = NULL;
1025 : }
1026 :
1027 : /* if ss was a success state, may need to remember location */
1028 11444694 : if ((ss->flags & POSTSTATE) && ss->lastseen != d->lastpost &&
1029 36 : (d->lastpost == NULL || d->lastpost < ss->lastseen))
1030 36 : d->lastpost = ss->lastseen;
1031 :
1032 : /* likewise for a no-progress state */
1033 11444694 : if ((ss->flags & NOPROGRESS) && ss->lastseen != d->lastnopr &&
1034 12 : (d->lastnopr == NULL || d->lastnopr < ss->lastseen))
1035 12 : d->lastnopr = ss->lastseen;
1036 :
1037 11444694 : return ss;
1038 : }
1039 :
1040 : /*
1041 : * pickss - pick the next stateset to be used
1042 : */
1043 : static struct sset *
1044 11444694 : pickss(struct vars *v,
1045 : struct dfa *d,
1046 : chr *cp,
1047 : chr *start)
1048 : {
1049 : int i;
1050 : struct sset *ss;
1051 : struct sset *end;
1052 : chr *ancient;
1053 :
1054 : /* shortcut for cases where cache isn't full */
1055 11444694 : if (d->nssused < d->nssets)
1056 : {
1057 11444562 : i = d->nssused;
1058 11444562 : d->nssused++;
1059 11444562 : ss = &d->ssets[i];
1060 : FDEBUG(("new c%d\n", i));
1061 : /* set up innards */
1062 11444562 : ss->states = &d->statesarea[i * d->wordsper];
1063 11444562 : ss->flags = 0;
1064 11444562 : ss->ins.ss = NULL;
1065 11444562 : ss->ins.co = WHITE; /* give it some value */
1066 11444562 : ss->outs = &d->outsarea[i * d->ncolors];
1067 11444562 : ss->inchain = &d->incarea[i * d->ncolors];
1068 99923618 : for (i = 0; i < d->ncolors; i++)
1069 : {
1070 88479056 : ss->outs[i] = NULL;
1071 88479056 : ss->inchain[i].ss = NULL;
1072 : }
1073 11444562 : return ss;
1074 : }
1075 :
1076 : /* look for oldest, or old enough anyway */
1077 132 : if (cp - start > d->nssets * 2 / 3) /* oldest 33% are expendable */
1078 132 : ancient = cp - d->nssets * 2 / 3;
1079 : else
1080 0 : ancient = start;
1081 144 : for (ss = d->search, end = &d->ssets[d->nssets]; ss < end; ss++)
1082 132 : if ((ss->lastseen == NULL || ss->lastseen < ancient) &&
1083 132 : !(ss->flags & LOCKED))
1084 : {
1085 120 : d->search = ss + 1;
1086 : FDEBUG(("replacing c%d\n", (int) (ss - d->ssets)));
1087 120 : return ss;
1088 : }
1089 24 : for (ss = d->ssets, end = d->search; ss < end; ss++)
1090 24 : if ((ss->lastseen == NULL || ss->lastseen < ancient) &&
1091 24 : !(ss->flags & LOCKED))
1092 : {
1093 12 : d->search = ss + 1;
1094 : FDEBUG(("replacing c%d\n", (int) (ss - d->ssets)));
1095 12 : return ss;
1096 : }
1097 :
1098 : /* nobody's old enough?!? -- something's really wrong */
1099 : FDEBUG(("cannot find victim to replace!\n"));
1100 0 : ERR(REG_ASSERT);
1101 0 : return NULL;
1102 : }
|
