Line data Source code
1 : /*
2 : * NFA utilities.
3 : * This file is #included by regcomp.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/regc_nfa.c
32 : *
33 : *
34 : * One or two things that technically ought to be in here
35 : * are actually in color.c, thanks to some incestuous relationships in
36 : * the color chains.
37 : */
38 :
39 : #define NISERR() VISERR(nfa->v)
40 : #define NERR(e) VERR(nfa->v, (e))
41 :
42 :
43 : /*
44 : * newnfa - set up an NFA
45 : */
46 : static struct nfa * /* the NFA, or NULL */
47 18852 : newnfa(struct vars *v,
48 : struct colormap *cm,
49 : struct nfa *parent) /* NULL if primary NFA */
50 : {
51 : struct nfa *nfa;
52 :
53 18852 : nfa = (struct nfa *) MALLOC(sizeof(struct nfa));
54 18852 : if (nfa == NULL)
55 : {
56 0 : ERR(REG_ESPACE);
57 0 : return NULL;
58 : }
59 :
60 : /* Make the NFA minimally valid, so freenfa() will behave sanely */
61 18852 : nfa->states = NULL;
62 18852 : nfa->slast = NULL;
63 18852 : nfa->freestates = NULL;
64 18852 : nfa->freearcs = NULL;
65 18852 : nfa->lastsb = NULL;
66 18852 : nfa->lastab = NULL;
67 18852 : nfa->lastsbused = 0;
68 18852 : nfa->lastabused = 0;
69 18852 : nfa->nstates = 0;
70 18852 : nfa->cm = cm;
71 18852 : nfa->v = v;
72 18852 : nfa->bos[0] = nfa->bos[1] = COLORLESS;
73 18852 : nfa->eos[0] = nfa->eos[1] = COLORLESS;
74 18852 : nfa->flags = 0;
75 18852 : nfa->minmatchall = nfa->maxmatchall = -1;
76 18852 : nfa->parent = parent; /* Precedes newfstate so parent is valid. */
77 :
78 : /* Create required infrastructure */
79 18852 : nfa->post = newfstate(nfa, '@'); /* number 0 */
80 18852 : nfa->pre = newfstate(nfa, '>'); /* number 1 */
81 18852 : nfa->init = newstate(nfa); /* may become invalid later */
82 18852 : nfa->final = newstate(nfa);
83 18852 : if (ISERR())
84 : {
85 0 : freenfa(nfa);
86 0 : return NULL;
87 : }
88 18852 : rainbow(nfa, nfa->cm, PLAIN, COLORLESS, nfa->pre, nfa->init);
89 18852 : newarc(nfa, '^', 1, nfa->pre, nfa->init);
90 18852 : newarc(nfa, '^', 0, nfa->pre, nfa->init);
91 18852 : rainbow(nfa, nfa->cm, PLAIN, COLORLESS, nfa->final, nfa->post);
92 18852 : newarc(nfa, '$', 1, nfa->final, nfa->post);
93 18852 : newarc(nfa, '$', 0, nfa->final, nfa->post);
94 :
95 18852 : if (ISERR())
96 : {
97 0 : freenfa(nfa);
98 0 : return NULL;
99 : }
100 18852 : return nfa;
101 : }
102 :
103 : /*
104 : * freenfa - free an entire NFA
105 : */
106 : static void
107 18852 : freenfa(struct nfa *nfa)
108 : {
109 : struct statebatch *sb;
110 : struct statebatch *sbnext;
111 : struct arcbatch *ab;
112 : struct arcbatch *abnext;
113 :
114 39358 : for (sb = nfa->lastsb; sb != NULL; sb = sbnext)
115 : {
116 20506 : sbnext = sb->next;
117 20506 : nfa->v->spaceused -= STATEBATCHSIZE(sb->nstates);
118 20506 : FREE(sb);
119 : }
120 18852 : nfa->lastsb = NULL;
121 53886 : for (ab = nfa->lastab; ab != NULL; ab = abnext)
122 : {
123 35034 : abnext = ab->next;
124 35034 : nfa->v->spaceused -= ARCBATCHSIZE(ab->narcs);
125 35034 : FREE(ab);
126 : }
127 18852 : nfa->lastab = NULL;
128 :
129 18852 : nfa->nstates = -1;
130 18852 : FREE(nfa);
131 18852 : }
132 :
133 : /*
134 : * newstate - allocate an NFA state, with zero flag value
135 : */
136 : static struct state * /* NULL on error */
137 494096 : newstate(struct nfa *nfa)
138 : {
139 : struct state *s;
140 :
141 : /*
142 : * This is a handy place to check for operation cancel during regex
143 : * compilation, since no code path will go very long without making a new
144 : * state or arc.
145 : */
146 494096 : INTERRUPT(nfa->v->re);
147 :
148 : /* first, recycle anything that's on the freelist */
149 494096 : if (nfa->freestates != NULL)
150 : {
151 35370 : s = nfa->freestates;
152 35370 : nfa->freestates = s->next;
153 : }
154 : /* otherwise, is there anything left in the last statebatch? */
155 458726 : else if (nfa->lastsb != NULL && nfa->lastsbused < nfa->lastsb->nstates)
156 : {
157 438220 : s = &nfa->lastsb->s[nfa->lastsbused++];
158 : }
159 : /* otherwise, need to allocate a new statebatch */
160 : else
161 : {
162 : struct statebatch *newSb;
163 : size_t nstates;
164 :
165 20506 : if (nfa->v->spaceused >= REG_MAX_COMPILE_SPACE)
166 : {
167 0 : NERR(REG_ETOOBIG);
168 0 : return NULL;
169 : }
170 20506 : nstates = (nfa->lastsb != NULL) ? nfa->lastsb->nstates * 2 : FIRSTSBSIZE;
171 20506 : if (nstates > MAXSBSIZE)
172 50 : nstates = MAXSBSIZE;
173 20506 : newSb = (struct statebatch *) MALLOC(STATEBATCHSIZE(nstates));
174 20506 : if (newSb == NULL)
175 : {
176 0 : NERR(REG_ESPACE);
177 0 : return NULL;
178 : }
179 20506 : nfa->v->spaceused += STATEBATCHSIZE(nstates);
180 20506 : newSb->nstates = nstates;
181 20506 : newSb->next = nfa->lastsb;
182 20506 : nfa->lastsb = newSb;
183 20506 : nfa->lastsbused = 1;
184 20506 : s = &newSb->s[0];
185 : }
186 :
187 : assert(nfa->nstates >= 0);
188 494096 : s->no = nfa->nstates++;
189 494096 : s->flag = 0;
190 494096 : if (nfa->states == NULL)
191 18852 : nfa->states = s;
192 494096 : s->nins = 0;
193 494096 : s->ins = NULL;
194 494096 : s->nouts = 0;
195 494096 : s->outs = NULL;
196 494096 : s->tmp = NULL;
197 494096 : s->next = NULL;
198 494096 : if (nfa->slast != NULL)
199 : {
200 : assert(nfa->slast->next == NULL);
201 475244 : nfa->slast->next = s;
202 : }
203 494096 : s->prev = nfa->slast;
204 494096 : nfa->slast = s;
205 494096 : return s;
206 : }
207 :
208 : /*
209 : * newfstate - allocate an NFA state with a specified flag value
210 : */
211 : static struct state * /* NULL on error */
212 37704 : newfstate(struct nfa *nfa, int flag)
213 : {
214 : struct state *s;
215 :
216 37704 : s = newstate(nfa);
217 37704 : if (s != NULL)
218 37704 : s->flag = (char) flag;
219 37704 : return s;
220 : }
221 :
222 : /*
223 : * dropstate - delete a state's inarcs and outarcs and free it
224 : */
225 : static void
226 214042 : dropstate(struct nfa *nfa,
227 : struct state *s)
228 : {
229 : struct arc *a;
230 :
231 252790 : while ((a = s->ins) != NULL)
232 38748 : freearc(nfa, a);
233 351388 : while ((a = s->outs) != NULL)
234 137346 : freearc(nfa, a);
235 214042 : freestate(nfa, s);
236 214042 : }
237 :
238 : /*
239 : * freestate - free a state, which has no in-arcs or out-arcs
240 : */
241 : static void
242 223024 : freestate(struct nfa *nfa,
243 : struct state *s)
244 : {
245 : assert(s != NULL);
246 : assert(s->nins == 0 && s->nouts == 0);
247 :
248 223024 : s->no = FREESTATE;
249 223024 : s->flag = 0;
250 223024 : if (s->next != NULL)
251 208558 : s->next->prev = s->prev;
252 : else
253 : {
254 : assert(s == nfa->slast);
255 14466 : nfa->slast = s->prev;
256 : }
257 223024 : if (s->prev != NULL)
258 223024 : s->prev->next = s->next;
259 : else
260 : {
261 : assert(s == nfa->states);
262 0 : nfa->states = s->next;
263 : }
264 223024 : s->prev = NULL;
265 223024 : s->next = nfa->freestates; /* don't delete it, put it on the free list */
266 223024 : nfa->freestates = s;
267 223024 : }
268 :
269 : /*
270 : * newarc - set up a new arc within an NFA
271 : *
272 : * This function checks to make sure that no duplicate arcs are created.
273 : * In general we never want duplicates.
274 : *
275 : * However: in principle, a RAINBOW arc is redundant with any plain arc
276 : * (unless that arc is for a pseudocolor). But we don't try to recognize
277 : * that redundancy, either here or in allied operations such as moveins().
278 : * The pseudocolor consideration makes that more costly than it seems worth.
279 : */
280 : static void
281 1878602 : newarc(struct nfa *nfa,
282 : int t,
283 : color co,
284 : struct state *from,
285 : struct state *to)
286 : {
287 : struct arc *a;
288 :
289 : assert(from != NULL && to != NULL);
290 :
291 : /*
292 : * This is a handy place to check for operation cancel during regex
293 : * compilation, since no code path will go very long without making a new
294 : * state or arc.
295 : */
296 1878602 : INTERRUPT(nfa->v->re);
297 :
298 : /* check for duplicate arc, using whichever chain is shorter */
299 1878602 : if (from->nouts <= to->nins)
300 : {
301 5973062 : for (a = from->outs; a != NULL; a = a->outchain)
302 5108438 : if (a->to == to && a->co == co && a->type == t)
303 169938 : return;
304 : }
305 : else
306 : {
307 76819570 : for (a = to->ins; a != NULL; a = a->inchain)
308 76146594 : if (a->from == from && a->co == co && a->type == t)
309 171064 : return;
310 : }
311 :
312 : /* no dup, so create the arc */
313 1537600 : createarc(nfa, t, co, from, to);
314 : }
315 :
316 : /*
317 : * createarc - create a new arc within an NFA
318 : *
319 : * This function must *only* be used after verifying that there is no existing
320 : * identical arc (same type/color/from/to).
321 : */
322 : static void
323 17426000 : createarc(struct nfa *nfa,
324 : int t,
325 : color co,
326 : struct state *from,
327 : struct state *to)
328 : {
329 : struct arc *a;
330 :
331 17426000 : a = allocarc(nfa);
332 17426000 : if (NISERR())
333 11418 : return;
334 : assert(a != NULL);
335 :
336 17414582 : a->type = t;
337 17414582 : a->co = co;
338 17414582 : a->to = to;
339 17414582 : a->from = from;
340 :
341 : /*
342 : * Put the new arc on the beginning, not the end, of the chains; it's
343 : * simpler here, and freearc() is the same cost either way. See also the
344 : * logic in moveins() and its cohorts, as well as fixempties().
345 : */
346 17414582 : a->inchain = to->ins;
347 17414582 : a->inchainRev = NULL;
348 17414582 : if (to->ins)
349 16831438 : to->ins->inchainRev = a;
350 17414582 : to->ins = a;
351 17414582 : a->outchain = from->outs;
352 17414582 : a->outchainRev = NULL;
353 17414582 : if (from->outs)
354 16906536 : from->outs->outchainRev = a;
355 17414582 : from->outs = a;
356 :
357 17414582 : from->nouts++;
358 17414582 : to->nins++;
359 :
360 17414582 : if (COLORED(a) && nfa->parent == NULL)
361 1591164 : colorchain(nfa->cm, a);
362 : }
363 :
364 : /*
365 : * allocarc - allocate a new arc within an NFA
366 : */
367 : static struct arc * /* NULL for failure */
368 17426000 : allocarc(struct nfa *nfa)
369 : {
370 : struct arc *a;
371 :
372 : /* first, recycle anything that's on the freelist */
373 17426000 : if (nfa->freearcs != NULL)
374 : {
375 1386302 : a = nfa->freearcs;
376 1386302 : nfa->freearcs = a->freechain;
377 : }
378 : /* otherwise, is there anything left in the last arcbatch? */
379 16039698 : else if (nfa->lastab != NULL && nfa->lastabused < nfa->lastab->narcs)
380 : {
381 15993246 : a = &nfa->lastab->a[nfa->lastabused++];
382 : }
383 : /* otherwise, need to allocate a new arcbatch */
384 : else
385 : {
386 : struct arcbatch *newAb;
387 : size_t narcs;
388 :
389 46452 : if (nfa->v->spaceused >= REG_MAX_COMPILE_SPACE)
390 : {
391 11418 : NERR(REG_ETOOBIG);
392 11418 : return NULL;
393 : }
394 35034 : narcs = (nfa->lastab != NULL) ? nfa->lastab->narcs * 2 : FIRSTABSIZE;
395 35034 : if (narcs > MAXABSIZE)
396 15058 : narcs = MAXABSIZE;
397 35034 : newAb = (struct arcbatch *) MALLOC(ARCBATCHSIZE(narcs));
398 35034 : if (newAb == NULL)
399 : {
400 0 : NERR(REG_ESPACE);
401 0 : return NULL;
402 : }
403 35034 : nfa->v->spaceused += ARCBATCHSIZE(narcs);
404 35034 : newAb->narcs = narcs;
405 35034 : newAb->next = nfa->lastab;
406 35034 : nfa->lastab = newAb;
407 35034 : nfa->lastabused = 1;
408 35034 : a = &newAb->a[0];
409 : }
410 :
411 17414582 : return a;
412 : }
413 :
414 : /*
415 : * freearc - free an arc
416 : */
417 : static void
418 1604434 : freearc(struct nfa *nfa,
419 : struct arc *victim)
420 : {
421 1604434 : struct state *from = victim->from;
422 1604434 : struct state *to = victim->to;
423 : struct arc *predecessor;
424 :
425 : assert(victim->type != 0);
426 :
427 : /* take it off color chain if necessary */
428 1604434 : if (COLORED(victim) && nfa->parent == NULL)
429 737306 : uncolorchain(nfa->cm, victim);
430 :
431 : /* take it off source's out-chain */
432 : assert(from != NULL);
433 1604434 : predecessor = victim->outchainRev;
434 1604434 : if (predecessor == NULL)
435 : {
436 : assert(from->outs == victim);
437 465486 : from->outs = victim->outchain;
438 : }
439 : else
440 : {
441 : assert(predecessor->outchain == victim);
442 1138948 : predecessor->outchain = victim->outchain;
443 : }
444 1604434 : if (victim->outchain != NULL)
445 : {
446 : assert(victim->outchain->outchainRev == victim);
447 1035188 : victim->outchain->outchainRev = predecessor;
448 : }
449 1604434 : from->nouts--;
450 :
451 : /* take it off target's in-chain */
452 : assert(to != NULL);
453 1604434 : predecessor = victim->inchainRev;
454 1604434 : if (predecessor == NULL)
455 : {
456 : assert(to->ins == victim);
457 766050 : to->ins = victim->inchain;
458 : }
459 : else
460 : {
461 : assert(predecessor->inchain == victim);
462 838384 : predecessor->inchain = victim->inchain;
463 : }
464 1604434 : if (victim->inchain != NULL)
465 : {
466 : assert(victim->inchain->inchainRev == victim);
467 1066410 : victim->inchain->inchainRev = predecessor;
468 : }
469 1604434 : to->nins--;
470 :
471 : /* clean up and place on NFA's free list */
472 1604434 : victim->type = 0;
473 1604434 : victim->from = NULL; /* precautions... */
474 1604434 : victim->to = NULL;
475 1604434 : victim->inchain = NULL;
476 1604434 : victim->inchainRev = NULL;
477 1604434 : victim->outchain = NULL;
478 1604434 : victim->outchainRev = NULL;
479 1604434 : victim->freechain = nfa->freearcs;
480 1604434 : nfa->freearcs = victim;
481 1604434 : }
482 :
483 : /*
484 : * changearcsource - flip an arc to have a different from state
485 : *
486 : * Caller must have verified that there is no pre-existing duplicate arc.
487 : */
488 : static void
489 588 : changearcsource(struct arc *a, struct state *newfrom)
490 : {
491 588 : struct state *oldfrom = a->from;
492 : struct arc *predecessor;
493 :
494 : assert(oldfrom != newfrom);
495 :
496 : /* take it off old source's out-chain */
497 : assert(oldfrom != NULL);
498 588 : predecessor = a->outchainRev;
499 588 : if (predecessor == NULL)
500 : {
501 : assert(oldfrom->outs == a);
502 588 : oldfrom->outs = a->outchain;
503 : }
504 : else
505 : {
506 : assert(predecessor->outchain == a);
507 0 : predecessor->outchain = a->outchain;
508 : }
509 588 : if (a->outchain != NULL)
510 : {
511 : assert(a->outchain->outchainRev == a);
512 574 : a->outchain->outchainRev = predecessor;
513 : }
514 588 : oldfrom->nouts--;
515 :
516 588 : a->from = newfrom;
517 :
518 : /* prepend it to new source's out-chain */
519 588 : a->outchain = newfrom->outs;
520 588 : a->outchainRev = NULL;
521 588 : if (newfrom->outs)
522 588 : newfrom->outs->outchainRev = a;
523 588 : newfrom->outs = a;
524 588 : newfrom->nouts++;
525 588 : }
526 :
527 : /*
528 : * changearctarget - flip an arc to have a different to state
529 : *
530 : * Caller must have verified that there is no pre-existing duplicate arc.
531 : */
532 : static void
533 320 : changearctarget(struct arc *a, struct state *newto)
534 : {
535 320 : struct state *oldto = a->to;
536 : struct arc *predecessor;
537 :
538 : assert(oldto != newto);
539 :
540 : /* take it off old target's in-chain */
541 : assert(oldto != NULL);
542 320 : predecessor = a->inchainRev;
543 320 : if (predecessor == NULL)
544 : {
545 : assert(oldto->ins == a);
546 320 : oldto->ins = a->inchain;
547 : }
548 : else
549 : {
550 : assert(predecessor->inchain == a);
551 0 : predecessor->inchain = a->inchain;
552 : }
553 320 : if (a->inchain != NULL)
554 : {
555 : assert(a->inchain->inchainRev == a);
556 312 : a->inchain->inchainRev = predecessor;
557 : }
558 320 : oldto->nins--;
559 :
560 320 : a->to = newto;
561 :
562 : /* prepend it to new target's in-chain */
563 320 : a->inchain = newto->ins;
564 320 : a->inchainRev = NULL;
565 320 : if (newto->ins)
566 320 : newto->ins->inchainRev = a;
567 320 : newto->ins = a;
568 320 : newto->nins++;
569 320 : }
570 :
571 : /*
572 : * hasnonemptyout - Does state have a non-EMPTY out arc?
573 : */
574 : static int
575 216580 : hasnonemptyout(struct state *s)
576 : {
577 : struct arc *a;
578 :
579 244766 : for (a = s->outs; a != NULL; a = a->outchain)
580 : {
581 240818 : if (a->type != EMPTY)
582 212632 : return 1;
583 : }
584 3948 : return 0;
585 : }
586 :
587 : /*
588 : * findarc - find arc, if any, from given source with given type and color
589 : * If there is more than one such arc, the result is random.
590 : */
591 : static struct arc *
592 1100 : findarc(struct state *s,
593 : int type,
594 : color co)
595 : {
596 : struct arc *a;
597 :
598 2746 : for (a = s->outs; a != NULL; a = a->outchain)
599 1648 : if (a->type == type && a->co == co)
600 2 : return a;
601 1098 : return NULL;
602 : }
603 :
604 : /*
605 : * cparc - allocate a new arc within an NFA, copying details from old one
606 : */
607 : static void
608 1491314 : cparc(struct nfa *nfa,
609 : struct arc *oa,
610 : struct state *from,
611 : struct state *to)
612 : {
613 1491314 : newarc(nfa, oa->type, oa->co, from, to);
614 1491314 : }
615 :
616 : /*
617 : * sortins - sort the in arcs of a state by from/color/type
618 : */
619 : static void
620 30614 : sortins(struct nfa *nfa,
621 : struct state *s)
622 : {
623 : struct arc **sortarray;
624 : struct arc *a;
625 30614 : int n = s->nins;
626 : int i;
627 :
628 30614 : if (n <= 1)
629 4 : return; /* nothing to do */
630 : /* make an array of arc pointers ... */
631 30610 : sortarray = (struct arc **) MALLOC(n * sizeof(struct arc *));
632 30610 : if (sortarray == NULL)
633 : {
634 0 : NERR(REG_ESPACE);
635 0 : return;
636 : }
637 30610 : i = 0;
638 135904 : for (a = s->ins; a != NULL; a = a->inchain)
639 105294 : sortarray[i++] = a;
640 : assert(i == n);
641 : /* ... sort the array */
642 30610 : qsort(sortarray, n, sizeof(struct arc *), sortins_cmp);
643 : /* ... and rebuild arc list in order */
644 : /* it seems worth special-casing first and last items to simplify loop */
645 30610 : a = sortarray[0];
646 30610 : s->ins = a;
647 30610 : a->inchain = sortarray[1];
648 30610 : a->inchainRev = NULL;
649 74684 : for (i = 1; i < n - 1; i++)
650 : {
651 44074 : a = sortarray[i];
652 44074 : a->inchain = sortarray[i + 1];
653 44074 : a->inchainRev = sortarray[i - 1];
654 : }
655 30610 : a = sortarray[i];
656 30610 : a->inchain = NULL;
657 30610 : a->inchainRev = sortarray[i - 1];
658 30610 : FREE(sortarray);
659 : }
660 :
661 : static int
662 75799544 : sortins_cmp(const void *a, const void *b)
663 : {
664 75799544 : const struct arc *aa = *((const struct arc *const *) a);
665 75799544 : const struct arc *bb = *((const struct arc *const *) b);
666 :
667 : /* we check the fields in the order they are most likely to be different */
668 75799544 : if (aa->from->no < bb->from->no)
669 60278346 : return -1;
670 15521198 : if (aa->from->no > bb->from->no)
671 15047776 : return 1;
672 473422 : if (aa->co < bb->co)
673 247422 : return -1;
674 226000 : if (aa->co > bb->co)
675 222808 : return 1;
676 3192 : if (aa->type < bb->type)
677 112 : return -1;
678 3080 : if (aa->type > bb->type)
679 58 : return 1;
680 3022 : return 0;
681 : }
682 :
683 : /*
684 : * sortouts - sort the out arcs of a state by to/color/type
685 : */
686 : static void
687 32 : sortouts(struct nfa *nfa,
688 : struct state *s)
689 : {
690 : struct arc **sortarray;
691 : struct arc *a;
692 32 : int n = s->nouts;
693 : int i;
694 :
695 32 : if (n <= 1)
696 0 : return; /* nothing to do */
697 : /* make an array of arc pointers ... */
698 32 : sortarray = (struct arc **) MALLOC(n * sizeof(struct arc *));
699 32 : if (sortarray == NULL)
700 : {
701 0 : NERR(REG_ESPACE);
702 0 : return;
703 : }
704 32 : i = 0;
705 968 : for (a = s->outs; a != NULL; a = a->outchain)
706 936 : sortarray[i++] = a;
707 : assert(i == n);
708 : /* ... sort the array */
709 32 : qsort(sortarray, n, sizeof(struct arc *), sortouts_cmp);
710 : /* ... and rebuild arc list in order */
711 : /* it seems worth special-casing first and last items to simplify loop */
712 32 : a = sortarray[0];
713 32 : s->outs = a;
714 32 : a->outchain = sortarray[1];
715 32 : a->outchainRev = NULL;
716 904 : for (i = 1; i < n - 1; i++)
717 : {
718 872 : a = sortarray[i];
719 872 : a->outchain = sortarray[i + 1];
720 872 : a->outchainRev = sortarray[i - 1];
721 : }
722 32 : a = sortarray[i];
723 32 : a->outchain = NULL;
724 32 : a->outchainRev = sortarray[i - 1];
725 32 : FREE(sortarray);
726 : }
727 :
728 : static int
729 5492 : sortouts_cmp(const void *a, const void *b)
730 : {
731 5492 : const struct arc *aa = *((const struct arc *const *) a);
732 5492 : const struct arc *bb = *((const struct arc *const *) b);
733 :
734 : /* we check the fields in the order they are most likely to be different */
735 5492 : if (aa->to->no < bb->to->no)
736 610 : return -1;
737 4882 : if (aa->to->no > bb->to->no)
738 160 : return 1;
739 4722 : if (aa->co < bb->co)
740 2562 : return -1;
741 2160 : if (aa->co > bb->co)
742 2116 : return 1;
743 44 : if (aa->type < bb->type)
744 2 : return -1;
745 42 : if (aa->type > bb->type)
746 14 : return 1;
747 28 : return 0;
748 : }
749 :
750 : /*
751 : * Common decision logic about whether to use arc-by-arc operations or
752 : * sort/merge. If there's just a few source arcs we cannot recoup the
753 : * cost of sorting the destination arc list, no matter how large it is.
754 : * Otherwise, limit the number of arc-by-arc comparisons to about 1000
755 : * (a somewhat arbitrary choice, but the breakeven point would probably
756 : * be machine dependent anyway).
757 : */
758 : #define BULK_ARC_OP_USE_SORT(nsrcarcs, ndestarcs) \
759 : ((nsrcarcs) < 4 ? 0 : ((nsrcarcs) > 32 || (ndestarcs) > 32))
760 :
761 : /*
762 : * moveins - move all in arcs of a state to another state
763 : *
764 : * You might think this could be done better by just updating the
765 : * existing arcs, and you would be right if it weren't for the need
766 : * for duplicate suppression, which makes it easier to just make new
767 : * ones to exploit the suppression built into newarc.
768 : *
769 : * However, if we have a whole lot of arcs to deal with, retail duplicate
770 : * checks become too slow. In that case we proceed by sorting and merging
771 : * the arc lists, and then we can indeed just update the arcs in-place.
772 : *
773 : * On the other hand, it's also true that this is frequently called with
774 : * a brand-new newState that has no existing in-arcs. In that case,
775 : * de-duplication is unnecessary, so we can just blindly move all the arcs.
776 : */
777 : static void
778 250102 : moveins(struct nfa *nfa,
779 : struct state *oldState,
780 : struct state *newState)
781 : {
782 : assert(oldState != newState);
783 :
784 250102 : if (newState->nins == 0)
785 : {
786 : /* No need for de-duplication */
787 : struct arc *a;
788 :
789 203092 : while ((a = oldState->ins) != NULL)
790 : {
791 105754 : createarc(nfa, a->type, a->co, a->from, newState);
792 105754 : freearc(nfa, a);
793 : }
794 : }
795 152764 : else if (!BULK_ARC_OP_USE_SORT(oldState->nins, newState->nins))
796 152672 : {
797 : /* With not too many arcs, just do them one at a time */
798 : struct arc *a;
799 :
800 375888 : while ((a = oldState->ins) != NULL)
801 : {
802 223216 : cparc(nfa, a, a->from, newState);
803 223216 : freearc(nfa, a);
804 : }
805 : }
806 : else
807 : {
808 : /*
809 : * With many arcs, use a sort-merge approach. Note changearctarget()
810 : * will put the arc onto the front of newState's chain, so it does not
811 : * break our walk through the sorted part of the chain.
812 : */
813 : struct arc *oa;
814 : struct arc *na;
815 :
816 : /*
817 : * Because we bypass newarc() in this code path, we'd better include a
818 : * cancel check.
819 : */
820 92 : INTERRUPT(nfa->v->re);
821 :
822 92 : sortins(nfa, oldState);
823 92 : sortins(nfa, newState);
824 92 : if (NISERR())
825 0 : return; /* might have failed to sort */
826 92 : oa = oldState->ins;
827 92 : na = newState->ins;
828 3310 : while (oa != NULL && na != NULL)
829 : {
830 3214 : struct arc *a = oa;
831 :
832 3214 : switch (sortins_cmp(&oa, &na))
833 : {
834 166 : case -1:
835 : /* newState does not have anything matching oa */
836 166 : oa = oa->inchain;
837 :
838 : /*
839 : * Rather than doing createarc+freearc, we can just unlink
840 : * and relink the existing arc struct.
841 : */
842 166 : changearctarget(a, newState);
843 166 : break;
844 472 : case 0:
845 : /* match, advance in both lists */
846 472 : oa = oa->inchain;
847 472 : na = na->inchain;
848 : /* ... and drop duplicate arc from oldState */
849 472 : freearc(nfa, a);
850 472 : break;
851 2576 : case +1:
852 : /* advance only na; oa might have a match later */
853 2576 : na = na->inchain;
854 2576 : break;
855 3306 : default:
856 : assert(NOTREACHED);
857 : }
858 : }
859 246 : while (oa != NULL)
860 : {
861 : /* newState does not have anything matching oa */
862 154 : struct arc *a = oa;
863 :
864 154 : oa = oa->inchain;
865 154 : changearctarget(a, newState);
866 : }
867 : }
868 :
869 : assert(oldState->nins == 0);
870 : assert(oldState->ins == NULL);
871 : }
872 :
873 : /*
874 : * copyins - copy in arcs of a state to another state
875 : *
876 : * The comments for moveins() apply here as well. However, in current
877 : * usage, this is *only* called with brand-new target states, so that
878 : * only the "no need for de-duplication" code path is ever reached.
879 : * We keep the rest #ifdef'd out in case it's needed in the future.
880 : */
881 : static void
882 19366 : copyins(struct nfa *nfa,
883 : struct state *oldState,
884 : struct state *newState)
885 : {
886 : assert(oldState != newState);
887 : assert(newState->nins == 0); /* see comment above */
888 :
889 19366 : if (newState->nins == 0)
890 : {
891 : /* No need for de-duplication */
892 : struct arc *a;
893 :
894 261900 : for (a = oldState->ins; a != NULL; a = a->inchain)
895 242534 : createarc(nfa, a->type, a->co, a->from, newState);
896 : }
897 : #ifdef NOT_USED /* see comment above */
898 : else if (!BULK_ARC_OP_USE_SORT(oldState->nins, newState->nins))
899 : {
900 : /* With not too many arcs, just do them one at a time */
901 : struct arc *a;
902 :
903 : for (a = oldState->ins; a != NULL; a = a->inchain)
904 : cparc(nfa, a, a->from, newState);
905 : }
906 : else
907 : {
908 : /*
909 : * With many arcs, use a sort-merge approach. Note that createarc()
910 : * will put new arcs onto the front of newState's chain, so it does
911 : * not break our walk through the sorted part of the chain.
912 : */
913 : struct arc *oa;
914 : struct arc *na;
915 :
916 : /*
917 : * Because we bypass newarc() in this code path, we'd better include a
918 : * cancel check.
919 : */
920 : INTERRUPT(nfa->v->re);
921 :
922 : sortins(nfa, oldState);
923 : sortins(nfa, newState);
924 : if (NISERR())
925 : return; /* might have failed to sort */
926 : oa = oldState->ins;
927 : na = newState->ins;
928 : while (oa != NULL && na != NULL)
929 : {
930 : struct arc *a = oa;
931 :
932 : switch (sortins_cmp(&oa, &na))
933 : {
934 : case -1:
935 : /* newState does not have anything matching oa */
936 : oa = oa->inchain;
937 : createarc(nfa, a->type, a->co, a->from, newState);
938 : break;
939 : case 0:
940 : /* match, advance in both lists */
941 : oa = oa->inchain;
942 : na = na->inchain;
943 : break;
944 : case +1:
945 : /* advance only na; oa might have a match later */
946 : na = na->inchain;
947 : break;
948 : default:
949 : assert(NOTREACHED);
950 : }
951 : }
952 : while (oa != NULL)
953 : {
954 : /* newState does not have anything matching oa */
955 : struct arc *a = oa;
956 :
957 : oa = oa->inchain;
958 : createarc(nfa, a->type, a->co, a->from, newState);
959 : }
960 : }
961 : #endif /* NOT_USED */
962 19366 : }
963 :
964 : /*
965 : * mergeins - merge a list of inarcs into a state
966 : *
967 : * This is much like copyins, but the source arcs are listed in an array,
968 : * and are not guaranteed unique. It's okay to clobber the array contents.
969 : */
970 : static void
971 249848 : mergeins(struct nfa *nfa,
972 : struct state *s,
973 : struct arc **arcarray,
974 : int arccount)
975 : {
976 : struct arc *na;
977 : int i;
978 : int j;
979 :
980 249848 : if (arccount <= 0)
981 219418 : return;
982 :
983 : /*
984 : * Because we bypass newarc() in this code path, we'd better include a
985 : * cancel check.
986 : */
987 30430 : INTERRUPT(nfa->v->re);
988 :
989 : /* Sort existing inarcs as well as proposed new ones */
990 30430 : sortins(nfa, s);
991 30430 : if (NISERR())
992 0 : return; /* might have failed to sort */
993 :
994 30430 : qsort(arcarray, arccount, sizeof(struct arc *), sortins_cmp);
995 :
996 : /*
997 : * arcarray very likely includes dups, so we must eliminate them. (This
998 : * could be folded into the next loop, but it's not worth the trouble.)
999 : */
1000 30430 : j = 0;
1001 15030586 : for (i = 1; i < arccount; i++)
1002 : {
1003 15000156 : switch (sortins_cmp(&arcarray[j], &arcarray[i]))
1004 : {
1005 14998944 : case -1:
1006 : /* non-dup */
1007 14998944 : arcarray[++j] = arcarray[i];
1008 14998944 : break;
1009 1212 : case 0:
1010 : /* dup */
1011 1212 : break;
1012 15000156 : default:
1013 : /* trouble */
1014 : assert(NOTREACHED);
1015 : }
1016 : }
1017 30430 : arccount = j + 1;
1018 :
1019 : /*
1020 : * Now merge into s' inchain. Note that createarc() will put new arcs
1021 : * onto the front of s's chain, so it does not break our walk through the
1022 : * sorted part of the chain.
1023 : */
1024 30430 : i = 0;
1025 30430 : na = s->ins;
1026 15060202 : while (i < arccount && na != NULL)
1027 : {
1028 15029772 : struct arc *a = arcarray[i];
1029 :
1030 15029772 : switch (sortins_cmp(&a, &na))
1031 : {
1032 14986990 : case -1:
1033 : /* s does not have anything matching a */
1034 14986990 : createarc(nfa, a->type, a->co, a->from, s);
1035 14986990 : i++;
1036 14986990 : break;
1037 20 : case 0:
1038 : /* match, advance in both lists */
1039 20 : i++;
1040 20 : na = na->inchain;
1041 20 : break;
1042 42762 : case +1:
1043 : /* advance only na; array might have a match later */
1044 42762 : na = na->inchain;
1045 42762 : break;
1046 15029772 : default:
1047 : assert(NOTREACHED);
1048 : }
1049 : }
1050 72794 : while (i < arccount)
1051 : {
1052 : /* s does not have anything matching a */
1053 42364 : struct arc *a = arcarray[i];
1054 :
1055 42364 : createarc(nfa, a->type, a->co, a->from, s);
1056 42364 : i++;
1057 : }
1058 : }
1059 :
1060 : /*
1061 : * moveouts - move all out arcs of a state to another state
1062 : *
1063 : * See comments for moveins()
1064 : */
1065 : static void
1066 72174 : moveouts(struct nfa *nfa,
1067 : struct state *oldState,
1068 : struct state *newState)
1069 : {
1070 : assert(oldState != newState);
1071 :
1072 72174 : if (newState->nouts == 0)
1073 : {
1074 : /* No need for de-duplication */
1075 : struct arc *a;
1076 :
1077 59204 : while ((a = oldState->outs) != NULL)
1078 : {
1079 33396 : createarc(nfa, a->type, a->co, newState, a->to);
1080 33396 : freearc(nfa, a);
1081 : }
1082 : }
1083 46366 : else if (!BULK_ARC_OP_USE_SORT(oldState->nouts, newState->nouts))
1084 46350 : {
1085 : /* With not too many arcs, just do them one at a time */
1086 : struct arc *a;
1087 :
1088 113180 : while ((a = oldState->outs) != NULL)
1089 : {
1090 66830 : cparc(nfa, a, newState, a->to);
1091 66830 : freearc(nfa, a);
1092 : }
1093 : }
1094 : else
1095 : {
1096 : /*
1097 : * With many arcs, use a sort-merge approach. Note changearcsource()
1098 : * will put the arc onto the front of newState's chain, so it does not
1099 : * break our walk through the sorted part of the chain.
1100 : */
1101 : struct arc *oa;
1102 : struct arc *na;
1103 :
1104 : /*
1105 : * Because we bypass newarc() in this code path, we'd better include a
1106 : * cancel check.
1107 : */
1108 16 : INTERRUPT(nfa->v->re);
1109 :
1110 16 : sortouts(nfa, oldState);
1111 16 : sortouts(nfa, newState);
1112 16 : if (NISERR())
1113 0 : return; /* might have failed to sort */
1114 16 : oa = oldState->outs;
1115 16 : na = newState->outs;
1116 578 : while (oa != NULL && na != NULL)
1117 : {
1118 560 : struct arc *a = oa;
1119 :
1120 560 : switch (sortouts_cmp(&oa, &na))
1121 : {
1122 516 : case -1:
1123 : /* newState does not have anything matching oa */
1124 516 : oa = oa->outchain;
1125 :
1126 : /*
1127 : * Rather than doing createarc+freearc, we can just unlink
1128 : * and relink the existing arc struct.
1129 : */
1130 516 : changearcsource(a, newState);
1131 516 : break;
1132 28 : case 0:
1133 : /* match, advance in both lists */
1134 28 : oa = oa->outchain;
1135 28 : na = na->outchain;
1136 : /* ... and drop duplicate arc from oldState */
1137 28 : freearc(nfa, a);
1138 28 : break;
1139 16 : case +1:
1140 : /* advance only na; oa might have a match later */
1141 16 : na = na->outchain;
1142 16 : break;
1143 576 : default:
1144 : assert(NOTREACHED);
1145 : }
1146 : }
1147 88 : while (oa != NULL)
1148 : {
1149 : /* newState does not have anything matching oa */
1150 72 : struct arc *a = oa;
1151 :
1152 72 : oa = oa->outchain;
1153 72 : changearcsource(a, newState);
1154 : }
1155 : }
1156 :
1157 : assert(oldState->nouts == 0);
1158 : assert(oldState->outs == NULL);
1159 : }
1160 :
1161 : /*
1162 : * copyouts - copy out arcs of a state to another state
1163 : *
1164 : * See comments for copyins()
1165 : */
1166 : static void
1167 15282 : copyouts(struct nfa *nfa,
1168 : struct state *oldState,
1169 : struct state *newState)
1170 : {
1171 : assert(oldState != newState);
1172 : assert(newState->nouts == 0); /* see comment above */
1173 :
1174 15282 : if (newState->nouts == 0)
1175 : {
1176 : /* No need for de-duplication */
1177 : struct arc *a;
1178 :
1179 492644 : for (a = oldState->outs; a != NULL; a = a->outchain)
1180 477362 : createarc(nfa, a->type, a->co, newState, a->to);
1181 : }
1182 : #ifdef NOT_USED /* see comment above */
1183 : else if (!BULK_ARC_OP_USE_SORT(oldState->nouts, newState->nouts))
1184 : {
1185 : /* With not too many arcs, just do them one at a time */
1186 : struct arc *a;
1187 :
1188 : for (a = oldState->outs; a != NULL; a = a->outchain)
1189 : cparc(nfa, a, newState, a->to);
1190 : }
1191 : else
1192 : {
1193 : /*
1194 : * With many arcs, use a sort-merge approach. Note that createarc()
1195 : * will put new arcs onto the front of newState's chain, so it does
1196 : * not break our walk through the sorted part of the chain.
1197 : */
1198 : struct arc *oa;
1199 : struct arc *na;
1200 :
1201 : /*
1202 : * Because we bypass newarc() in this code path, we'd better include a
1203 : * cancel check.
1204 : */
1205 : INTERRUPT(nfa->v->re);
1206 :
1207 : sortouts(nfa, oldState);
1208 : sortouts(nfa, newState);
1209 : if (NISERR())
1210 : return; /* might have failed to sort */
1211 : oa = oldState->outs;
1212 : na = newState->outs;
1213 : while (oa != NULL && na != NULL)
1214 : {
1215 : struct arc *a = oa;
1216 :
1217 : switch (sortouts_cmp(&oa, &na))
1218 : {
1219 : case -1:
1220 : /* newState does not have anything matching oa */
1221 : oa = oa->outchain;
1222 : createarc(nfa, a->type, a->co, newState, a->to);
1223 : break;
1224 : case 0:
1225 : /* match, advance in both lists */
1226 : oa = oa->outchain;
1227 : na = na->outchain;
1228 : break;
1229 : case +1:
1230 : /* advance only na; oa might have a match later */
1231 : na = na->outchain;
1232 : break;
1233 : default:
1234 : assert(NOTREACHED);
1235 : }
1236 : }
1237 : while (oa != NULL)
1238 : {
1239 : /* newState does not have anything matching oa */
1240 : struct arc *a = oa;
1241 :
1242 : oa = oa->outchain;
1243 : createarc(nfa, a->type, a->co, newState, a->to);
1244 : }
1245 : }
1246 : #endif /* NOT_USED */
1247 15282 : }
1248 :
1249 : /*
1250 : * cloneouts - copy out arcs of a state to another state pair, modifying type
1251 : *
1252 : * This is only used to convert PLAIN arcs to AHEAD/BEHIND arcs, which share
1253 : * the same interpretation of "co". It wouldn't be sensible with LACONs.
1254 : */
1255 : static void
1256 342 : cloneouts(struct nfa *nfa,
1257 : struct state *old,
1258 : struct state *from,
1259 : struct state *to,
1260 : int type)
1261 : {
1262 : struct arc *a;
1263 :
1264 : assert(old != from);
1265 : assert(type == AHEAD || type == BEHIND);
1266 :
1267 1190 : for (a = old->outs; a != NULL; a = a->outchain)
1268 : {
1269 : assert(a->type == PLAIN);
1270 848 : newarc(nfa, type, a->co, from, to);
1271 : }
1272 342 : }
1273 :
1274 : /*
1275 : * delsub - delete a sub-NFA, updating subre pointers if necessary
1276 : *
1277 : * This uses a recursive traversal of the sub-NFA, marking already-seen
1278 : * states using their tmp pointer.
1279 : */
1280 : static void
1281 8804 : delsub(struct nfa *nfa,
1282 : struct state *lp, /* the sub-NFA goes from here... */
1283 : struct state *rp) /* ...to here, *not* inclusive */
1284 : {
1285 : assert(lp != rp);
1286 :
1287 8804 : rp->tmp = rp; /* mark end */
1288 :
1289 8804 : deltraverse(nfa, lp, lp);
1290 8804 : if (NISERR())
1291 0 : return; /* asserts might not hold after failure */
1292 : assert(lp->nouts == 0 && rp->nins == 0); /* did the job */
1293 : assert(lp->no != FREESTATE && rp->no != FREESTATE); /* no more */
1294 :
1295 8804 : rp->tmp = NULL; /* unmark end */
1296 8804 : lp->tmp = NULL; /* and begin, marked by deltraverse */
1297 : }
1298 :
1299 : /*
1300 : * deltraverse - the recursive heart of delsub
1301 : * This routine's basic job is to destroy all out-arcs of the state.
1302 : */
1303 : static void
1304 25766 : deltraverse(struct nfa *nfa,
1305 : struct state *leftend,
1306 : struct state *s)
1307 : {
1308 : struct arc *a;
1309 : struct state *to;
1310 :
1311 : /* Since this is recursive, it could be driven to stack overflow */
1312 25766 : if (STACK_TOO_DEEP(nfa->v->re))
1313 : {
1314 0 : NERR(REG_ETOOBIG);
1315 0 : return;
1316 : }
1317 :
1318 25766 : if (s->nouts == 0)
1319 202 : return; /* nothing to do */
1320 25564 : if (s->tmp != NULL)
1321 8632 : return; /* already in progress */
1322 :
1323 16932 : s->tmp = s; /* mark as in progress */
1324 :
1325 33894 : while ((a = s->outs) != NULL)
1326 : {
1327 16962 : to = a->to;
1328 16962 : deltraverse(nfa, leftend, to);
1329 16962 : if (NISERR())
1330 0 : return; /* asserts might not hold after failure */
1331 : assert(to->nouts == 0 || to->tmp != NULL);
1332 16962 : freearc(nfa, a);
1333 16962 : if (to->nins == 0 && to->tmp == NULL)
1334 : {
1335 : assert(to->nouts == 0);
1336 8128 : freestate(nfa, to);
1337 : }
1338 : }
1339 :
1340 : assert(s->no != FREESTATE); /* we're still here */
1341 : assert(s == leftend || s->nins != 0); /* and still reachable */
1342 : assert(s->nouts == 0); /* but have no outarcs */
1343 :
1344 16932 : s->tmp = NULL; /* we're done here */
1345 : }
1346 :
1347 : /*
1348 : * dupnfa - duplicate sub-NFA
1349 : *
1350 : * Another recursive traversal, this time using tmp to point to duplicates
1351 : * as well as mark already-seen states. (You knew there was a reason why
1352 : * it's a state pointer, didn't you? :-))
1353 : */
1354 : static void
1355 14310 : dupnfa(struct nfa *nfa,
1356 : struct state *start, /* duplicate of subNFA starting here */
1357 : struct state *stop, /* and stopping here */
1358 : struct state *from, /* stringing duplicate from here */
1359 : struct state *to) /* to here */
1360 : {
1361 14310 : if (start == stop)
1362 : {
1363 0 : newarc(nfa, EMPTY, 0, from, to);
1364 0 : return;
1365 : }
1366 :
1367 14310 : stop->tmp = to;
1368 14310 : duptraverse(nfa, start, from);
1369 : /* done, except for clearing out the tmp pointers */
1370 :
1371 14310 : stop->tmp = NULL;
1372 14310 : cleartraverse(nfa, start);
1373 : }
1374 :
1375 : /*
1376 : * duptraverse - recursive heart of dupnfa
1377 : */
1378 : static void
1379 372316 : duptraverse(struct nfa *nfa,
1380 : struct state *s,
1381 : struct state *stmp) /* s's duplicate, or NULL */
1382 : {
1383 : struct arc *a;
1384 :
1385 : /* Since this is recursive, it could be driven to stack overflow */
1386 372316 : if (STACK_TOO_DEEP(nfa->v->re))
1387 : {
1388 0 : NERR(REG_ETOOBIG);
1389 0 : return;
1390 : }
1391 :
1392 372316 : if (s->tmp != NULL)
1393 121174 : return; /* already done */
1394 :
1395 251142 : s->tmp = (stmp == NULL) ? newstate(nfa) : stmp;
1396 251142 : if (s->tmp == NULL)
1397 : {
1398 : assert(NISERR());
1399 0 : return;
1400 : }
1401 :
1402 609148 : for (a = s->outs; a != NULL && !NISERR(); a = a->outchain)
1403 : {
1404 358006 : duptraverse(nfa, a->to, (struct state *) NULL);
1405 358006 : if (NISERR())
1406 0 : break;
1407 : assert(a->to->tmp != NULL);
1408 358006 : cparc(nfa, a, s->tmp, a->to->tmp);
1409 : }
1410 : }
1411 :
1412 : /*
1413 : * removeconstraints - remove any constraints in an NFA
1414 : *
1415 : * Constraint arcs are replaced by empty arcs, essentially treating all
1416 : * constraints as automatically satisfied.
1417 : */
1418 : static void
1419 202 : removeconstraints(struct nfa *nfa,
1420 : struct state *start, /* process subNFA starting here */
1421 : struct state *stop) /* and stopping here */
1422 : {
1423 202 : if (start == stop)
1424 0 : return;
1425 :
1426 202 : stop->tmp = stop;
1427 202 : removetraverse(nfa, start);
1428 : /* done, except for clearing out the tmp pointers */
1429 :
1430 202 : stop->tmp = NULL;
1431 202 : cleartraverse(nfa, start);
1432 : }
1433 :
1434 : /*
1435 : * removetraverse - recursive heart of removeconstraints
1436 : */
1437 : static void
1438 562 : removetraverse(struct nfa *nfa,
1439 : struct state *s)
1440 : {
1441 : struct arc *a;
1442 : struct arc *oa;
1443 :
1444 : /* Since this is recursive, it could be driven to stack overflow */
1445 562 : if (STACK_TOO_DEEP(nfa->v->re))
1446 : {
1447 0 : NERR(REG_ETOOBIG);
1448 0 : return;
1449 : }
1450 :
1451 562 : if (s->tmp != NULL)
1452 252 : return; /* already done */
1453 :
1454 310 : s->tmp = s;
1455 670 : for (a = s->outs; a != NULL && !NISERR(); a = oa)
1456 : {
1457 360 : removetraverse(nfa, a->to);
1458 360 : if (NISERR())
1459 0 : break;
1460 360 : oa = a->outchain;
1461 360 : switch (a->type)
1462 : {
1463 346 : case PLAIN:
1464 : case EMPTY:
1465 : /* nothing to do */
1466 346 : break;
1467 14 : case AHEAD:
1468 : case BEHIND:
1469 : case '^':
1470 : case '$':
1471 : case LACON:
1472 : /* replace it */
1473 14 : newarc(nfa, EMPTY, 0, s, a->to);
1474 14 : freearc(nfa, a);
1475 14 : break;
1476 0 : default:
1477 0 : NERR(REG_ASSERT);
1478 0 : break;
1479 : }
1480 : }
1481 : }
1482 :
1483 : /*
1484 : * cleartraverse - recursive cleanup for algorithms that leave tmp ptrs set
1485 : */
1486 : static void
1487 2116702 : cleartraverse(struct nfa *nfa,
1488 : struct state *s)
1489 : {
1490 : struct arc *a;
1491 :
1492 : /* Since this is recursive, it could be driven to stack overflow */
1493 2116702 : if (STACK_TOO_DEEP(nfa->v->re))
1494 : {
1495 0 : NERR(REG_ETOOBIG);
1496 0 : return;
1497 : }
1498 :
1499 2116702 : if (s->tmp == NULL)
1500 1248810 : return;
1501 867892 : s->tmp = NULL;
1502 :
1503 2932872 : for (a = s->outs; a != NULL; a = a->outchain)
1504 2064980 : cleartraverse(nfa, a->to);
1505 : }
1506 :
1507 : /*
1508 : * single_color_transition - does getting from s1 to s2 cross one PLAIN arc?
1509 : *
1510 : * If traversing from s1 to s2 requires a single PLAIN match (possibly of any
1511 : * of a set of colors), return a state whose outarc list contains only PLAIN
1512 : * arcs of those color(s). Otherwise return NULL.
1513 : *
1514 : * This is used before optimizing the NFA, so there may be EMPTY arcs, which
1515 : * we should ignore; the possibility of an EMPTY is why the result state could
1516 : * be different from s1.
1517 : *
1518 : * It's worth troubling to handle multiple parallel PLAIN arcs here because a
1519 : * bracket construct such as [abc] might yield either one or several parallel
1520 : * PLAIN arcs depending on earlier atoms in the expression. We'd rather that
1521 : * that implementation detail not create user-visible performance differences.
1522 : */
1523 : static struct state *
1524 254 : single_color_transition(struct state *s1, struct state *s2)
1525 : {
1526 : struct arc *a;
1527 :
1528 : /* Ignore leading EMPTY arc, if any */
1529 254 : if (s1->nouts == 1 && s1->outs->type == EMPTY)
1530 254 : s1 = s1->outs->to;
1531 : /* Likewise for any trailing EMPTY arc */
1532 254 : if (s2->nins == 1 && s2->ins->type == EMPTY)
1533 254 : s2 = s2->ins->from;
1534 : /* Perhaps we could have a single-state loop in between, if so reject */
1535 254 : if (s1 == s2)
1536 0 : return NULL;
1537 : /* s1 must have at least one outarc... */
1538 254 : if (s1->outs == NULL)
1539 0 : return NULL;
1540 : /* ... and they must all be PLAIN arcs to s2 */
1541 436 : for (a = s1->outs; a != NULL; a = a->outchain)
1542 : {
1543 268 : if (a->type != PLAIN || a->to != s2)
1544 86 : return NULL;
1545 : }
1546 : /* OK, return s1 as the possessor of the relevant outarcs */
1547 168 : return s1;
1548 : }
1549 :
1550 : /*
1551 : * specialcolors - fill in special colors for an NFA
1552 : */
1553 : static void
1554 18614 : specialcolors(struct nfa *nfa)
1555 : {
1556 : /* false colors for BOS, BOL, EOS, EOL */
1557 18614 : if (nfa->parent == NULL)
1558 : {
1559 7416 : nfa->bos[0] = pseudocolor(nfa->cm);
1560 7416 : nfa->bos[1] = pseudocolor(nfa->cm);
1561 7416 : nfa->eos[0] = pseudocolor(nfa->cm);
1562 7416 : nfa->eos[1] = pseudocolor(nfa->cm);
1563 : }
1564 : else
1565 : {
1566 : assert(nfa->parent->bos[0] != COLORLESS);
1567 11198 : nfa->bos[0] = nfa->parent->bos[0];
1568 : assert(nfa->parent->bos[1] != COLORLESS);
1569 11198 : nfa->bos[1] = nfa->parent->bos[1];
1570 : assert(nfa->parent->eos[0] != COLORLESS);
1571 11198 : nfa->eos[0] = nfa->parent->eos[0];
1572 : assert(nfa->parent->eos[1] != COLORLESS);
1573 11198 : nfa->eos[1] = nfa->parent->eos[1];
1574 : }
1575 18614 : }
1576 :
1577 : /*
1578 : * optimize - optimize an NFA
1579 : *
1580 : * The main goal of this function is not so much "optimization" (though it
1581 : * does try to get rid of useless NFA states) as reducing the NFA to a form
1582 : * the regex executor can handle. The executor, and indeed the cNFA format
1583 : * that is its input, can only handle PLAIN and LACON arcs. The output of
1584 : * the regex parser also includes EMPTY (do-nothing) arcs, as well as
1585 : * ^, $, AHEAD, and BEHIND constraint arcs, which we must get rid of here.
1586 : * We first get rid of EMPTY arcs and then deal with the constraint arcs.
1587 : * The hardest part of either job is to get rid of circular loops of the
1588 : * target arc type. We would have to do that in any case, though, as such a
1589 : * loop would otherwise allow the executor to cycle through the loop endlessly
1590 : * without making any progress in the input string.
1591 : */
1592 : static long /* re_info bits */
1593 18608 : optimize(struct nfa *nfa,
1594 : FILE *f) /* for debug output; NULL none */
1595 : {
1596 : #ifdef REG_DEBUG
1597 : int verbose = (f != NULL) ? 1 : 0;
1598 :
1599 : if (verbose)
1600 : fprintf(f, "\ninitial cleanup:\n");
1601 : #endif
1602 18608 : cleanup(nfa); /* may simplify situation */
1603 : #ifdef REG_DEBUG
1604 : if (verbose)
1605 : dumpnfa(nfa, f);
1606 : if (verbose)
1607 : fprintf(f, "\nempties:\n");
1608 : #endif
1609 18608 : fixempties(nfa, f); /* get rid of EMPTY arcs */
1610 : #ifdef REG_DEBUG
1611 : if (verbose)
1612 : fprintf(f, "\nconstraints:\n");
1613 : #endif
1614 18608 : fixconstraintloops(nfa, f); /* get rid of constraint loops */
1615 18608 : pullback(nfa, f); /* pull back constraints backward */
1616 18608 : pushfwd(nfa, f); /* push fwd constraints forward */
1617 : #ifdef REG_DEBUG
1618 : if (verbose)
1619 : fprintf(f, "\nfinal cleanup:\n");
1620 : #endif
1621 18608 : cleanup(nfa); /* final tidying */
1622 : #ifdef REG_DEBUG
1623 : if (verbose)
1624 : dumpnfa(nfa, f);
1625 : #endif
1626 18608 : return analyze(nfa); /* and analysis */
1627 : }
1628 :
1629 : /*
1630 : * pullback - pull back constraints backward to eliminate them
1631 : */
1632 : static void
1633 29512 : pullback(struct nfa *nfa,
1634 : FILE *f) /* for debug output; NULL none */
1635 : {
1636 : struct state *s;
1637 : struct state *nexts;
1638 : struct arc *a;
1639 : struct arc *nexta;
1640 : struct state *intermediates;
1641 : int progress;
1642 :
1643 : /* find and pull until there are no more */
1644 : do
1645 : {
1646 29512 : progress = 0;
1647 430054 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
1648 : {
1649 400542 : nexts = s->next;
1650 400542 : intermediates = NULL;
1651 1844602 : for (a = s->outs; a != NULL && !NISERR(); a = nexta)
1652 : {
1653 1444060 : nexta = a->outchain;
1654 1444060 : if (a->type == '^' || a->type == BEHIND)
1655 98370 : if (pull(nfa, a, &intermediates))
1656 38650 : progress = 1;
1657 : }
1658 : /* clear tmp fields of intermediate states created here */
1659 402994 : while (intermediates != NULL)
1660 : {
1661 2452 : struct state *ns = intermediates->tmp;
1662 :
1663 2452 : intermediates->tmp = NULL;
1664 2452 : intermediates = ns;
1665 : }
1666 : /* if s is now useless, get rid of it */
1667 400542 : if ((s->nins == 0 || s->nouts == 0) && !s->flag)
1668 33434 : dropstate(nfa, s);
1669 : }
1670 29512 : if (progress && f != NULL)
1671 0 : dumpnfa(nfa, f);
1672 29512 : } while (progress && !NISERR());
1673 18608 : if (NISERR())
1674 6 : return;
1675 :
1676 : /*
1677 : * Any ^ constraints we were able to pull to the start state can now be
1678 : * replaced by PLAIN arcs referencing the BOS or BOL colors. There should
1679 : * be no other ^ or BEHIND arcs left in the NFA, though we do not check
1680 : * that here (compact() will fail if so).
1681 : */
1682 71230 : for (a = nfa->pre->outs; a != NULL; a = nexta)
1683 : {
1684 52628 : nexta = a->outchain;
1685 52628 : if (a->type == '^')
1686 : {
1687 : assert(a->co == 0 || a->co == 1);
1688 37288 : newarc(nfa, PLAIN, nfa->bos[a->co], a->from, a->to);
1689 37288 : freearc(nfa, a);
1690 : }
1691 : }
1692 : }
1693 :
1694 : /*
1695 : * pull - pull a back constraint backward past its source state
1696 : *
1697 : * Returns 1 if successful (which it always is unless the source is the
1698 : * start state or we have an internal error), 0 if nothing happened.
1699 : *
1700 : * A significant property of this function is that it deletes no pre-existing
1701 : * states, and no outarcs of the constraint's from state other than the given
1702 : * constraint arc. This makes the loops in pullback() safe, at the cost that
1703 : * we may leave useless states behind. Therefore, we leave it to pullback()
1704 : * to delete such states.
1705 : *
1706 : * If the from state has multiple back-constraint outarcs, and/or multiple
1707 : * compatible constraint inarcs, we only need to create one new intermediate
1708 : * state per combination of predecessor and successor states. *intermediates
1709 : * points to a list of such intermediate states for this from state (chained
1710 : * through their tmp fields).
1711 : */
1712 : static int
1713 98370 : pull(struct nfa *nfa,
1714 : struct arc *con,
1715 : struct state **intermediates)
1716 : {
1717 98370 : struct state *from = con->from;
1718 98370 : struct state *to = con->to;
1719 : struct arc *a;
1720 : struct arc *nexta;
1721 : struct state *s;
1722 :
1723 : assert(from != to); /* should have gotten rid of this earlier */
1724 98370 : if (from->flag) /* can't pull back beyond start */
1725 59720 : return 0;
1726 38650 : if (from->nins == 0)
1727 : { /* unreachable */
1728 7304 : freearc(nfa, con);
1729 7304 : return 1;
1730 : }
1731 :
1732 : /*
1733 : * First, clone from state if necessary to avoid other outarcs. This may
1734 : * seem wasteful, but it simplifies the logic, and we'll get rid of the
1735 : * clone state again at the bottom.
1736 : */
1737 31346 : if (from->nouts > 1)
1738 : {
1739 19366 : s = newstate(nfa);
1740 19366 : if (NISERR())
1741 0 : return 0;
1742 19366 : copyins(nfa, from, s); /* duplicate inarcs */
1743 19366 : cparc(nfa, con, s, to); /* move constraint arc */
1744 19366 : freearc(nfa, con);
1745 19366 : if (NISERR())
1746 0 : return 0;
1747 19366 : from = s;
1748 19366 : con = from->outs;
1749 : }
1750 : assert(from->nouts == 1);
1751 :
1752 : /* propagate the constraint into the from state's inarcs */
1753 317852 : for (a = from->ins; a != NULL && !NISERR(); a = nexta)
1754 : {
1755 286506 : nexta = a->inchain;
1756 286506 : switch (combine(nfa, con, a))
1757 : {
1758 87780 : case INCOMPATIBLE: /* destroy the arc */
1759 87780 : freearc(nfa, a);
1760 87780 : break;
1761 16536 : case SATISFIED: /* no action needed */
1762 16536 : break;
1763 179686 : case COMPATIBLE: /* swap the two arcs, more or less */
1764 : /* need an intermediate state, but might have one already */
1765 203272 : for (s = *intermediates; s != NULL; s = s->tmp)
1766 : {
1767 : assert(s->nins > 0 && s->nouts > 0);
1768 200820 : if (s->ins->from == a->from && s->outs->to == to)
1769 177234 : break;
1770 : }
1771 179686 : if (s == NULL)
1772 : {
1773 2452 : s = newstate(nfa);
1774 2452 : if (NISERR())
1775 0 : return 0;
1776 2452 : s->tmp = *intermediates;
1777 2452 : *intermediates = s;
1778 : }
1779 179686 : cparc(nfa, con, a->from, s);
1780 179686 : cparc(nfa, a, s, to);
1781 179686 : freearc(nfa, a);
1782 179686 : break;
1783 2504 : case REPLACEARC: /* replace arc's color */
1784 2504 : newarc(nfa, a->type, con->co, a->from, to);
1785 2504 : freearc(nfa, a);
1786 2504 : break;
1787 0 : default:
1788 : assert(NOTREACHED);
1789 0 : break;
1790 : }
1791 : }
1792 :
1793 : /* remaining inarcs, if any, incorporate the constraint */
1794 31346 : moveins(nfa, from, to);
1795 31346 : freearc(nfa, con);
1796 : /* from state is now useless, but we leave it to pullback() to clean up */
1797 31346 : return 1;
1798 : }
1799 :
1800 : /*
1801 : * pushfwd - push forward constraints forward to eliminate them
1802 : */
1803 : static void
1804 27758 : pushfwd(struct nfa *nfa,
1805 : FILE *f) /* for debug output; NULL none */
1806 : {
1807 : struct state *s;
1808 : struct state *nexts;
1809 : struct arc *a;
1810 : struct arc *nexta;
1811 : struct state *intermediates;
1812 : int progress;
1813 :
1814 : /* find and push until there are no more */
1815 : do
1816 : {
1817 27758 : progress = 0;
1818 374188 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
1819 : {
1820 346430 : nexts = s->next;
1821 346430 : intermediates = NULL;
1822 1641996 : for (a = s->ins; a != NULL && !NISERR(); a = nexta)
1823 : {
1824 1295566 : nexta = a->inchain;
1825 1295566 : if (a->type == '$' || a->type == AHEAD)
1826 69926 : if (push(nfa, a, &intermediates))
1827 22286 : progress = 1;
1828 : }
1829 : /* clear tmp fields of intermediate states created here */
1830 346434 : while (intermediates != NULL)
1831 : {
1832 4 : struct state *ns = intermediates->tmp;
1833 :
1834 4 : intermediates->tmp = NULL;
1835 4 : intermediates = ns;
1836 : }
1837 : /* if s is now useless, get rid of it */
1838 346430 : if ((s->nins == 0 || s->nouts == 0) && !s->flag)
1839 22846 : dropstate(nfa, s);
1840 : }
1841 27758 : if (progress && f != NULL)
1842 0 : dumpnfa(nfa, f);
1843 27758 : } while (progress && !NISERR());
1844 18608 : if (NISERR())
1845 6 : return;
1846 :
1847 : /*
1848 : * Any $ constraints we were able to push to the post state can now be
1849 : * replaced by PLAIN arcs referencing the EOS or EOL colors. There should
1850 : * be no other $ or AHEAD arcs left in the NFA, though we do not check
1851 : * that here (compact() will fail if so).
1852 : */
1853 59514 : for (a = nfa->post->ins; a != NULL; a = nexta)
1854 : {
1855 40912 : nexta = a->inchain;
1856 40912 : if (a->type == '$')
1857 : {
1858 : assert(a->co == 0 || a->co == 1);
1859 28564 : newarc(nfa, PLAIN, nfa->eos[a->co], a->from, a->to);
1860 28564 : freearc(nfa, a);
1861 : }
1862 : }
1863 : }
1864 :
1865 : /*
1866 : * push - push a forward constraint forward past its destination state
1867 : *
1868 : * Returns 1 if successful (which it always is unless the destination is the
1869 : * post state or we have an internal error), 0 if nothing happened.
1870 : *
1871 : * A significant property of this function is that it deletes no pre-existing
1872 : * states, and no inarcs of the constraint's to state other than the given
1873 : * constraint arc. This makes the loops in pushfwd() safe, at the cost that
1874 : * we may leave useless states behind. Therefore, we leave it to pushfwd()
1875 : * to delete such states.
1876 : *
1877 : * If the to state has multiple forward-constraint inarcs, and/or multiple
1878 : * compatible constraint outarcs, we only need to create one new intermediate
1879 : * state per combination of predecessor and successor states. *intermediates
1880 : * points to a list of such intermediate states for this to state (chained
1881 : * through their tmp fields).
1882 : */
1883 : static int
1884 69926 : push(struct nfa *nfa,
1885 : struct arc *con,
1886 : struct state **intermediates)
1887 : {
1888 69926 : struct state *from = con->from;
1889 69926 : struct state *to = con->to;
1890 : struct arc *a;
1891 : struct arc *nexta;
1892 : struct state *s;
1893 :
1894 : assert(to != from); /* should have gotten rid of this earlier */
1895 69926 : if (to->flag) /* can't push forward beyond end */
1896 47640 : return 0;
1897 22286 : if (to->nouts == 0)
1898 : { /* dead end */
1899 820 : freearc(nfa, con);
1900 820 : return 1;
1901 : }
1902 :
1903 : /*
1904 : * First, clone to state if necessary to avoid other inarcs. This may
1905 : * seem wasteful, but it simplifies the logic, and we'll get rid of the
1906 : * clone state again at the bottom.
1907 : */
1908 21466 : if (to->nins > 1)
1909 : {
1910 11834 : s = newstate(nfa);
1911 11834 : if (NISERR())
1912 0 : return 0;
1913 11834 : copyouts(nfa, to, s); /* duplicate outarcs */
1914 11834 : cparc(nfa, con, from, s); /* move constraint arc */
1915 11834 : freearc(nfa, con);
1916 11834 : if (NISERR())
1917 0 : return 0;
1918 11834 : to = s;
1919 11834 : con = to->ins;
1920 : }
1921 : assert(to->nins == 1);
1922 :
1923 : /* propagate the constraint into the to state's outarcs */
1924 130344 : for (a = to->outs; a != NULL && !NISERR(); a = nexta)
1925 : {
1926 108878 : nexta = a->outchain;
1927 108878 : switch (combine(nfa, con, a))
1928 : {
1929 91592 : case INCOMPATIBLE: /* destroy the arc */
1930 91592 : freearc(nfa, a);
1931 91592 : break;
1932 13928 : case SATISFIED: /* no action needed */
1933 13928 : break;
1934 16 : case COMPATIBLE: /* swap the two arcs, more or less */
1935 : /* need an intermediate state, but might have one already */
1936 16 : for (s = *intermediates; s != NULL; s = s->tmp)
1937 : {
1938 : assert(s->nins > 0 && s->nouts > 0);
1939 12 : if (s->ins->from == from && s->outs->to == a->to)
1940 12 : break;
1941 : }
1942 16 : if (s == NULL)
1943 : {
1944 4 : s = newstate(nfa);
1945 4 : if (NISERR())
1946 0 : return 0;
1947 4 : s->tmp = *intermediates;
1948 4 : *intermediates = s;
1949 : }
1950 16 : cparc(nfa, con, s, a->to);
1951 16 : cparc(nfa, a, from, s);
1952 16 : freearc(nfa, a);
1953 16 : break;
1954 3342 : case REPLACEARC: /* replace arc's color */
1955 3342 : newarc(nfa, a->type, con->co, from, a->to);
1956 3342 : freearc(nfa, a);
1957 3342 : break;
1958 0 : default:
1959 : assert(NOTREACHED);
1960 0 : break;
1961 : }
1962 : }
1963 :
1964 : /* remaining outarcs, if any, incorporate the constraint */
1965 21466 : moveouts(nfa, to, from);
1966 21466 : freearc(nfa, con);
1967 : /* to state is now useless, but we leave it to pushfwd() to clean up */
1968 21466 : return 1;
1969 : }
1970 :
1971 : /*
1972 : * combine - constraint lands on an arc, what happens?
1973 : *
1974 : * #def INCOMPATIBLE 1 // destroys arc
1975 : * #def SATISFIED 2 // constraint satisfied
1976 : * #def COMPATIBLE 3 // compatible but not satisfied yet
1977 : * #def REPLACEARC 4 // replace arc's color with constraint color
1978 : */
1979 : static int
1980 395384 : combine(struct nfa *nfa,
1981 : struct arc *con,
1982 : struct arc *a)
1983 : {
1984 : #define CA(ct,at) (((ct)<<CHAR_BIT) | (at))
1985 :
1986 395384 : switch (CA(con->type, a->type))
1987 : {
1988 88652 : case CA('^', PLAIN): /* newlines are handled separately */
1989 : case CA('$', PLAIN):
1990 88652 : return INCOMPATIBLE;
1991 : break;
1992 37312 : case CA(AHEAD, PLAIN): /* color constraints meet colors */
1993 : case CA(BEHIND, PLAIN):
1994 37312 : if (con->co == a->co)
1995 1562 : return SATISFIED;
1996 35750 : if (con->co == RAINBOW)
1997 : {
1998 : /* con is satisfied unless arc's color is a pseudocolor */
1999 4 : if (!(nfa->cm->cd[a->co].flags & PSEUDO))
2000 4 : return SATISFIED;
2001 : }
2002 35746 : else if (a->co == RAINBOW)
2003 : {
2004 : /* con is incompatible if it's for a pseudocolor */
2005 : /* (this is hypothetical; we make no such constraints today) */
2006 5838 : if (nfa->cm->cd[con->co].flags & PSEUDO)
2007 0 : return INCOMPATIBLE;
2008 : /* otherwise, constraint constrains arc to be only its color */
2009 5838 : return REPLACEARC;
2010 : }
2011 29908 : return INCOMPATIBLE;
2012 : break;
2013 49142 : case CA('^', '^'): /* collision, similar constraints */
2014 : case CA('$', '$'):
2015 49142 : if (con->co == a->co) /* true duplication */
2016 27976 : return SATISFIED;
2017 21166 : return INCOMPATIBLE;
2018 : break;
2019 24620 : case CA(AHEAD, AHEAD): /* collision, similar constraints */
2020 : case CA(BEHIND, BEHIND):
2021 24620 : if (con->co == a->co) /* true duplication */
2022 918 : return SATISFIED;
2023 23702 : if (con->co == RAINBOW)
2024 : {
2025 : /* con is satisfied unless arc's color is a pseudocolor */
2026 4 : if (!(nfa->cm->cd[a->co].flags & PSEUDO))
2027 4 : return SATISFIED;
2028 : }
2029 23698 : else if (a->co == RAINBOW)
2030 : {
2031 : /* con is incompatible if it's for a pseudocolor */
2032 : /* (this is hypothetical; we make no such constraints today) */
2033 8 : if (nfa->cm->cd[con->co].flags & PSEUDO)
2034 0 : return INCOMPATIBLE;
2035 : /* otherwise, constraint constrains arc to be only its color */
2036 8 : return REPLACEARC;
2037 : }
2038 23690 : return INCOMPATIBLE;
2039 : break;
2040 15956 : case CA('^', BEHIND): /* collision, dissimilar constraints */
2041 : case CA(BEHIND, '^'):
2042 : case CA('$', AHEAD):
2043 : case CA(AHEAD, '$'):
2044 15956 : return INCOMPATIBLE;
2045 : break;
2046 179702 : case CA('^', '$'): /* constraints passing each other */
2047 : case CA('^', AHEAD):
2048 : case CA(BEHIND, '$'):
2049 : case CA(BEHIND, AHEAD):
2050 : case CA('$', '^'):
2051 : case CA('$', BEHIND):
2052 : case CA(AHEAD, '^'):
2053 : case CA(AHEAD, BEHIND):
2054 : case CA('^', LACON):
2055 : case CA(BEHIND, LACON):
2056 : case CA('$', LACON):
2057 : case CA(AHEAD, LACON):
2058 179702 : return COMPATIBLE;
2059 : break;
2060 : }
2061 : assert(NOTREACHED);
2062 0 : return INCOMPATIBLE; /* for benefit of blind compilers */
2063 : }
2064 :
2065 : /*
2066 : * fixempties - get rid of EMPTY arcs
2067 : */
2068 : static void
2069 18608 : fixempties(struct nfa *nfa,
2070 : FILE *f) /* for debug output; NULL none */
2071 : {
2072 : struct state *s;
2073 : struct state *s2;
2074 : struct state *nexts;
2075 : struct arc *a;
2076 : struct arc *nexta;
2077 : int totalinarcs;
2078 : struct arc **inarcsorig;
2079 : struct arc **arcarray;
2080 : int arccount;
2081 : int prevnins;
2082 : int nskip;
2083 :
2084 : /*
2085 : * First, get rid of any states whose sole out-arc is an EMPTY, since
2086 : * they're basically just aliases for their successor. The parsing
2087 : * algorithm creates enough of these that it's worth special-casing this.
2088 : */
2089 423684 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
2090 : {
2091 405076 : nexts = s->next;
2092 405076 : if (s->flag || s->nouts != 1)
2093 111530 : continue;
2094 293546 : a = s->outs;
2095 : assert(a != NULL && a->outchain == NULL);
2096 293546 : if (a->type != EMPTY)
2097 172128 : continue;
2098 121418 : if (s != a->to)
2099 121418 : moveins(nfa, s, a->to);
2100 121418 : dropstate(nfa, s);
2101 : }
2102 :
2103 : /*
2104 : * Similarly, get rid of any state with a single EMPTY in-arc, by folding
2105 : * it into its predecessor.
2106 : */
2107 302266 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
2108 : {
2109 283658 : nexts = s->next;
2110 : /* while we're at it, ensure tmp fields are clear for next step */
2111 : assert(s->tmp == NULL);
2112 283658 : if (s->flag || s->nins != 1)
2113 105930 : continue;
2114 177728 : a = s->ins;
2115 : assert(a != NULL && a->inchain == NULL);
2116 177728 : if (a->type != EMPTY)
2117 152828 : continue;
2118 24900 : if (s != a->from)
2119 24900 : moveouts(nfa, s, a->from);
2120 24900 : dropstate(nfa, s);
2121 : }
2122 :
2123 18608 : if (NISERR())
2124 0 : return;
2125 :
2126 : /*
2127 : * For each remaining NFA state, find all other states from which it is
2128 : * reachable by a chain of one or more EMPTY arcs. Then generate new arcs
2129 : * that eliminate the need for each such chain.
2130 : *
2131 : * We could replace a chain of EMPTY arcs that leads from a "from" state
2132 : * to a "to" state either by pushing non-EMPTY arcs forward (linking
2133 : * directly from "from"'s predecessors to "to") or by pulling them back
2134 : * (linking directly from "from" to "to"'s successors). We choose to
2135 : * always do the former; this choice is somewhat arbitrary, but the
2136 : * approach below requires that we uniformly do one or the other.
2137 : *
2138 : * Suppose we have a chain of N successive EMPTY arcs (where N can easily
2139 : * approach the size of the NFA). All of the intermediate states must
2140 : * have additional inarcs and outarcs, else they'd have been removed by
2141 : * the steps above. Assuming their inarcs are mostly not empties, we will
2142 : * add O(N^2) arcs to the NFA, since a non-EMPTY inarc leading to any one
2143 : * state in the chain must be duplicated to lead to all its successor
2144 : * states as well. So there is no hope of doing less than O(N^2) work;
2145 : * however, we should endeavor to keep the big-O cost from being even
2146 : * worse than that, which it can easily become without care. In
2147 : * particular, suppose we were to copy all S1's inarcs forward to S2, and
2148 : * then also to S3, and then later we consider pushing S2's inarcs forward
2149 : * to S3. If we include the arcs already copied from S1 in that, we'd be
2150 : * doing O(N^3) work. (The duplicate-arc elimination built into newarc()
2151 : * and its cohorts would get rid of the extra arcs, but not without cost.)
2152 : *
2153 : * We can avoid this cost by treating only arcs that existed at the start
2154 : * of this phase as candidates to be pushed forward. To identify those,
2155 : * we remember the first inarc each state had to start with. We rely on
2156 : * the fact that newarc() and friends put new arcs on the front of their
2157 : * to-states' inchains, and that this phase never deletes arcs, so that
2158 : * the original arcs must be the last arcs in their to-states' inchains.
2159 : *
2160 : * So the process here is that, for each state in the NFA, we gather up
2161 : * all non-EMPTY inarcs of states that can reach the target state via
2162 : * EMPTY arcs. We then sort, de-duplicate, and merge these arcs into the
2163 : * target state's inchain. (We can safely use sort-merge for this as long
2164 : * as we update each state's original-arcs pointer after we add arcs to
2165 : * it; the sort step of mergeins probably changed the order of the old
2166 : * arcs.)
2167 : *
2168 : * Another refinement worth making is that, because we only add non-EMPTY
2169 : * arcs during this phase, and all added arcs have the same from-state as
2170 : * the non-EMPTY arc they were cloned from, we know ahead of time that any
2171 : * states having only EMPTY outarcs will be useless for lack of outarcs
2172 : * after we drop the EMPTY arcs. (They cannot gain non-EMPTY outarcs if
2173 : * they had none to start with.) So we need not bother to update the
2174 : * inchains of such states at all.
2175 : */
2176 :
2177 : /* Remember the states' first original inarcs */
2178 : /* ... and while at it, count how many old inarcs there are altogether */
2179 18608 : inarcsorig = (struct arc **) MALLOC(nfa->nstates * sizeof(struct arc *));
2180 18608 : if (inarcsorig == NULL)
2181 : {
2182 0 : NERR(REG_ESPACE);
2183 0 : return;
2184 : }
2185 18608 : totalinarcs = 0;
2186 277366 : for (s = nfa->states; s != NULL; s = s->next)
2187 : {
2188 258758 : inarcsorig[s->no] = s->ins;
2189 258758 : totalinarcs += s->nins;
2190 : }
2191 :
2192 : /*
2193 : * Create a workspace for accumulating the inarcs to be added to the
2194 : * current target state. totalinarcs is probably a considerable
2195 : * overestimate of the space needed, but the NFA is unlikely to be large
2196 : * enough at this point to make it worth being smarter.
2197 : */
2198 18608 : arcarray = (struct arc **) MALLOC(totalinarcs * sizeof(struct arc *));
2199 18608 : if (arcarray == NULL)
2200 : {
2201 0 : NERR(REG_ESPACE);
2202 0 : FREE(inarcsorig);
2203 0 : return;
2204 : }
2205 :
2206 : /* And iterate over the target states */
2207 272404 : for (s = nfa->states; s != NULL && !NISERR(); s = s->next)
2208 : {
2209 : /* Ignore target states without non-EMPTY outarcs, per note above */
2210 253796 : if (!s->flag && !hasnonemptyout(s))
2211 3948 : continue;
2212 :
2213 : /* Find predecessor states and accumulate their original inarcs */
2214 249848 : arccount = 0;
2215 15236758 : for (s2 = emptyreachable(nfa, s, s, inarcsorig); s2 != s; s2 = nexts)
2216 : {
2217 : /* Add s2's original inarcs to arcarray[], but ignore empties */
2218 45002510 : for (a = inarcsorig[s2->no]; a != NULL; a = a->inchain)
2219 : {
2220 30015600 : if (a->type != EMPTY)
2221 15030586 : arcarray[arccount++] = a;
2222 : }
2223 :
2224 : /* Reset the tmp fields as we walk back */
2225 14986910 : nexts = s2->tmp;
2226 14986910 : s2->tmp = NULL;
2227 : }
2228 249848 : s->tmp = NULL;
2229 : assert(arccount <= totalinarcs);
2230 :
2231 : /* Remember how many original inarcs this state has */
2232 249848 : prevnins = s->nins;
2233 :
2234 : /* Add non-duplicate inarcs to target state */
2235 249848 : mergeins(nfa, s, arcarray, arccount);
2236 :
2237 : /* Now we must update the state's inarcsorig pointer */
2238 249848 : nskip = s->nins - prevnins;
2239 249848 : a = s->ins;
2240 15267784 : while (nskip-- > 0)
2241 15017936 : a = a->inchain;
2242 249848 : inarcsorig[s->no] = a;
2243 : }
2244 :
2245 18608 : FREE(arcarray);
2246 18608 : FREE(inarcsorig);
2247 :
2248 18608 : if (NISERR())
2249 6 : return;
2250 :
2251 : /*
2252 : * Now remove all the EMPTY arcs, since we don't need them anymore.
2253 : */
2254 259348 : for (s = nfa->states; s != NULL; s = s->next)
2255 : {
2256 1474252 : for (a = s->outs; a != NULL; a = nexta)
2257 : {
2258 1233506 : nexta = a->outchain;
2259 1233506 : if (a->type == EMPTY)
2260 25278 : freearc(nfa, a);
2261 : }
2262 : }
2263 :
2264 : /*
2265 : * And remove any states that have become useless. (This cleanup is not
2266 : * very thorough, and would be even less so if we tried to combine it with
2267 : * the previous step; but cleanup() will take care of anything we miss.)
2268 : */
2269 259348 : for (s = nfa->states; s != NULL; s = nexts)
2270 : {
2271 240746 : nexts = s->next;
2272 240746 : if ((s->nins == 0 || s->nouts == 0) && !s->flag)
2273 3948 : dropstate(nfa, s);
2274 : }
2275 :
2276 18602 : if (f != NULL)
2277 0 : dumpnfa(nfa, f);
2278 : }
2279 :
2280 : /*
2281 : * emptyreachable - recursively find all states that can reach s by EMPTY arcs
2282 : *
2283 : * The return value is the last such state found. Its tmp field links back
2284 : * to the next-to-last such state, and so on back to s, so that all these
2285 : * states can be located without searching the whole NFA.
2286 : *
2287 : * Since this is only used in fixempties(), we pass in the inarcsorig[] array
2288 : * maintained by that function. This lets us skip over all new inarcs, which
2289 : * are certainly not EMPTY arcs.
2290 : *
2291 : * The maximum recursion depth here is equal to the length of the longest
2292 : * loop-free chain of EMPTY arcs, which is surely no more than the size of
2293 : * the NFA ... but that could still be enough to cause trouble.
2294 : */
2295 : static struct state *
2296 15236758 : emptyreachable(struct nfa *nfa,
2297 : struct state *s,
2298 : struct state *lastfound,
2299 : struct arc **inarcsorig)
2300 : {
2301 : struct arc *a;
2302 :
2303 : /* Since this is recursive, it could be driven to stack overflow */
2304 15236758 : if (STACK_TOO_DEEP(nfa->v->re))
2305 : {
2306 0 : NERR(REG_ETOOBIG);
2307 0 : return lastfound;
2308 : }
2309 :
2310 15236758 : s->tmp = lastfound;
2311 15236758 : lastfound = s;
2312 45672928 : for (a = inarcsorig[s->no]; a != NULL; a = a->inchain)
2313 : {
2314 30436170 : if (a->type == EMPTY && a->from->tmp == NULL)
2315 14986910 : lastfound = emptyreachable(nfa, a->from, lastfound, inarcsorig);
2316 : }
2317 15236758 : return lastfound;
2318 : }
2319 :
2320 : /*
2321 : * isconstraintarc - detect whether an arc is of a constraint type
2322 : */
2323 : static inline int
2324 2548890 : isconstraintarc(struct arc *a)
2325 : {
2326 2548890 : switch (a->type)
2327 : {
2328 322462 : case '^':
2329 : case '$':
2330 : case BEHIND:
2331 : case AHEAD:
2332 : case LACON:
2333 322462 : return 1;
2334 : }
2335 2226428 : return 0;
2336 : }
2337 :
2338 : /*
2339 : * hasconstraintout - does state have a constraint out arc?
2340 : */
2341 : static int
2342 25262 : hasconstraintout(struct state *s)
2343 : {
2344 : struct arc *a;
2345 :
2346 47920 : for (a = s->outs; a != NULL; a = a->outchain)
2347 : {
2348 38444 : if (isconstraintarc(a))
2349 15786 : return 1;
2350 : }
2351 9476 : return 0;
2352 : }
2353 :
2354 : /*
2355 : * fixconstraintloops - get rid of loops containing only constraint arcs
2356 : *
2357 : * A loop of states that contains only constraint arcs is useless, since
2358 : * passing around the loop represents no forward progress. Moreover, it
2359 : * would cause infinite looping in pullback/pushfwd, so we need to get rid
2360 : * of such loops before doing that.
2361 : */
2362 : static void
2363 18608 : fixconstraintloops(struct nfa *nfa,
2364 : FILE *f) /* for debug output; NULL none */
2365 : {
2366 : struct state *s;
2367 : struct state *nexts;
2368 : struct arc *a;
2369 : struct arc *nexta;
2370 : int hasconstraints;
2371 :
2372 : /*
2373 : * In the trivial case of a state that loops to itself, we can just drop
2374 : * the constraint arc altogether. This is worth special-casing because
2375 : * such loops are far more common than loops containing multiple states.
2376 : * While we're at it, note whether any constraint arcs survive.
2377 : */
2378 18608 : hasconstraints = 0;
2379 255406 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
2380 : {
2381 236798 : nexts = s->next;
2382 : /* while we're at it, ensure tmp fields are clear for next step */
2383 : assert(s->tmp == NULL);
2384 1441080 : for (a = s->outs; a != NULL && !NISERR(); a = nexta)
2385 : {
2386 1204282 : nexta = a->outchain;
2387 1204282 : if (isconstraintarc(a))
2388 : {
2389 119034 : if (a->to == s)
2390 352 : freearc(nfa, a);
2391 : else
2392 118682 : hasconstraints = 1;
2393 : }
2394 : }
2395 : /* If we removed all the outarcs, the state is useless. */
2396 236798 : if (s->nouts == 0 && !s->flag)
2397 0 : dropstate(nfa, s);
2398 : }
2399 :
2400 : /* Nothing to do if no remaining constraint arcs */
2401 18608 : if (NISERR() || !hasconstraints)
2402 6 : return;
2403 :
2404 : /*
2405 : * Starting from each remaining NFA state, search outwards for a
2406 : * constraint loop. If we find a loop, break the loop, then start the
2407 : * search over. (We could possibly retain some state from the first scan,
2408 : * but it would complicate things greatly, and multi-state constraint
2409 : * loops are rare enough that it's not worth optimizing the case.)
2410 : */
2411 18602 : restart:
2412 270334 : for (s = nfa->states; s != NULL && !NISERR(); s = s->next)
2413 : {
2414 251732 : if (findconstraintloop(nfa, s))
2415 412 : goto restart;
2416 : }
2417 :
2418 18602 : if (NISERR())
2419 0 : return;
2420 :
2421 : /*
2422 : * Now remove any states that have become useless. (This cleanup is not
2423 : * very thorough, and would be even less so if we tried to combine it with
2424 : * the previous step; but cleanup() will take care of anything we miss.)
2425 : *
2426 : * Because findconstraintloop intentionally doesn't reset all tmp fields,
2427 : * we have to clear them after it's done. This is a convenient place to
2428 : * do that, too.
2429 : */
2430 257122 : for (s = nfa->states; s != NULL; s = nexts)
2431 : {
2432 238520 : nexts = s->next;
2433 238520 : s->tmp = NULL;
2434 238520 : if ((s->nins == 0 || s->nouts == 0) && !s->flag)
2435 426 : dropstate(nfa, s);
2436 : }
2437 :
2438 18602 : if (f != NULL)
2439 0 : dumpnfa(nfa, f);
2440 : }
2441 :
2442 : /*
2443 : * findconstraintloop - recursively find a loop of constraint arcs
2444 : *
2445 : * If we find a loop, break it by calling breakconstraintloop(), then
2446 : * return 1; otherwise return 0.
2447 : *
2448 : * State tmp fields are guaranteed all NULL on a success return, because
2449 : * breakconstraintloop does that. After a failure return, any state that
2450 : * is known not to be part of a loop is marked with s->tmp == s; this allows
2451 : * us not to have to re-prove that fact on later calls. (This convention is
2452 : * workable because we already eliminated single-state loops.)
2453 : *
2454 : * Note that the found loop doesn't necessarily include the first state we
2455 : * are called on. Any loop reachable from that state will do.
2456 : *
2457 : * The maximum recursion depth here is one more than the length of the longest
2458 : * loop-free chain of constraint arcs, which is surely no more than the size
2459 : * of the NFA ... but that could still be enough to cause trouble.
2460 : */
2461 : static int
2462 410542 : findconstraintloop(struct nfa *nfa, struct state *s)
2463 : {
2464 : struct arc *a;
2465 :
2466 : /* Since this is recursive, it could be driven to stack overflow */
2467 410542 : if (STACK_TOO_DEEP(nfa->v->re))
2468 : {
2469 0 : NERR(REG_ETOOBIG);
2470 0 : return 1; /* to exit as quickly as possible */
2471 : }
2472 :
2473 410542 : if (s->tmp != NULL)
2474 : {
2475 : /* Already proven uninteresting? */
2476 154070 : if (s->tmp == s)
2477 153658 : return 0;
2478 : /* Found a loop involving s */
2479 412 : breakconstraintloop(nfa, s);
2480 : /* The tmp fields have been cleaned up by breakconstraintloop */
2481 412 : return 1;
2482 : }
2483 1531444 : for (a = s->outs; a != NULL; a = a->outchain)
2484 : {
2485 1276290 : if (isconstraintarc(a))
2486 : {
2487 158810 : struct state *sto = a->to;
2488 :
2489 : assert(sto != s);
2490 158810 : s->tmp = sto;
2491 158810 : if (findconstraintloop(nfa, sto))
2492 1318 : return 1;
2493 : }
2494 : }
2495 :
2496 : /*
2497 : * If we get here, no constraint loop exists leading out from s. Mark it
2498 : * with s->tmp == s so we need not rediscover that fact again later.
2499 : */
2500 255154 : s->tmp = s;
2501 255154 : return 0;
2502 : }
2503 :
2504 : /*
2505 : * breakconstraintloop - break a loop of constraint arcs
2506 : *
2507 : * sinitial is any one member state of the loop. Each loop member's tmp
2508 : * field links to its successor within the loop. (Note that this function
2509 : * will reset all the tmp fields to NULL.)
2510 : *
2511 : * We can break the loop by, for any one state S1 in the loop, cloning its
2512 : * loop successor state S2 (and possibly following states), and then moving
2513 : * all S1->S2 constraint arcs to point to the cloned S2. The cloned S2 should
2514 : * copy any non-constraint outarcs of S2. Constraint outarcs should be
2515 : * dropped if they point back to S1, else they need to be copied as arcs to
2516 : * similarly cloned states S3, S4, etc. In general, each cloned state copies
2517 : * non-constraint outarcs, drops constraint outarcs that would lead to itself
2518 : * or any earlier cloned state, and sends other constraint outarcs to newly
2519 : * cloned states. No cloned state will have any inarcs that aren't constraint
2520 : * arcs or do not lead from S1 or earlier-cloned states. It's okay to drop
2521 : * constraint back-arcs since they would not take us to any state we've not
2522 : * already been in; therefore, no new constraint loop is created. In this way
2523 : * we generate a modified NFA that can still represent every useful state
2524 : * sequence, but not sequences that represent state loops with no consumption
2525 : * of input data. Note that the set of cloned states will certainly include
2526 : * all of the loop member states other than S1, and it may also include
2527 : * non-loop states that are reachable from S2 via constraint arcs. This is
2528 : * important because there is no guarantee that findconstraintloop found a
2529 : * maximal loop (and searching for one would be NP-hard, so don't try).
2530 : * Frequently the "non-loop states" are actually part of a larger loop that
2531 : * we didn't notice, and indeed there may be several overlapping loops.
2532 : * This technique ensures convergence in such cases, while considering only
2533 : * the originally-found loop does not.
2534 : *
2535 : * If there is only one S1->S2 constraint arc, then that constraint is
2536 : * certainly satisfied when we enter any of the clone states. This means that
2537 : * in the common case where many of the constraint arcs are identically
2538 : * labeled, we can merge together clone states linked by a similarly-labeled
2539 : * constraint: if we can get to the first one we can certainly get to the
2540 : * second, so there's no need to distinguish. This greatly reduces the number
2541 : * of new states needed, so we preferentially break the given loop at a state
2542 : * pair where this is true.
2543 : *
2544 : * Furthermore, it's fairly common to find that a cloned successor state has
2545 : * no outarcs, especially if we're a bit aggressive about removing unnecessary
2546 : * outarcs. If that happens, then there is simply not any interesting state
2547 : * that can be reached through the predecessor's loop arcs, which means we can
2548 : * break the loop just by removing those loop arcs, with no new states added.
2549 : */
2550 : static void
2551 412 : breakconstraintloop(struct nfa *nfa, struct state *sinitial)
2552 : {
2553 : struct state *s;
2554 : struct state *shead;
2555 : struct state *stail;
2556 : struct state *sclone;
2557 : struct state *nexts;
2558 : struct arc *refarc;
2559 : struct arc *a;
2560 : struct arc *nexta;
2561 :
2562 : /*
2563 : * Start by identifying which loop step we want to break at.
2564 : * Preferentially this is one with only one constraint arc. (XXX are
2565 : * there any other secondary heuristics we want to use here?) Set refarc
2566 : * to point to the selected lone constraint arc, if there is one.
2567 : */
2568 412 : refarc = NULL;
2569 412 : s = sinitial;
2570 : do
2571 : {
2572 1060 : nexts = s->tmp;
2573 : assert(nexts != s); /* should not see any one-element loops */
2574 1060 : if (refarc == NULL)
2575 : {
2576 664 : int narcs = 0;
2577 :
2578 7600 : for (a = s->outs; a != NULL; a = a->outchain)
2579 : {
2580 6936 : if (a->to == nexts && isconstraintarc(a))
2581 : {
2582 2592 : refarc = a;
2583 2592 : narcs++;
2584 : }
2585 : }
2586 : assert(narcs > 0);
2587 664 : if (narcs > 1)
2588 372 : refarc = NULL; /* multiple constraint arcs here, no good */
2589 : }
2590 1060 : s = nexts;
2591 1060 : } while (s != sinitial);
2592 :
2593 412 : if (refarc)
2594 : {
2595 : /* break at the refarc */
2596 292 : shead = refarc->from;
2597 292 : stail = refarc->to;
2598 : assert(stail == shead->tmp);
2599 : }
2600 : else
2601 : {
2602 : /* for lack of a better idea, break after sinitial */
2603 120 : shead = sinitial;
2604 120 : stail = sinitial->tmp;
2605 : }
2606 :
2607 : /*
2608 : * Reset the tmp fields so that we can use them for local storage in
2609 : * clonesuccessorstates. (findconstraintloop won't mind, since it's just
2610 : * going to abandon its search anyway.)
2611 : */
2612 34722 : for (s = nfa->states; s != NULL; s = s->next)
2613 34310 : s->tmp = NULL;
2614 :
2615 : /*
2616 : * Recursively build clone state(s) as needed.
2617 : */
2618 412 : sclone = newstate(nfa);
2619 412 : if (sclone == NULL)
2620 : {
2621 : assert(NISERR());
2622 0 : return;
2623 : }
2624 :
2625 412 : clonesuccessorstates(nfa, stail, sclone, shead, refarc,
2626 : NULL, NULL, nfa->nstates);
2627 :
2628 412 : if (NISERR())
2629 0 : return;
2630 :
2631 : /*
2632 : * It's possible that sclone has no outarcs at all, in which case it's
2633 : * useless. (We don't try extremely hard to get rid of useless states
2634 : * here, but this is an easy and fairly common case.)
2635 : */
2636 412 : if (sclone->nouts == 0)
2637 : {
2638 98 : freestate(nfa, sclone);
2639 98 : sclone = NULL;
2640 : }
2641 :
2642 : /*
2643 : * Move shead's constraint-loop arcs to point to sclone, or just drop them
2644 : * if we discovered we don't need sclone.
2645 : */
2646 4512 : for (a = shead->outs; a != NULL; a = nexta)
2647 : {
2648 4100 : nexta = a->outchain;
2649 4100 : if (a->to == stail && isconstraintarc(a))
2650 : {
2651 978 : if (sclone)
2652 824 : cparc(nfa, a, shead, sclone);
2653 978 : freearc(nfa, a);
2654 978 : if (NISERR())
2655 0 : break;
2656 : }
2657 : }
2658 : }
2659 :
2660 : /*
2661 : * clonesuccessorstates - create a tree of constraint-arc successor states
2662 : *
2663 : * ssource is the state to be cloned, and sclone is the state to copy its
2664 : * outarcs into. sclone's inarcs, if any, should already be set up.
2665 : *
2666 : * spredecessor is the original predecessor state that we are trying to build
2667 : * successors for (it may not be the immediate predecessor of ssource).
2668 : * refarc, if not NULL, is the original constraint arc that is known to have
2669 : * been traversed out of spredecessor to reach the successor(s).
2670 : *
2671 : * For each cloned successor state, we transiently create a "donemap" that is
2672 : * a boolean array showing which source states we've already visited for this
2673 : * clone state. This prevents infinite recursion as well as useless repeat
2674 : * visits to the same state subtree (which can add up fast, since typical NFAs
2675 : * have multiple redundant arc pathways). Each donemap is a char array
2676 : * indexed by state number. The donemaps are all of the same size "nstates",
2677 : * which is nfa->nstates as of the start of the recursion. This is enough to
2678 : * have entries for all pre-existing states, but *not* entries for clone
2679 : * states created during the recursion. That's okay since we have no need to
2680 : * mark those.
2681 : *
2682 : * curdonemap is NULL when recursing to a new sclone state, or sclone's
2683 : * donemap when we are recursing without having created a new state (which we
2684 : * do when we decide we can merge a successor state into the current clone
2685 : * state). outerdonemap is NULL at the top level and otherwise the parent
2686 : * clone state's donemap.
2687 : *
2688 : * The successor states we create and fill here form a strict tree structure,
2689 : * with each state having exactly one predecessor, except that the toplevel
2690 : * state has no inarcs as yet (breakconstraintloop will add its inarcs from
2691 : * spredecessor after we're done). Thus, we can examine sclone's inarcs back
2692 : * to the root, plus refarc if any, to identify the set of constraints already
2693 : * known valid at the current point. This allows us to avoid generating extra
2694 : * successor states.
2695 : */
2696 : static void
2697 3622 : clonesuccessorstates(struct nfa *nfa,
2698 : struct state *ssource,
2699 : struct state *sclone,
2700 : struct state *spredecessor,
2701 : struct arc *refarc,
2702 : char *curdonemap,
2703 : char *outerdonemap,
2704 : int nstates)
2705 : {
2706 : char *donemap;
2707 : struct arc *a;
2708 :
2709 : /* Since this is recursive, it could be driven to stack overflow */
2710 3622 : if (STACK_TOO_DEEP(nfa->v->re))
2711 : {
2712 0 : NERR(REG_ETOOBIG);
2713 0 : return;
2714 : }
2715 :
2716 : /* If this state hasn't already got a donemap, create one */
2717 3622 : donemap = curdonemap;
2718 3622 : if (donemap == NULL)
2719 : {
2720 1820 : donemap = (char *) MALLOC(nstates * sizeof(char));
2721 1820 : if (donemap == NULL)
2722 : {
2723 0 : NERR(REG_ESPACE);
2724 0 : return;
2725 : }
2726 :
2727 1820 : if (outerdonemap != NULL)
2728 : {
2729 : /*
2730 : * Not at outermost recursion level, so copy the outer level's
2731 : * donemap; this ensures that we see states in process of being
2732 : * visited at outer levels, or already merged into predecessor
2733 : * states, as ones we shouldn't traverse back to.
2734 : */
2735 1408 : memcpy(donemap, outerdonemap, nstates * sizeof(char));
2736 : }
2737 : else
2738 : {
2739 : /* At outermost level, only spredecessor is off-limits */
2740 412 : memset(donemap, 0, nstates * sizeof(char));
2741 : assert(spredecessor->no < nstates);
2742 412 : donemap[spredecessor->no] = 1;
2743 : }
2744 : }
2745 :
2746 : /* Mark ssource as visited in the donemap */
2747 : assert(ssource->no < nstates);
2748 : assert(donemap[ssource->no] == 0);
2749 3622 : donemap[ssource->no] = 1;
2750 :
2751 : /*
2752 : * We proceed by first cloning all of ssource's outarcs, creating new
2753 : * clone states as needed but not doing more with them than that. Then in
2754 : * a second pass, recurse to process the child clone states. This allows
2755 : * us to have only one child clone state per reachable source state, even
2756 : * when there are multiple outarcs leading to the same state. Also, when
2757 : * we do visit a child state, its set of inarcs is known exactly, which
2758 : * makes it safe to apply the constraint-is-already-checked optimization.
2759 : * Also, this ensures that we've merged all the states we can into the
2760 : * current clone before we recurse to any children, thus possibly saving
2761 : * them from making extra images of those states.
2762 : *
2763 : * While this function runs, child clone states of the current state are
2764 : * marked by setting their tmp fields to point to the original state they
2765 : * were cloned from. This makes it possible to detect multiple outarcs
2766 : * leading to the same state, and also makes it easy to distinguish clone
2767 : * states from original states (which will have tmp == NULL).
2768 : */
2769 29926 : for (a = ssource->outs; a != NULL && !NISERR(); a = a->outchain)
2770 : {
2771 26304 : struct state *sto = a->to;
2772 :
2773 : /*
2774 : * We do not consider cloning successor states that have no constraint
2775 : * outarcs; just link to them as-is. They cannot be part of a
2776 : * constraint loop so there is no need to make copies. In particular,
2777 : * this rule keeps us from trying to clone the post state, which would
2778 : * be a bad idea.
2779 : */
2780 26304 : if (isconstraintarc(a) && hasconstraintout(sto))
2781 10970 : {
2782 : struct state *prevclone;
2783 : int canmerge;
2784 : struct arc *a2;
2785 :
2786 : /*
2787 : * Back-link constraint arcs must not be followed. Nor is there a
2788 : * need to revisit states previously merged into this clone.
2789 : */
2790 : assert(sto->no < nstates);
2791 15786 : if (donemap[sto->no] != 0)
2792 4816 : continue;
2793 :
2794 : /*
2795 : * Check whether we already have a child clone state for this
2796 : * source state.
2797 : */
2798 10970 : prevclone = NULL;
2799 37524 : for (a2 = sclone->outs; a2 != NULL; a2 = a2->outchain)
2800 : {
2801 34314 : if (a2->to->tmp == sto)
2802 : {
2803 7760 : prevclone = a2->to;
2804 7760 : break;
2805 : }
2806 : }
2807 :
2808 : /*
2809 : * If this arc is labeled the same as refarc, or the same as any
2810 : * arc we must have traversed to get to sclone, then no additional
2811 : * constraints need to be met to get to sto, so we should just
2812 : * merge its outarcs into sclone.
2813 : */
2814 10970 : if (refarc && a->type == refarc->type && a->co == refarc->co)
2815 1802 : canmerge = 1;
2816 : else
2817 : {
2818 : struct state *s;
2819 :
2820 9168 : canmerge = 0;
2821 45892 : for (s = sclone; s->ins; s = s->ins->from)
2822 : {
2823 36724 : if (s->nins == 1 &&
2824 36 : a->type == s->ins->type && a->co == s->ins->co)
2825 : {
2826 0 : canmerge = 1;
2827 0 : break;
2828 : }
2829 : }
2830 : }
2831 :
2832 10970 : if (canmerge)
2833 : {
2834 : /*
2835 : * We can merge into sclone. If we previously made a child
2836 : * clone state, drop it; there's no need to visit it. (This
2837 : * can happen if ssource has multiple pathways to sto, and we
2838 : * only just now found one that is provably a no-op.)
2839 : */
2840 1802 : if (prevclone)
2841 0 : dropstate(nfa, prevclone); /* kills our outarc, too */
2842 :
2843 : /* Recurse to merge sto's outarcs into sclone */
2844 1802 : clonesuccessorstates(nfa,
2845 : sto,
2846 : sclone,
2847 : spredecessor,
2848 : refarc,
2849 : donemap,
2850 : outerdonemap,
2851 : nstates);
2852 : /* sto should now be marked as previously visited */
2853 : assert(NISERR() || donemap[sto->no] == 1);
2854 : }
2855 9168 : else if (prevclone)
2856 : {
2857 : /*
2858 : * We already have a clone state for this successor, so just
2859 : * make another arc to it.
2860 : */
2861 7760 : cparc(nfa, a, sclone, prevclone);
2862 : }
2863 : else
2864 : {
2865 : /*
2866 : * We need to create a new successor clone state.
2867 : */
2868 : struct state *stoclone;
2869 :
2870 1408 : stoclone = newstate(nfa);
2871 1408 : if (stoclone == NULL)
2872 : {
2873 : assert(NISERR());
2874 0 : break;
2875 : }
2876 : /* Mark it as to what it's a clone of */
2877 1408 : stoclone->tmp = sto;
2878 : /* ... and add the outarc leading to it */
2879 1408 : cparc(nfa, a, sclone, stoclone);
2880 : }
2881 : }
2882 : else
2883 : {
2884 : /*
2885 : * Non-constraint outarcs just get copied to sclone, as do outarcs
2886 : * leading to states with no constraint outarc.
2887 : */
2888 10518 : cparc(nfa, a, sclone, sto);
2889 : }
2890 : }
2891 :
2892 : /*
2893 : * If we are at outer level for this clone state, recurse to all its child
2894 : * clone states, clearing their tmp fields as we go. (If we're not
2895 : * outermost for sclone, leave this to be done by the outer call level.)
2896 : * Note that if we have multiple outarcs leading to the same clone state,
2897 : * it will only be recursed-to once.
2898 : */
2899 3622 : if (curdonemap == NULL)
2900 : {
2901 16166 : for (a = sclone->outs; a != NULL && !NISERR(); a = a->outchain)
2902 : {
2903 14346 : struct state *stoclone = a->to;
2904 14346 : struct state *sto = stoclone->tmp;
2905 :
2906 14346 : if (sto != NULL)
2907 : {
2908 1408 : stoclone->tmp = NULL;
2909 1408 : clonesuccessorstates(nfa,
2910 : sto,
2911 : stoclone,
2912 : spredecessor,
2913 : refarc,
2914 : NULL,
2915 : donemap,
2916 : nstates);
2917 : }
2918 : }
2919 :
2920 : /* Don't forget to free sclone's donemap when done with it */
2921 1820 : FREE(donemap);
2922 : }
2923 : }
2924 :
2925 : /*
2926 : * cleanup - clean up NFA after optimizations
2927 : */
2928 : static void
2929 37216 : cleanup(struct nfa *nfa)
2930 : {
2931 : struct state *s;
2932 : struct state *nexts;
2933 : int n;
2934 :
2935 37216 : if (NISERR())
2936 6 : return;
2937 :
2938 : /* clear out unreachable or dead-end states */
2939 : /* use pre to mark reachable, then post to mark can-reach-post */
2940 37210 : markreachable(nfa, nfa->pre, (struct state *) NULL, nfa->pre);
2941 37210 : markcanreach(nfa, nfa->post, nfa->pre, nfa->post);
2942 659984 : for (s = nfa->states; s != NULL && !NISERR(); s = nexts)
2943 : {
2944 622774 : nexts = s->next;
2945 622774 : if (s->tmp != nfa->post && !s->flag)
2946 6240 : dropstate(nfa, s);
2947 : }
2948 : assert(NISERR() || nfa->post->nins == 0 || nfa->post->tmp == nfa->post);
2949 37210 : cleartraverse(nfa, nfa->pre);
2950 : assert(NISERR() || nfa->post->nins == 0 || nfa->post->tmp == NULL);
2951 : /* the nins==0 (final unreachable) case will be caught later */
2952 :
2953 : /* renumber surviving states */
2954 37210 : n = 0;
2955 653744 : for (s = nfa->states; s != NULL; s = s->next)
2956 616534 : s->no = n++;
2957 37210 : nfa->nstates = n;
2958 : }
2959 :
2960 : /*
2961 : * markreachable - recursive marking of reachable states
2962 : */
2963 : static void
2964 1744078 : markreachable(struct nfa *nfa,
2965 : struct state *s,
2966 : struct state *okay, /* consider only states with this mark */
2967 : struct state *mark) /* the value to mark with */
2968 : {
2969 : struct arc *a;
2970 :
2971 : /* Since this is recursive, it could be driven to stack overflow */
2972 1744078 : if (STACK_TOO_DEEP(nfa->v->re))
2973 : {
2974 0 : NERR(REG_ETOOBIG);
2975 0 : return;
2976 : }
2977 :
2978 1744078 : if (s->tmp != okay)
2979 1127552 : return;
2980 616526 : s->tmp = mark;
2981 :
2982 2323394 : for (a = s->outs; a != NULL; a = a->outchain)
2983 1706868 : markreachable(nfa, a->to, okay, mark);
2984 : }
2985 :
2986 : /*
2987 : * markcanreach - recursive marking of states which can reach here
2988 : */
2989 : static void
2990 1744984 : markcanreach(struct nfa *nfa,
2991 : struct state *s,
2992 : struct state *okay, /* consider only states with this mark */
2993 : struct state *mark) /* the value to mark with */
2994 : {
2995 : struct arc *a;
2996 :
2997 : /* Since this is recursive, it could be driven to stack overflow */
2998 1744984 : if (STACK_TOO_DEEP(nfa->v->re))
2999 : {
3000 0 : NERR(REG_ETOOBIG);
3001 0 : return;
3002 : }
3003 :
3004 1744984 : if (s->tmp != okay)
3005 1128638 : return;
3006 616346 : s->tmp = mark;
3007 :
3008 2324120 : for (a = s->ins; a != NULL; a = a->inchain)
3009 1707774 : markcanreach(nfa, a->from, okay, mark);
3010 : }
3011 :
3012 : /*
3013 : * analyze - ascertain potentially-useful facts about an optimized NFA
3014 : */
3015 : static long /* re_info bits to be ORed in */
3016 18608 : analyze(struct nfa *nfa)
3017 : {
3018 : struct arc *a;
3019 : struct arc *aa;
3020 :
3021 18608 : if (NISERR())
3022 6 : return 0;
3023 :
3024 : /* Detect whether NFA can't match anything */
3025 18602 : if (nfa->pre->outs == NULL)
3026 94 : return REG_UIMPOSSIBLE;
3027 :
3028 : /* Detect whether NFA matches all strings (possibly with length bounds) */
3029 18508 : checkmatchall(nfa);
3030 :
3031 : /* Detect whether NFA can possibly match a zero-length string */
3032 55696 : for (a = nfa->pre->outs; a != NULL; a = a->outchain)
3033 1542160 : for (aa = a->to->outs; aa != NULL; aa = aa->outchain)
3034 1504972 : if (aa->to == nfa->post)
3035 3178 : return REG_UEMPTYMATCH;
3036 15330 : return 0;
3037 : }
3038 :
3039 : /*
3040 : * checkmatchall - does the NFA represent no more than a string length test?
3041 : *
3042 : * If so, set nfa->minmatchall and nfa->maxmatchall correctly (they are -1
3043 : * to begin with) and set the MATCHALL bit in nfa->flags.
3044 : *
3045 : * To succeed, we require all arcs to be PLAIN RAINBOW arcs, except for those
3046 : * for pseudocolors (i.e., BOS/BOL/EOS/EOL). We must be able to reach the
3047 : * post state via RAINBOW arcs, and if there are any loops in the graph, they
3048 : * must be loop-to-self arcs, ensuring that each loop iteration consumes
3049 : * exactly one character. (Longer loops are problematic because they create
3050 : * non-consecutive possible match lengths; we have no good way to represent
3051 : * that situation for lengths beyond the DUPINF limit.)
3052 : *
3053 : * Pseudocolor arcs complicate things a little. We know that they can only
3054 : * appear as pre-state outarcs (for BOS/BOL) or post-state inarcs (for
3055 : * EOS/EOL). There, they must exactly replicate the parallel RAINBOW arcs,
3056 : * e.g. if the pre state has one RAINBOW outarc to state 2, it must have BOS
3057 : * and BOL outarcs to state 2, and no others. Missing or extra pseudocolor
3058 : * arcs can occur, meaning that the NFA involves some constraint on the
3059 : * adjacent characters, which makes it not a matchall NFA.
3060 : */
3061 : static void
3062 18508 : checkmatchall(struct nfa *nfa)
3063 : {
3064 : bool **haspaths;
3065 : struct state *s;
3066 : int i;
3067 :
3068 : /*
3069 : * If there are too many states, don't bother trying to detect matchall.
3070 : * This limit serves to bound the time and memory we could consume below.
3071 : * Note that even if the graph is all-RAINBOW, if there are significantly
3072 : * more than DUPINF states then it's likely that there are paths of length
3073 : * more than DUPINF, which would force us to fail anyhow. In practice,
3074 : * plausible ways of writing a matchall regex with maximum finite path
3075 : * length K tend not to have very many more than K states.
3076 : */
3077 18508 : if (nfa->nstates > DUPINF * 2)
3078 12 : return;
3079 :
3080 : /*
3081 : * First, scan all the states to verify that only RAINBOW arcs appear,
3082 : * plus pseudocolor arcs adjacent to the pre and post states. This lets
3083 : * us quickly eliminate most cases that aren't matchall NFAs.
3084 : */
3085 67480 : for (s = nfa->states; s != NULL; s = s->next)
3086 : {
3087 : struct arc *a;
3088 :
3089 194256 : for (a = s->outs; a != NULL; a = a->outchain)
3090 : {
3091 145272 : if (a->type != PLAIN)
3092 90 : return; /* any LACONs make it non-matchall */
3093 145182 : if (a->co != RAINBOW)
3094 : {
3095 61082 : if (nfa->cm->cd[a->co].flags & PSEUDO)
3096 : {
3097 : /*
3098 : * Pseudocolor arc: verify it's in a valid place (this
3099 : * seems quite unlikely to fail, but let's be sure).
3100 : */
3101 44222 : if (s == nfa->pre &&
3102 31938 : (a->co == nfa->bos[0] || a->co == nfa->bos[1]))
3103 : /* okay BOS/BOL arc */ ;
3104 12284 : else if (a->to == nfa->post &&
3105 12284 : (a->co == nfa->eos[0] || a->co == nfa->eos[1]))
3106 : /* okay EOS/EOL arc */ ;
3107 : else
3108 0 : return; /* unexpected pseudocolor arc */
3109 : /* We'll check these arcs some more below. */
3110 : }
3111 : else
3112 16860 : return; /* any other color makes it non-matchall */
3113 : }
3114 : }
3115 : /* Also, assert that the tmp fields are available for use. */
3116 : assert(s->tmp == NULL);
3117 : }
3118 :
3119 : /*
3120 : * The next cheapest check we can make is to verify that the BOS/BOL
3121 : * outarcs of the pre state reach the same states as its RAINBOW outarcs.
3122 : * If they don't, the NFA expresses some constraints on the character
3123 : * before the matched string, making it non-matchall. Likewise, the
3124 : * EOS/EOL inarcs of the post state must match its RAINBOW inarcs.
3125 : */
3126 1546 : if (!check_out_colors_match(nfa->pre, RAINBOW, nfa->bos[0]) ||
3127 1540 : !check_out_colors_match(nfa->pre, RAINBOW, nfa->bos[1]) ||
3128 1204 : !check_in_colors_match(nfa->post, RAINBOW, nfa->eos[0]) ||
3129 1196 : !check_in_colors_match(nfa->post, RAINBOW, nfa->eos[1]))
3130 558 : return;
3131 :
3132 : /*
3133 : * Initialize an array of path-length arrays, in which
3134 : * checkmatchall_recurse will return per-state results. This lets us
3135 : * memo-ize the recursive search and avoid exponential time consumption.
3136 : */
3137 988 : haspaths = (bool **) MALLOC(nfa->nstates * sizeof(bool *));
3138 988 : if (haspaths == NULL)
3139 0 : return; /* fail quietly */
3140 988 : memset(haspaths, 0, nfa->nstates * sizeof(bool *));
3141 :
3142 : /*
3143 : * Recursively search the graph for all-RAINBOW paths to the "post" state,
3144 : * starting at the "pre" state, and computing the lengths of the paths.
3145 : * (Given the preceding checks, there should be at least one such path.
3146 : * However we could get back a false result anyway, in case there are
3147 : * multi-state loops, paths exceeding DUPINF+1 length, or non-algorithmic
3148 : * failures such as ENOMEM.)
3149 : */
3150 988 : if (checkmatchall_recurse(nfa, nfa->pre, haspaths))
3151 : {
3152 : /* The useful result is the path length array for the pre state */
3153 964 : bool *haspath = haspaths[nfa->pre->no];
3154 : int minmatch,
3155 : maxmatch,
3156 : morematch;
3157 :
3158 : assert(haspath != NULL);
3159 :
3160 : /*
3161 : * haspath[] now represents the set of possible path lengths; but we
3162 : * want to reduce that to a min and max value, because it doesn't seem
3163 : * worth complicating regexec.c to deal with nonconsecutive possible
3164 : * match lengths. Find min and max of first run of lengths, then
3165 : * verify there are no nonconsecutive lengths.
3166 : */
3167 3966 : for (minmatch = 0; minmatch <= DUPINF + 1; minmatch++)
3168 : {
3169 3966 : if (haspath[minmatch])
3170 964 : break;
3171 : }
3172 : assert(minmatch <= DUPINF + 1); /* else checkmatchall_recurse lied */
3173 63566 : for (maxmatch = minmatch; maxmatch < DUPINF + 1; maxmatch++)
3174 : {
3175 63326 : if (!haspath[maxmatch + 1])
3176 724 : break;
3177 : }
3178 180060 : for (morematch = maxmatch + 1; morematch <= DUPINF + 1; morematch++)
3179 : {
3180 179108 : if (haspath[morematch])
3181 : {
3182 12 : haspath = NULL; /* fail, there are nonconsecutive lengths */
3183 12 : break;
3184 : }
3185 : }
3186 :
3187 964 : if (haspath != NULL)
3188 : {
3189 : /*
3190 : * Success, so record the info. Here we have a fine point: the
3191 : * path length from the pre state includes the pre-to-initial
3192 : * transition, so it's one more than the actually matched string
3193 : * length. (We avoided counting the final-to-post transition
3194 : * within checkmatchall_recurse, but not this one.) This is why
3195 : * checkmatchall_recurse allows one more level of path length than
3196 : * might seem necessary. This decrement also takes care of
3197 : * converting checkmatchall_recurse's definition of "infinity" as
3198 : * "DUPINF+1" to our normal representation as "DUPINF".
3199 : */
3200 : assert(minmatch > 0); /* else pre and post states were adjacent */
3201 952 : nfa->minmatchall = minmatch - 1;
3202 952 : nfa->maxmatchall = maxmatch - 1;
3203 952 : nfa->flags |= MATCHALL;
3204 : }
3205 : }
3206 :
3207 : /* Clean up */
3208 10396 : for (i = 0; i < nfa->nstates; i++)
3209 : {
3210 9408 : if (haspaths[i] != NULL)
3211 8420 : FREE(haspaths[i]);
3212 : }
3213 988 : FREE(haspaths);
3214 : }
3215 :
3216 : /*
3217 : * checkmatchall_recurse - recursive search for checkmatchall
3218 : *
3219 : * s is the state to be examined in this recursion level.
3220 : * haspaths[] is an array of per-state exit path length arrays.
3221 : *
3222 : * We return true if the search was performed successfully, false if
3223 : * we had to fail because of multi-state loops or other internal reasons.
3224 : * (Because "dead" states that can't reach the post state have been
3225 : * eliminated, and we already verified that only RAINBOW and matching
3226 : * pseudocolor arcs exist, every state should have RAINBOW path(s) to
3227 : * the post state. Hence we take a false result from recursive calls
3228 : * as meaning that we'd better fail altogether, not just that that
3229 : * particular state can't reach the post state.)
3230 : *
3231 : * On success, we store a malloc'd result array in haspaths[s->no],
3232 : * showing the possible path lengths from s to the post state.
3233 : * Each state's haspath[] array is of length DUPINF+2. The entries from
3234 : * k = 0 to DUPINF are true if there is an all-RAINBOW path of length k
3235 : * from this state to the string end. haspath[DUPINF+1] is true if all
3236 : * path lengths >= DUPINF+1 are possible. (Situations that cannot be
3237 : * represented under these rules cause failure.)
3238 : *
3239 : * checkmatchall is responsible for eventually freeing the haspath[] arrays.
3240 : */
3241 : static bool
3242 8420 : checkmatchall_recurse(struct nfa *nfa, struct state *s, bool **haspaths)
3243 : {
3244 8420 : bool result = false;
3245 8420 : bool foundloop = false;
3246 : bool *haspath;
3247 : struct arc *a;
3248 :
3249 : /*
3250 : * Since this is recursive, it could be driven to stack overflow. But we
3251 : * need not treat that as a hard failure; just deem the NFA non-matchall.
3252 : */
3253 8420 : if (STACK_TOO_DEEP(nfa->v->re))
3254 0 : return false;
3255 :
3256 : /* In case the search takes a long time, check for cancel */
3257 8420 : INTERRUPT(nfa->v->re);
3258 :
3259 : /* Create a haspath array for this state */
3260 8420 : haspath = (bool *) MALLOC((DUPINF + 2) * sizeof(bool));
3261 8420 : if (haspath == NULL)
3262 0 : return false; /* again, treat as non-matchall */
3263 8420 : memset(haspath, 0, (DUPINF + 2) * sizeof(bool));
3264 :
3265 : /* Mark this state as being visited */
3266 : assert(s->tmp == NULL);
3267 8420 : s->tmp = s;
3268 :
3269 82884 : for (a = s->outs; a != NULL; a = a->outchain)
3270 : {
3271 74560 : if (a->co != RAINBOW)
3272 6420 : continue; /* ignore pseudocolor arcs */
3273 68140 : if (a->to == nfa->post)
3274 : {
3275 : /* We found an all-RAINBOW path to the post state */
3276 976 : result = true;
3277 :
3278 : /*
3279 : * Mark this state as being zero steps away from the string end
3280 : * (the transition to the post state isn't counted).
3281 : */
3282 976 : haspath[0] = true;
3283 : }
3284 67164 : else if (a->to == s)
3285 : {
3286 : /* We found a cycle of length 1, which we'll deal with below. */
3287 246 : foundloop = true;
3288 : }
3289 66918 : else if (a->to->tmp != NULL)
3290 : {
3291 : /* It's busy, so we found a cycle of length > 1, so fail. */
3292 12 : result = false;
3293 12 : break;
3294 : }
3295 : else
3296 : {
3297 : /* Consider paths forward through this to-state. */
3298 : bool *nexthaspath;
3299 : int i;
3300 :
3301 : /* If to-state was not already visited, recurse */
3302 66906 : if (haspaths[a->to->no] == NULL)
3303 : {
3304 7432 : result = checkmatchall_recurse(nfa, a->to, haspaths);
3305 : /* Fail if any recursive path fails */
3306 7432 : if (!result)
3307 72 : break;
3308 : }
3309 : else
3310 : {
3311 : /* The previous visit must have found path(s) to the end */
3312 59474 : result = true;
3313 : }
3314 : assert(a->to->tmp == NULL);
3315 66834 : nexthaspath = haspaths[a->to->no];
3316 : assert(nexthaspath != NULL);
3317 :
3318 : /*
3319 : * Now, for every path of length i from a->to to the string end,
3320 : * there is a path of length i + 1 from s to the string end.
3321 : */
3322 66834 : if (nexthaspath[DUPINF] != nexthaspath[DUPINF + 1])
3323 : {
3324 : /*
3325 : * a->to has a path of length exactly DUPINF, but not longer;
3326 : * or it has paths of all lengths > DUPINF but not one of
3327 : * exactly that length. In either case, we cannot represent
3328 : * the possible path lengths from s correctly, so fail.
3329 : */
3330 12 : result = false;
3331 12 : break;
3332 : }
3333 : /* Merge knowledge of these path lengths into what we have */
3334 17173254 : for (i = 0; i < DUPINF; i++)
3335 17106432 : haspath[i + 1] |= nexthaspath[i];
3336 : /* Infinity + 1 is still infinity */
3337 66822 : haspath[DUPINF + 1] |= nexthaspath[DUPINF + 1];
3338 : }
3339 : }
3340 :
3341 8420 : if (result && foundloop)
3342 : {
3343 : /*
3344 : * If there is a length-1 loop at this state, then find the shortest
3345 : * known path length to the end. The loop means that every larger
3346 : * path length is possible, too. (It doesn't matter whether any of
3347 : * the longer lengths were already known possible.)
3348 : */
3349 : int i;
3350 :
3351 312 : for (i = 0; i <= DUPINF; i++)
3352 : {
3353 312 : if (haspath[i])
3354 246 : break;
3355 : }
3356 63402 : for (i++; i <= DUPINF + 1; i++)
3357 63156 : haspath[i] = true;
3358 : }
3359 :
3360 : /* Report out the completed path length map */
3361 : assert(s->no < nfa->nstates);
3362 : assert(haspaths[s->no] == NULL);
3363 8420 : haspaths[s->no] = haspath;
3364 :
3365 : /* Mark state no longer busy */
3366 8420 : s->tmp = NULL;
3367 :
3368 8420 : return result;
3369 : }
3370 :
3371 : /*
3372 : * check_out_colors_match - subroutine for checkmatchall
3373 : *
3374 : * Check whether the set of states reachable from s by arcs of color co1
3375 : * is equivalent to the set reachable by arcs of color co2.
3376 : * checkmatchall already verified that all of the NFA's arcs are PLAIN,
3377 : * so we need not examine arc types here.
3378 : */
3379 : static bool
3380 3086 : check_out_colors_match(struct state *s, color co1, color co2)
3381 : {
3382 3086 : bool result = true;
3383 : struct arc *a;
3384 :
3385 : /*
3386 : * To do this in linear time, we assume that the NFA contains no duplicate
3387 : * arcs. Run through the out-arcs, marking states reachable by arcs of
3388 : * color co1. Run through again, un-marking states reachable by arcs of
3389 : * color co2; if we see a not-marked state, we know this co2 arc is
3390 : * unmatched. Then run through again, checking for still-marked states,
3391 : * and in any case leaving all the tmp fields reset to NULL.
3392 : */
3393 18906 : for (a = s->outs; a != NULL; a = a->outchain)
3394 : {
3395 15820 : if (a->co == co1)
3396 : {
3397 : assert(a->to->tmp == NULL);
3398 5028 : a->to->tmp = a->to;
3399 : }
3400 : }
3401 18906 : for (a = s->outs; a != NULL; a = a->outchain)
3402 : {
3403 15820 : if (a->co == co2)
3404 : {
3405 5396 : if (a->to->tmp != NULL)
3406 5024 : a->to->tmp = NULL;
3407 : else
3408 372 : result = false; /* unmatched co2 arc */
3409 : }
3410 : }
3411 18906 : for (a = s->outs; a != NULL; a = a->outchain)
3412 : {
3413 15820 : if (a->co == co1)
3414 : {
3415 5028 : if (a->to->tmp != NULL)
3416 : {
3417 4 : result = false; /* unmatched co1 arc */
3418 4 : a->to->tmp = NULL;
3419 : }
3420 : }
3421 : }
3422 3086 : return result;
3423 : }
3424 :
3425 : /*
3426 : * check_in_colors_match - subroutine for checkmatchall
3427 : *
3428 : * Check whether the set of states that can reach s by arcs of color co1
3429 : * is equivalent to the set that can reach s by arcs of color co2.
3430 : * checkmatchall already verified that all of the NFA's arcs are PLAIN,
3431 : * so we need not examine arc types here.
3432 : */
3433 : static bool
3434 2400 : check_in_colors_match(struct state *s, color co1, color co2)
3435 : {
3436 2400 : bool result = true;
3437 : struct arc *a;
3438 :
3439 : /*
3440 : * Identical algorithm to check_out_colors_match, except examine the
3441 : * from-states of s' inarcs.
3442 : */
3443 8876 : for (a = s->ins; a != NULL; a = a->inchain)
3444 : {
3445 6476 : if (a->co == co1)
3446 : {
3447 : assert(a->from->tmp == NULL);
3448 2008 : a->from->tmp = a->from;
3449 : }
3450 : }
3451 8876 : for (a = s->ins; a != NULL; a = a->inchain)
3452 : {
3453 6476 : if (a->co == co2)
3454 : {
3455 2234 : if (a->from->tmp != NULL)
3456 2004 : a->from->tmp = NULL;
3457 : else
3458 230 : result = false; /* unmatched co2 arc */
3459 : }
3460 : }
3461 8876 : for (a = s->ins; a != NULL; a = a->inchain)
3462 : {
3463 6476 : if (a->co == co1)
3464 : {
3465 2008 : if (a->from->tmp != NULL)
3466 : {
3467 4 : result = false; /* unmatched co1 arc */
3468 4 : a->from->tmp = NULL;
3469 : }
3470 : }
3471 : }
3472 2400 : return result;
3473 : }
3474 :
3475 : /*
3476 : * compact - construct the compact representation of an NFA
3477 : */
3478 : static void
3479 18602 : compact(struct nfa *nfa,
3480 : struct cnfa *cnfa)
3481 : {
3482 : struct state *s;
3483 : struct arc *a;
3484 : size_t nstates;
3485 : size_t narcs;
3486 : struct carc *ca;
3487 : struct carc *first;
3488 :
3489 : assert(!NISERR());
3490 :
3491 18602 : nstates = 0;
3492 18602 : narcs = 0;
3493 233508 : for (s = nfa->states; s != NULL; s = s->next)
3494 : {
3495 214906 : nstates++;
3496 214906 : narcs += s->nouts + 1; /* need one extra for endmarker */
3497 : }
3498 :
3499 18602 : cnfa->stflags = (char *) MALLOC(nstates * sizeof(char));
3500 18602 : cnfa->states = (struct carc **) MALLOC(nstates * sizeof(struct carc *));
3501 18602 : cnfa->arcs = (struct carc *) MALLOC(narcs * sizeof(struct carc));
3502 18602 : if (cnfa->stflags == NULL || cnfa->states == NULL || cnfa->arcs == NULL)
3503 : {
3504 0 : if (cnfa->stflags != NULL)
3505 0 : FREE(cnfa->stflags);
3506 0 : if (cnfa->states != NULL)
3507 0 : FREE(cnfa->states);
3508 0 : if (cnfa->arcs != NULL)
3509 0 : FREE(cnfa->arcs);
3510 0 : NERR(REG_ESPACE);
3511 0 : return;
3512 : }
3513 18602 : cnfa->nstates = nstates;
3514 18602 : cnfa->pre = nfa->pre->no;
3515 18602 : cnfa->post = nfa->post->no;
3516 18602 : cnfa->bos[0] = nfa->bos[0];
3517 18602 : cnfa->bos[1] = nfa->bos[1];
3518 18602 : cnfa->eos[0] = nfa->eos[0];
3519 18602 : cnfa->eos[1] = nfa->eos[1];
3520 18602 : cnfa->ncolors = maxcolor(nfa->cm) + 1;
3521 18602 : cnfa->flags = nfa->flags;
3522 18602 : cnfa->minmatchall = nfa->minmatchall;
3523 18602 : cnfa->maxmatchall = nfa->maxmatchall;
3524 :
3525 18602 : ca = cnfa->arcs;
3526 233508 : for (s = nfa->states; s != NULL; s = s->next)
3527 : {
3528 : assert((size_t) s->no < nstates);
3529 214906 : cnfa->stflags[s->no] = 0;
3530 214906 : cnfa->states[s->no] = ca;
3531 214906 : first = ca;
3532 1745312 : for (a = s->outs; a != NULL; a = a->outchain)
3533 1530406 : switch (a->type)
3534 : {
3535 1530208 : case PLAIN:
3536 1530208 : ca->co = a->co;
3537 1530208 : ca->to = a->to->no;
3538 1530208 : ca++;
3539 1530208 : break;
3540 198 : case LACON:
3541 : assert(s->no != cnfa->pre);
3542 : assert(a->co >= 0);
3543 198 : ca->co = (color) (cnfa->ncolors + a->co);
3544 198 : ca->to = a->to->no;
3545 198 : ca++;
3546 198 : cnfa->flags |= HASLACONS;
3547 198 : break;
3548 0 : default:
3549 0 : NERR(REG_ASSERT);
3550 0 : return;
3551 : }
3552 214906 : carcsort(first, ca - first);
3553 214906 : ca->co = COLORLESS;
3554 214906 : ca->to = 0;
3555 214906 : ca++;
3556 : }
3557 : assert(ca == &cnfa->arcs[narcs]);
3558 : assert(cnfa->nstates != 0);
3559 :
3560 : /* mark no-progress states */
3561 80164 : for (a = nfa->pre->outs; a != NULL; a = a->outchain)
3562 61562 : cnfa->stflags[a->to->no] = CNFA_NOPROGRESS;
3563 18602 : cnfa->stflags[nfa->pre->no] = CNFA_NOPROGRESS;
3564 : }
3565 :
3566 : /*
3567 : * carcsort - sort compacted-NFA arcs by color
3568 : */
3569 : static void
3570 214906 : carcsort(struct carc *first, size_t n)
3571 : {
3572 214906 : if (n > 1)
3573 46784 : qsort(first, n, sizeof(struct carc), carc_cmp);
3574 214906 : }
3575 :
3576 : static int
3577 15458278 : carc_cmp(const void *a, const void *b)
3578 : {
3579 15458278 : const struct carc *aa = (const struct carc *) a;
3580 15458278 : const struct carc *bb = (const struct carc *) b;
3581 :
3582 15458278 : if (aa->co < bb->co)
3583 138126 : return -1;
3584 15320152 : if (aa->co > bb->co)
3585 210088 : return +1;
3586 15110064 : if (aa->to < bb->to)
3587 10362984 : return -1;
3588 4747080 : if (aa->to > bb->to)
3589 4747080 : return +1;
3590 : /* This is unreached, since there should be no duplicate arcs now: */
3591 0 : return 0;
3592 : }
3593 :
3594 : /*
3595 : * freecnfa - free a compacted NFA
3596 : */
3597 : static void
3598 4288 : freecnfa(struct cnfa *cnfa)
3599 : {
3600 : assert(!NULLCNFA(*cnfa)); /* not empty already */
3601 4288 : FREE(cnfa->stflags);
3602 4288 : FREE(cnfa->states);
3603 4288 : FREE(cnfa->arcs);
3604 4288 : ZAPCNFA(*cnfa);
3605 4288 : }
3606 :
3607 : /*
3608 : * dumpnfa - dump an NFA in human-readable form
3609 : */
3610 : static void
3611 0 : dumpnfa(struct nfa *nfa,
3612 : FILE *f)
3613 : {
3614 : #ifdef REG_DEBUG
3615 : struct state *s;
3616 : int nstates = 0;
3617 : int narcs = 0;
3618 :
3619 : fprintf(f, "pre %d, post %d", nfa->pre->no, nfa->post->no);
3620 : if (nfa->bos[0] != COLORLESS)
3621 : fprintf(f, ", bos [%ld]", (long) nfa->bos[0]);
3622 : if (nfa->bos[1] != COLORLESS)
3623 : fprintf(f, ", bol [%ld]", (long) nfa->bos[1]);
3624 : if (nfa->eos[0] != COLORLESS)
3625 : fprintf(f, ", eos [%ld]", (long) nfa->eos[0]);
3626 : if (nfa->eos[1] != COLORLESS)
3627 : fprintf(f, ", eol [%ld]", (long) nfa->eos[1]);
3628 : if (nfa->flags & HASLACONS)
3629 : fprintf(f, ", haslacons");
3630 : if (nfa->flags & MATCHALL)
3631 : {
3632 : fprintf(f, ", minmatchall %d", nfa->minmatchall);
3633 : if (nfa->maxmatchall == DUPINF)
3634 : fprintf(f, ", maxmatchall inf");
3635 : else
3636 : fprintf(f, ", maxmatchall %d", nfa->maxmatchall);
3637 : }
3638 : fprintf(f, "\n");
3639 : for (s = nfa->states; s != NULL; s = s->next)
3640 : {
3641 : dumpstate(s, f);
3642 : nstates++;
3643 : narcs += s->nouts;
3644 : }
3645 : fprintf(f, "total of %d states, %d arcs\n", nstates, narcs);
3646 : if (nfa->parent == NULL)
3647 : dumpcolors(nfa->cm, f);
3648 : fflush(f);
3649 : #endif
3650 0 : }
3651 :
3652 : #ifdef REG_DEBUG /* subordinates of dumpnfa */
3653 :
3654 : /*
3655 : * dumpstate - dump an NFA state in human-readable form
3656 : */
3657 : static void
3658 : dumpstate(struct state *s,
3659 : FILE *f)
3660 : {
3661 : struct arc *a;
3662 :
3663 : fprintf(f, "%d%s%c", s->no, (s->tmp != NULL) ? "T" : "",
3664 : (s->flag) ? s->flag : '.');
3665 : if (s->prev != NULL && s->prev->next != s)
3666 : fprintf(f, "\tstate chain bad\n");
3667 : if (s->nouts == 0)
3668 : fprintf(f, "\tno out arcs\n");
3669 : else
3670 : dumparcs(s, f);
3671 : for (a = s->ins; a != NULL; a = a->inchain)
3672 : {
3673 : if (a->to != s)
3674 : fprintf(f, "\tlink from %d to %d on %d's in-chain\n",
3675 : a->from->no, a->to->no, s->no);
3676 : }
3677 : fflush(f);
3678 : }
3679 :
3680 : /*
3681 : * dumparcs - dump out-arcs in human-readable form
3682 : */
3683 : static void
3684 : dumparcs(struct state *s,
3685 : FILE *f)
3686 : {
3687 : int pos;
3688 : struct arc *a;
3689 :
3690 : /* printing oldest arcs first is usually clearer */
3691 : a = s->outs;
3692 : assert(a != NULL);
3693 : while (a->outchain != NULL)
3694 : a = a->outchain;
3695 : pos = 1;
3696 : do
3697 : {
3698 : dumparc(a, s, f);
3699 : if (pos == 5)
3700 : {
3701 : fprintf(f, "\n");
3702 : pos = 1;
3703 : }
3704 : else
3705 : pos++;
3706 : a = a->outchainRev;
3707 : } while (a != NULL);
3708 : if (pos != 1)
3709 : fprintf(f, "\n");
3710 : }
3711 :
3712 : /*
3713 : * dumparc - dump one outarc in readable form, including prefixing tab
3714 : */
3715 : static void
3716 : dumparc(struct arc *a,
3717 : struct state *s,
3718 : FILE *f)
3719 : {
3720 : struct arc *aa;
3721 :
3722 : fprintf(f, "\t");
3723 : switch (a->type)
3724 : {
3725 : case PLAIN:
3726 : if (a->co == RAINBOW)
3727 : fprintf(f, "[*]");
3728 : else
3729 : fprintf(f, "[%ld]", (long) a->co);
3730 : break;
3731 : case AHEAD:
3732 : if (a->co == RAINBOW)
3733 : fprintf(f, ">*>");
3734 : else
3735 : fprintf(f, ">%ld>", (long) a->co);
3736 : break;
3737 : case BEHIND:
3738 : if (a->co == RAINBOW)
3739 : fprintf(f, "<*<");
3740 : else
3741 : fprintf(f, "<%ld<", (long) a->co);
3742 : break;
3743 : case LACON:
3744 : fprintf(f, ":%ld:", (long) a->co);
3745 : break;
3746 : case '^':
3747 : case '$':
3748 : fprintf(f, "%c%d", a->type, (int) a->co);
3749 : break;
3750 : case EMPTY:
3751 : break;
3752 : default:
3753 : fprintf(f, "0x%x/0%lo", a->type, (long) a->co);
3754 : break;
3755 : }
3756 : if (a->from != s)
3757 : fprintf(f, "?%d?", a->from->no);
3758 : for (aa = a->from->outs; aa != NULL; aa = aa->outchain)
3759 : if (aa == a)
3760 : break; /* NOTE BREAK OUT */
3761 : if (aa == NULL)
3762 : fprintf(f, "?!?"); /* missing from out-chain */
3763 : fprintf(f, "->");
3764 : if (a->to == NULL)
3765 : {
3766 : fprintf(f, "NULL");
3767 : return;
3768 : }
3769 : fprintf(f, "%d", a->to->no);
3770 : for (aa = a->to->ins; aa != NULL; aa = aa->inchain)
3771 : if (aa == a)
3772 : break; /* NOTE BREAK OUT */
3773 : if (aa == NULL)
3774 : fprintf(f, "?!?"); /* missing from in-chain */
3775 : }
3776 : #endif /* REG_DEBUG */
3777 :
3778 : /*
3779 : * dumpcnfa - dump a compacted NFA in human-readable form
3780 : */
3781 : #ifdef REG_DEBUG
3782 : static void
3783 : dumpcnfa(struct cnfa *cnfa,
3784 : FILE *f)
3785 : {
3786 : int st;
3787 :
3788 : fprintf(f, "pre %d, post %d", cnfa->pre, cnfa->post);
3789 : if (cnfa->bos[0] != COLORLESS)
3790 : fprintf(f, ", bos [%ld]", (long) cnfa->bos[0]);
3791 : if (cnfa->bos[1] != COLORLESS)
3792 : fprintf(f, ", bol [%ld]", (long) cnfa->bos[1]);
3793 : if (cnfa->eos[0] != COLORLESS)
3794 : fprintf(f, ", eos [%ld]", (long) cnfa->eos[0]);
3795 : if (cnfa->eos[1] != COLORLESS)
3796 : fprintf(f, ", eol [%ld]", (long) cnfa->eos[1]);
3797 : if (cnfa->flags & HASLACONS)
3798 : fprintf(f, ", haslacons");
3799 : if (cnfa->flags & MATCHALL)
3800 : {
3801 : fprintf(f, ", minmatchall %d", cnfa->minmatchall);
3802 : if (cnfa->maxmatchall == DUPINF)
3803 : fprintf(f, ", maxmatchall inf");
3804 : else
3805 : fprintf(f, ", maxmatchall %d", cnfa->maxmatchall);
3806 : }
3807 : fprintf(f, "\n");
3808 : for (st = 0; st < cnfa->nstates; st++)
3809 : dumpcstate(st, cnfa, f);
3810 : fflush(f);
3811 : }
3812 : #endif
3813 :
3814 : #ifdef REG_DEBUG /* subordinates of dumpcnfa */
3815 :
3816 : /*
3817 : * dumpcstate - dump a compacted-NFA state in human-readable form
3818 : */
3819 : static void
3820 : dumpcstate(int st,
3821 : struct cnfa *cnfa,
3822 : FILE *f)
3823 : {
3824 : struct carc *ca;
3825 : int pos;
3826 :
3827 : fprintf(f, "%d%s", st, (cnfa->stflags[st] & CNFA_NOPROGRESS) ? ":" : ".");
3828 : pos = 1;
3829 : for (ca = cnfa->states[st]; ca->co != COLORLESS; ca++)
3830 : {
3831 : if (ca->co == RAINBOW)
3832 : fprintf(f, "\t[*]->%d", ca->to);
3833 : else if (ca->co < cnfa->ncolors)
3834 : fprintf(f, "\t[%ld]->%d", (long) ca->co, ca->to);
3835 : else
3836 : fprintf(f, "\t:%ld:->%d", (long) (ca->co - cnfa->ncolors), ca->to);
3837 : if (pos == 5)
3838 : {
3839 : fprintf(f, "\n");
3840 : pos = 1;
3841 : }
3842 : else
3843 : pos++;
3844 : }
3845 : if (ca == cnfa->states[st] || pos != 1)
3846 : fprintf(f, "\n");
3847 : fflush(f);
3848 : }
3849 :
3850 : #endif /* REG_DEBUG */
|