1 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
2 * confused with the original package (see point 3 below). Thanks, Henry!
5 /* Additional note: this code is very heavily munged from Henry's version
6 * in places. In some spots I've traded clarity for efficiency, so don't
7 * blame Henry for some of the lack of readability.
10 /* $Header: regcomp.c,v 4.0 91/03/20 01:39:01 lwall Locked $
13 * Revision 4.0 91/03/20 01:39:01 lwall
19 * regcomp and regexec -- regsub and regerror are not used in perl
21 * Copyright (c) 1986 by University of Toronto.
22 * Written by Henry Spencer. Not derived from licensed software.
24 * Permission is granted to anyone to use this software for any
25 * purpose on any computer system, and to redistribute it freely,
26 * subject to the following restrictions:
28 * 1. The author is not responsible for the consequences of use of
29 * this software, no matter how awful, even if they arise
32 * 2. The origin of this software must not be misrepresented, either
33 * by explicit claim or by omission.
35 * 3. Altered versions must be plainly marked as such, and must not
36 * be misrepresented as being the original software.
39 **** Alterations to Henry's code are...
41 **** Copyright (c) 1989, Larry Wall
43 **** You may distribute under the terms of the GNU General Public License
44 **** as specified in the README file that comes with the perl 3.0 kit.
46 * Beware that some of this code is subtly aware of the way operator
47 * precedence is structured in regular expressions. Serious changes in
48 * regular-expression syntax might require a total rethink.
56 # if defined(BUGGY_MSC6)
57 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
58 # pragma optimize("a",off)
59 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
60 # pragma optimize("w",on )
61 # endif /* BUGGY_MSC6 */
68 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
69 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
70 ((*s) == '{' && regcurly(s)))
71 #define META "^$.[()|?+*\\"
74 * Flags to be passed up and down.
76 #define HASWIDTH 01 /* Known never to match null string. */
77 #define SIMPLE 02 /* Simple enough to be STAR/PLUS operand. */
78 #define SPSTART 04 /* Starts with * or +. */
79 #define WORST 0 /* Worst case. */
82 * Global work variables for regcomp().
84 static char *regprecomp; /* uncompiled string. */
85 static char *regparse; /* Input-scan pointer. */
86 static char *regxend; /* End of input for compile */
87 static int regnpar; /* () count. */
88 static char *regcode; /* Code-emit pointer; ®dummy = don't. */
89 static long regsize; /* Code size. */
91 static int regsawbracket; /* Did we do {d,d} trick? */
94 * Forward declarations for regcomp()'s friends.
96 STATIC int regcurly();
98 STATIC char *regbranch();
99 STATIC char *regpiece();
100 STATIC char *regatom();
101 STATIC char *regclass();
102 STATIC char *regnode();
103 STATIC char *reganode();
105 STATIC void reginsert();
106 STATIC void regtail();
107 STATIC void regoptail();
110 - regcomp - compile a regular expression into internal code
112 * We can't allocate space until we know how big the compiled form will be,
113 * but we can't compile it (and thus know how big it is) until we've got a
114 * place to put the code. So we cheat: we compile it twice, once with code
115 * generation turned off and size counting turned on, and once "for real".
116 * This also means that we don't allocate space until we are sure that the
117 * thing really will compile successfully, and we never have to move the
118 * code and thus invalidate pointers into it. (Note that it has to be in
119 * one piece because free() must be able to free it all.) [NB: not true in perl]
121 * Beware that the optimization-preparation code in here knows about some
122 * of the structure of the compiled regexp. [I'll say.]
125 regcomp(exp,xend,fold)
132 register STR *longish;
135 register char *first;
140 extern char *safemalloc();
141 extern char *savestr();
145 fatal("NULL regexp argument");
147 /* First pass: determine size, legality. */
151 regprecomp = nsavestr(exp,xend-exp);
157 if (reg(0, &flags) == NULL) {
158 Safefree(regprecomp);
163 /* Small enough for pointer-storage convention? */
164 if (regsize >= 32767L) /* Probably could be 65535L. */
165 FAIL("regexp too big");
167 /* Allocate space. */
168 Newc(1001, r, sizeof(regexp) + (unsigned)regsize, char, regexp);
170 FAIL("regexp out of space");
172 /* Second pass: emit code. */
174 bcopy(regprecomp,exp,xend-exp);
175 r->precomp = regprecomp;
179 regcode = r->program;
181 if (reg(0, &flags) == NULL)
184 /* Dig out information for optimizations. */
185 r->regstart = Nullstr; /* Worst-case defaults. */
187 r->regmust = Nullstr;
189 r->regstclass = Nullch;
190 scan = r->program+1; /* First BRANCH. */
191 if (OP(regnext(scan)) == END) {/* Only one top-level choice. */
192 scan = NEXTOPER(scan);
195 while (OP(first) == OPEN ||
196 (OP(first) == BRANCH && OP(regnext(first)) != BRANCH) ||
197 (OP(first) == PLUS) ||
198 (OP(first) == CURLY && ARG1(first) > 0) ) {
199 if (OP(first) == PLUS)
202 first += regarglen[OP(first)];
203 first = NEXTOPER(first);
206 /* Starting-point info. */
207 if (OP(first) == EXACTLY) {
209 str_make(OPERAND(first)+1,*OPERAND(first));
210 if (r->regstart->str_cur > !(sawstudy|fold))
211 fbmcompile(r->regstart,fold);
213 else if ((exp = index(simple,OP(first))) && exp > simple)
214 r->regstclass = first;
215 else if (OP(first) == BOUND || OP(first) == NBOUND)
216 r->regstclass = first;
217 else if (OP(first) == BOL ||
218 (OP(first) == STAR && OP(NEXTOPER(first)) == ANY) )
219 r->reganch = 1; /* kinda turn .* into ^.* */
220 r->reganch |= sawplus;
224 fprintf(stderr,"first %d next %d offset %d\n",
225 OP(first), OP(NEXTOPER(first)), first - scan);
228 * If there's something expensive in the r.e., find the
229 * longest literal string that must appear and make it the
230 * regmust. Resolve ties in favor of later strings, since
231 * the regstart check works with the beginning of the r.e.
232 * and avoiding duplication strengthens checking. Not a
233 * strong reason, but sufficient in the absence of others.
234 * [Now we resolve ties in favor of the earlier string if
235 * it happens that curback has been invalidated, since the
236 * earlier string may buy us something the later one won't.]
238 longish = str_make("",0);
239 longest = str_make("",0);
244 while (OP(scan) != END) {
245 if (OP(scan) == BRANCH) {
246 if (OP(regnext(scan)) == BRANCH) {
248 while (OP(scan) == BRANCH)
249 scan = regnext(scan);
251 else /* single branch is ok */
252 scan = NEXTOPER(scan);
254 if (OP(scan) == EXACTLY) {
258 while (OP(t = regnext(scan)) == CLOSE)
260 if (curback - backish == len) {
261 str_ncat(longish, OPERAND(first)+1,
263 len += *OPERAND(first);
264 curback += *OPERAND(first);
265 first = regnext(scan);
267 else if (*OPERAND(first) >= len + (curback >= 0)) {
268 len = *OPERAND(first);
269 str_nset(longish, OPERAND(first)+1,len);
272 first = regnext(scan);
275 curback += *OPERAND(first);
277 else if (index(varies,OP(scan))) {
280 if (longish->str_cur > longest->str_cur) {
281 str_sset(longest,longish);
284 str_nset(longish,"",0);
286 else if (index(simple,OP(scan))) {
289 if (longish->str_cur > longest->str_cur) {
290 str_sset(longest,longish);
293 str_nset(longish,"",0);
295 scan = regnext(scan);
298 /* Prefer earlier on tie, unless we can tail match latter */
300 if (longish->str_cur + (OP(first) == EOL) > longest->str_cur) {
301 str_sset(longest,longish);
305 str_nset(longish,"",0);
310 !fbminstr(r->regstart->str_ptr,
311 r->regstart->str_ptr + r->regstart->str_cur,
316 r->regmust = longest;
319 r->regback = backest;
321 > !(sawstudy || fold || OP(first) == EOL) )
322 fbmcompile(r->regmust,fold);
323 r->regmust->str_u.str_useful = 100;
324 if (OP(first) == EOL && longish->str_cur)
325 r->regmust->str_pok |= SP_TAIL;
334 r->do_folding = fold;
335 r->nparens = regnpar - 1;
336 New(1002, r->startp, regnpar, char*);
337 New(1002, r->endp, regnpar, char*);
346 - reg - regular expression, i.e. main body or parenthesized thing
348 * Caller must absorb opening parenthesis.
350 * Combining parenthesis handling with the base level of regular expression
351 * is a trifle forced, but the need to tie the tails of the branches to what
352 * follows makes it hard to avoid.
356 int paren; /* Parenthesized? */
361 register char *ender;
365 *flagp = HASWIDTH; /* Tentatively. */
367 /* Make an OPEN node, if parenthesized. */
371 ret = reganode(OPEN, parno);
375 /* Pick up the branches, linking them together. */
376 br = regbranch(&flags);
380 regtail(ret, br); /* OPEN -> first. */
383 if (!(flags&HASWIDTH))
385 *flagp |= flags&SPSTART;
386 while (*regparse == '|') {
388 br = regbranch(&flags);
391 regtail(ret, br); /* BRANCH -> BRANCH. */
392 if (!(flags&HASWIDTH))
394 *flagp |= flags&SPSTART;
397 /* Make a closing node, and hook it on the end. */
399 ender = reganode(CLOSE, parno);
401 ender = regnode(END);
404 /* Hook the tails of the branches to the closing node. */
405 for (br = ret; br != NULL; br = regnext(br))
406 regoptail(br, ender);
408 /* Check for proper termination. */
409 if (paren && *regparse++ != ')') {
410 FAIL("unmatched () in regexp");
411 } else if (!paren && regparse < regxend) {
412 if (*regparse == ')') {
413 FAIL("unmatched () in regexp");
415 FAIL("junk on end of regexp"); /* "Can't happen". */
423 - regbranch - one alternative of an | operator
425 * Implements the concatenation operator.
432 register char *chain;
433 register char *latest;
436 *flagp = WORST; /* Tentatively. */
438 ret = regnode(BRANCH);
440 while (regparse < regxend && *regparse != '|' && *regparse != ')') {
441 latest = regpiece(&flags);
444 *flagp |= flags&HASWIDTH;
445 if (chain == NULL) /* First piece. */
446 *flagp |= flags&SPSTART;
448 regtail(chain, latest);
451 if (chain == NULL) /* Loop ran zero times. */
452 (void) regnode(NOTHING);
458 - regpiece - something followed by possible [*+?]
460 * Note that the branching code sequences used for ? and the general cases
461 * of * and + are somewhat optimized: they use the same NOTHING node as
462 * both the endmarker for their branch list and the body of the last branch.
463 * It might seem that this node could be dispensed with entirely, but the
464 * endmarker role is not redundant.
474 char *origparse = regparse;
475 int orignpar = regnpar;
480 ret = regatom(&flags);
486 /* Here's a total kludge: if after the atom there's a {\d+,?\d*}
487 * then we decrement the first number by one and reset our
488 * parsing back to the beginning of the same atom. If the first number
489 * is down to 0, decrement the second number instead and fake up
490 * a ? after it. Given the way this compiler doesn't keep track
491 * of offsets on the first pass, this is the only way to replicate
492 * a piece of code. Sigh.
494 if (op == '{' && regcurly(regparse)) {
497 while (isdigit(*next) || *next == ',') {
506 if (*next == '}') { /* got one */
510 iter = atoi(regparse);
511 if (flags&SIMPLE) { /* we can do it right after all */
514 reginsert(CURLY, ret);
516 *flagp = (WORST|HASWIDTH);
522 if (tmp && tmp < iter)
523 fatal("Can't do {n,m} with n > m");
524 if (regcode != ®dummy) {
526 *(unsigned short *)(ret+3) = iter;
527 *(unsigned short *)(ret+5) = tmp;
529 ret[3] = iter >> 8; ret[4] = iter & 0377;
530 ret[5] = tmp >> 8; ret[6] = tmp & 0377;
536 regsawbracket++; /* remember we clobbered exp */
539 sprintf(regparse,"%.*d", max-regparse, iter - 1);
541 if (*max == ',' && max[1] != '}') {
542 if (atoi(max+1) <= 0)
543 fatal("Can't do {n,m} with n > m");
545 sprintf(max+1,"%.*d", next-(max+1), atoi(max+1) - 1);
548 if (iter != 1 || *max == ',') {
549 regparse = origparse; /* back up input pointer */
550 regnpar = orignpar; /* don't make more parens */
562 if (max == next) { /* any number more? */
564 op = '*'; /* fake up one with a star */
567 op = '?'; /* fake up optional atom */
569 sprintf(max,"%.*d", next-max, iter - 1);
574 regparse = origparse - 1; /* offset ++ below */
579 fatal("Can't do {n,0}");
582 fatal("Can't do {0}");
591 if (!(flags&HASWIDTH) && op != '?')
592 FAIL("regexp *+ operand could be empty");
593 *flagp = (op != '+') ? (WORST|SPSTART) : (WORST|HASWIDTH);
595 if (op == '*' && (flags&SIMPLE))
596 reginsert(STAR, ret);
597 else if (op == '*') {
598 /* Emit x* as (x&|), where & means "self". */
599 reginsert(BRANCH, ret); /* Either x */
600 regoptail(ret, regnode(BACK)); /* and loop */
601 regoptail(ret, ret); /* back */
602 regtail(ret, regnode(BRANCH)); /* or */
603 regtail(ret, regnode(NOTHING)); /* null. */
604 } else if (op == '+' && (flags&SIMPLE))
605 reginsert(PLUS, ret);
606 else if (op == '+') {
607 /* Emit x+ as x(&|), where & means "self". */
608 next = regnode(BRANCH); /* Either */
610 regtail(regnode(BACK), ret); /* loop back */
611 regtail(next, regnode(BRANCH)); /* or */
612 regtail(ret, regnode(NOTHING)); /* null. */
613 } else if (op == '?') {
614 /* Emit x? as (x|) */
615 reginsert(BRANCH, ret); /* Either x */
616 regtail(ret, regnode(BRANCH)); /* or */
617 next = regnode(NOTHING); /* null. */
619 regoptail(ret, next);
623 if (ISMULT2(regparse))
624 FAIL("nested *?+ in regexp");
630 - regatom - the lowest level
632 * Optimization: gobbles an entire sequence of ordinary characters so that
633 * it can turn them into a single node, which is smaller to store and
634 * faster to run. Backslashed characters are exceptions, each becoming a
635 * separate node; the code is simpler that way and it's not worth fixing.
637 * [Yes, it is worth fixing, some scripts can run twice the speed.]
646 *flagp = WORST; /* Tentatively. */
648 switch (*regparse++) {
657 *flagp |= HASWIDTH|SIMPLE;
661 *flagp |= HASWIDTH|SIMPLE;
664 ret = reg(1, &flags);
667 *flagp |= flags&(HASWIDTH|SPSTART);
671 FAIL("internal urp in regexp"); /* Supposed to be caught earlier. */
676 FAIL("?+* follows nothing in regexp");
681 ret = regnode(ALNUM);
682 *flagp |= HASWIDTH|SIMPLE;
686 ret = regnode(NALNUM);
687 *flagp |= HASWIDTH|SIMPLE;
691 ret = regnode(BOUND);
696 ret = regnode(NBOUND);
701 ret = regnode(SPACE);
702 *flagp |= HASWIDTH|SIMPLE;
706 ret = regnode(NSPACE);
707 *flagp |= HASWIDTH|SIMPLE;
711 ret = regnode(DIGIT);
712 *flagp |= HASWIDTH|SIMPLE;
716 ret = regnode(NDIGIT);
717 *flagp |= HASWIDTH|SIMPLE;
730 case '1': case '2': case '3': case '4':
731 case '5': case '6': case '7': case '8': case '9':
733 int num = atoi(regparse);
735 if (num > 9 && num >= regnpar)
738 ret = reganode(REF, num);
739 while (isascii(*regparse) && isdigit(*regparse))
746 if (regparse >= regxend)
747 FAIL("trailing \\ in regexp");
761 ret = regnode(EXACTLY);
762 regc(0); /* save spot for len */
763 for (len=0, p=regparse-1;
764 len < 127 && p < regxend;
814 ender = scanhex(++p, 2, &numlen);
821 ender = toupper(ender);
824 case '0': case '1': case '2': case '3':case '4':
825 case '5': case '6': case '7': case '8':case '9':
827 (isdigit(p[1]) && atoi(p) >= regnpar) ) {
828 ender = scanoct(p, 3, &numlen);
838 FAIL("trailing \\ in regexp");
849 if (regfold && isupper(ender))
850 ender = tolower(ender);
851 if (ISMULT2(p)) { /* Back off on ?+*. */
865 FAIL("internal disaster in regexp");
869 if (regcode != ®dummy)
885 if (regcode == ®dummy)
889 bits[c >> 3] &= ~(1 << (c & 7));
891 bits[c >> 3] |= (1 << (c & 7));
899 register int lastclass;
900 register int range = 0;
905 ret = regnode(ANYOF);
906 if (*regparse == '^') { /* Complement of range. */
913 for (class = 0; class < 32; class++)
915 if (*regparse == ']' || *regparse == '-')
916 goto skipcond; /* allow 1st char to be ] or - */
917 while (regparse < regxend && *regparse != ']') {
919 class = UCHARAT(regparse++);
921 class = UCHARAT(regparse++);
924 for (class = 'a'; class <= 'z'; class++)
925 regset(bits,def,class);
926 for (class = 'A'; class <= 'Z'; class++)
927 regset(bits,def,class);
928 for (class = '0'; class <= '9'; class++)
929 regset(bits,def,class);
930 regset(bits,def,'_');
934 regset(bits,def,' ');
935 regset(bits,def,'\t');
936 regset(bits,def,'\r');
937 regset(bits,def,'\f');
938 regset(bits,def,'\n');
942 for (class = '0'; class <= '9'; class++)
943 regset(bits,def,class);
968 class = scanhex(regparse, 2, &numlen);
974 class = toupper(class);
977 case '0': case '1': case '2': case '3': case '4':
978 case '5': case '6': case '7': case '8': case '9':
979 class = scanoct(--regparse, 3, &numlen);
985 if (lastclass > class)
986 FAIL("invalid [] range in regexp");
991 if (*regparse == '-' && regparse+1 < regxend &&
992 regparse[1] != ']') {
995 continue; /* do it next time */
998 for ( ; lastclass <= class; lastclass++) {
999 regset(bits,def,lastclass);
1000 if (regfold && isupper(lastclass))
1001 regset(bits,def,tolower(lastclass));
1005 if (*regparse != ']')
1006 FAIL("unmatched [] in regexp");
1012 - regnode - emit a node
1014 static char * /* Location. */
1022 if (ret == ®dummy) {
1033 if (!((long)ret & 1))
1039 *ptr++ = '\0'; /* Null "next" pointer. */
1047 - reganode - emit a node with an argument
1049 static char * /* Location. */
1058 if (ret == ®dummy) {
1069 if (!((long)ret & 1))
1075 *ptr++ = '\0'; /* Null "next" pointer. */
1078 *(unsigned short *)(ret+3) = arg;
1080 ret[3] = arg >> 8; ret[4] = arg & 0377;
1089 - regc - emit (if appropriate) a byte of code
1095 if (regcode != ®dummy)
1102 - reginsert - insert an operator in front of already-emitted operand
1104 * Means relocating the operand.
1113 register char *place;
1114 register offset = (op == CURLY ? 4 : 0);
1116 if (regcode == ®dummy) {
1118 regsize += 4 + offset;
1120 regsize += 3 + offset;
1127 regcode += 4 + offset;
1129 regcode += 3 + offset;
1135 place = opnd; /* Op node, where operand used to be. */
1139 while (offset-- > 0)
1144 - regtail - set the next-pointer at the end of a node chain
1151 register char *scan;
1152 register char *temp;
1153 register int offset;
1158 /* Find last node. */
1161 temp = regnext(scan);
1168 offset = val - scan;
1170 *(short*)(scan+1) = offset;
1175 if (OP(scan) == BACK)
1176 offset = scan - val;
1178 offset = val - scan;
1179 *(scan+1) = (offset>>8)&0377;
1180 *(scan+2) = offset&0377;
1185 - regoptail - regtail on operand of first argument; nop if operandless
1192 /* "Operandless" and "op != BRANCH" are synonymous in practice. */
1193 if (p == NULL || p == ®dummy || OP(p) != BRANCH)
1195 regtail(NEXTOPER(p), val);
1199 - regcurly - a little FSA that accepts {\d+,?\d*}
1223 - regdump - dump a regexp onto stderr in vaguely comprehensible form
1230 register char op = EXACTLY; /* Arbitrary non-END op. */
1231 register char *next;
1235 while (op != END) { /* While that wasn't END last time... */
1241 fprintf(stderr,"%2d%s", s-r->program, regprop(s)); /* Where, what. */
1244 if (next == NULL) /* Next ptr. */
1245 fprintf(stderr,"(0)");
1247 fprintf(stderr,"(%d)", (s-r->program)+(next-s));
1252 if (op == EXACTLY) {
1253 /* Literal string, where present. */
1255 while (*s != '\0') {
1261 (void)putchar('\n');
1264 /* Header fields of interest. */
1266 fprintf(stderr,"start `%s' ", r->regstart->str_ptr);
1268 fprintf(stderr,"stclass `%s' ", regprop(r->regstclass));
1270 fprintf(stderr,"anchored ");
1272 fprintf(stderr,"plus ");
1273 if (r->regmust != NULL)
1274 fprintf(stderr,"must have \"%s\" back %d ", r->regmust->str_ptr,
1276 fprintf(stderr,"\n");
1280 - regprop - printable representation of opcode
1288 (void) strcpy(buf, ":");
1343 (void)sprintf(buf+strlen(buf), "CURLY {%d,%d}",
1348 (void)sprintf(buf+strlen(buf), "REF%d", ARG1(op));
1352 (void)sprintf(buf+strlen(buf), "OPEN%d", ARG1(op));
1356 (void)sprintf(buf+strlen(buf), "CLOSE%d", ARG1(op));
1366 FAIL("corrupted regexp opcode");
1369 (void) strcat(buf, p);
1372 #endif /* DEBUGGING */
1378 Safefree(r->precomp);
1379 r->precomp = Nullch;
1382 Safefree(r->subbase);
1383 r->subbase = Nullch;
1386 str_free(r->regmust);
1387 r->regmust = Nullstr;
1390 str_free(r->regstart);
1391 r->regstart = Nullstr;
1393 Safefree(r->startp);