5 * "A fair jaw-cracker dwarf-language must be." --Samwise Gamgee
8 /* This file contains functions for compiling a regular expression. See
9 * also regexec.c which funnily enough, contains functions for executing
10 * a regular expression.
12 * This file is also copied at build time to ext/re/re_comp.c, where
13 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
14 * This causes the main functions to be compiled under new names and with
15 * debugging support added, which makes "use re 'debug'" work.
18 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
19 * confused with the original package (see point 3 below). Thanks, Henry!
22 /* Additional note: this code is very heavily munged from Henry's version
23 * in places. In some spots I've traded clarity for efficiency, so don't
24 * blame Henry for some of the lack of readability.
27 /* The names of the functions have been changed from regcomp and
28 * regexec to pregcomp and pregexec in order to avoid conflicts
29 * with the POSIX routines of the same names.
32 #ifdef PERL_EXT_RE_BUILD
37 * pregcomp and pregexec -- regsub and regerror are not used in perl
39 * Copyright (c) 1986 by University of Toronto.
40 * Written by Henry Spencer. Not derived from licensed software.
42 * Permission is granted to anyone to use this software for any
43 * purpose on any computer system, and to redistribute it freely,
44 * subject to the following restrictions:
46 * 1. The author is not responsible for the consequences of use of
47 * this software, no matter how awful, even if they arise
50 * 2. The origin of this software must not be misrepresented, either
51 * by explicit claim or by omission.
53 * 3. Altered versions must be plainly marked as such, and must not
54 * be misrepresented as being the original software.
57 **** Alterations to Henry's code are...
59 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
60 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, by Larry Wall and others
62 **** You may distribute under the terms of either the GNU General Public
63 **** License or the Artistic License, as specified in the README file.
66 * Beware that some of this code is subtly aware of the way operator
67 * precedence is structured in regular expressions. Serious changes in
68 * regular-expression syntax might require a total rethink.
71 #define PERL_IN_REGCOMP_C
74 #ifndef PERL_IN_XSUB_RE
79 #ifdef PERL_IN_XSUB_RE
90 # if defined(BUGGY_MSC6)
91 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
92 # pragma optimize("a",off)
93 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
94 # pragma optimize("w",on )
95 # endif /* BUGGY_MSC6 */
102 typedef struct RExC_state_t {
103 U32 flags; /* are we folding, multilining? */
104 char *precomp; /* uncompiled string. */
106 char *start; /* Start of input for compile */
107 char *end; /* End of input for compile */
108 char *parse; /* Input-scan pointer. */
109 I32 whilem_seen; /* number of WHILEM in this expr */
110 regnode *emit_start; /* Start of emitted-code area */
111 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
112 I32 naughty; /* How bad is this pattern? */
113 I32 sawback; /* Did we see \1, ...? */
115 I32 size; /* Code size. */
116 I32 npar; /* () count. */
120 regnode **parens; /* offsets of each paren */
122 HV *charnames; /* cache of named sequences */
123 HV *paren_names; /* Paren names */
125 char *starttry; /* -Dr: where regtry was called. */
126 #define RExC_starttry (pRExC_state->starttry)
129 const char *lastparse;
131 #define RExC_lastparse (pRExC_state->lastparse)
132 #define RExC_lastnum (pRExC_state->lastnum)
136 #define RExC_flags (pRExC_state->flags)
137 #define RExC_precomp (pRExC_state->precomp)
138 #define RExC_rx (pRExC_state->rx)
139 #define RExC_start (pRExC_state->start)
140 #define RExC_end (pRExC_state->end)
141 #define RExC_parse (pRExC_state->parse)
142 #define RExC_whilem_seen (pRExC_state->whilem_seen)
143 #define RExC_offsets (pRExC_state->rx->offsets) /* I am not like the others */
144 #define RExC_emit (pRExC_state->emit)
145 #define RExC_emit_start (pRExC_state->emit_start)
146 #define RExC_naughty (pRExC_state->naughty)
147 #define RExC_sawback (pRExC_state->sawback)
148 #define RExC_seen (pRExC_state->seen)
149 #define RExC_size (pRExC_state->size)
150 #define RExC_npar (pRExC_state->npar)
151 #define RExC_extralen (pRExC_state->extralen)
152 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
153 #define RExC_seen_evals (pRExC_state->seen_evals)
154 #define RExC_utf8 (pRExC_state->utf8)
155 #define RExC_charnames (pRExC_state->charnames)
156 #define RExC_parens (pRExC_state->parens)
157 #define RExC_paren_names (pRExC_state->paren_names)
159 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
160 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
161 ((*s) == '{' && regcurly(s)))
164 #undef SPSTART /* dratted cpp namespace... */
167 * Flags to be passed up and down.
169 #define WORST 0 /* Worst case. */
170 #define HASWIDTH 0x1 /* Known to match non-null strings. */
171 #define SIMPLE 0x2 /* Simple enough to be STAR/PLUS operand. */
172 #define SPSTART 0x4 /* Starts with * or +. */
173 #define TRYAGAIN 0x8 /* Weeded out a declaration. */
175 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
177 /* whether trie related optimizations are enabled */
178 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
179 #define TRIE_STUDY_OPT
180 #define FULL_TRIE_STUDY
185 /* About scan_data_t.
187 During optimisation we recurse through the regexp program performing
188 various inplace (keyhole style) optimisations. In addition study_chunk
189 and scan_commit populate this data structure with information about
190 what strings MUST appear in the pattern. We look for the longest
191 string that must appear for at a fixed location, and we look for the
192 longest string that may appear at a floating location. So for instance
197 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
198 strings (because they follow a .* construct). study_chunk will identify
199 both FOO and BAR as being the longest fixed and floating strings respectively.
201 The strings can be composites, for instance
205 will result in a composite fixed substring 'foo'.
207 For each string some basic information is maintained:
209 - offset or min_offset
210 This is the position the string must appear at, or not before.
211 It also implicitly (when combined with minlenp) tells us how many
212 character must match before the string we are searching.
213 Likewise when combined with minlenp and the length of the string
214 tells us how many characters must appear after the string we have
218 Only used for floating strings. This is the rightmost point that
219 the string can appear at. Ifset to I32 max it indicates that the
220 string can occur infinitely far to the right.
223 A pointer to the minimum length of the pattern that the string
224 was found inside. This is important as in the case of positive
225 lookahead or positive lookbehind we can have multiple patterns
230 The minimum length of the pattern overall is 3, the minimum length
231 of the lookahead part is 3, but the minimum length of the part that
232 will actually match is 1. So 'FOO's minimum length is 3, but the
233 minimum length for the F is 1. This is important as the minimum length
234 is used to determine offsets in front of and behind the string being
235 looked for. Since strings can be composites this is the length of the
236 pattern at the time it was commited with a scan_commit. Note that
237 the length is calculated by study_chunk, so that the minimum lengths
238 are not known until the full pattern has been compiled, thus the
239 pointer to the value.
243 In the case of lookbehind the string being searched for can be
244 offset past the start point of the final matching string.
245 If this value was just blithely removed from the min_offset it would
246 invalidate some of the calculations for how many chars must match
247 before or after (as they are derived from min_offset and minlen and
248 the length of the string being searched for).
249 When the final pattern is compiled and the data is moved from the
250 scan_data_t structure into the regexp structure the information
251 about lookbehind is factored in, with the information that would
252 have been lost precalculated in the end_shift field for the
255 The fields pos_min and pos_delta are used to store the minimum offset
256 and the delta to the maximum offset at the current point in the pattern.
260 typedef struct scan_data_t {
261 /*I32 len_min; unused */
262 /*I32 len_delta; unused */
266 I32 last_end; /* min value, <0 unless valid. */
269 SV **longest; /* Either &l_fixed, or &l_float. */
270 SV *longest_fixed; /* longest fixed string found in pattern */
271 I32 offset_fixed; /* offset where it starts */
272 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
273 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
274 SV *longest_float; /* longest floating string found in pattern */
275 I32 offset_float_min; /* earliest point in string it can appear */
276 I32 offset_float_max; /* latest point in string it can appear */
277 I32 *minlen_float; /* pointer to the minlen relevent to the string */
278 I32 lookbehind_float; /* is the position of the string modified by LB */
282 struct regnode_charclass_class *start_class;
286 * Forward declarations for pregcomp()'s friends.
289 static const scan_data_t zero_scan_data =
290 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
292 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
293 #define SF_BEFORE_SEOL 0x0001
294 #define SF_BEFORE_MEOL 0x0002
295 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
296 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
299 # define SF_FIX_SHIFT_EOL (0+2)
300 # define SF_FL_SHIFT_EOL (0+4)
302 # define SF_FIX_SHIFT_EOL (+2)
303 # define SF_FL_SHIFT_EOL (+4)
306 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
307 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
309 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
310 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
311 #define SF_IS_INF 0x0040
312 #define SF_HAS_PAR 0x0080
313 #define SF_IN_PAR 0x0100
314 #define SF_HAS_EVAL 0x0200
315 #define SCF_DO_SUBSTR 0x0400
316 #define SCF_DO_STCLASS_AND 0x0800
317 #define SCF_DO_STCLASS_OR 0x1000
318 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
319 #define SCF_WHILEM_VISITED_POS 0x2000
321 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
324 #define UTF (RExC_utf8 != 0)
325 #define LOC ((RExC_flags & PMf_LOCALE) != 0)
326 #define FOLD ((RExC_flags & PMf_FOLD) != 0)
328 #define OOB_UNICODE 12345678
329 #define OOB_NAMEDCLASS -1
331 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
332 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
335 /* length of regex to show in messages that don't mark a position within */
336 #define RegexLengthToShowInErrorMessages 127
339 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
340 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
341 * op/pragma/warn/regcomp.
343 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
344 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
346 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
349 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
350 * arg. Show regex, up to a maximum length. If it's too long, chop and add
353 #define FAIL(msg) STMT_START { \
354 const char *ellipses = ""; \
355 IV len = RExC_end - RExC_precomp; \
358 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
359 if (len > RegexLengthToShowInErrorMessages) { \
360 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
361 len = RegexLengthToShowInErrorMessages - 10; \
364 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
365 msg, (int)len, RExC_precomp, ellipses); \
369 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
371 #define Simple_vFAIL(m) STMT_START { \
372 const IV offset = RExC_parse - RExC_precomp; \
373 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
374 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
378 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
380 #define vFAIL(m) STMT_START { \
382 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
387 * Like Simple_vFAIL(), but accepts two arguments.
389 #define Simple_vFAIL2(m,a1) STMT_START { \
390 const IV offset = RExC_parse - RExC_precomp; \
391 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
392 (int)offset, RExC_precomp, RExC_precomp + offset); \
396 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
398 #define vFAIL2(m,a1) STMT_START { \
400 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
401 Simple_vFAIL2(m, a1); \
406 * Like Simple_vFAIL(), but accepts three arguments.
408 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
409 const IV offset = RExC_parse - RExC_precomp; \
410 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
411 (int)offset, RExC_precomp, RExC_precomp + offset); \
415 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
417 #define vFAIL3(m,a1,a2) STMT_START { \
419 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
420 Simple_vFAIL3(m, a1, a2); \
424 * Like Simple_vFAIL(), but accepts four arguments.
426 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
427 const IV offset = RExC_parse - RExC_precomp; \
428 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
429 (int)offset, RExC_precomp, RExC_precomp + offset); \
432 #define vWARN(loc,m) STMT_START { \
433 const IV offset = loc - RExC_precomp; \
434 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
435 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
438 #define vWARNdep(loc,m) STMT_START { \
439 const IV offset = loc - RExC_precomp; \
440 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
441 "%s" REPORT_LOCATION, \
442 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
446 #define vWARN2(loc, m, a1) STMT_START { \
447 const IV offset = loc - RExC_precomp; \
448 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
449 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
452 #define vWARN3(loc, m, a1, a2) STMT_START { \
453 const IV offset = loc - RExC_precomp; \
454 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
455 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
458 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
459 const IV offset = loc - RExC_precomp; \
460 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
461 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
464 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
465 const IV offset = loc - RExC_precomp; \
466 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
467 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
471 /* Allow for side effects in s */
472 #define REGC(c,s) STMT_START { \
473 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
476 /* Macros for recording node offsets. 20001227 mjd@plover.com
477 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
478 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
479 * Element 0 holds the number n.
480 * Position is 1 indexed.
483 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
485 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
486 __LINE__, (node), (int)(byte))); \
488 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
490 RExC_offsets[2*(node)-1] = (byte); \
495 #define Set_Node_Offset(node,byte) \
496 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
497 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
499 #define Set_Node_Length_To_R(node,len) STMT_START { \
501 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
502 __LINE__, (int)(node), (int)(len))); \
504 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
506 RExC_offsets[2*(node)] = (len); \
511 #define Set_Node_Length(node,len) \
512 Set_Node_Length_To_R((node)-RExC_emit_start, len)
513 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
514 #define Set_Node_Cur_Length(node) \
515 Set_Node_Length(node, RExC_parse - parse_start)
517 /* Get offsets and lengths */
518 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
519 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
521 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
522 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
523 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
527 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
528 #define EXPERIMENTAL_INPLACESCAN
531 #define DEBUG_STUDYDATA(data,depth) \
532 DEBUG_OPTIMISE_MORE_r(if(data){ \
533 PerlIO_printf(Perl_debug_log, \
534 "%*s"/* Len:%"IVdf"/%"IVdf" */" Pos:%"IVdf"/%"IVdf \
535 " Flags: %"IVdf" Whilem_c: %"IVdf" Lcp: %"IVdf" ", \
536 (int)(depth)*2, "", \
537 (IV)((data)->pos_min), \
538 (IV)((data)->pos_delta), \
539 (IV)((data)->flags), \
540 (IV)((data)->whilem_c), \
541 (IV)((data)->last_closep ? *((data)->last_closep) : -1) \
543 if ((data)->last_found) \
544 PerlIO_printf(Perl_debug_log, \
545 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
546 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
547 SvPVX_const((data)->last_found), \
548 (IV)((data)->last_end), \
549 (IV)((data)->last_start_min), \
550 (IV)((data)->last_start_max), \
551 ((data)->longest && \
552 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
553 SvPVX_const((data)->longest_fixed), \
554 (IV)((data)->offset_fixed), \
555 ((data)->longest && \
556 (data)->longest==&((data)->longest_float)) ? "*" : "", \
557 SvPVX_const((data)->longest_float), \
558 (IV)((data)->offset_float_min), \
559 (IV)((data)->offset_float_max) \
561 PerlIO_printf(Perl_debug_log,"\n"); \
564 static void clear_re(pTHX_ void *r);
566 /* Mark that we cannot extend a found fixed substring at this point.
567 Update the longest found anchored substring and the longest found
568 floating substrings if needed. */
571 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp)
573 const STRLEN l = CHR_SVLEN(data->last_found);
574 const STRLEN old_l = CHR_SVLEN(*data->longest);
575 GET_RE_DEBUG_FLAGS_DECL;
577 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
578 SvSetMagicSV(*data->longest, data->last_found);
579 if (*data->longest == data->longest_fixed) {
580 data->offset_fixed = l ? data->last_start_min : data->pos_min;
581 if (data->flags & SF_BEFORE_EOL)
583 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
585 data->flags &= ~SF_FIX_BEFORE_EOL;
586 data->minlen_fixed=minlenp;
587 data->lookbehind_fixed=0;
590 data->offset_float_min = l ? data->last_start_min : data->pos_min;
591 data->offset_float_max = (l
592 ? data->last_start_max
593 : data->pos_min + data->pos_delta);
594 if ((U32)data->offset_float_max > (U32)I32_MAX)
595 data->offset_float_max = I32_MAX;
596 if (data->flags & SF_BEFORE_EOL)
598 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
600 data->flags &= ~SF_FL_BEFORE_EOL;
601 data->minlen_float=minlenp;
602 data->lookbehind_float=0;
605 SvCUR_set(data->last_found, 0);
607 SV * const sv = data->last_found;
608 if (SvUTF8(sv) && SvMAGICAL(sv)) {
609 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
615 data->flags &= ~SF_BEFORE_EOL;
616 DEBUG_STUDYDATA(data,0);
619 /* Can match anything (initialization) */
621 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
623 ANYOF_CLASS_ZERO(cl);
624 ANYOF_BITMAP_SETALL(cl);
625 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
627 cl->flags |= ANYOF_LOCALE;
630 /* Can match anything (initialization) */
632 S_cl_is_anything(const struct regnode_charclass_class *cl)
636 for (value = 0; value <= ANYOF_MAX; value += 2)
637 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
639 if (!(cl->flags & ANYOF_UNICODE_ALL))
641 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
646 /* Can match anything (initialization) */
648 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
650 Zero(cl, 1, struct regnode_charclass_class);
652 cl_anything(pRExC_state, cl);
656 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
658 Zero(cl, 1, struct regnode_charclass_class);
660 cl_anything(pRExC_state, cl);
662 cl->flags |= ANYOF_LOCALE;
665 /* 'And' a given class with another one. Can create false positives */
666 /* We assume that cl is not inverted */
668 S_cl_and(struct regnode_charclass_class *cl,
669 const struct regnode_charclass_class *and_with)
671 if (!(and_with->flags & ANYOF_CLASS)
672 && !(cl->flags & ANYOF_CLASS)
673 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
674 && !(and_with->flags & ANYOF_FOLD)
675 && !(cl->flags & ANYOF_FOLD)) {
678 if (and_with->flags & ANYOF_INVERT)
679 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
680 cl->bitmap[i] &= ~and_with->bitmap[i];
682 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
683 cl->bitmap[i] &= and_with->bitmap[i];
684 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
685 if (!(and_with->flags & ANYOF_EOS))
686 cl->flags &= ~ANYOF_EOS;
688 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
689 !(and_with->flags & ANYOF_INVERT)) {
690 cl->flags &= ~ANYOF_UNICODE_ALL;
691 cl->flags |= ANYOF_UNICODE;
692 ARG_SET(cl, ARG(and_with));
694 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
695 !(and_with->flags & ANYOF_INVERT))
696 cl->flags &= ~ANYOF_UNICODE_ALL;
697 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
698 !(and_with->flags & ANYOF_INVERT))
699 cl->flags &= ~ANYOF_UNICODE;
702 /* 'OR' a given class with another one. Can create false positives */
703 /* We assume that cl is not inverted */
705 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
707 if (or_with->flags & ANYOF_INVERT) {
709 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
710 * <= (B1 | !B2) | (CL1 | !CL2)
711 * which is wasteful if CL2 is small, but we ignore CL2:
712 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
713 * XXXX Can we handle case-fold? Unclear:
714 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
715 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
717 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
718 && !(or_with->flags & ANYOF_FOLD)
719 && !(cl->flags & ANYOF_FOLD) ) {
722 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
723 cl->bitmap[i] |= ~or_with->bitmap[i];
724 } /* XXXX: logic is complicated otherwise */
726 cl_anything(pRExC_state, cl);
729 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
730 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
731 && (!(or_with->flags & ANYOF_FOLD)
732 || (cl->flags & ANYOF_FOLD)) ) {
735 /* OR char bitmap and class bitmap separately */
736 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
737 cl->bitmap[i] |= or_with->bitmap[i];
738 if (or_with->flags & ANYOF_CLASS) {
739 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
740 cl->classflags[i] |= or_with->classflags[i];
741 cl->flags |= ANYOF_CLASS;
744 else { /* XXXX: logic is complicated, leave it along for a moment. */
745 cl_anything(pRExC_state, cl);
748 if (or_with->flags & ANYOF_EOS)
749 cl->flags |= ANYOF_EOS;
751 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
752 ARG(cl) != ARG(or_with)) {
753 cl->flags |= ANYOF_UNICODE_ALL;
754 cl->flags &= ~ANYOF_UNICODE;
756 if (or_with->flags & ANYOF_UNICODE_ALL) {
757 cl->flags |= ANYOF_UNICODE_ALL;
758 cl->flags &= ~ANYOF_UNICODE;
762 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
763 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
764 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
765 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
771 dump_trie_interim_list(trie,next_alloc)
772 dump_trie_interim_table(trie,next_alloc)
774 These routines dump out a trie in a somewhat readable format.
775 The _interim_ variants are used for debugging the interim
776 tables that are used to generate the final compressed
777 representation which is what dump_trie expects.
779 Part of the reason for their existance is to provide a form
780 of documentation as to how the different representations function.
786 Dumps the final compressed table form of the trie to Perl_debug_log.
787 Used for debugging make_trie().
791 S_dump_trie(pTHX_ const struct _reg_trie_data *trie,U32 depth)
794 SV *sv=sv_newmortal();
795 int colwidth= trie->widecharmap ? 6 : 4;
796 GET_RE_DEBUG_FLAGS_DECL;
799 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
800 (int)depth * 2 + 2,"",
801 "Match","Base","Ofs" );
803 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
804 SV ** const tmp = av_fetch( trie->revcharmap, state, 0);
806 PerlIO_printf( Perl_debug_log, "%*s",
808 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
809 PL_colors[0], PL_colors[1],
810 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
811 PERL_PV_ESCAPE_FIRSTCHAR
816 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
817 (int)depth * 2 + 2,"");
819 for( state = 0 ; state < trie->uniquecharcount ; state++ )
820 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
821 PerlIO_printf( Perl_debug_log, "\n");
823 for( state = 1 ; state < trie->statecount ; state++ ) {
824 const U32 base = trie->states[ state ].trans.base;
826 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
828 if ( trie->states[ state ].wordnum ) {
829 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
831 PerlIO_printf( Perl_debug_log, "%6s", "" );
834 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
839 while( ( base + ofs < trie->uniquecharcount ) ||
840 ( base + ofs - trie->uniquecharcount < trie->lasttrans
841 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
844 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
846 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
847 if ( ( base + ofs >= trie->uniquecharcount ) &&
848 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
849 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
851 PerlIO_printf( Perl_debug_log, "%*"UVXf,
853 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
855 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
859 PerlIO_printf( Perl_debug_log, "]");
862 PerlIO_printf( Perl_debug_log, "\n" );
866 dump_trie_interim_list(trie,next_alloc)
867 Dumps a fully constructed but uncompressed trie in list form.
868 List tries normally only are used for construction when the number of
869 possible chars (trie->uniquecharcount) is very high.
870 Used for debugging make_trie().
873 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie, U32 next_alloc,U32 depth)
876 SV *sv=sv_newmortal();
877 int colwidth= trie->widecharmap ? 6 : 4;
878 GET_RE_DEBUG_FLAGS_DECL;
879 /* print out the table precompression. */
880 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
881 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
882 "------:-----+-----------------\n" );
884 for( state=1 ; state < next_alloc ; state ++ ) {
887 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
888 (int)depth * 2 + 2,"", (UV)state );
889 if ( ! trie->states[ state ].wordnum ) {
890 PerlIO_printf( Perl_debug_log, "%5s| ","");
892 PerlIO_printf( Perl_debug_log, "W%4x| ",
893 trie->states[ state ].wordnum
896 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
897 SV ** const tmp = av_fetch( trie->revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
899 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
901 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
902 PL_colors[0], PL_colors[1],
903 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
904 PERL_PV_ESCAPE_FIRSTCHAR
906 TRIE_LIST_ITEM(state,charid).forid,
907 (UV)TRIE_LIST_ITEM(state,charid).newstate
910 PerlIO_printf(Perl_debug_log, "\n%*s| ",
911 (int)((depth * 2) + 14), "");
914 PerlIO_printf( Perl_debug_log, "\n");
919 dump_trie_interim_table(trie,next_alloc)
920 Dumps a fully constructed but uncompressed trie in table form.
921 This is the normal DFA style state transition table, with a few
922 twists to facilitate compression later.
923 Used for debugging make_trie().
926 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie, U32 next_alloc, U32 depth)
930 SV *sv=sv_newmortal();
931 int colwidth= trie->widecharmap ? 6 : 4;
932 GET_RE_DEBUG_FLAGS_DECL;
935 print out the table precompression so that we can do a visual check
936 that they are identical.
939 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
941 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
942 SV ** const tmp = av_fetch( trie->revcharmap, charid, 0);
944 PerlIO_printf( Perl_debug_log, "%*s",
946 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
947 PL_colors[0], PL_colors[1],
948 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
949 PERL_PV_ESCAPE_FIRSTCHAR
955 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
957 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
958 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
961 PerlIO_printf( Perl_debug_log, "\n" );
963 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
965 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
966 (int)depth * 2 + 2,"",
967 (UV)TRIE_NODENUM( state ) );
969 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
970 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
972 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
974 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
976 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
977 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
979 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
980 trie->states[ TRIE_NODENUM( state ) ].wordnum );
987 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
988 startbranch: the first branch in the whole branch sequence
989 first : start branch of sequence of branch-exact nodes.
990 May be the same as startbranch
991 last : Thing following the last branch.
992 May be the same as tail.
993 tail : item following the branch sequence
994 count : words in the sequence
995 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
998 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1000 A trie is an N'ary tree where the branches are determined by digital
1001 decomposition of the key. IE, at the root node you look up the 1st character and
1002 follow that branch repeat until you find the end of the branches. Nodes can be
1003 marked as "accepting" meaning they represent a complete word. Eg:
1007 would convert into the following structure. Numbers represent states, letters
1008 following numbers represent valid transitions on the letter from that state, if
1009 the number is in square brackets it represents an accepting state, otherwise it
1010 will be in parenthesis.
1012 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1016 (1) +-i->(6)-+-s->[7]
1018 +-s->(3)-+-h->(4)-+-e->[5]
1020 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1022 This shows that when matching against the string 'hers' we will begin at state 1
1023 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1024 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1025 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1026 single traverse. We store a mapping from accepting to state to which word was
1027 matched, and then when we have multiple possibilities we try to complete the
1028 rest of the regex in the order in which they occured in the alternation.
1030 The only prior NFA like behaviour that would be changed by the TRIE support is
1031 the silent ignoring of duplicate alternations which are of the form:
1033 / (DUPE|DUPE) X? (?{ ... }) Y /x
1035 Thus EVAL blocks follwing a trie may be called a different number of times with
1036 and without the optimisation. With the optimisations dupes will be silently
1037 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1038 the following demonstrates:
1040 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1042 which prints out 'word' three times, but
1044 'words'=~/(word|word|word)(?{ print $1 })S/
1046 which doesnt print it out at all. This is due to other optimisations kicking in.
1048 Example of what happens on a structural level:
1050 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1052 1: CURLYM[1] {1,32767}(18)
1063 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1064 and should turn into:
1066 1: CURLYM[1] {1,32767}(18)
1068 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1076 Cases where tail != last would be like /(?foo|bar)baz/:
1086 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1087 and would end up looking like:
1090 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1097 d = uvuni_to_utf8_flags(d, uv, 0);
1099 is the recommended Unicode-aware way of saying
1104 #define TRIE_STORE_REVCHAR \
1106 SV *tmp = newSVpvs(""); \
1107 if (UTF) SvUTF8_on(tmp); \
1108 Perl_sv_catpvf( aTHX_ tmp, "%c", (int)uvc ); \
1109 av_push( TRIE_REVCHARMAP(trie), tmp ); \
1112 #define TRIE_READ_CHAR STMT_START { \
1116 if ( foldlen > 0 ) { \
1117 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1122 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1123 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1124 foldlen -= UNISKIP( uvc ); \
1125 scan = foldbuf + UNISKIP( uvc ); \
1128 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1138 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1139 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1140 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1141 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1143 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1144 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1145 TRIE_LIST_CUR( state )++; \
1148 #define TRIE_LIST_NEW(state) STMT_START { \
1149 Newxz( trie->states[ state ].trans.list, \
1150 4, reg_trie_trans_le ); \
1151 TRIE_LIST_CUR( state ) = 1; \
1152 TRIE_LIST_LEN( state ) = 4; \
1155 #define TRIE_HANDLE_WORD(state) STMT_START { \
1156 U16 dupe= trie->states[ state ].wordnum; \
1157 regnode * const noper_next = regnext( noper ); \
1159 if (trie->wordlen) \
1160 trie->wordlen[ curword ] = wordlen; \
1162 /* store the word for dumping */ \
1164 if (OP(noper) != NOTHING) \
1165 tmp = newSVpvn(STRING(noper), STR_LEN(noper)); \
1167 tmp = newSVpvn( "", 0 ); \
1168 if ( UTF ) SvUTF8_on( tmp ); \
1169 av_push( trie->words, tmp ); \
1174 if ( noper_next < tail ) { \
1176 Newxz( trie->jump, word_count + 1, U16); \
1177 trie->jump[curword] = (U16)(noper_next - convert); \
1179 jumper = noper_next; \
1181 nextbranch= regnext(cur); \
1185 /* So it's a dupe. This means we need to maintain a */\
1186 /* linked-list from the first to the next. */\
1187 /* we only allocate the nextword buffer when there */\
1188 /* a dupe, so first time we have to do the allocation */\
1189 if (!trie->nextword) \
1190 Newxz( trie->nextword, word_count + 1, U16); \
1191 while ( trie->nextword[dupe] ) \
1192 dupe= trie->nextword[dupe]; \
1193 trie->nextword[dupe]= curword; \
1195 /* we haven't inserted this word yet. */ \
1196 trie->states[ state ].wordnum = curword; \
1201 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1202 ( ( base + charid >= ucharcount \
1203 && base + charid < ubound \
1204 && state == trie->trans[ base - ucharcount + charid ].check \
1205 && trie->trans[ base - ucharcount + charid ].next ) \
1206 ? trie->trans[ base - ucharcount + charid ].next \
1207 : ( state==1 ? special : 0 ) \
1211 #define MADE_JUMP_TRIE 2
1212 #define MADE_EXACT_TRIE 4
1215 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1218 /* first pass, loop through and scan words */
1219 reg_trie_data *trie;
1221 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1226 regnode *jumper = NULL;
1227 regnode *nextbranch = NULL;
1228 regnode *convert = NULL;
1229 /* we just use folder as a flag in utf8 */
1230 const U8 * const folder = ( flags == EXACTF
1232 : ( flags == EXACTFL
1238 const U32 data_slot = add_data( pRExC_state, 1, "t" );
1239 SV *re_trie_maxbuff;
1241 /* these are only used during construction but are useful during
1242 * debugging so we store them in the struct when debugging.
1244 STRLEN trie_charcount=0;
1245 AV *trie_revcharmap;
1247 GET_RE_DEBUG_FLAGS_DECL;
1249 PERL_UNUSED_ARG(depth);
1252 Newxz( trie, 1, reg_trie_data );
1254 trie->startstate = 1;
1255 trie->wordcount = word_count;
1256 RExC_rx->data->data[ data_slot ] = (void*)trie;
1257 Newxz( trie->charmap, 256, U16 );
1258 if (!(UTF && folder))
1259 Newxz( trie->bitmap, ANYOF_BITMAP_SIZE, char );
1261 trie->words = newAV();
1263 TRIE_REVCHARMAP(trie) = newAV();
1265 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1266 if (!SvIOK(re_trie_maxbuff)) {
1267 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1270 PerlIO_printf( Perl_debug_log,
1271 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1272 (int)depth * 2 + 2, "",
1273 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1274 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1278 /* Find the node we are going to overwrite */
1279 if ( first == startbranch && OP( last ) != BRANCH ) {
1280 /* whole branch chain */
1283 /* branch sub-chain */
1284 convert = NEXTOPER( first );
1287 /* -- First loop and Setup --
1289 We first traverse the branches and scan each word to determine if it
1290 contains widechars, and how many unique chars there are, this is
1291 important as we have to build a table with at least as many columns as we
1294 We use an array of integers to represent the character codes 0..255
1295 (trie->charmap) and we use a an HV* to store unicode characters. We use the
1296 native representation of the character value as the key and IV's for the
1299 *TODO* If we keep track of how many times each character is used we can
1300 remap the columns so that the table compression later on is more
1301 efficient in terms of memory by ensuring most common value is in the
1302 middle and the least common are on the outside. IMO this would be better
1303 than a most to least common mapping as theres a decent chance the most
1304 common letter will share a node with the least common, meaning the node
1305 will not be compressable. With a middle is most common approach the worst
1306 case is when we have the least common nodes twice.
1310 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1311 regnode * const noper = NEXTOPER( cur );
1312 const U8 *uc = (U8*)STRING( noper );
1313 const U8 * const e = uc + STR_LEN( noper );
1315 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1316 const U8 *scan = (U8*)NULL;
1317 U32 wordlen = 0; /* required init */
1320 if (OP(noper) == NOTHING) {
1325 TRIE_BITMAP_SET(trie,*uc);
1326 if ( folder ) TRIE_BITMAP_SET(trie,folder[ *uc ]);
1328 for ( ; uc < e ; uc += len ) {
1329 TRIE_CHARCOUNT(trie)++;
1333 if ( !trie->charmap[ uvc ] ) {
1334 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1336 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1341 if ( !trie->widecharmap )
1342 trie->widecharmap = newHV();
1344 svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1347 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1349 if ( !SvTRUE( *svpp ) ) {
1350 sv_setiv( *svpp, ++trie->uniquecharcount );
1355 if( cur == first ) {
1358 } else if (chars < trie->minlen) {
1360 } else if (chars > trie->maxlen) {
1364 } /* end first pass */
1365 DEBUG_TRIE_COMPILE_r(
1366 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1367 (int)depth * 2 + 2,"",
1368 ( trie->widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1369 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1370 (int)trie->minlen, (int)trie->maxlen )
1372 Newxz( trie->wordlen, word_count, U32 );
1375 We now know what we are dealing with in terms of unique chars and
1376 string sizes so we can calculate how much memory a naive
1377 representation using a flat table will take. If it's over a reasonable
1378 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1379 conservative but potentially much slower representation using an array
1382 At the end we convert both representations into the same compressed
1383 form that will be used in regexec.c for matching with. The latter
1384 is a form that cannot be used to construct with but has memory
1385 properties similar to the list form and access properties similar
1386 to the table form making it both suitable for fast searches and
1387 small enough that its feasable to store for the duration of a program.
1389 See the comment in the code where the compressed table is produced
1390 inplace from the flat tabe representation for an explanation of how
1391 the compression works.
1396 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1398 Second Pass -- Array Of Lists Representation
1400 Each state will be represented by a list of charid:state records
1401 (reg_trie_trans_le) the first such element holds the CUR and LEN
1402 points of the allocated array. (See defines above).
1404 We build the initial structure using the lists, and then convert
1405 it into the compressed table form which allows faster lookups
1406 (but cant be modified once converted).
1409 STRLEN transcount = 1;
1411 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1412 "%*sCompiling trie using list compiler\n",
1413 (int)depth * 2 + 2, ""));
1415 Newxz( trie->states, TRIE_CHARCOUNT(trie) + 2, reg_trie_state );
1419 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1421 regnode * const noper = NEXTOPER( cur );
1422 U8 *uc = (U8*)STRING( noper );
1423 const U8 * const e = uc + STR_LEN( noper );
1424 U32 state = 1; /* required init */
1425 U16 charid = 0; /* sanity init */
1426 U8 *scan = (U8*)NULL; /* sanity init */
1427 STRLEN foldlen = 0; /* required init */
1428 U32 wordlen = 0; /* required init */
1429 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1431 if (OP(noper) != NOTHING) {
1432 for ( ; uc < e ; uc += len ) {
1437 charid = trie->charmap[ uvc ];
1439 SV** const svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 0);
1443 charid=(U16)SvIV( *svpp );
1446 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1453 if ( !trie->states[ state ].trans.list ) {
1454 TRIE_LIST_NEW( state );
1456 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1457 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1458 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1463 newstate = next_alloc++;
1464 TRIE_LIST_PUSH( state, charid, newstate );
1469 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1473 TRIE_HANDLE_WORD(state);
1475 } /* end second pass */
1477 /* next alloc is the NEXT state to be allocated */
1478 trie->statecount = next_alloc;
1479 Renew( trie->states, next_alloc, reg_trie_state );
1481 /* and now dump it out before we compress it */
1482 DEBUG_TRIE_COMPILE_MORE_r(
1483 dump_trie_interim_list(trie,next_alloc,depth+1)
1486 Newxz( trie->trans, transcount ,reg_trie_trans );
1493 for( state=1 ; state < next_alloc ; state ++ ) {
1497 DEBUG_TRIE_COMPILE_MORE_r(
1498 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1502 if (trie->states[state].trans.list) {
1503 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1507 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1508 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1509 if ( forid < minid ) {
1511 } else if ( forid > maxid ) {
1515 if ( transcount < tp + maxid - minid + 1) {
1517 Renew( trie->trans, transcount, reg_trie_trans );
1518 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1520 base = trie->uniquecharcount + tp - minid;
1521 if ( maxid == minid ) {
1523 for ( ; zp < tp ; zp++ ) {
1524 if ( ! trie->trans[ zp ].next ) {
1525 base = trie->uniquecharcount + zp - minid;
1526 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1527 trie->trans[ zp ].check = state;
1533 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1534 trie->trans[ tp ].check = state;
1539 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1540 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1541 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1542 trie->trans[ tid ].check = state;
1544 tp += ( maxid - minid + 1 );
1546 Safefree(trie->states[ state ].trans.list);
1549 DEBUG_TRIE_COMPILE_MORE_r(
1550 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1553 trie->states[ state ].trans.base=base;
1555 trie->lasttrans = tp + 1;
1559 Second Pass -- Flat Table Representation.
1561 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1562 We know that we will need Charcount+1 trans at most to store the data
1563 (one row per char at worst case) So we preallocate both structures
1564 assuming worst case.
1566 We then construct the trie using only the .next slots of the entry
1569 We use the .check field of the first entry of the node temporarily to
1570 make compression both faster and easier by keeping track of how many non
1571 zero fields are in the node.
1573 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1576 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1577 number representing the first entry of the node, and state as a
1578 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1579 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1580 are 2 entrys per node. eg:
1588 The table is internally in the right hand, idx form. However as we also
1589 have to deal with the states array which is indexed by nodenum we have to
1590 use TRIE_NODENUM() to convert.
1593 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1594 "%*sCompiling trie using table compiler\n",
1595 (int)depth * 2 + 2, ""));
1597 Newxz( trie->trans, ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1,
1599 Newxz( trie->states, TRIE_CHARCOUNT(trie) + 2, reg_trie_state );
1600 next_alloc = trie->uniquecharcount + 1;
1603 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1605 regnode * const noper = NEXTOPER( cur );
1606 const U8 *uc = (U8*)STRING( noper );
1607 const U8 * const e = uc + STR_LEN( noper );
1609 U32 state = 1; /* required init */
1611 U16 charid = 0; /* sanity init */
1612 U32 accept_state = 0; /* sanity init */
1613 U8 *scan = (U8*)NULL; /* sanity init */
1615 STRLEN foldlen = 0; /* required init */
1616 U32 wordlen = 0; /* required init */
1617 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1619 if ( OP(noper) != NOTHING ) {
1620 for ( ; uc < e ; uc += len ) {
1625 charid = trie->charmap[ uvc ];
1627 SV* const * const svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 0);
1628 charid = svpp ? (U16)SvIV(*svpp) : 0;
1632 if ( !trie->trans[ state + charid ].next ) {
1633 trie->trans[ state + charid ].next = next_alloc;
1634 trie->trans[ state ].check++;
1635 next_alloc += trie->uniquecharcount;
1637 state = trie->trans[ state + charid ].next;
1639 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1641 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1644 accept_state = TRIE_NODENUM( state );
1645 TRIE_HANDLE_WORD(accept_state);
1647 } /* end second pass */
1649 /* and now dump it out before we compress it */
1650 DEBUG_TRIE_COMPILE_MORE_r(
1651 dump_trie_interim_table(trie,next_alloc,depth+1)
1656 * Inplace compress the table.*
1658 For sparse data sets the table constructed by the trie algorithm will
1659 be mostly 0/FAIL transitions or to put it another way mostly empty.
1660 (Note that leaf nodes will not contain any transitions.)
1662 This algorithm compresses the tables by eliminating most such
1663 transitions, at the cost of a modest bit of extra work during lookup:
1665 - Each states[] entry contains a .base field which indicates the
1666 index in the state[] array wheres its transition data is stored.
1668 - If .base is 0 there are no valid transitions from that node.
1670 - If .base is nonzero then charid is added to it to find an entry in
1673 -If trans[states[state].base+charid].check!=state then the
1674 transition is taken to be a 0/Fail transition. Thus if there are fail
1675 transitions at the front of the node then the .base offset will point
1676 somewhere inside the previous nodes data (or maybe even into a node
1677 even earlier), but the .check field determines if the transition is
1681 The following process inplace converts the table to the compressed
1682 table: We first do not compress the root node 1,and mark its all its
1683 .check pointers as 1 and set its .base pointer as 1 as well. This
1684 allows to do a DFA construction from the compressed table later, and
1685 ensures that any .base pointers we calculate later are greater than
1688 - We set 'pos' to indicate the first entry of the second node.
1690 - We then iterate over the columns of the node, finding the first and
1691 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1692 and set the .check pointers accordingly, and advance pos
1693 appropriately and repreat for the next node. Note that when we copy
1694 the next pointers we have to convert them from the original
1695 NODEIDX form to NODENUM form as the former is not valid post
1698 - If a node has no transitions used we mark its base as 0 and do not
1699 advance the pos pointer.
1701 - If a node only has one transition we use a second pointer into the
1702 structure to fill in allocated fail transitions from other states.
1703 This pointer is independent of the main pointer and scans forward
1704 looking for null transitions that are allocated to a state. When it
1705 finds one it writes the single transition into the "hole". If the
1706 pointer doesnt find one the single transition is appended as normal.
1708 - Once compressed we can Renew/realloc the structures to release the
1711 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1712 specifically Fig 3.47 and the associated pseudocode.
1716 const U32 laststate = TRIE_NODENUM( next_alloc );
1719 trie->statecount = laststate;
1721 for ( state = 1 ; state < laststate ; state++ ) {
1723 const U32 stateidx = TRIE_NODEIDX( state );
1724 const U32 o_used = trie->trans[ stateidx ].check;
1725 U32 used = trie->trans[ stateidx ].check;
1726 trie->trans[ stateidx ].check = 0;
1728 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1729 if ( flag || trie->trans[ stateidx + charid ].next ) {
1730 if ( trie->trans[ stateidx + charid ].next ) {
1732 for ( ; zp < pos ; zp++ ) {
1733 if ( ! trie->trans[ zp ].next ) {
1737 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1738 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1739 trie->trans[ zp ].check = state;
1740 if ( ++zp > pos ) pos = zp;
1747 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1749 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1750 trie->trans[ pos ].check = state;
1755 trie->lasttrans = pos + 1;
1756 Renew( trie->states, laststate, reg_trie_state);
1757 DEBUG_TRIE_COMPILE_MORE_r(
1758 PerlIO_printf( Perl_debug_log,
1759 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1760 (int)depth * 2 + 2,"",
1761 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1764 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1767 } /* end table compress */
1769 DEBUG_TRIE_COMPILE_MORE_r(
1770 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1771 (int)depth * 2 + 2, "",
1772 (UV)trie->statecount,
1773 (UV)trie->lasttrans)
1775 /* resize the trans array to remove unused space */
1776 Renew( trie->trans, trie->lasttrans, reg_trie_trans);
1778 /* and now dump out the compressed format */
1779 DEBUG_TRIE_COMPILE_r(
1780 dump_trie(trie,depth+1)
1783 { /* Modify the program and insert the new TRIE node*/
1784 U8 nodetype =(U8)(flags & 0xFF);
1788 regnode *optimize = NULL;
1790 U32 mjd_nodelen = 0;
1793 This means we convert either the first branch or the first Exact,
1794 depending on whether the thing following (in 'last') is a branch
1795 or not and whther first is the startbranch (ie is it a sub part of
1796 the alternation or is it the whole thing.)
1797 Assuming its a sub part we conver the EXACT otherwise we convert
1798 the whole branch sequence, including the first.
1800 /* Find the node we are going to overwrite */
1801 if ( first != startbranch || OP( last ) == BRANCH ) {
1802 /* branch sub-chain */
1803 NEXT_OFF( first ) = (U16)(last - first);
1805 mjd_offset= Node_Offset((convert));
1806 mjd_nodelen= Node_Length((convert));
1808 /* whole branch chain */
1811 const regnode *nop = NEXTOPER( convert );
1812 mjd_offset= Node_Offset((nop));
1813 mjd_nodelen= Node_Length((nop));
1818 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1819 (int)depth * 2 + 2, "",
1820 (UV)mjd_offset, (UV)mjd_nodelen)
1823 /* But first we check to see if there is a common prefix we can
1824 split out as an EXACT and put in front of the TRIE node. */
1825 trie->startstate= 1;
1826 if ( trie->bitmap && !trie->widecharmap && !trie->jump ) {
1828 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1832 const U32 base = trie->states[ state ].trans.base;
1834 if ( trie->states[state].wordnum )
1837 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1838 if ( ( base + ofs >= trie->uniquecharcount ) &&
1839 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1840 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1842 if ( ++count > 1 ) {
1843 SV **tmp = av_fetch( TRIE_REVCHARMAP(trie), ofs, 0);
1844 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1845 if ( state == 1 ) break;
1847 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1849 PerlIO_printf(Perl_debug_log,
1850 "%*sNew Start State=%"UVuf" Class: [",
1851 (int)depth * 2 + 2, "",
1854 SV ** const tmp = av_fetch( TRIE_REVCHARMAP(trie), idx, 0);
1855 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1857 TRIE_BITMAP_SET(trie,*ch);
1859 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1861 PerlIO_printf(Perl_debug_log, (char*)ch)
1865 TRIE_BITMAP_SET(trie,*ch);
1867 TRIE_BITMAP_SET(trie,folder[ *ch ]);
1868 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
1874 SV **tmp = av_fetch( TRIE_REVCHARMAP(trie), idx, 0);
1875 const char *ch = SvPV_nolen_const( *tmp );
1877 PerlIO_printf( Perl_debug_log,
1878 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
1879 (int)depth * 2 + 2, "",
1880 (UV)state, (UV)idx, ch)
1883 OP( convert ) = nodetype;
1884 str=STRING(convert);
1893 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
1899 regnode *n = convert+NODE_SZ_STR(convert);
1900 NEXT_OFF(convert) = NODE_SZ_STR(convert);
1901 trie->startstate = state;
1902 trie->minlen -= (state - 1);
1903 trie->maxlen -= (state - 1);
1905 regnode *fix = convert;
1907 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
1908 while( ++fix < n ) {
1909 Set_Node_Offset_Length(fix, 0, 0);
1915 NEXT_OFF(convert) = (U16)(tail - convert);
1916 DEBUG_r(optimize= n);
1922 if ( trie->maxlen ) {
1923 NEXT_OFF( convert ) = (U16)(tail - convert);
1924 ARG_SET( convert, data_slot );
1925 /* Store the offset to the first unabsorbed branch in
1926 jump[0], which is otherwise unused by the jump logic.
1927 We use this when dumping a trie and during optimisation. */
1929 trie->jump[0] = (U16)(nextbranch - convert);
1932 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
1933 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
1935 OP( convert ) = TRIEC;
1936 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
1937 Safefree(trie->bitmap);
1940 OP( convert ) = TRIE;
1942 /* store the type in the flags */
1943 convert->flags = nodetype;
1947 + regarglen[ OP( convert ) ];
1949 /* XXX We really should free up the resource in trie now,
1950 as we won't use them - (which resources?) dmq */
1952 /* needed for dumping*/
1953 DEBUG_r(if (optimize) {
1954 regnode *opt = convert;
1955 while ( ++opt < optimize) {
1956 Set_Node_Offset_Length(opt,0,0);
1959 Try to clean up some of the debris left after the
1962 while( optimize < jumper ) {
1963 mjd_nodelen += Node_Length((optimize));
1964 OP( optimize ) = OPTIMIZED;
1965 Set_Node_Offset_Length(optimize,0,0);
1968 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
1970 } /* end node insert */
1972 SvREFCNT_dec(TRIE_REVCHARMAP(trie));
1976 : trie->startstate>1
1982 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
1984 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
1986 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
1987 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
1990 We find the fail state for each state in the trie, this state is the longest proper
1991 suffix of the current states 'word' that is also a proper prefix of another word in our
1992 trie. State 1 represents the word '' and is the thus the default fail state. This allows
1993 the DFA not to have to restart after its tried and failed a word at a given point, it
1994 simply continues as though it had been matching the other word in the first place.
1996 'abcdgu'=~/abcdefg|cdgu/
1997 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
1998 fail, which would bring use to the state representing 'd' in the second word where we would
1999 try 'g' and succeed, prodceding to match 'cdgu'.
2001 /* add a fail transition */
2002 reg_trie_data *trie=(reg_trie_data *)RExC_rx->data->data[ARG(source)];
2004 const U32 ucharcount = trie->uniquecharcount;
2005 const U32 numstates = trie->statecount;
2006 const U32 ubound = trie->lasttrans + ucharcount;
2010 U32 base = trie->states[ 1 ].trans.base;
2013 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2014 GET_RE_DEBUG_FLAGS_DECL;
2016 PERL_UNUSED_ARG(depth);
2020 ARG_SET( stclass, data_slot );
2021 Newxz( aho, 1, reg_ac_data );
2022 RExC_rx->data->data[ data_slot ] = (void*)aho;
2024 aho->states=(reg_trie_state *)savepvn((const char*)trie->states,
2025 numstates * sizeof(reg_trie_state));
2026 Newxz( q, numstates, U32);
2027 Newxz( aho->fail, numstates, U32 );
2030 /* initialize fail[0..1] to be 1 so that we always have
2031 a valid final fail state */
2032 fail[ 0 ] = fail[ 1 ] = 1;
2034 for ( charid = 0; charid < ucharcount ; charid++ ) {
2035 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2037 q[ q_write ] = newstate;
2038 /* set to point at the root */
2039 fail[ q[ q_write++ ] ]=1;
2042 while ( q_read < q_write) {
2043 const U32 cur = q[ q_read++ % numstates ];
2044 base = trie->states[ cur ].trans.base;
2046 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2047 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2049 U32 fail_state = cur;
2052 fail_state = fail[ fail_state ];
2053 fail_base = aho->states[ fail_state ].trans.base;
2054 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2056 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2057 fail[ ch_state ] = fail_state;
2058 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2060 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2062 q[ q_write++ % numstates] = ch_state;
2066 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2067 when we fail in state 1, this allows us to use the
2068 charclass scan to find a valid start char. This is based on the principle
2069 that theres a good chance the string being searched contains lots of stuff
2070 that cant be a start char.
2072 fail[ 0 ] = fail[ 1 ] = 0;
2073 DEBUG_TRIE_COMPILE_r({
2074 PerlIO_printf(Perl_debug_log, "%*sStclass Failtable (%"UVuf" states): 0",
2075 (int)(depth * 2), "", numstates
2077 for( q_read=1; q_read<numstates; q_read++ ) {
2078 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2080 PerlIO_printf(Perl_debug_log, "\n");
2083 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2088 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2089 * These need to be revisited when a newer toolchain becomes available.
2091 #if defined(__sparc64__) && defined(__GNUC__)
2092 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2093 # undef SPARC64_GCC_WORKAROUND
2094 # define SPARC64_GCC_WORKAROUND 1
2098 #define DEBUG_PEEP(str,scan,depth) \
2099 DEBUG_OPTIMISE_r({ \
2100 SV * const mysv=sv_newmortal(); \
2101 regnode *Next = regnext(scan); \
2102 regprop(RExC_rx, mysv, scan); \
2103 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2104 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2105 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2112 #define JOIN_EXACT(scan,min,flags) \
2113 if (PL_regkind[OP(scan)] == EXACT) \
2114 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2117 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2118 /* Merge several consecutive EXACTish nodes into one. */
2119 regnode *n = regnext(scan);
2121 regnode *next = scan + NODE_SZ_STR(scan);
2125 regnode *stop = scan;
2126 GET_RE_DEBUG_FLAGS_DECL;
2128 PERL_UNUSED_ARG(depth);
2130 #ifndef EXPERIMENTAL_INPLACESCAN
2131 PERL_UNUSED_ARG(flags);
2132 PERL_UNUSED_ARG(val);
2134 DEBUG_PEEP("join",scan,depth);
2136 /* Skip NOTHING, merge EXACT*. */
2138 ( PL_regkind[OP(n)] == NOTHING ||
2139 (stringok && (OP(n) == OP(scan))))
2141 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2143 if (OP(n) == TAIL || n > next)
2145 if (PL_regkind[OP(n)] == NOTHING) {
2146 DEBUG_PEEP("skip:",n,depth);
2147 NEXT_OFF(scan) += NEXT_OFF(n);
2148 next = n + NODE_STEP_REGNODE;
2155 else if (stringok) {
2156 const unsigned int oldl = STR_LEN(scan);
2157 regnode * const nnext = regnext(n);
2159 DEBUG_PEEP("merg",n,depth);
2162 if (oldl + STR_LEN(n) > U8_MAX)
2164 NEXT_OFF(scan) += NEXT_OFF(n);
2165 STR_LEN(scan) += STR_LEN(n);
2166 next = n + NODE_SZ_STR(n);
2167 /* Now we can overwrite *n : */
2168 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2176 #ifdef EXPERIMENTAL_INPLACESCAN
2177 if (flags && !NEXT_OFF(n)) {
2178 DEBUG_PEEP("atch", val, depth);
2179 if (reg_off_by_arg[OP(n)]) {
2180 ARG_SET(n, val - n);
2183 NEXT_OFF(n) = val - n;
2190 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2192 Two problematic code points in Unicode casefolding of EXACT nodes:
2194 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2195 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2201 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2202 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2204 This means that in case-insensitive matching (or "loose matching",
2205 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2206 length of the above casefolded versions) can match a target string
2207 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2208 This would rather mess up the minimum length computation.
2210 What we'll do is to look for the tail four bytes, and then peek
2211 at the preceding two bytes to see whether we need to decrease
2212 the minimum length by four (six minus two).
2214 Thanks to the design of UTF-8, there cannot be false matches:
2215 A sequence of valid UTF-8 bytes cannot be a subsequence of
2216 another valid sequence of UTF-8 bytes.
2219 char * const s0 = STRING(scan), *s, *t;
2220 char * const s1 = s0 + STR_LEN(scan) - 1;
2221 char * const s2 = s1 - 4;
2222 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2223 const char t0[] = "\xaf\x49\xaf\x42";
2225 const char t0[] = "\xcc\x88\xcc\x81";
2227 const char * const t1 = t0 + 3;
2230 s < s2 && (t = ninstr(s, s1, t0, t1));
2233 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2234 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2236 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2237 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2245 n = scan + NODE_SZ_STR(scan);
2247 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2254 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2258 /* REx optimizer. Converts nodes into quickier variants "in place".
2259 Finds fixed substrings. */
2261 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2262 to the position after last scanned or to NULL. */
2267 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2268 I32 *minlenp, I32 *deltap,
2269 regnode *last, scan_data_t *data, U32 flags, U32 depth)
2270 /* scanp: Start here (read-write). */
2271 /* deltap: Write maxlen-minlen here. */
2272 /* last: Stop before this one. */
2275 I32 min = 0, pars = 0, code;
2276 regnode *scan = *scanp, *next;
2278 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2279 int is_inf_internal = 0; /* The studied chunk is infinite */
2280 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2281 scan_data_t data_fake;
2282 struct regnode_charclass_class and_with; /* Valid if flags & SCF_DO_STCLASS_OR */
2283 SV *re_trie_maxbuff = NULL;
2284 regnode *first_non_open = scan;
2287 GET_RE_DEBUG_FLAGS_DECL;
2289 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2292 while (first_non_open && OP(first_non_open) == OPEN)
2293 first_non_open=regnext(first_non_open);
2297 while (scan && OP(scan) != END && scan < last) {
2298 /* Peephole optimizer: */
2299 DEBUG_STUDYDATA(data,depth);
2300 DEBUG_PEEP("Peep",scan,depth);
2301 JOIN_EXACT(scan,&min,0);
2303 /* Follow the next-chain of the current node and optimize
2304 away all the NOTHINGs from it. */
2305 if (OP(scan) != CURLYX) {
2306 const int max = (reg_off_by_arg[OP(scan)]
2308 /* I32 may be smaller than U16 on CRAYs! */
2309 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2310 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2314 /* Skip NOTHING and LONGJMP. */
2315 while ((n = regnext(n))
2316 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2317 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2318 && off + noff < max)
2320 if (reg_off_by_arg[OP(scan)])
2323 NEXT_OFF(scan) = off;
2328 /* The principal pseudo-switch. Cannot be a switch, since we
2329 look into several different things. */
2330 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2331 || OP(scan) == IFTHEN || OP(scan) == SUSPEND) {
2332 next = regnext(scan);
2334 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2336 if (OP(next) == code || code == IFTHEN || code == SUSPEND) {
2337 /* NOTE - There is similar code to this block below for handling
2338 TRIE nodes on a re-study. If you change stuff here check there
2340 I32 max1 = 0, min1 = I32_MAX, num = 0;
2341 struct regnode_charclass_class accum;
2342 regnode * const startbranch=scan;
2344 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
2345 scan_commit(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2346 if (flags & SCF_DO_STCLASS)
2347 cl_init_zero(pRExC_state, &accum);
2349 while (OP(scan) == code) {
2350 I32 deltanext, minnext, f = 0, fake;
2351 struct regnode_charclass_class this_class;
2354 data_fake.flags = 0;
2356 data_fake.whilem_c = data->whilem_c;
2357 data_fake.last_closep = data->last_closep;
2360 data_fake.last_closep = &fake;
2361 next = regnext(scan);
2362 scan = NEXTOPER(scan);
2364 scan = NEXTOPER(scan);
2365 if (flags & SCF_DO_STCLASS) {
2366 cl_init(pRExC_state, &this_class);
2367 data_fake.start_class = &this_class;
2368 f = SCF_DO_STCLASS_AND;
2370 if (flags & SCF_WHILEM_VISITED_POS)
2371 f |= SCF_WHILEM_VISITED_POS;
2373 /* we suppose the run is continuous, last=next...*/
2374 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2375 next, &data_fake, f,depth+1);
2378 if (max1 < minnext + deltanext)
2379 max1 = minnext + deltanext;
2380 if (deltanext == I32_MAX)
2381 is_inf = is_inf_internal = 1;
2383 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2386 if (data_fake.flags & SF_HAS_EVAL)
2387 data->flags |= SF_HAS_EVAL;
2388 data->whilem_c = data_fake.whilem_c;
2390 if (flags & SCF_DO_STCLASS)
2391 cl_or(pRExC_state, &accum, &this_class);
2392 if (code == SUSPEND)
2395 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2397 if (flags & SCF_DO_SUBSTR) {
2398 data->pos_min += min1;
2399 data->pos_delta += max1 - min1;
2400 if (max1 != min1 || is_inf)
2401 data->longest = &(data->longest_float);
2404 delta += max1 - min1;
2405 if (flags & SCF_DO_STCLASS_OR) {
2406 cl_or(pRExC_state, data->start_class, &accum);
2408 cl_and(data->start_class, &and_with);
2409 flags &= ~SCF_DO_STCLASS;
2412 else if (flags & SCF_DO_STCLASS_AND) {
2414 cl_and(data->start_class, &accum);
2415 flags &= ~SCF_DO_STCLASS;
2418 /* Switch to OR mode: cache the old value of
2419 * data->start_class */
2420 StructCopy(data->start_class, &and_with,
2421 struct regnode_charclass_class);
2422 flags &= ~SCF_DO_STCLASS_AND;
2423 StructCopy(&accum, data->start_class,
2424 struct regnode_charclass_class);
2425 flags |= SCF_DO_STCLASS_OR;
2426 data->start_class->flags |= ANYOF_EOS;
2430 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2433 Assuming this was/is a branch we are dealing with: 'scan' now
2434 points at the item that follows the branch sequence, whatever
2435 it is. We now start at the beginning of the sequence and look
2442 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2444 If we can find such a subseqence we need to turn the first
2445 element into a trie and then add the subsequent branch exact
2446 strings to the trie.
2450 1. patterns where the whole set of branch can be converted.
2452 2. patterns where only a subset can be converted.
2454 In case 1 we can replace the whole set with a single regop
2455 for the trie. In case 2 we need to keep the start and end
2458 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2459 becomes BRANCH TRIE; BRANCH X;
2461 There is an additional case, that being where there is a
2462 common prefix, which gets split out into an EXACT like node
2463 preceding the TRIE node.
2465 If x(1..n)==tail then we can do a simple trie, if not we make
2466 a "jump" trie, such that when we match the appropriate word
2467 we "jump" to the appopriate tail node. Essentailly we turn
2468 a nested if into a case structure of sorts.
2473 if (!re_trie_maxbuff) {
2474 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2475 if (!SvIOK(re_trie_maxbuff))
2476 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2478 if ( SvIV(re_trie_maxbuff)>=0 ) {
2480 regnode *first = (regnode *)NULL;
2481 regnode *last = (regnode *)NULL;
2482 regnode *tail = scan;
2487 SV * const mysv = sv_newmortal(); /* for dumping */
2489 /* var tail is used because there may be a TAIL
2490 regop in the way. Ie, the exacts will point to the
2491 thing following the TAIL, but the last branch will
2492 point at the TAIL. So we advance tail. If we
2493 have nested (?:) we may have to move through several
2497 while ( OP( tail ) == TAIL ) {
2498 /* this is the TAIL generated by (?:) */
2499 tail = regnext( tail );
2504 regprop(RExC_rx, mysv, tail );
2505 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2506 (int)depth * 2 + 2, "",
2507 "Looking for TRIE'able sequences. Tail node is: ",
2508 SvPV_nolen_const( mysv )
2514 step through the branches, cur represents each
2515 branch, noper is the first thing to be matched
2516 as part of that branch and noper_next is the
2517 regnext() of that node. if noper is an EXACT
2518 and noper_next is the same as scan (our current
2519 position in the regex) then the EXACT branch is
2520 a possible optimization target. Once we have
2521 two or more consequetive such branches we can
2522 create a trie of the EXACT's contents and stich
2523 it in place. If the sequence represents all of
2524 the branches we eliminate the whole thing and
2525 replace it with a single TRIE. If it is a
2526 subsequence then we need to stitch it in. This
2527 means the first branch has to remain, and needs
2528 to be repointed at the item on the branch chain
2529 following the last branch optimized. This could
2530 be either a BRANCH, in which case the
2531 subsequence is internal, or it could be the
2532 item following the branch sequence in which
2533 case the subsequence is at the end.
2537 /* dont use tail as the end marker for this traverse */
2538 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2539 regnode * const noper = NEXTOPER( cur );
2540 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2541 regnode * const noper_next = regnext( noper );
2545 regprop(RExC_rx, mysv, cur);
2546 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2547 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2549 regprop(RExC_rx, mysv, noper);
2550 PerlIO_printf( Perl_debug_log, " -> %s",
2551 SvPV_nolen_const(mysv));
2554 regprop(RExC_rx, mysv, noper_next );
2555 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2556 SvPV_nolen_const(mysv));
2558 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2559 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2561 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2562 : PL_regkind[ OP( noper ) ] == EXACT )
2563 || OP(noper) == NOTHING )
2565 && noper_next == tail
2570 if ( !first || optype == NOTHING ) {
2571 if (!first) first = cur;
2572 optype = OP( noper );
2578 make_trie( pRExC_state,
2579 startbranch, first, cur, tail, count,
2582 if ( PL_regkind[ OP( noper ) ] == EXACT
2584 && noper_next == tail
2589 optype = OP( noper );
2599 regprop(RExC_rx, mysv, cur);
2600 PerlIO_printf( Perl_debug_log,
2601 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2602 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2606 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2607 #ifdef TRIE_STUDY_OPT
2608 if ( ((made == MADE_EXACT_TRIE &&
2609 startbranch == first)
2610 || ( first_non_open == first )) &&
2612 flags |= SCF_TRIE_RESTUDY;
2620 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2621 scan = NEXTOPER(NEXTOPER(scan));
2622 } else /* single branch is optimized. */
2623 scan = NEXTOPER(scan);
2626 else if (OP(scan) == EXACT) {
2627 I32 l = STR_LEN(scan);
2630 const U8 * const s = (U8*)STRING(scan);
2631 l = utf8_length(s, s + l);
2632 uc = utf8_to_uvchr(s, NULL);
2634 uc = *((U8*)STRING(scan));
2637 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2638 /* The code below prefers earlier match for fixed
2639 offset, later match for variable offset. */
2640 if (data->last_end == -1) { /* Update the start info. */
2641 data->last_start_min = data->pos_min;
2642 data->last_start_max = is_inf
2643 ? I32_MAX : data->pos_min + data->pos_delta;
2645 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2647 SvUTF8_on(data->last_found);
2649 SV * const sv = data->last_found;
2650 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2651 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2652 if (mg && mg->mg_len >= 0)
2653 mg->mg_len += utf8_length((U8*)STRING(scan),
2654 (U8*)STRING(scan)+STR_LEN(scan));
2656 data->last_end = data->pos_min + l;
2657 data->pos_min += l; /* As in the first entry. */
2658 data->flags &= ~SF_BEFORE_EOL;
2660 if (flags & SCF_DO_STCLASS_AND) {
2661 /* Check whether it is compatible with what we know already! */
2665 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2666 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2667 && (!(data->start_class->flags & ANYOF_FOLD)
2668 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2671 ANYOF_CLASS_ZERO(data->start_class);
2672 ANYOF_BITMAP_ZERO(data->start_class);
2674 ANYOF_BITMAP_SET(data->start_class, uc);
2675 data->start_class->flags &= ~ANYOF_EOS;
2677 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2679 else if (flags & SCF_DO_STCLASS_OR) {
2680 /* false positive possible if the class is case-folded */
2682 ANYOF_BITMAP_SET(data->start_class, uc);
2684 data->start_class->flags |= ANYOF_UNICODE_ALL;
2685 data->start_class->flags &= ~ANYOF_EOS;
2686 cl_and(data->start_class, &and_with);
2688 flags &= ~SCF_DO_STCLASS;
2690 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
2691 I32 l = STR_LEN(scan);
2692 UV uc = *((U8*)STRING(scan));
2694 /* Search for fixed substrings supports EXACT only. */
2695 if (flags & SCF_DO_SUBSTR) {
2697 scan_commit(pRExC_state, data, minlenp);
2700 const U8 * const s = (U8 *)STRING(scan);
2701 l = utf8_length(s, s + l);
2702 uc = utf8_to_uvchr(s, NULL);
2705 if (flags & SCF_DO_SUBSTR)
2707 if (flags & SCF_DO_STCLASS_AND) {
2708 /* Check whether it is compatible with what we know already! */
2712 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2713 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2714 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2716 ANYOF_CLASS_ZERO(data->start_class);
2717 ANYOF_BITMAP_ZERO(data->start_class);
2719 ANYOF_BITMAP_SET(data->start_class, uc);
2720 data->start_class->flags &= ~ANYOF_EOS;
2721 data->start_class->flags |= ANYOF_FOLD;
2722 if (OP(scan) == EXACTFL)
2723 data->start_class->flags |= ANYOF_LOCALE;
2726 else if (flags & SCF_DO_STCLASS_OR) {
2727 if (data->start_class->flags & ANYOF_FOLD) {
2728 /* false positive possible if the class is case-folded.
2729 Assume that the locale settings are the same... */
2731 ANYOF_BITMAP_SET(data->start_class, uc);
2732 data->start_class->flags &= ~ANYOF_EOS;
2734 cl_and(data->start_class, &and_with);
2736 flags &= ~SCF_DO_STCLASS;
2738 else if (strchr((const char*)PL_varies,OP(scan))) {
2739 I32 mincount, maxcount, minnext, deltanext, fl = 0;
2740 I32 f = flags, pos_before = 0;
2741 regnode * const oscan = scan;
2742 struct regnode_charclass_class this_class;
2743 struct regnode_charclass_class *oclass = NULL;
2744 I32 next_is_eval = 0;
2746 switch (PL_regkind[OP(scan)]) {
2747 case WHILEM: /* End of (?:...)* . */
2748 scan = NEXTOPER(scan);
2751 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
2752 next = NEXTOPER(scan);
2753 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
2755 maxcount = REG_INFTY;
2756 next = regnext(scan);
2757 scan = NEXTOPER(scan);
2761 if (flags & SCF_DO_SUBSTR)
2766 if (flags & SCF_DO_STCLASS) {
2768 maxcount = REG_INFTY;
2769 next = regnext(scan);
2770 scan = NEXTOPER(scan);
2773 is_inf = is_inf_internal = 1;
2774 scan = regnext(scan);
2775 if (flags & SCF_DO_SUBSTR) {
2776 scan_commit(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
2777 data->longest = &(data->longest_float);
2779 goto optimize_curly_tail;
2781 mincount = ARG1(scan);
2782 maxcount = ARG2(scan);
2783 next = regnext(scan);
2784 if (OP(scan) == CURLYX) {
2785 I32 lp = (data ? *(data->last_closep) : 0);
2786 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
2788 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
2789 next_is_eval = (OP(scan) == EVAL);
2791 if (flags & SCF_DO_SUBSTR) {
2792 if (mincount == 0) scan_commit(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
2793 pos_before = data->pos_min;
2797 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
2799 data->flags |= SF_IS_INF;
2801 if (flags & SCF_DO_STCLASS) {
2802 cl_init(pRExC_state, &this_class);
2803 oclass = data->start_class;
2804 data->start_class = &this_class;
2805 f |= SCF_DO_STCLASS_AND;
2806 f &= ~SCF_DO_STCLASS_OR;
2808 /* These are the cases when once a subexpression
2809 fails at a particular position, it cannot succeed
2810 even after backtracking at the enclosing scope.
2812 XXXX what if minimal match and we are at the
2813 initial run of {n,m}? */
2814 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
2815 f &= ~SCF_WHILEM_VISITED_POS;
2817 /* This will finish on WHILEM, setting scan, or on NULL: */
2818 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext, last, data,
2820 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
2822 if (flags & SCF_DO_STCLASS)
2823 data->start_class = oclass;
2824 if (mincount == 0 || minnext == 0) {
2825 if (flags & SCF_DO_STCLASS_OR) {
2826 cl_or(pRExC_state, data->start_class, &this_class);
2828 else if (flags & SCF_DO_STCLASS_AND) {
2829 /* Switch to OR mode: cache the old value of
2830 * data->start_class */
2831 StructCopy(data->start_class, &and_with,
2832 struct regnode_charclass_class);
2833 flags &= ~SCF_DO_STCLASS_AND;
2834 StructCopy(&this_class, data->start_class,
2835 struct regnode_charclass_class);
2836 flags |= SCF_DO_STCLASS_OR;
2837 data->start_class->flags |= ANYOF_EOS;
2839 } else { /* Non-zero len */
2840 if (flags & SCF_DO_STCLASS_OR) {
2841 cl_or(pRExC_state, data->start_class, &this_class);
2842 cl_and(data->start_class, &and_with);
2844 else if (flags & SCF_DO_STCLASS_AND)
2845 cl_and(data->start_class, &this_class);
2846 flags &= ~SCF_DO_STCLASS;
2848 if (!scan) /* It was not CURLYX, but CURLY. */
2850 if ( /* ? quantifier ok, except for (?{ ... }) */
2851 (next_is_eval || !(mincount == 0 && maxcount == 1))
2852 && (minnext == 0) && (deltanext == 0)
2853 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
2854 && maxcount <= REG_INFTY/3 /* Complement check for big count */
2855 && ckWARN(WARN_REGEXP))
2858 "Quantifier unexpected on zero-length expression");
2861 min += minnext * mincount;
2862 is_inf_internal |= ((maxcount == REG_INFTY
2863 && (minnext + deltanext) > 0)
2864 || deltanext == I32_MAX);
2865 is_inf |= is_inf_internal;
2866 delta += (minnext + deltanext) * maxcount - minnext * mincount;
2868 /* Try powerful optimization CURLYX => CURLYN. */
2869 if ( OP(oscan) == CURLYX && data
2870 && data->flags & SF_IN_PAR
2871 && !(data->flags & SF_HAS_EVAL)
2872 && !deltanext && minnext == 1 ) {
2873 /* Try to optimize to CURLYN. */
2874 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
2875 regnode * const nxt1 = nxt;
2882 if (!strchr((const char*)PL_simple,OP(nxt))
2883 && !(PL_regkind[OP(nxt)] == EXACT
2884 && STR_LEN(nxt) == 1))
2890 if (OP(nxt) != CLOSE)
2892 /* Now we know that nxt2 is the only contents: */
2893 oscan->flags = (U8)ARG(nxt);
2895 OP(nxt1) = NOTHING; /* was OPEN. */
2897 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
2898 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
2899 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
2900 OP(nxt) = OPTIMIZED; /* was CLOSE. */
2901 OP(nxt + 1) = OPTIMIZED; /* was count. */
2902 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
2907 /* Try optimization CURLYX => CURLYM. */
2908 if ( OP(oscan) == CURLYX && data
2909 && !(data->flags & SF_HAS_PAR)
2910 && !(data->flags & SF_HAS_EVAL)
2911 && !deltanext /* atom is fixed width */
2912 && minnext != 0 /* CURLYM can't handle zero width */
2914 /* XXXX How to optimize if data == 0? */
2915 /* Optimize to a simpler form. */
2916 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
2920 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
2921 && (OP(nxt2) != WHILEM))
2923 OP(nxt2) = SUCCEED; /* Whas WHILEM */
2924 /* Need to optimize away parenths. */
2925 if (data->flags & SF_IN_PAR) {
2926 /* Set the parenth number. */
2927 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
2929 if (OP(nxt) != CLOSE)
2930 FAIL("Panic opt close");
2931 oscan->flags = (U8)ARG(nxt);
2932 OP(nxt1) = OPTIMIZED; /* was OPEN. */
2933 OP(nxt) = OPTIMIZED; /* was CLOSE. */
2935 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
2936 OP(nxt + 1) = OPTIMIZED; /* was count. */
2937 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
2938 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
2941 while ( nxt1 && (OP(nxt1) != WHILEM)) {
2942 regnode *nnxt = regnext(nxt1);
2945 if (reg_off_by_arg[OP(nxt1)])
2946 ARG_SET(nxt1, nxt2 - nxt1);
2947 else if (nxt2 - nxt1 < U16_MAX)
2948 NEXT_OFF(nxt1) = nxt2 - nxt1;
2950 OP(nxt) = NOTHING; /* Cannot beautify */
2955 /* Optimize again: */
2956 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
2962 else if ((OP(oscan) == CURLYX)
2963 && (flags & SCF_WHILEM_VISITED_POS)
2964 /* See the comment on a similar expression above.
2965 However, this time it not a subexpression
2966 we care about, but the expression itself. */
2967 && (maxcount == REG_INFTY)
2968 && data && ++data->whilem_c < 16) {
2969 /* This stays as CURLYX, we can put the count/of pair. */
2970 /* Find WHILEM (as in regexec.c) */
2971 regnode *nxt = oscan + NEXT_OFF(oscan);
2973 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
2975 PREVOPER(nxt)->flags = (U8)(data->whilem_c
2976 | (RExC_whilem_seen << 4)); /* On WHILEM */
2978 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
2980 if (flags & SCF_DO_SUBSTR) {
2981 SV *last_str = NULL;
2982 int counted = mincount != 0;
2984 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
2985 #if defined(SPARC64_GCC_WORKAROUND)
2988 const char *s = NULL;
2991 if (pos_before >= data->last_start_min)
2994 b = data->last_start_min;
2997 s = SvPV_const(data->last_found, l);
2998 old = b - data->last_start_min;
3001 I32 b = pos_before >= data->last_start_min
3002 ? pos_before : data->last_start_min;
3004 const char * const s = SvPV_const(data->last_found, l);
3005 I32 old = b - data->last_start_min;
3009 old = utf8_hop((U8*)s, old) - (U8*)s;
3012 /* Get the added string: */
3013 last_str = newSVpvn(s + old, l);
3015 SvUTF8_on(last_str);
3016 if (deltanext == 0 && pos_before == b) {
3017 /* What was added is a constant string */
3019 SvGROW(last_str, (mincount * l) + 1);
3020 repeatcpy(SvPVX(last_str) + l,
3021 SvPVX_const(last_str), l, mincount - 1);
3022 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3023 /* Add additional parts. */
3024 SvCUR_set(data->last_found,
3025 SvCUR(data->last_found) - l);
3026 sv_catsv(data->last_found, last_str);
3028 SV * sv = data->last_found;
3030 SvUTF8(sv) && SvMAGICAL(sv) ?
3031 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3032 if (mg && mg->mg_len >= 0)
3033 mg->mg_len += CHR_SVLEN(last_str);
3035 data->last_end += l * (mincount - 1);
3038 /* start offset must point into the last copy */
3039 data->last_start_min += minnext * (mincount - 1);
3040 data->last_start_max += is_inf ? I32_MAX
3041 : (maxcount - 1) * (minnext + data->pos_delta);
3044 /* It is counted once already... */
3045 data->pos_min += minnext * (mincount - counted);
3046 data->pos_delta += - counted * deltanext +
3047 (minnext + deltanext) * maxcount - minnext * mincount;
3048 if (mincount != maxcount) {
3049 /* Cannot extend fixed substrings found inside
3051 scan_commit(pRExC_state,data,minlenp);
3052 if (mincount && last_str) {
3053 SV * const sv = data->last_found;
3054 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3055 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3059 sv_setsv(sv, last_str);
3060 data->last_end = data->pos_min;
3061 data->last_start_min =
3062 data->pos_min - CHR_SVLEN(last_str);
3063 data->last_start_max = is_inf
3065 : data->pos_min + data->pos_delta
3066 - CHR_SVLEN(last_str);
3068 data->longest = &(data->longest_float);
3070 SvREFCNT_dec(last_str);
3072 if (data && (fl & SF_HAS_EVAL))
3073 data->flags |= SF_HAS_EVAL;
3074 optimize_curly_tail:
3075 if (OP(oscan) != CURLYX) {
3076 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3078 NEXT_OFF(oscan) += NEXT_OFF(next);
3081 default: /* REF and CLUMP only? */
3082 if (flags & SCF_DO_SUBSTR) {
3083 scan_commit(pRExC_state,data,minlenp); /* Cannot expect anything... */
3084 data->longest = &(data->longest_float);
3086 is_inf = is_inf_internal = 1;
3087 if (flags & SCF_DO_STCLASS_OR)
3088 cl_anything(pRExC_state, data->start_class);
3089 flags &= ~SCF_DO_STCLASS;
3093 else if (strchr((const char*)PL_simple,OP(scan))) {
3096 if (flags & SCF_DO_SUBSTR) {
3097 scan_commit(pRExC_state,data,minlenp);
3101 if (flags & SCF_DO_STCLASS) {
3102 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3104 /* Some of the logic below assumes that switching
3105 locale on will only add false positives. */
3106 switch (PL_regkind[OP(scan)]) {
3110 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3111 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3112 cl_anything(pRExC_state, data->start_class);
3115 if (OP(scan) == SANY)
3117 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3118 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3119 || (data->start_class->flags & ANYOF_CLASS));
3120 cl_anything(pRExC_state, data->start_class);
3122 if (flags & SCF_DO_STCLASS_AND || !value)
3123 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3126 if (flags & SCF_DO_STCLASS_AND)
3127 cl_and(data->start_class,
3128 (struct regnode_charclass_class*)scan);
3130 cl_or(pRExC_state, data->start_class,
3131 (struct regnode_charclass_class*)scan);
3134 if (flags & SCF_DO_STCLASS_AND) {
3135 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3136 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3137 for (value = 0; value < 256; value++)
3138 if (!isALNUM(value))
3139 ANYOF_BITMAP_CLEAR(data->start_class, value);
3143 if (data->start_class->flags & ANYOF_LOCALE)
3144 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3146 for (value = 0; value < 256; value++)
3148 ANYOF_BITMAP_SET(data->start_class, value);
3153 if (flags & SCF_DO_STCLASS_AND) {
3154 if (data->start_class->flags & ANYOF_LOCALE)
3155 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3158 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3159 data->start_class->flags |= ANYOF_LOCALE;
3163 if (flags & SCF_DO_STCLASS_AND) {
3164 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3165 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3166 for (value = 0; value < 256; value++)
3168 ANYOF_BITMAP_CLEAR(data->start_class, value);
3172 if (data->start_class->flags & ANYOF_LOCALE)
3173 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3175 for (value = 0; value < 256; value++)
3176 if (!isALNUM(value))
3177 ANYOF_BITMAP_SET(data->start_class, value);
3182 if (flags & SCF_DO_STCLASS_AND) {
3183 if (data->start_class->flags & ANYOF_LOCALE)
3184 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3187 data->start_class->flags |= ANYOF_LOCALE;
3188 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3192 if (flags & SCF_DO_STCLASS_AND) {
3193 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3194 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3195 for (value = 0; value < 256; value++)
3196 if (!isSPACE(value))
3197 ANYOF_BITMAP_CLEAR(data->start_class, value);
3201 if (data->start_class->flags & ANYOF_LOCALE)
3202 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3204 for (value = 0; value < 256; value++)
3206 ANYOF_BITMAP_SET(data->start_class, value);
3211 if (flags & SCF_DO_STCLASS_AND) {
3212 if (data->start_class->flags & ANYOF_LOCALE)
3213 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3216 data->start_class->flags |= ANYOF_LOCALE;
3217 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3221 if (flags & SCF_DO_STCLASS_AND) {
3222 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3223 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3224 for (value = 0; value < 256; value++)
3226 ANYOF_BITMAP_CLEAR(data->start_class, value);
3230 if (data->start_class->flags & ANYOF_LOCALE)
3231 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3233 for (value = 0; value < 256; value++)
3234 if (!isSPACE(value))
3235 ANYOF_BITMAP_SET(data->start_class, value);
3240 if (flags & SCF_DO_STCLASS_AND) {
3241 if (data->start_class->flags & ANYOF_LOCALE) {
3242 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3243 for (value = 0; value < 256; value++)
3244 if (!isSPACE(value))
3245 ANYOF_BITMAP_CLEAR(data->start_class, value);
3249 data->start_class->flags |= ANYOF_LOCALE;
3250 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3254 if (flags & SCF_DO_STCLASS_AND) {
3255 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3256 for (value = 0; value < 256; value++)
3257 if (!isDIGIT(value))
3258 ANYOF_BITMAP_CLEAR(data->start_class, value);
3261 if (data->start_class->flags & ANYOF_LOCALE)
3262 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3264 for (value = 0; value < 256; value++)
3266 ANYOF_BITMAP_SET(data->start_class, value);
3271 if (flags & SCF_DO_STCLASS_AND) {
3272 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3273 for (value = 0; value < 256; value++)
3275 ANYOF_BITMAP_CLEAR(data->start_class, value);
3278 if (data->start_class->flags & ANYOF_LOCALE)
3279 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3281 for (value = 0; value < 256; value++)
3282 if (!isDIGIT(value))
3283 ANYOF_BITMAP_SET(data->start_class, value);
3288 if (flags & SCF_DO_STCLASS_OR)
3289 cl_and(data->start_class, &and_with);
3290 flags &= ~SCF_DO_STCLASS;
3293 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3294 data->flags |= (OP(scan) == MEOL
3298 else if ( PL_regkind[OP(scan)] == BRANCHJ
3299 /* Lookbehind, or need to calculate parens/evals/stclass: */
3300 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3301 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3302 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3303 || OP(scan) == UNLESSM )
3305 /* Negative Lookahead/lookbehind
3306 In this case we can't do fixed string optimisation.
3309 I32 deltanext, minnext, fake = 0;
3311 struct regnode_charclass_class intrnl;
3314 data_fake.flags = 0;
3316 data_fake.whilem_c = data->whilem_c;
3317 data_fake.last_closep = data->last_closep;
3320 data_fake.last_closep = &fake;
3321 if ( flags & SCF_DO_STCLASS && !scan->flags
3322 && OP(scan) == IFMATCH ) { /* Lookahead */
3323 cl_init(pRExC_state, &intrnl);
3324 data_fake.start_class = &intrnl;
3325 f |= SCF_DO_STCLASS_AND;
3327 if (flags & SCF_WHILEM_VISITED_POS)
3328 f |= SCF_WHILEM_VISITED_POS;
3329 next = regnext(scan);
3330 nscan = NEXTOPER(NEXTOPER(scan));
3331 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext, last, &data_fake, f,depth+1);
3334 vFAIL("Variable length lookbehind not implemented");
3336 else if (minnext > (I32)U8_MAX) {
3337 vFAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3339 scan->flags = (U8)minnext;
3342 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3344 if (data_fake.flags & SF_HAS_EVAL)
3345 data->flags |= SF_HAS_EVAL;
3346 data->whilem_c = data_fake.whilem_c;
3348 if (f & SCF_DO_STCLASS_AND) {
3349 const int was = (data->start_class->flags & ANYOF_EOS);
3351 cl_and(data->start_class, &intrnl);
3353 data->start_class->flags |= ANYOF_EOS;
3356 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3358 /* Positive Lookahead/lookbehind
3359 In this case we can do fixed string optimisation,
3360 but we must be careful about it. Note in the case of
3361 lookbehind the positions will be offset by the minimum
3362 length of the pattern, something we won't know about
3363 until after the recurse.
3365 I32 deltanext, fake = 0;
3367 struct regnode_charclass_class intrnl;
3369 /* We use SAVEFREEPV so that when the full compile
3370 is finished perl will clean up the allocated
3371 minlens when its all done. This was we don't
3372 have to worry about freeing them when we know
3373 they wont be used, which would be a pain.
3376 Newx( minnextp, 1, I32 );
3377 SAVEFREEPV(minnextp);
3380 StructCopy(data, &data_fake, scan_data_t);
3381 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3384 scan_commit(pRExC_state, &data_fake,minlenp);
3385 data_fake.last_found=newSVsv(data->last_found);
3389 data_fake.last_closep = &fake;
3390 data_fake.flags = 0;
3392 data_fake.flags |= SF_IS_INF;
3393 if ( flags & SCF_DO_STCLASS && !scan->flags
3394 && OP(scan) == IFMATCH ) { /* Lookahead */
3395 cl_init(pRExC_state, &intrnl);
3396 data_fake.start_class = &intrnl;
3397 f |= SCF_DO_STCLASS_AND;
3399 if (flags & SCF_WHILEM_VISITED_POS)
3400 f |= SCF_WHILEM_VISITED_POS;
3401 next = regnext(scan);
3402 nscan = NEXTOPER(NEXTOPER(scan));
3404 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext, last, &data_fake, f,depth+1);
3407 vFAIL("Variable length lookbehind not implemented");
3409 else if (*minnextp > (I32)U8_MAX) {
3410 vFAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3412 scan->flags = (U8)*minnextp;
3418 if (f & SCF_DO_STCLASS_AND) {
3419 const int was = (data->start_class->flags & ANYOF_EOS);
3421 cl_and(data->start_class, &intrnl);
3423 data->start_class->flags |= ANYOF_EOS;
3426 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3428 if (data_fake.flags & SF_HAS_EVAL)
3429 data->flags |= SF_HAS_EVAL;
3430 data->whilem_c = data_fake.whilem_c;
3431 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3432 if (RExC_rx->minlen<*minnextp)
3433 RExC_rx->minlen=*minnextp;
3434 scan_commit(pRExC_state, &data_fake, minnextp);
3435 SvREFCNT_dec(data_fake.last_found);
3437 if ( data_fake.minlen_fixed != minlenp )
3439 data->offset_fixed= data_fake.offset_fixed;
3440 data->minlen_fixed= data_fake.minlen_fixed;
3441 data->lookbehind_fixed+= scan->flags;
3443 if ( data_fake.minlen_float != minlenp )
3445 data->minlen_float= data_fake.minlen_float;
3446 data->offset_float_min=data_fake.offset_float_min;
3447 data->offset_float_max=data_fake.offset_float_max;
3448 data->lookbehind_float+= scan->flags;
3457 else if (OP(scan) == OPEN) {
3460 else if (OP(scan) == CLOSE) {
3461 if ((I32)ARG(scan) == is_par) {
3462 next = regnext(scan);
3464 if ( next && (OP(next) != WHILEM) && next < last)
3465 is_par = 0; /* Disable optimization */
3468 *(data->last_closep) = ARG(scan);
3470 else if (OP(scan) == EVAL) {
3472 data->flags |= SF_HAS_EVAL;
3474 else if ( (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
3475 || OP(scan)==RECURSE) /* recursion */
3477 if (OP(scan)==RECURSE) {
3478 ARG2L_SET( scan, RExC_parens[ARG(scan)-1] - scan );
3480 if (flags & SCF_DO_SUBSTR) {
3481 scan_commit(pRExC_state,data,minlenp);
3482 data->longest = &(data->longest_float);
3484 is_inf = is_inf_internal = 1;
3485 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3486 cl_anything(pRExC_state, data->start_class);
3487 flags &= ~SCF_DO_STCLASS;
3489 #ifdef TRIE_STUDY_OPT
3490 #ifdef FULL_TRIE_STUDY
3491 else if (PL_regkind[OP(scan)] == TRIE) {
3492 /* NOTE - There is similar code to this block above for handling
3493 BRANCH nodes on the initial study. If you change stuff here
3495 regnode *trie_node= scan;
3496 regnode *tail= regnext(scan);
3497 reg_trie_data *trie = (reg_trie_data*)RExC_rx->data->data[ ARG(scan) ];
3498 I32 max1 = 0, min1 = I32_MAX;
3499 struct regnode_charclass_class accum;
3501 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
3502 scan_commit(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
3503 if (flags & SCF_DO_STCLASS)
3504 cl_init_zero(pRExC_state, &accum);
3510 const regnode *nextbranch= NULL;
3513 for ( word=1 ; word <= trie->wordcount ; word++)
3515 I32 deltanext=0, minnext=0, f = 0, fake;
3516 struct regnode_charclass_class this_class;
3518 data_fake.flags = 0;
3520 data_fake.whilem_c = data->whilem_c;
3521 data_fake.last_closep = data->last_closep;
3524 data_fake.last_closep = &fake;
3526 if (flags & SCF_DO_STCLASS) {
3527 cl_init(pRExC_state, &this_class);
3528 data_fake.start_class = &this_class;
3529 f = SCF_DO_STCLASS_AND;
3531 if (flags & SCF_WHILEM_VISITED_POS)
3532 f |= SCF_WHILEM_VISITED_POS;
3534 if (trie->jump[word]) {
3536 nextbranch = trie_node + trie->jump[0];
3537 scan= trie_node + trie->jump[word];
3538 /* We go from the jump point to the branch that follows
3539 it. Note this means we need the vestigal unused branches
3540 even though they arent otherwise used.
3542 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3543 (regnode *)nextbranch, &data_fake, f,depth+1);
3545 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
3546 nextbranch= regnext((regnode*)nextbranch);
3548 if (min1 > (I32)(minnext + trie->minlen))
3549 min1 = minnext + trie->minlen;
3550 if (max1 < (I32)(minnext + deltanext + trie->maxlen))
3551 max1 = minnext + deltanext + trie->maxlen;
3552 if (deltanext == I32_MAX)
3553 is_inf = is_inf_internal = 1;
3555 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3559 if (data_fake.flags & SF_HAS_EVAL)
3560 data->flags |= SF_HAS_EVAL;
3561 data->whilem_c = data_fake.whilem_c;
3563 if (flags & SCF_DO_STCLASS)
3564 cl_or(pRExC_state, &accum, &this_class);
3567 if (flags & SCF_DO_SUBSTR) {
3568 data->pos_min += min1;
3569 data->pos_delta += max1 - min1;
3570 if (max1 != min1 || is_inf)
3571 data->longest = &(data->longest_float);
3574 delta += max1 - min1;
3575 if (flags & SCF_DO_STCLASS_OR) {
3576 cl_or(pRExC_state, data->start_class, &accum);
3578 cl_and(data->start_class, &and_with);
3579 flags &= ~SCF_DO_STCLASS;
3582 else if (flags & SCF_DO_STCLASS_AND) {
3584 cl_and(data->start_class, &accum);
3585 flags &= ~SCF_DO_STCLASS;
3588 /* Switch to OR mode: cache the old value of
3589 * data->start_class */
3590 StructCopy(data->start_class, &and_with,
3591 struct regnode_charclass_class);
3592 flags &= ~SCF_DO_STCLASS_AND;
3593 StructCopy(&accum, data->start_class,
3594 struct regnode_charclass_class);
3595 flags |= SCF_DO_STCLASS_OR;
3596 data->start_class->flags |= ANYOF_EOS;
3603 else if (PL_regkind[OP(scan)] == TRIE) {
3604 reg_trie_data *trie = (reg_trie_data*)RExC_rx->data->data[ ARG(scan) ];
3607 min += trie->minlen;
3608 delta += (trie->maxlen - trie->minlen);
3609 flags &= ~SCF_DO_STCLASS; /* xxx */
3610 if (flags & SCF_DO_SUBSTR) {
3611 scan_commit(pRExC_state,data,minlenp); /* Cannot expect anything... */
3612 data->pos_min += trie->minlen;
3613 data->pos_delta += (trie->maxlen - trie->minlen);
3614 if (trie->maxlen != trie->minlen)
3615 data->longest = &(data->longest_float);
3617 if (trie->jump) /* no more substrings -- for now /grr*/
3618 flags &= ~SCF_DO_SUBSTR;
3620 #endif /* old or new */
3621 #endif /* TRIE_STUDY_OPT */
3622 /* Else: zero-length, ignore. */
3623 scan = regnext(scan);
3628 *deltap = is_inf_internal ? I32_MAX : delta;
3629 if (flags & SCF_DO_SUBSTR && is_inf)
3630 data->pos_delta = I32_MAX - data->pos_min;
3631 if (is_par > (I32)U8_MAX)
3633 if (is_par && pars==1 && data) {
3634 data->flags |= SF_IN_PAR;
3635 data->flags &= ~SF_HAS_PAR;
3637 else if (pars && data) {
3638 data->flags |= SF_HAS_PAR;
3639 data->flags &= ~SF_IN_PAR;
3641 if (flags & SCF_DO_STCLASS_OR)
3642 cl_and(data->start_class, &and_with);
3643 if (flags & SCF_TRIE_RESTUDY)
3644 data->flags |= SCF_TRIE_RESTUDY;
3646 DEBUG_STUDYDATA(data,depth);
3652 S_add_data(RExC_state_t *pRExC_state, I32 n, const char *s)
3654 if (RExC_rx->data) {
3655 const U32 count = RExC_rx->data->count;
3656 Renewc(RExC_rx->data,
3657 sizeof(*RExC_rx->data) + sizeof(void*) * (count + n - 1),
3658 char, struct reg_data);
3659 Renew(RExC_rx->data->what, count + n, U8);
3660 RExC_rx->data->count += n;
3663 Newxc(RExC_rx->data, sizeof(*RExC_rx->data) + sizeof(void*) * (n - 1),
3664 char, struct reg_data);
3665 Newx(RExC_rx->data->what, n, U8);
3666 RExC_rx->data->count = n;
3668 Copy(s, RExC_rx->data->what + RExC_rx->data->count - n, n, U8);
3669 return RExC_rx->data->count - n;
3672 #ifndef PERL_IN_XSUB_RE
3674 Perl_reginitcolors(pTHX)
3677 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
3679 char *t = savepv(s);
3683 t = strchr(t, '\t');
3689 PL_colors[i] = t = (char *)"";
3694 PL_colors[i++] = (char *)"";
3701 #ifdef TRIE_STUDY_OPT
3702 #define CHECK_RESTUDY_GOTO \
3704 (data.flags & SCF_TRIE_RESTUDY) \
3708 #define CHECK_RESTUDY_GOTO
3712 - pregcomp - compile a regular expression into internal code
3714 * We can't allocate space until we know how big the compiled form will be,
3715 * but we can't compile it (and thus know how big it is) until we've got a
3716 * place to put the code. So we cheat: we compile it twice, once with code
3717 * generation turned off and size counting turned on, and once "for real".
3718 * This also means that we don't allocate space until we are sure that the
3719 * thing really will compile successfully, and we never have to move the
3720 * code and thus invalidate pointers into it. (Note that it has to be in
3721 * one piece because free() must be able to free it all.) [NB: not true in perl]
3723 * Beware that the optimization-preparation code in here knows about some
3724 * of the structure of the compiled regexp. [I'll say.]
3729 #ifndef PERL_IN_XSUB_RE
3730 #define RE_ENGINE_PTR &PL_core_reg_engine
3732 extern const struct regexp_engine my_reg_engine;
3733 #define RE_ENGINE_PTR &my_reg_engine
3735 /* these make a few things look better, to avoid indentation */
3736 #define BEGIN_BLOCK {
3740 Perl_pregcomp(pTHX_ char *exp, char *xend, PMOP *pm)
3743 GET_RE_DEBUG_FLAGS_DECL;
3744 DEBUG_r(if (!PL_colorset) reginitcolors());
3745 #ifndef PERL_IN_XSUB_RE
3747 /* Dispatch a request to compile a regexp to correct
3749 HV * const table = GvHV(PL_hintgv);
3751 SV **ptr= hv_fetchs(table, "regcomp", FALSE);
3752 if (ptr && SvIOK(*ptr) && SvIV(*ptr)) {
3753 const regexp_engine *eng=INT2PTR(regexp_engine*,SvIV(*ptr));
3755 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
3758 return CALLREGCOMP_ENG(eng, exp, xend, pm);
3772 RExC_state_t RExC_state;
3773 RExC_state_t * const pRExC_state = &RExC_state;
3774 #ifdef TRIE_STUDY_OPT
3776 RExC_state_t copyRExC_state;
3779 FAIL("NULL regexp argument");
3781 RExC_utf8 = pm->op_pmdynflags & PMdf_CMP_UTF8;
3785 SV *dsv= sv_newmortal();
3786 RE_PV_QUOTED_DECL(s, RExC_utf8,
3787 dsv, RExC_precomp, (xend - exp), 60);
3788 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
3789 PL_colors[4],PL_colors[5],s);
3791 RExC_flags = pm->op_pmflags;
3795 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
3796 RExC_seen_evals = 0;
3799 /* First pass: determine size, legality. */
3806 RExC_emit = &PL_regdummy;
3807 RExC_whilem_seen = 0;
3808 RExC_charnames = NULL;
3810 RExC_paren_names = NULL;
3812 #if 0 /* REGC() is (currently) a NOP at the first pass.
3813 * Clever compilers notice this and complain. --jhi */
3814 REGC((U8)REG_MAGIC, (char*)RExC_emit);
3816 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n"));
3817 if (reg(pRExC_state, 0, &flags,1) == NULL) {
3818 RExC_precomp = NULL;
3822 PerlIO_printf(Perl_debug_log,
3823 "Required size %"IVdf" nodes\n"
3824 "Starting second pass (creation)\n",
3827 RExC_lastparse=NULL;
3829 /* Small enough for pointer-storage convention?
3830 If extralen==0, this means that we will not need long jumps. */
3831 if (RExC_size >= 0x10000L && RExC_extralen)
3832 RExC_size += RExC_extralen;
3835 if (RExC_whilem_seen > 15)
3836 RExC_whilem_seen = 15;
3838 /* Allocate space and zero-initialize. Note, the two step process
3839 of zeroing when in debug mode, thus anything assigned has to
3840 happen after that */
3841 Newxc(r, sizeof(regexp) + (unsigned)RExC_size * sizeof(regnode),
3844 FAIL("Regexp out of space");
3846 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
3847 Zero(r, sizeof(regexp) + (unsigned)RExC_size * sizeof(regnode), char);
3849 /* initialization begins here */
3850 r->engine= RE_ENGINE_PTR;
3852 r->prelen = xend - exp;
3853 r->precomp = savepvn(RExC_precomp, r->prelen);
3855 #ifdef PERL_OLD_COPY_ON_WRITE
3856 r->saved_copy = NULL;
3858 r->reganch = pm->op_pmflags & PMf_COMPILETIME;
3859 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
3860 r->lastparen = 0; /* mg.c reads this. */
3862 r->substrs = 0; /* Useful during FAIL. */
3863 r->startp = 0; /* Useful during FAIL. */
3867 if (RExC_seen & REG_SEEN_RECURSE) {
3868 Newxz(RExC_parens, RExC_npar,regnode *);
3869 SAVEFREEPV(RExC_parens);
3872 /* Useful during FAIL. */
3873 Newxz(r->offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
3875 r->offsets[0] = RExC_size;
3877 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
3878 "%s %"UVuf" bytes for offset annotations.\n",
3879 r->offsets ? "Got" : "Couldn't get",
3880 (UV)((2*RExC_size+1) * sizeof(U32))));
3884 /* Second pass: emit code. */
3885 RExC_flags = pm->op_pmflags; /* don't let top level (?i) bleed */
3890 RExC_emit_start = r->program;
3891 RExC_emit = r->program;
3892 /* Store the count of eval-groups for security checks: */
3893 RExC_emit->next_off = (RExC_seen_evals > (I32)U16_MAX) ?