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->laststate ; 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
911 PerlIO_printf( Perl_debug_log, "\n");
916 dump_trie_interim_table(trie,next_alloc)
917 Dumps a fully constructed but uncompressed trie in table form.
918 This is the normal DFA style state transition table, with a few
919 twists to facilitate compression later.
920 Used for debugging make_trie().
923 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie, U32 next_alloc, U32 depth)
927 SV *sv=sv_newmortal();
928 int colwidth= trie->widecharmap ? 6 : 4;
929 GET_RE_DEBUG_FLAGS_DECL;
932 print out the table precompression so that we can do a visual check
933 that they are identical.
936 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
938 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
939 SV ** const tmp = av_fetch( trie->revcharmap, charid, 0);
941 PerlIO_printf( Perl_debug_log, "%*s",
943 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
944 PL_colors[0], PL_colors[1],
945 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
946 PERL_PV_ESCAPE_FIRSTCHAR
952 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
954 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
955 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
958 PerlIO_printf( Perl_debug_log, "\n" );
960 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
962 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
963 (int)depth * 2 + 2,"",
964 (UV)TRIE_NODENUM( state ) );
966 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
967 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
969 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
971 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
973 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
974 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
976 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
977 trie->states[ TRIE_NODENUM( state ) ].wordnum );
984 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
985 startbranch: the first branch in the whole branch sequence
986 first : start branch of sequence of branch-exact nodes.
987 May be the same as startbranch
988 last : Thing following the last branch.
989 May be the same as tail.
990 tail : item following the branch sequence
991 count : words in the sequence
992 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
995 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
997 A trie is an N'ary tree where the branches are determined by digital
998 decomposition of the key. IE, at the root node you look up the 1st character and
999 follow that branch repeat until you find the end of the branches. Nodes can be
1000 marked as "accepting" meaning they represent a complete word. Eg:
1004 would convert into the following structure. Numbers represent states, letters
1005 following numbers represent valid transitions on the letter from that state, if
1006 the number is in square brackets it represents an accepting state, otherwise it
1007 will be in parenthesis.
1009 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1013 (1) +-i->(6)-+-s->[7]
1015 +-s->(3)-+-h->(4)-+-e->[5]
1017 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1019 This shows that when matching against the string 'hers' we will begin at state 1
1020 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1021 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1022 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1023 single traverse. We store a mapping from accepting to state to which word was
1024 matched, and then when we have multiple possibilities we try to complete the
1025 rest of the regex in the order in which they occured in the alternation.
1027 The only prior NFA like behaviour that would be changed by the TRIE support is
1028 the silent ignoring of duplicate alternations which are of the form:
1030 / (DUPE|DUPE) X? (?{ ... }) Y /x
1032 Thus EVAL blocks follwing a trie may be called a different number of times with
1033 and without the optimisation. With the optimisations dupes will be silently
1034 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1035 the following demonstrates:
1037 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1039 which prints out 'word' three times, but
1041 'words'=~/(word|word|word)(?{ print $1 })S/
1043 which doesnt print it out at all. This is due to other optimisations kicking in.
1045 Example of what happens on a structural level:
1047 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1049 1: CURLYM[1] {1,32767}(18)
1060 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1061 and should turn into:
1063 1: CURLYM[1] {1,32767}(18)
1065 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1073 Cases where tail != last would be like /(?foo|bar)baz/:
1083 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1084 and would end up looking like:
1087 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1094 d = uvuni_to_utf8_flags(d, uv, 0);
1096 is the recommended Unicode-aware way of saying
1101 #define TRIE_STORE_REVCHAR \
1103 SV *tmp = Perl_newSVpvf_nocontext( "%c", (int)uvc ); \
1104 if (UTF) SvUTF8_on(tmp); \
1105 av_push( TRIE_REVCHARMAP(trie), tmp ); \
1108 #define TRIE_READ_CHAR STMT_START { \
1112 if ( foldlen > 0 ) { \
1113 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1118 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1119 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1120 foldlen -= UNISKIP( uvc ); \
1121 scan = foldbuf + UNISKIP( uvc ); \
1124 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1134 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1135 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1136 TRIE_LIST_LEN( state ) *= 2; \
1137 Renew( trie->states[ state ].trans.list, \
1138 TRIE_LIST_LEN( state ), reg_trie_trans_le ); \
1140 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1141 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1142 TRIE_LIST_CUR( state )++; \
1145 #define TRIE_LIST_NEW(state) STMT_START { \
1146 Newxz( trie->states[ state ].trans.list, \
1147 4, reg_trie_trans_le ); \
1148 TRIE_LIST_CUR( state ) = 1; \
1149 TRIE_LIST_LEN( state ) = 4; \
1152 #define TRIE_HANDLE_WORD(state) STMT_START { \
1153 U16 dupe= trie->states[ state ].wordnum; \
1154 regnode * const noper_next = regnext( noper ); \
1156 if (trie->wordlen) \
1157 trie->wordlen[ curword ] = wordlen; \
1159 /* store the word for dumping */ \
1161 if (OP(noper) != NOTHING) \
1162 tmp = newSVpvn(STRING(noper), STR_LEN(noper)); \
1164 tmp = newSVpvn( "", 0 ); \
1165 if ( UTF ) SvUTF8_on( tmp ); \
1166 av_push( trie->words, tmp ); \
1171 if ( noper_next < tail ) { \
1173 Newxz( trie->jump, word_count + 1, U16); \
1174 trie->jump[curword] = (U16)(tail - noper_next); \
1176 jumper = noper_next; \
1178 nextbranch= regnext(cur); \
1182 /* So it's a dupe. This means we need to maintain a */\
1183 /* linked-list from the first to the next. */\
1184 /* we only allocate the nextword buffer when there */\
1185 /* a dupe, so first time we have to do the allocation */\
1186 if (!trie->nextword) \
1187 Newxz( trie->nextword, word_count + 1, U16); \
1188 while ( trie->nextword[dupe] ) \
1189 dupe= trie->nextword[dupe]; \
1190 trie->nextword[dupe]= curword; \
1192 /* we haven't inserted this word yet. */ \
1193 trie->states[ state ].wordnum = curword; \
1198 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1199 ( ( base + charid >= ucharcount \
1200 && base + charid < ubound \
1201 && state == trie->trans[ base - ucharcount + charid ].check \
1202 && trie->trans[ base - ucharcount + charid ].next ) \
1203 ? trie->trans[ base - ucharcount + charid ].next \
1204 : ( state==1 ? special : 0 ) \
1208 #define MADE_JUMP_TRIE 2
1209 #define MADE_EXACT_TRIE 4
1212 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1215 /* first pass, loop through and scan words */
1216 reg_trie_data *trie;
1218 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1223 regnode *jumper = NULL;
1224 regnode *nextbranch = NULL;
1225 /* we just use folder as a flag in utf8 */
1226 const U8 * const folder = ( flags == EXACTF
1228 : ( flags == EXACTFL
1234 const U32 data_slot = add_data( pRExC_state, 1, "t" );
1235 SV *re_trie_maxbuff;
1237 /* these are only used during construction but are useful during
1238 * debugging so we store them in the struct when debugging.
1240 STRLEN trie_charcount=0;
1241 AV *trie_revcharmap;
1243 GET_RE_DEBUG_FLAGS_DECL;
1245 PERL_UNUSED_ARG(depth);
1248 Newxz( trie, 1, reg_trie_data );
1250 trie->startstate = 1;
1251 trie->wordcount = word_count;
1252 RExC_rx->data->data[ data_slot ] = (void*)trie;
1253 Newxz( trie->charmap, 256, U16 );
1254 if (!(UTF && folder))
1255 Newxz( trie->bitmap, ANYOF_BITMAP_SIZE, char );
1257 trie->words = newAV();
1259 TRIE_REVCHARMAP(trie) = newAV();
1261 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1262 if (!SvIOK(re_trie_maxbuff)) {
1263 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1266 PerlIO_printf( Perl_debug_log,
1267 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1268 (int)depth * 2 + 2, "",
1269 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1270 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1273 /* -- First loop and Setup --
1275 We first traverse the branches and scan each word to determine if it
1276 contains widechars, and how many unique chars there are, this is
1277 important as we have to build a table with at least as many columns as we
1280 We use an array of integers to represent the character codes 0..255
1281 (trie->charmap) and we use a an HV* to store unicode characters. We use the
1282 native representation of the character value as the key and IV's for the
1285 *TODO* If we keep track of how many times each character is used we can
1286 remap the columns so that the table compression later on is more
1287 efficient in terms of memory by ensuring most common value is in the
1288 middle and the least common are on the outside. IMO this would be better
1289 than a most to least common mapping as theres a decent chance the most
1290 common letter will share a node with the least common, meaning the node
1291 will not be compressable. With a middle is most common approach the worst
1292 case is when we have the least common nodes twice.
1296 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1297 regnode * const noper = NEXTOPER( cur );
1298 const U8 *uc = (U8*)STRING( noper );
1299 const U8 * const e = uc + STR_LEN( noper );
1301 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1302 const U8 *scan = (U8*)NULL;
1303 U32 wordlen = 0; /* required init */
1306 if (OP(noper) == NOTHING) {
1311 TRIE_BITMAP_SET(trie,*uc);
1312 if ( folder ) TRIE_BITMAP_SET(trie,folder[ *uc ]);
1314 for ( ; uc < e ; uc += len ) {
1315 TRIE_CHARCOUNT(trie)++;
1319 if ( !trie->charmap[ uvc ] ) {
1320 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1322 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1327 if ( !trie->widecharmap )
1328 trie->widecharmap = newHV();
1330 svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1333 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1335 if ( !SvTRUE( *svpp ) ) {
1336 sv_setiv( *svpp, ++trie->uniquecharcount );
1341 if( cur == first ) {
1344 } else if (chars < trie->minlen) {
1346 } else if (chars > trie->maxlen) {
1350 } /* end first pass */
1351 DEBUG_TRIE_COMPILE_r(
1352 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1353 (int)depth * 2 + 2,"",
1354 ( trie->widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1355 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1356 (int)trie->minlen, (int)trie->maxlen )
1358 Newxz( trie->wordlen, word_count, U32 );
1361 We now know what we are dealing with in terms of unique chars and
1362 string sizes so we can calculate how much memory a naive
1363 representation using a flat table will take. If it's over a reasonable
1364 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1365 conservative but potentially much slower representation using an array
1368 At the end we convert both representations into the same compressed
1369 form that will be used in regexec.c for matching with. The latter
1370 is a form that cannot be used to construct with but has memory
1371 properties similar to the list form and access properties similar
1372 to the table form making it both suitable for fast searches and
1373 small enough that its feasable to store for the duration of a program.
1375 See the comment in the code where the compressed table is produced
1376 inplace from the flat tabe representation for an explanation of how
1377 the compression works.
1382 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1384 Second Pass -- Array Of Lists Representation
1386 Each state will be represented by a list of charid:state records
1387 (reg_trie_trans_le) the first such element holds the CUR and LEN
1388 points of the allocated array. (See defines above).
1390 We build the initial structure using the lists, and then convert
1391 it into the compressed table form which allows faster lookups
1392 (but cant be modified once converted).
1395 STRLEN transcount = 1;
1397 Newxz( trie->states, TRIE_CHARCOUNT(trie) + 2, reg_trie_state );
1401 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1403 regnode * const noper = NEXTOPER( cur );
1404 U8 *uc = (U8*)STRING( noper );
1405 const U8 * const e = uc + STR_LEN( noper );
1406 U32 state = 1; /* required init */
1407 U16 charid = 0; /* sanity init */
1408 U8 *scan = (U8*)NULL; /* sanity init */
1409 STRLEN foldlen = 0; /* required init */
1410 U32 wordlen = 0; /* required init */
1411 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1413 if (OP(noper) != NOTHING) {
1414 for ( ; uc < e ; uc += len ) {
1419 charid = trie->charmap[ uvc ];
1421 SV** const svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 0);
1425 charid=(U16)SvIV( *svpp );
1428 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1435 if ( !trie->states[ state ].trans.list ) {
1436 TRIE_LIST_NEW( state );
1438 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1439 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1440 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1445 newstate = next_alloc++;
1446 TRIE_LIST_PUSH( state, charid, newstate );
1451 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1455 TRIE_HANDLE_WORD(state);
1457 } /* end second pass */
1459 trie->laststate = next_alloc;
1460 Renew( trie->states, next_alloc, reg_trie_state );
1462 /* and now dump it out before we compress it */
1463 DEBUG_TRIE_COMPILE_MORE_r(
1464 dump_trie_interim_list(trie,next_alloc,depth+1)
1467 Newxz( trie->trans, transcount ,reg_trie_trans );
1474 for( state=1 ; state < next_alloc ; state ++ ) {
1478 DEBUG_TRIE_COMPILE_MORE_r(
1479 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1483 if (trie->states[state].trans.list) {
1484 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1488 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1489 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1490 if ( forid < minid ) {
1492 } else if ( forid > maxid ) {
1496 if ( transcount < tp + maxid - minid + 1) {
1498 Renew( trie->trans, transcount, reg_trie_trans );
1499 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1501 base = trie->uniquecharcount + tp - minid;
1502 if ( maxid == minid ) {
1504 for ( ; zp < tp ; zp++ ) {
1505 if ( ! trie->trans[ zp ].next ) {
1506 base = trie->uniquecharcount + zp - minid;
1507 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1508 trie->trans[ zp ].check = state;
1514 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1515 trie->trans[ tp ].check = state;
1520 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1521 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1522 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1523 trie->trans[ tid ].check = state;
1525 tp += ( maxid - minid + 1 );
1527 Safefree(trie->states[ state ].trans.list);
1530 DEBUG_TRIE_COMPILE_MORE_r(
1531 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1534 trie->states[ state ].trans.base=base;
1536 trie->lasttrans = tp + 1;
1540 Second Pass -- Flat Table Representation.
1542 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1543 We know that we will need Charcount+1 trans at most to store the data
1544 (one row per char at worst case) So we preallocate both structures
1545 assuming worst case.
1547 We then construct the trie using only the .next slots of the entry
1550 We use the .check field of the first entry of the node temporarily to
1551 make compression both faster and easier by keeping track of how many non
1552 zero fields are in the node.
1554 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1557 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1558 number representing the first entry of the node, and state as a
1559 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1560 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1561 are 2 entrys per node. eg:
1569 The table is internally in the right hand, idx form. However as we also
1570 have to deal with the states array which is indexed by nodenum we have to
1571 use TRIE_NODENUM() to convert.
1576 Newxz( trie->trans, ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1,
1578 Newxz( trie->states, TRIE_CHARCOUNT(trie) + 2, reg_trie_state );
1579 next_alloc = trie->uniquecharcount + 1;
1582 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1584 regnode * const noper = NEXTOPER( cur );
1585 const U8 *uc = (U8*)STRING( noper );
1586 const U8 * const e = uc + STR_LEN( noper );
1588 U32 state = 1; /* required init */
1590 U16 charid = 0; /* sanity init */
1591 U32 accept_state = 0; /* sanity init */
1592 U8 *scan = (U8*)NULL; /* sanity init */
1594 STRLEN foldlen = 0; /* required init */
1595 U32 wordlen = 0; /* required init */
1596 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1598 if ( OP(noper) != NOTHING ) {
1599 for ( ; uc < e ; uc += len ) {
1604 charid = trie->charmap[ uvc ];
1606 SV* const * const svpp = hv_fetch( trie->widecharmap, (char*)&uvc, sizeof( UV ), 0);
1607 charid = svpp ? (U16)SvIV(*svpp) : 0;
1611 if ( !trie->trans[ state + charid ].next ) {
1612 trie->trans[ state + charid ].next = next_alloc;
1613 trie->trans[ state ].check++;
1614 next_alloc += trie->uniquecharcount;
1616 state = trie->trans[ state + charid ].next;
1618 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1620 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1623 accept_state = TRIE_NODENUM( state );
1624 TRIE_HANDLE_WORD(accept_state);
1626 } /* end second pass */
1628 /* and now dump it out before we compress it */
1629 DEBUG_TRIE_COMPILE_MORE_r(
1630 dump_trie_interim_table(trie,next_alloc,depth+1)
1635 * Inplace compress the table.*
1637 For sparse data sets the table constructed by the trie algorithm will
1638 be mostly 0/FAIL transitions or to put it another way mostly empty.
1639 (Note that leaf nodes will not contain any transitions.)
1641 This algorithm compresses the tables by eliminating most such
1642 transitions, at the cost of a modest bit of extra work during lookup:
1644 - Each states[] entry contains a .base field which indicates the
1645 index in the state[] array wheres its transition data is stored.
1647 - If .base is 0 there are no valid transitions from that node.
1649 - If .base is nonzero then charid is added to it to find an entry in
1652 -If trans[states[state].base+charid].check!=state then the
1653 transition is taken to be a 0/Fail transition. Thus if there are fail
1654 transitions at the front of the node then the .base offset will point
1655 somewhere inside the previous nodes data (or maybe even into a node
1656 even earlier), but the .check field determines if the transition is
1660 The following process inplace converts the table to the compressed
1661 table: We first do not compress the root node 1,and mark its all its
1662 .check pointers as 1 and set its .base pointer as 1 as well. This
1663 allows to do a DFA construction from the compressed table later, and
1664 ensures that any .base pointers we calculate later are greater than
1667 - We set 'pos' to indicate the first entry of the second node.
1669 - We then iterate over the columns of the node, finding the first and
1670 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1671 and set the .check pointers accordingly, and advance pos
1672 appropriately and repreat for the next node. Note that when we copy
1673 the next pointers we have to convert them from the original
1674 NODEIDX form to NODENUM form as the former is not valid post
1677 - If a node has no transitions used we mark its base as 0 and do not
1678 advance the pos pointer.
1680 - If a node only has one transition we use a second pointer into the
1681 structure to fill in allocated fail transitions from other states.
1682 This pointer is independent of the main pointer and scans forward
1683 looking for null transitions that are allocated to a state. When it
1684 finds one it writes the single transition into the "hole". If the
1685 pointer doesnt find one the single transition is appended as normal.
1687 - Once compressed we can Renew/realloc the structures to release the
1690 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1691 specifically Fig 3.47 and the associated pseudocode.
1695 const U32 laststate = TRIE_NODENUM( next_alloc );
1698 trie->laststate = laststate;
1700 for ( state = 1 ; state < laststate ; state++ ) {
1702 const U32 stateidx = TRIE_NODEIDX( state );
1703 const U32 o_used = trie->trans[ stateidx ].check;
1704 U32 used = trie->trans[ stateidx ].check;
1705 trie->trans[ stateidx ].check = 0;
1707 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1708 if ( flag || trie->trans[ stateidx + charid ].next ) {
1709 if ( trie->trans[ stateidx + charid ].next ) {
1711 for ( ; zp < pos ; zp++ ) {
1712 if ( ! trie->trans[ zp ].next ) {
1716 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1717 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1718 trie->trans[ zp ].check = state;
1719 if ( ++zp > pos ) pos = zp;
1726 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1728 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1729 trie->trans[ pos ].check = state;
1734 trie->lasttrans = pos + 1;
1735 Renew( trie->states, laststate + 1, reg_trie_state);
1736 DEBUG_TRIE_COMPILE_MORE_r(
1737 PerlIO_printf( Perl_debug_log,
1738 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1739 (int)depth * 2 + 2,"",
1740 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1743 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1746 } /* end table compress */
1748 /* resize the trans array to remove unused space */
1749 Renew( trie->trans, trie->lasttrans, reg_trie_trans);
1751 /* and now dump out the compressed format */
1752 DEBUG_TRIE_COMPILE_r(
1753 dump_trie(trie,depth+1)
1756 { /* Modify the program and insert the new TRIE node*/
1758 U8 nodetype =(U8)(flags & 0xFF);
1762 regnode *optimize = NULL;
1764 U32 mjd_nodelen = 0;
1767 This means we convert either the first branch or the first Exact,
1768 depending on whether the thing following (in 'last') is a branch
1769 or not and whther first is the startbranch (ie is it a sub part of
1770 the alternation or is it the whole thing.)
1771 Assuming its a sub part we conver the EXACT otherwise we convert
1772 the whole branch sequence, including the first.
1774 /* Find the node we are going to overwrite */
1775 if ( first == startbranch && OP( last ) != BRANCH ) {
1776 /* whole branch chain */
1779 const regnode *nop = NEXTOPER( convert );
1780 mjd_offset= Node_Offset((nop));
1781 mjd_nodelen= Node_Length((nop));
1784 /* branch sub-chain */
1785 convert = NEXTOPER( first );
1786 NEXT_OFF( first ) = (U16)(last - first);
1788 mjd_offset= Node_Offset((convert));
1789 mjd_nodelen= Node_Length((convert));
1793 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1794 (int)depth * 2 + 2, "",
1795 (UV)mjd_offset, (UV)mjd_nodelen)
1798 /* But first we check to see if there is a common prefix we can
1799 split out as an EXACT and put in front of the TRIE node. */
1800 trie->startstate= 1;
1801 if ( trie->bitmap && !trie->widecharmap && !trie->jump ) {
1804 PerlIO_printf(Perl_debug_log, "%*sLaststate:%"UVuf"\n",
1805 (int)depth * 2 + 2, "",
1806 (UV)trie->laststate)
1808 for ( state = 1 ; state < trie->laststate-1 ; state++ ) {
1812 const U32 base = trie->states[ state ].trans.base;
1814 if ( trie->states[state].wordnum )
1817 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1818 if ( ( base + ofs >= trie->uniquecharcount ) &&
1819 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1820 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1822 if ( ++count > 1 ) {
1823 SV **tmp = av_fetch( TRIE_REVCHARMAP(trie), ofs, 0);
1824 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1825 if ( state == 1 ) break;
1827 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1829 PerlIO_printf(Perl_debug_log,
1830 "%*sNew Start State=%"UVuf" Class: [",
1831 (int)depth * 2 + 2, "",
1834 SV ** const tmp = av_fetch( TRIE_REVCHARMAP(trie), idx, 0);
1835 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1837 TRIE_BITMAP_SET(trie,*ch);
1839 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1841 PerlIO_printf(Perl_debug_log, (char*)ch)
1845 TRIE_BITMAP_SET(trie,*ch);
1847 TRIE_BITMAP_SET(trie,folder[ *ch ]);
1848 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
1854 SV **tmp = av_fetch( TRIE_REVCHARMAP(trie), idx, 0);
1855 const char *ch = SvPV_nolen_const( *tmp );
1857 PerlIO_printf( Perl_debug_log,
1858 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
1859 (int)depth * 2 + 2, "",
1860 (UV)state, (UV)idx, ch)
1863 OP( convert ) = nodetype;
1864 str=STRING(convert);
1873 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
1879 regnode *n = convert+NODE_SZ_STR(convert);
1880 NEXT_OFF(convert) = NODE_SZ_STR(convert);
1881 trie->startstate = state;
1882 trie->minlen -= (state - 1);
1883 trie->maxlen -= (state - 1);
1885 regnode *fix = convert;
1887 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
1888 while( ++fix < n ) {
1889 Set_Node_Offset_Length(fix, 0, 0);
1895 NEXT_OFF(convert) = (U16)(tail - convert);
1896 DEBUG_r(optimize= n);
1902 if ( trie->maxlen ) {
1903 NEXT_OFF( convert ) = (U16)(tail - convert);
1904 ARG_SET( convert, data_slot );
1905 /* Store the offset to the first unabsorbed branch in
1906 jump[0], which is otherwise unused by the jump logic.
1907 We use this when dumping a trie and during optimisation. */
1909 trie->jump[0] = (U16)(tail - nextbranch);
1912 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
1913 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
1915 OP( convert ) = TRIEC;
1916 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
1917 Safefree(trie->bitmap);
1920 OP( convert ) = TRIE;
1922 /* store the type in the flags */
1923 convert->flags = nodetype;
1927 + regarglen[ OP( convert ) ];
1929 /* XXX We really should free up the resource in trie now,
1930 as we won't use them - (which resources?) dmq */
1932 /* needed for dumping*/
1933 DEBUG_r(if (optimize) {
1934 regnode *opt = convert;
1935 while ( ++opt < optimize) {
1936 Set_Node_Offset_Length(opt,0,0);
1939 Try to clean up some of the debris left after the
1942 while( optimize < jumper ) {
1943 mjd_nodelen += Node_Length((optimize));
1944 OP( optimize ) = OPTIMIZED;
1945 Set_Node_Offset_Length(optimize,0,0);
1948 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
1950 } /* end node insert */
1952 SvREFCNT_dec(TRIE_REVCHARMAP(trie));
1956 : trie->startstate>1
1962 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
1964 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
1966 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
1967 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
1970 We find the fail state for each state in the trie, this state is the longest proper
1971 suffix of the current states 'word' that is also a proper prefix of another word in our
1972 trie. State 1 represents the word '' and is the thus the default fail state. This allows
1973 the DFA not to have to restart after its tried and failed a word at a given point, it
1974 simply continues as though it had been matching the other word in the first place.
1976 'abcdgu'=~/abcdefg|cdgu/
1977 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
1978 fail, which would bring use to the state representing 'd' in the second word where we would
1979 try 'g' and succeed, prodceding to match 'cdgu'.
1981 /* add a fail transition */
1982 reg_trie_data *trie=(reg_trie_data *)RExC_rx->data->data[ARG(source)];
1984 const U32 ucharcount = trie->uniquecharcount;
1985 const U32 numstates = trie->laststate;
1986 const U32 ubound = trie->lasttrans + ucharcount;
1990 U32 base = trie->states[ 1 ].trans.base;
1993 const U32 data_slot = add_data( pRExC_state, 1, "T" );
1994 GET_RE_DEBUG_FLAGS_DECL;
1996 PERL_UNUSED_ARG(depth);
2000 ARG_SET( stclass, data_slot );
2001 Newxz( aho, 1, reg_ac_data );
2002 RExC_rx->data->data[ data_slot ] = (void*)aho;
2004 aho->states=(reg_trie_state *)savepvn((const char*)trie->states,
2005 (trie->laststate+1)*sizeof(reg_trie_state));
2006 Newxz( q, numstates, U32);
2007 Newxz( aho->fail, numstates, U32 );
2010 /* initialize fail[0..1] to be 1 so that we always have
2011 a valid final fail state */
2012 fail[ 0 ] = fail[ 1 ] = 1;
2014 for ( charid = 0; charid < ucharcount ; charid++ ) {
2015 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2017 q[ q_write ] = newstate;
2018 /* set to point at the root */
2019 fail[ q[ q_write++ ] ]=1;
2022 while ( q_read < q_write) {
2023 const U32 cur = q[ q_read++ % numstates ];
2024 base = trie->states[ cur ].trans.base;
2026 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2027 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2029 U32 fail_state = cur;
2032 fail_state = fail[ fail_state ];
2033 fail_base = aho->states[ fail_state ].trans.base;
2034 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2036 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2037 fail[ ch_state ] = fail_state;
2038 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2040 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2042 q[ q_write++ % numstates] = ch_state;
2046 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2047 when we fail in state 1, this allows us to use the
2048 charclass scan to find a valid start char. This is based on the principle
2049 that theres a good chance the string being searched contains lots of stuff
2050 that cant be a start char.
2052 fail[ 0 ] = fail[ 1 ] = 0;
2053 DEBUG_TRIE_COMPILE_r({
2054 PerlIO_printf(Perl_debug_log, "%*sStclass Failtable: 0", (int)(depth * 2), "");
2055 for( q_read=1; q_read<numstates; q_read++ ) {
2056 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2058 PerlIO_printf(Perl_debug_log, "\n");
2061 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2066 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2067 * These need to be revisited when a newer toolchain becomes available.
2069 #if defined(__sparc64__) && defined(__GNUC__)
2070 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2071 # undef SPARC64_GCC_WORKAROUND
2072 # define SPARC64_GCC_WORKAROUND 1
2076 #define DEBUG_PEEP(str,scan,depth) \
2077 DEBUG_OPTIMISE_r({ \
2078 SV * const mysv=sv_newmortal(); \
2079 regnode *Next = regnext(scan); \
2080 regprop(RExC_rx, mysv, scan); \
2081 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s [%d]\n", \
2082 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2083 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2090 #define JOIN_EXACT(scan,min,flags) \
2091 if (PL_regkind[OP(scan)] == EXACT) \
2092 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2095 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2096 /* Merge several consecutive EXACTish nodes into one. */
2097 regnode *n = regnext(scan);
2099 regnode *next = scan + NODE_SZ_STR(scan);
2103 regnode *stop = scan;
2104 GET_RE_DEBUG_FLAGS_DECL;
2106 PERL_UNUSED_ARG(depth);
2108 #ifndef EXPERIMENTAL_INPLACESCAN
2109 PERL_UNUSED_ARG(flags);
2110 PERL_UNUSED_ARG(val);
2112 DEBUG_PEEP("join",scan,depth);
2114 /* Skip NOTHING, merge EXACT*. */
2116 ( PL_regkind[OP(n)] == NOTHING ||
2117 (stringok && (OP(n) == OP(scan))))
2119 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2121 if (OP(n) == TAIL || n > next)
2123 if (PL_regkind[OP(n)] == NOTHING) {
2124 DEBUG_PEEP("skip:",n,depth);
2125 NEXT_OFF(scan) += NEXT_OFF(n);
2126 next = n + NODE_STEP_REGNODE;
2133 else if (stringok) {
2134 const unsigned int oldl = STR_LEN(scan);
2135 regnode * const nnext = regnext(n);
2137 DEBUG_PEEP("merg",n,depth);
2140 if (oldl + STR_LEN(n) > U8_MAX)
2142 NEXT_OFF(scan) += NEXT_OFF(n);
2143 STR_LEN(scan) += STR_LEN(n);
2144 next = n + NODE_SZ_STR(n);
2145 /* Now we can overwrite *n : */
2146 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2154 #ifdef EXPERIMENTAL_INPLACESCAN
2155 if (flags && !NEXT_OFF(n)) {
2156 DEBUG_PEEP("atch", val, depth);
2157 if (reg_off_by_arg[OP(n)]) {
2158 ARG_SET(n, val - n);
2161 NEXT_OFF(n) = val - n;
2168 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2170 Two problematic code points in Unicode casefolding of EXACT nodes:
2172 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2173 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2179 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2180 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2182 This means that in case-insensitive matching (or "loose matching",
2183 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2184 length of the above casefolded versions) can match a target string
2185 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2186 This would rather mess up the minimum length computation.
2188 What we'll do is to look for the tail four bytes, and then peek
2189 at the preceding two bytes to see whether we need to decrease
2190 the minimum length by four (six minus two).
2192 Thanks to the design of UTF-8, there cannot be false matches:
2193 A sequence of valid UTF-8 bytes cannot be a subsequence of
2194 another valid sequence of UTF-8 bytes.
2197 char * const s0 = STRING(scan), *s, *t;
2198 char * const s1 = s0 + STR_LEN(scan) - 1;
2199 char * const s2 = s1 - 4;
2200 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2201 const char t0[] = "\xaf\x49\xaf\x42";
2203 const char t0[] = "\xcc\x88\xcc\x81";
2205 const char * const t1 = t0 + 3;
2208 s < s2 && (t = ninstr(s, s1, t0, t1));
2211 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2212 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2214 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2215 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2223 n = scan + NODE_SZ_STR(scan);
2225 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2232 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2236 /* REx optimizer. Converts nodes into quickier variants "in place".
2237 Finds fixed substrings. */
2239 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2240 to the position after last scanned or to NULL. */
2245 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2246 I32 *minlenp, I32 *deltap,
2247 regnode *last, scan_data_t *data, U32 flags, U32 depth)
2248 /* scanp: Start here (read-write). */
2249 /* deltap: Write maxlen-minlen here. */
2250 /* last: Stop before this one. */
2253 I32 min = 0, pars = 0, code;
2254 regnode *scan = *scanp, *next;
2256 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2257 int is_inf_internal = 0; /* The studied chunk is infinite */
2258 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2259 scan_data_t data_fake;
2260 struct regnode_charclass_class and_with; /* Valid if flags & SCF_DO_STCLASS_OR */
2261 SV *re_trie_maxbuff = NULL;
2262 regnode *first_non_open = scan;
2265 GET_RE_DEBUG_FLAGS_DECL;
2267 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2270 while (first_non_open && OP(first_non_open) == OPEN)
2271 first_non_open=regnext(first_non_open);
2275 while (scan && OP(scan) != END && scan < last) {
2276 /* Peephole optimizer: */
2277 DEBUG_STUDYDATA(data,depth);
2278 DEBUG_PEEP("Peep",scan,depth);
2279 JOIN_EXACT(scan,&min,0);
2281 /* Follow the next-chain of the current node and optimize
2282 away all the NOTHINGs from it. */
2283 if (OP(scan) != CURLYX) {
2284 const int max = (reg_off_by_arg[OP(scan)]
2286 /* I32 may be smaller than U16 on CRAYs! */
2287 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2288 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2292 /* Skip NOTHING and LONGJMP. */
2293 while ((n = regnext(n))
2294 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2295 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2296 && off + noff < max)
2298 if (reg_off_by_arg[OP(scan)])
2301 NEXT_OFF(scan) = off;
2306 /* The principal pseudo-switch. Cannot be a switch, since we
2307 look into several different things. */
2308 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2309 || OP(scan) == IFTHEN || OP(scan) == SUSPEND) {
2310 next = regnext(scan);
2312 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2314 if (OP(next) == code || code == IFTHEN || code == SUSPEND) {
2315 /* NOTE - There is similar code to this block below for handling
2316 TRIE nodes on a re-study. If you change stuff here check there
2318 I32 max1 = 0, min1 = I32_MAX, num = 0;
2319 struct regnode_charclass_class accum;
2320 regnode * const startbranch=scan;
2322 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
2323 scan_commit(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2324 if (flags & SCF_DO_STCLASS)
2325 cl_init_zero(pRExC_state, &accum);
2327 while (OP(scan) == code) {
2328 I32 deltanext, minnext, f = 0, fake;
2329 struct regnode_charclass_class this_class;
2332 data_fake.flags = 0;
2334 data_fake.whilem_c = data->whilem_c;
2335 data_fake.last_closep = data->last_closep;
2338 data_fake.last_closep = &fake;
2339 next = regnext(scan);
2340 scan = NEXTOPER(scan);
2342 scan = NEXTOPER(scan);
2343 if (flags & SCF_DO_STCLASS) {
2344 cl_init(pRExC_state, &this_class);
2345 data_fake.start_class = &this_class;
2346 f = SCF_DO_STCLASS_AND;
2348 if (flags & SCF_WHILEM_VISITED_POS)
2349 f |= SCF_WHILEM_VISITED_POS;
2351 /* we suppose the run is continuous, last=next...*/
2352 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2353 next, &data_fake, f,depth+1);
2356 if (max1 < minnext + deltanext)
2357 max1 = minnext + deltanext;
2358 if (deltanext == I32_MAX)
2359 is_inf = is_inf_internal = 1;
2361 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2364 if (data_fake.flags & SF_HAS_EVAL)
2365 data->flags |= SF_HAS_EVAL;
2366 data->whilem_c = data_fake.whilem_c;
2368 if (flags & SCF_DO_STCLASS)
2369 cl_or(pRExC_state, &accum, &this_class);
2370 if (code == SUSPEND)
2373 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2375 if (flags & SCF_DO_SUBSTR) {
2376 data->pos_min += min1;
2377 data->pos_delta += max1 - min1;
2378 if (max1 != min1 || is_inf)
2379 data->longest = &(data->longest_float);
2382 delta += max1 - min1;
2383 if (flags & SCF_DO_STCLASS_OR) {
2384 cl_or(pRExC_state, data->start_class, &accum);
2386 cl_and(data->start_class, &and_with);
2387 flags &= ~SCF_DO_STCLASS;
2390 else if (flags & SCF_DO_STCLASS_AND) {
2392 cl_and(data->start_class, &accum);
2393 flags &= ~SCF_DO_STCLASS;
2396 /* Switch to OR mode: cache the old value of
2397 * data->start_class */
2398 StructCopy(data->start_class, &and_with,
2399 struct regnode_charclass_class);
2400 flags &= ~SCF_DO_STCLASS_AND;
2401 StructCopy(&accum, data->start_class,
2402 struct regnode_charclass_class);
2403 flags |= SCF_DO_STCLASS_OR;
2404 data->start_class->flags |= ANYOF_EOS;
2408 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2411 Assuming this was/is a branch we are dealing with: 'scan' now
2412 points at the item that follows the branch sequence, whatever
2413 it is. We now start at the beginning of the sequence and look
2420 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2422 If we can find such a subseqence we need to turn the first
2423 element into a trie and then add the subsequent branch exact
2424 strings to the trie.
2428 1. patterns where the whole set of branch can be converted.
2430 2. patterns where only a subset can be converted.
2432 In case 1 we can replace the whole set with a single regop
2433 for the trie. In case 2 we need to keep the start and end
2436 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2437 becomes BRANCH TRIE; BRANCH X;
2439 There is an additional case, that being where there is a
2440 common prefix, which gets split out into an EXACT like node
2441 preceding the TRIE node.
2443 If x(1..n)==tail then we can do a simple trie, if not we make
2444 a "jump" trie, such that when we match the appropriate word
2445 we "jump" to the appopriate tail node. Essentailly we turn
2446 a nested if into a case structure of sorts.
2451 if (!re_trie_maxbuff) {
2452 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2453 if (!SvIOK(re_trie_maxbuff))
2454 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2456 if ( SvIV(re_trie_maxbuff)>=0 ) {
2458 regnode *first = (regnode *)NULL;
2459 regnode *last = (regnode *)NULL;
2460 regnode *tail = scan;
2465 SV * const mysv = sv_newmortal(); /* for dumping */
2467 /* var tail is used because there may be a TAIL
2468 regop in the way. Ie, the exacts will point to the
2469 thing following the TAIL, but the last branch will
2470 point at the TAIL. So we advance tail. If we
2471 have nested (?:) we may have to move through several
2475 while ( OP( tail ) == TAIL ) {
2476 /* this is the TAIL generated by (?:) */
2477 tail = regnext( tail );
2482 regprop(RExC_rx, mysv, tail );
2483 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2484 (int)depth * 2 + 2, "",
2485 "Looking for TRIE'able sequences. Tail node is: ",
2486 SvPV_nolen_const( mysv )
2492 step through the branches, cur represents each
2493 branch, noper is the first thing to be matched
2494 as part of that branch and noper_next is the
2495 regnext() of that node. if noper is an EXACT
2496 and noper_next is the same as scan (our current
2497 position in the regex) then the EXACT branch is
2498 a possible optimization target. Once we have
2499 two or more consequetive such branches we can
2500 create a trie of the EXACT's contents and stich
2501 it in place. If the sequence represents all of
2502 the branches we eliminate the whole thing and
2503 replace it with a single TRIE. If it is a
2504 subsequence then we need to stitch it in. This
2505 means the first branch has to remain, and needs
2506 to be repointed at the item on the branch chain
2507 following the last branch optimized. This could
2508 be either a BRANCH, in which case the
2509 subsequence is internal, or it could be the
2510 item following the branch sequence in which
2511 case the subsequence is at the end.
2515 /* dont use tail as the end marker for this traverse */
2516 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2517 regnode * const noper = NEXTOPER( cur );
2518 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2519 regnode * const noper_next = regnext( noper );
2523 regprop(RExC_rx, mysv, cur);
2524 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2525 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2527 regprop(RExC_rx, mysv, noper);
2528 PerlIO_printf( Perl_debug_log, " -> %s",
2529 SvPV_nolen_const(mysv));
2532 regprop(RExC_rx, mysv, noper_next );
2533 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2534 SvPV_nolen_const(mysv));
2536 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2537 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2539 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2540 : PL_regkind[ OP( noper ) ] == EXACT )
2541 || OP(noper) == NOTHING )
2543 && noper_next == tail
2548 if ( !first || optype == NOTHING ) {
2549 if (!first) first = cur;
2550 optype = OP( noper );
2556 make_trie( pRExC_state,
2557 startbranch, first, cur, tail, count,
2560 if ( PL_regkind[ OP( noper ) ] == EXACT
2562 && noper_next == tail
2567 optype = OP( noper );
2577 regprop(RExC_rx, mysv, cur);
2578 PerlIO_printf( Perl_debug_log,
2579 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2580 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2584 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2585 #ifdef TRIE_STUDY_OPT
2586 if ( ((made == MADE_EXACT_TRIE &&
2587 startbranch == first)
2588 || ( first_non_open == first )) &&
2590 flags |= SCF_TRIE_RESTUDY;
2598 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2599 scan = NEXTOPER(NEXTOPER(scan));
2600 } else /* single branch is optimized. */
2601 scan = NEXTOPER(scan);
2604 else if (OP(scan) == EXACT) {
2605 I32 l = STR_LEN(scan);
2608 const U8 * const s = (U8*)STRING(scan);
2609 l = utf8_length(s, s + l);
2610 uc = utf8_to_uvchr(s, NULL);
2612 uc = *((U8*)STRING(scan));
2615 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2616 /* The code below prefers earlier match for fixed
2617 offset, later match for variable offset. */
2618 if (data->last_end == -1) { /* Update the start info. */
2619 data->last_start_min = data->pos_min;
2620 data->last_start_max = is_inf
2621 ? I32_MAX : data->pos_min + data->pos_delta;
2623 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2625 SvUTF8_on(data->last_found);
2627 SV * const sv = data->last_found;
2628 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2629 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2630 if (mg && mg->mg_len >= 0)
2631 mg->mg_len += utf8_length((U8*)STRING(scan),
2632 (U8*)STRING(scan)+STR_LEN(scan));
2634 data->last_end = data->pos_min + l;
2635 data->pos_min += l; /* As in the first entry. */
2636 data->flags &= ~SF_BEFORE_EOL;
2638 if (flags & SCF_DO_STCLASS_AND) {
2639 /* Check whether it is compatible with what we know already! */
2643 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2644 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2645 && (!(data->start_class->flags & ANYOF_FOLD)
2646 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2649 ANYOF_CLASS_ZERO(data->start_class);
2650 ANYOF_BITMAP_ZERO(data->start_class);
2652 ANYOF_BITMAP_SET(data->start_class, uc);
2653 data->start_class->flags &= ~ANYOF_EOS;
2655 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2657 else if (flags & SCF_DO_STCLASS_OR) {
2658 /* false positive possible if the class is case-folded */
2660 ANYOF_BITMAP_SET(data->start_class, uc);
2662 data->start_class->flags |= ANYOF_UNICODE_ALL;
2663 data->start_class->flags &= ~ANYOF_EOS;
2664 cl_and(data->start_class, &and_with);
2666 flags &= ~SCF_DO_STCLASS;
2668 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
2669 I32 l = STR_LEN(scan);
2670 UV uc = *((U8*)STRING(scan));
2672 /* Search for fixed substrings supports EXACT only. */
2673 if (flags & SCF_DO_SUBSTR) {
2675 scan_commit(pRExC_state, data, minlenp);
2678 const U8 * const s = (U8 *)STRING(scan);
2679 l = utf8_length(s, s + l);
2680 uc = utf8_to_uvchr(s, NULL);
2683 if (flags & SCF_DO_SUBSTR)
2685 if (flags & SCF_DO_STCLASS_AND) {
2686 /* Check whether it is compatible with what we know already! */
2690 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2691 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2692 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2694 ANYOF_CLASS_ZERO(data->start_class);
2695 ANYOF_BITMAP_ZERO(data->start_class);
2697 ANYOF_BITMAP_SET(data->start_class, uc);
2698 data->start_class->flags &= ~ANYOF_EOS;
2699 data->start_class->flags |= ANYOF_FOLD;
2700 if (OP(scan) == EXACTFL)
2701 data->start_class->flags |= ANYOF_LOCALE;
2704 else if (flags & SCF_DO_STCLASS_OR) {
2705 if (data->start_class->flags & ANYOF_FOLD) {
2706 /* false positive possible if the class is case-folded.
2707 Assume that the locale settings are the same... */
2709 ANYOF_BITMAP_SET(data->start_class, uc);
2710 data->start_class->flags &= ~ANYOF_EOS;
2712 cl_and(data->start_class, &and_with);
2714 flags &= ~SCF_DO_STCLASS;
2716 else if (strchr((const char*)PL_varies,OP(scan))) {
2717 I32 mincount, maxcount, minnext, deltanext, fl = 0;
2718 I32 f = flags, pos_before = 0;
2719 regnode * const oscan = scan;
2720 struct regnode_charclass_class this_class;
2721 struct regnode_charclass_class *oclass = NULL;
2722 I32 next_is_eval = 0;
2724 switch (PL_regkind[OP(scan)]) {
2725 case WHILEM: /* End of (?:...)* . */
2726 scan = NEXTOPER(scan);
2729 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
2730 next = NEXTOPER(scan);
2731 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
2733 maxcount = REG_INFTY;
2734 next = regnext(scan);
2735 scan = NEXTOPER(scan);
2739 if (flags & SCF_DO_SUBSTR)
2744 if (flags & SCF_DO_STCLASS) {
2746 maxcount = REG_INFTY;
2747 next = regnext(scan);
2748 scan = NEXTOPER(scan);
2751 is_inf = is_inf_internal = 1;
2752 scan = regnext(scan);
2753 if (flags & SCF_DO_SUBSTR) {
2754 scan_commit(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
2755 data->longest = &(data->longest_float);
2757 goto optimize_curly_tail;
2759 mincount = ARG1(scan);
2760 maxcount = ARG2(scan);
2761 next = regnext(scan);
2762 if (OP(scan) == CURLYX) {
2763 I32 lp = (data ? *(data->last_closep) : 0);
2764 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
2766 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
2767 next_is_eval = (OP(scan) == EVAL);
2769 if (flags & SCF_DO_SUBSTR) {
2770 if (mincount == 0) scan_commit(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
2771 pos_before = data->pos_min;
2775 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
2777 data->flags |= SF_IS_INF;
2779 if (flags & SCF_DO_STCLASS) {
2780 cl_init(pRExC_state, &this_class);
2781 oclass = data->start_class;
2782 data->start_class = &this_class;
2783 f |= SCF_DO_STCLASS_AND;
2784 f &= ~SCF_DO_STCLASS_OR;
2786 /* These are the cases when once a subexpression
2787 fails at a particular position, it cannot succeed
2788 even after backtracking at the enclosing scope.
2790 XXXX what if minimal match and we are at the
2791 initial run of {n,m}? */
2792 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
2793 f &= ~SCF_WHILEM_VISITED_POS;
2795 /* This will finish on WHILEM, setting scan, or on NULL: */
2796 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext, last, data,
2798 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
2800 if (flags & SCF_DO_STCLASS)
2801 data->start_class = oclass;
2802 if (mincount == 0 || minnext == 0) {
2803 if (flags & SCF_DO_STCLASS_OR) {
2804 cl_or(pRExC_state, data->start_class, &this_class);
2806 else if (flags & SCF_DO_STCLASS_AND) {
2807 /* Switch to OR mode: cache the old value of
2808 * data->start_class */
2809 StructCopy(data->start_class, &and_with,
2810 struct regnode_charclass_class);
2811 flags &= ~SCF_DO_STCLASS_AND;
2812 StructCopy(&this_class, data->start_class,
2813 struct regnode_charclass_class);
2814 flags |= SCF_DO_STCLASS_OR;
2815 data->start_class->flags |= ANYOF_EOS;
2817 } else { /* Non-zero len */
2818 if (flags & SCF_DO_STCLASS_OR) {
2819 cl_or(pRExC_state, data->start_class, &this_class);
2820 cl_and(data->start_class, &and_with);
2822 else if (flags & SCF_DO_STCLASS_AND)
2823 cl_and(data->start_class, &this_class);
2824 flags &= ~SCF_DO_STCLASS;
2826 if (!scan) /* It was not CURLYX, but CURLY. */
2828 if ( /* ? quantifier ok, except for (?{ ... }) */
2829 (next_is_eval || !(mincount == 0 && maxcount == 1))
2830 && (minnext == 0) && (deltanext == 0)
2831 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
2832 && maxcount <= REG_INFTY/3 /* Complement check for big count */
2833 && ckWARN(WARN_REGEXP))
2836 "Quantifier unexpected on zero-length expression");
2839 min += minnext * mincount;
2840 is_inf_internal |= ((maxcount == REG_INFTY
2841 && (minnext + deltanext) > 0)
2842 || deltanext == I32_MAX);
2843 is_inf |= is_inf_internal;
2844 delta += (minnext + deltanext) * maxcount - minnext * mincount;
2846 /* Try powerful optimization CURLYX => CURLYN. */
2847 if ( OP(oscan) == CURLYX && data
2848 && data->flags & SF_IN_PAR
2849 && !(data->flags & SF_HAS_EVAL)
2850 && !deltanext && minnext == 1 ) {
2851 /* Try to optimize to CURLYN. */
2852 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
2853 regnode * const nxt1 = nxt;
2860 if (!strchr((const char*)PL_simple,OP(nxt))
2861 && !(PL_regkind[OP(nxt)] == EXACT
2862 && STR_LEN(nxt) == 1))
2868 if (OP(nxt) != CLOSE)
2870 /* Now we know that nxt2 is the only contents: */
2871 oscan->flags = (U8)ARG(nxt);
2873 OP(nxt1) = NOTHING; /* was OPEN. */
2875 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
2876 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
2877 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
2878 OP(nxt) = OPTIMIZED; /* was CLOSE. */
2879 OP(nxt + 1) = OPTIMIZED; /* was count. */
2880 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
2885 /* Try optimization CURLYX => CURLYM. */
2886 if ( OP(oscan) == CURLYX && data
2887 && !(data->flags & SF_HAS_PAR)
2888 && !(data->flags & SF_HAS_EVAL)
2889 && !deltanext /* atom is fixed width */
2890 && minnext != 0 /* CURLYM can't handle zero width */
2892 /* XXXX How to optimize if data == 0? */
2893 /* Optimize to a simpler form. */
2894 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
2898 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
2899 && (OP(nxt2) != WHILEM))
2901 OP(nxt2) = SUCCEED; /* Whas WHILEM */
2902 /* Need to optimize away parenths. */
2903 if (data->flags & SF_IN_PAR) {
2904 /* Set the parenth number. */
2905 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
2907 if (OP(nxt) != CLOSE)
2908 FAIL("Panic opt close");
2909 oscan->flags = (U8)ARG(nxt);
2910 OP(nxt1) = OPTIMIZED; /* was OPEN. */
2911 OP(nxt) = OPTIMIZED; /* was CLOSE. */
2913 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
2914 OP(nxt + 1) = OPTIMIZED; /* was count. */
2915 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
2916 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
2919 while ( nxt1 && (OP(nxt1) != WHILEM)) {
2920 regnode *nnxt = regnext(nxt1);
2923 if (reg_off_by_arg[OP(nxt1)])
2924 ARG_SET(nxt1, nxt2 - nxt1);
2925 else if (nxt2 - nxt1 < U16_MAX)
2926 NEXT_OFF(nxt1) = nxt2 - nxt1;
2928 OP(nxt) = NOTHING; /* Cannot beautify */
2933 /* Optimize again: */
2934 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
2940 else if ((OP(oscan) == CURLYX)
2941 && (flags & SCF_WHILEM_VISITED_POS)
2942 /* See the comment on a similar expression above.
2943 However, this time it not a subexpression
2944 we care about, but the expression itself. */
2945 && (maxcount == REG_INFTY)
2946 && data && ++data->whilem_c < 16) {
2947 /* This stays as CURLYX, we can put the count/of pair. */
2948 /* Find WHILEM (as in regexec.c) */
2949 regnode *nxt = oscan + NEXT_OFF(oscan);
2951 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
2953 PREVOPER(nxt)->flags = (U8)(data->whilem_c
2954 | (RExC_whilem_seen << 4)); /* On WHILEM */
2956 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
2958 if (flags & SCF_DO_SUBSTR) {
2959 SV *last_str = NULL;
2960 int counted = mincount != 0;
2962 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
2963 #if defined(SPARC64_GCC_WORKAROUND)
2966 const char *s = NULL;
2969 if (pos_before >= data->last_start_min)
2972 b = data->last_start_min;
2975 s = SvPV_const(data->last_found, l);
2976 old = b - data->last_start_min;
2979 I32 b = pos_before >= data->last_start_min
2980 ? pos_before : data->last_start_min;
2982 const char * const s = SvPV_const(data->last_found, l);
2983 I32 old = b - data->last_start_min;
2987 old = utf8_hop((U8*)s, old) - (U8*)s;
2990 /* Get the added string: */
2991 last_str = newSVpvn(s + old, l);
2993 SvUTF8_on(last_str);
2994 if (deltanext == 0 && pos_before == b) {
2995 /* What was added is a constant string */
2997 SvGROW(last_str, (mincount * l) + 1);
2998 repeatcpy(SvPVX(last_str) + l,
2999 SvPVX_const(last_str), l, mincount - 1);
3000 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3001 /* Add additional parts. */
3002 SvCUR_set(data->last_found,
3003 SvCUR(data->last_found) - l);
3004 sv_catsv(data->last_found, last_str);
3006 SV * sv = data->last_found;
3008 SvUTF8(sv) && SvMAGICAL(sv) ?
3009 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3010 if (mg && mg->mg_len >= 0)
3011 mg->mg_len += CHR_SVLEN(last_str);
3013 data->last_end += l * (mincount - 1);
3016 /* start offset must point into the last copy */
3017 data->last_start_min += minnext * (mincount - 1);
3018 data->last_start_max += is_inf ? I32_MAX
3019 : (maxcount - 1) * (minnext + data->pos_delta);
3022 /* It is counted once already... */
3023 data->pos_min += minnext * (mincount - counted);
3024 data->pos_delta += - counted * deltanext +
3025 (minnext + deltanext) * maxcount - minnext * mincount;
3026 if (mincount != maxcount) {
3027 /* Cannot extend fixed substrings found inside
3029 scan_commit(pRExC_state,data,minlenp);
3030 if (mincount && last_str) {
3031 SV * const sv = data->last_found;
3032 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3033 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3037 sv_setsv(sv, last_str);
3038 data->last_end = data->pos_min;
3039 data->last_start_min =
3040 data->pos_min - CHR_SVLEN(last_str);
3041 data->last_start_max = is_inf
3043 : data->pos_min + data->pos_delta
3044 - CHR_SVLEN(last_str);
3046 data->longest = &(data->longest_float);
3048 SvREFCNT_dec(last_str);
3050 if (data && (fl & SF_HAS_EVAL))
3051 data->flags |= SF_HAS_EVAL;
3052 optimize_curly_tail:
3053 if (OP(oscan) != CURLYX) {
3054 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3056 NEXT_OFF(oscan) += NEXT_OFF(next);
3059 default: /* REF and CLUMP only? */
3060 if (flags & SCF_DO_SUBSTR) {
3061 scan_commit(pRExC_state,data,minlenp); /* Cannot expect anything... */
3062 data->longest = &(data->longest_float);
3064 is_inf = is_inf_internal = 1;
3065 if (flags & SCF_DO_STCLASS_OR)
3066 cl_anything(pRExC_state, data->start_class);
3067 flags &= ~SCF_DO_STCLASS;
3071 else if (strchr((const char*)PL_simple,OP(scan))) {
3074 if (flags & SCF_DO_SUBSTR) {
3075 scan_commit(pRExC_state,data,minlenp);
3079 if (flags & SCF_DO_STCLASS) {
3080 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3082 /* Some of the logic below assumes that switching
3083 locale on will only add false positives. */
3084 switch (PL_regkind[OP(scan)]) {
3088 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3089 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3090 cl_anything(pRExC_state, data->start_class);
3093 if (OP(scan) == SANY)
3095 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3096 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3097 || (data->start_class->flags & ANYOF_CLASS));
3098 cl_anything(pRExC_state, data->start_class);
3100 if (flags & SCF_DO_STCLASS_AND || !value)
3101 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3104 if (flags & SCF_DO_STCLASS_AND)
3105 cl_and(data->start_class,
3106 (struct regnode_charclass_class*)scan);
3108 cl_or(pRExC_state, data->start_class,
3109 (struct regnode_charclass_class*)scan);
3112 if (flags & SCF_DO_STCLASS_AND) {
3113 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3114 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3115 for (value = 0; value < 256; value++)
3116 if (!isALNUM(value))
3117 ANYOF_BITMAP_CLEAR(data->start_class, value);
3121 if (data->start_class->flags & ANYOF_LOCALE)
3122 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3124 for (value = 0; value < 256; value++)
3126 ANYOF_BITMAP_SET(data->start_class, value);
3131 if (flags & SCF_DO_STCLASS_AND) {
3132 if (data->start_class->flags & ANYOF_LOCALE)
3133 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3136 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3137 data->start_class->flags |= ANYOF_LOCALE;
3141 if (flags & SCF_DO_STCLASS_AND) {
3142 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3143 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3144 for (value = 0; value < 256; value++)
3146 ANYOF_BITMAP_CLEAR(data->start_class, value);
3150 if (data->start_class->flags & ANYOF_LOCALE)
3151 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3153 for (value = 0; value < 256; value++)
3154 if (!isALNUM(value))
3155 ANYOF_BITMAP_SET(data->start_class, value);
3160 if (flags & SCF_DO_STCLASS_AND) {
3161 if (data->start_class->flags & ANYOF_LOCALE)
3162 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3165 data->start_class->flags |= ANYOF_LOCALE;
3166 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3170 if (flags & SCF_DO_STCLASS_AND) {
3171 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3172 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3173 for (value = 0; value < 256; value++)
3174 if (!isSPACE(value))
3175 ANYOF_BITMAP_CLEAR(data->start_class, value);
3179 if (data->start_class->flags & ANYOF_LOCALE)
3180 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3182 for (value = 0; value < 256; value++)
3184 ANYOF_BITMAP_SET(data->start_class, value);
3189 if (flags & SCF_DO_STCLASS_AND) {
3190 if (data->start_class->flags & ANYOF_LOCALE)
3191 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3194 data->start_class->flags |= ANYOF_LOCALE;
3195 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3199 if (flags & SCF_DO_STCLASS_AND) {
3200 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3201 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3202 for (value = 0; value < 256; value++)
3204 ANYOF_BITMAP_CLEAR(data->start_class, value);
3208 if (data->start_class->flags & ANYOF_LOCALE)
3209 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3211 for (value = 0; value < 256; value++)
3212 if (!isSPACE(value))
3213 ANYOF_BITMAP_SET(data->start_class, value);
3218 if (flags & SCF_DO_STCLASS_AND) {
3219 if (data->start_class->flags & ANYOF_LOCALE) {
3220 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3221 for (value = 0; value < 256; value++)
3222 if (!isSPACE(value))
3223 ANYOF_BITMAP_CLEAR(data->start_class, value);
3227 data->start_class->flags |= ANYOF_LOCALE;
3228 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3232 if (flags & SCF_DO_STCLASS_AND) {
3233 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3234 for (value = 0; value < 256; value++)
3235 if (!isDIGIT(value))
3236 ANYOF_BITMAP_CLEAR(data->start_class, value);
3239 if (data->start_class->flags & ANYOF_LOCALE)
3240 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3242 for (value = 0; value < 256; value++)
3244 ANYOF_BITMAP_SET(data->start_class, value);
3249 if (flags & SCF_DO_STCLASS_AND) {
3250 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3251 for (value = 0; value < 256; value++)
3253 ANYOF_BITMAP_CLEAR(data->start_class, value);
3256 if (data->start_class->flags & ANYOF_LOCALE)
3257 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3259 for (value = 0; value < 256; value++)
3260 if (!isDIGIT(value))
3261 ANYOF_BITMAP_SET(data->start_class, value);
3266 if (flags & SCF_DO_STCLASS_OR)
3267 cl_and(data->start_class, &and_with);
3268 flags &= ~SCF_DO_STCLASS;
3271 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3272 data->flags |= (OP(scan) == MEOL
3276 else if ( PL_regkind[OP(scan)] == BRANCHJ
3277 /* Lookbehind, or need to calculate parens/evals/stclass: */
3278 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3279 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3280 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3281 || OP(scan) == UNLESSM )
3283 /* Negative Lookahead/lookbehind
3284 In this case we can't do fixed string optimisation.
3287 I32 deltanext, minnext, fake = 0;
3289 struct regnode_charclass_class intrnl;
3292 data_fake.flags = 0;
3294 data_fake.whilem_c = data->whilem_c;
3295 data_fake.last_closep = data->last_closep;
3298 data_fake.last_closep = &fake;
3299 if ( flags & SCF_DO_STCLASS && !scan->flags
3300 && OP(scan) == IFMATCH ) { /* Lookahead */
3301 cl_init(pRExC_state, &intrnl);
3302 data_fake.start_class = &intrnl;
3303 f |= SCF_DO_STCLASS_AND;
3305 if (flags & SCF_WHILEM_VISITED_POS)
3306 f |= SCF_WHILEM_VISITED_POS;
3307 next = regnext(scan);
3308 nscan = NEXTOPER(NEXTOPER(scan));
3309 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext, last, &data_fake, f,depth+1);
3312 vFAIL("Variable length lookbehind not implemented");
3314 else if (minnext > (I32)U8_MAX) {
3315 vFAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3317 scan->flags = (U8)minnext;
3320 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3322 if (data_fake.flags & SF_HAS_EVAL)
3323 data->flags |= SF_HAS_EVAL;
3324 data->whilem_c = data_fake.whilem_c;
3326 if (f & SCF_DO_STCLASS_AND) {
3327 const int was = (data->start_class->flags & ANYOF_EOS);
3329 cl_and(data->start_class, &intrnl);
3331 data->start_class->flags |= ANYOF_EOS;
3334 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3336 /* Positive Lookahead/lookbehind
3337 In this case we can do fixed string optimisation,
3338 but we must be careful about it. Note in the case of
3339 lookbehind the positions will be offset by the minimum
3340 length of the pattern, something we won't know about
3341 until after the recurse.
3343 I32 deltanext, fake = 0;
3345 struct regnode_charclass_class intrnl;
3347 /* We use SAVEFREEPV so that when the full compile
3348 is finished perl will clean up the allocated
3349 minlens when its all done. This was we don't
3350 have to worry about freeing them when we know
3351 they wont be used, which would be a pain.
3354 Newx( minnextp, 1, I32 );
3355 SAVEFREEPV(minnextp);
3358 StructCopy(data, &data_fake, scan_data_t);
3359 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3362 scan_commit(pRExC_state, &data_fake,minlenp);
3363 data_fake.last_found=newSVsv(data->last_found);
3367 data_fake.last_closep = &fake;
3368 data_fake.flags = 0;
3370 data_fake.flags |= SF_IS_INF;
3371 if ( flags & SCF_DO_STCLASS && !scan->flags
3372 && OP(scan) == IFMATCH ) { /* Lookahead */
3373 cl_init(pRExC_state, &intrnl);
3374 data_fake.start_class = &intrnl;
3375 f |= SCF_DO_STCLASS_AND;
3377 if (flags & SCF_WHILEM_VISITED_POS)
3378 f |= SCF_WHILEM_VISITED_POS;
3379 next = regnext(scan);
3380 nscan = NEXTOPER(NEXTOPER(scan));
3382 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext, last, &data_fake, f,depth+1);
3385 vFAIL("Variable length lookbehind not implemented");
3387 else if (*minnextp > (I32)U8_MAX) {
3388 vFAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3390 scan->flags = (U8)*minnextp;
3396 if (f & SCF_DO_STCLASS_AND) {
3397 const int was = (data->start_class->flags & ANYOF_EOS);
3399 cl_and(data->start_class, &intrnl);
3401 data->start_class->flags |= ANYOF_EOS;
3404 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3406 if (data_fake.flags & SF_HAS_EVAL)
3407 data->flags |= SF_HAS_EVAL;
3408 data->whilem_c = data_fake.whilem_c;
3409 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3410 if (RExC_rx->minlen<*minnextp)
3411 RExC_rx->minlen=*minnextp;
3412 scan_commit(pRExC_state, &data_fake, minnextp);
3413 SvREFCNT_dec(data_fake.last_found);
3415 if ( data_fake.minlen_fixed != minlenp )
3417 data->offset_fixed= data_fake.offset_fixed;
3418 data->minlen_fixed= data_fake.minlen_fixed;
3419 data->lookbehind_fixed+= scan->flags;
3421 if ( data_fake.minlen_float != minlenp )
3423 data->minlen_float= data_fake.minlen_float;
3424 data->offset_float_min=data_fake.offset_float_min;
3425 data->offset_float_max=data_fake.offset_float_max;
3426 data->lookbehind_float+= scan->flags;
3435 else if (OP(scan) == OPEN) {
3438 else if (OP(scan) == CLOSE) {
3439 if ((I32)ARG(scan) == is_par) {
3440 next = regnext(scan);
3442 if ( next && (OP(next) != WHILEM) && next < last)
3443 is_par = 0; /* Disable optimization */
3446 *(data->last_closep) = ARG(scan);
3448 else if (OP(scan) == EVAL) {
3450 data->flags |= SF_HAS_EVAL;
3452 else if ( (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
3453 || OP(scan)==RECURSE) /* recursion */
3455 if (OP(scan)==RECURSE) {
3456 ARG2L_SET( scan, RExC_parens[ARG(scan)-1] - scan );
3458 if (flags & SCF_DO_SUBSTR) {
3459 scan_commit(pRExC_state,data,minlenp);
3460 data->longest = &(data->longest_float);
3462 is_inf = is_inf_internal = 1;
3463 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3464 cl_anything(pRExC_state, data->start_class);
3465 flags &= ~SCF_DO_STCLASS;
3467 #ifdef TRIE_STUDY_OPT
3468 #ifdef FULL_TRIE_STUDY
3469 else if (PL_regkind[OP(scan)] == TRIE) {
3470 /* NOTE - There is similar code to this block above for handling
3471 BRANCH nodes on the initial study. If you change stuff here
3473 regnode *tail= regnext(scan);
3474 reg_trie_data *trie = (reg_trie_data*)RExC_rx->data->data[ ARG(scan) ];
3475 I32 max1 = 0, min1 = I32_MAX;
3476 struct regnode_charclass_class accum;
3478 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
3479 scan_commit(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
3480 if (flags & SCF_DO_STCLASS)
3481 cl_init_zero(pRExC_state, &accum);
3487 const regnode *nextbranch= NULL;
3490 for ( word=1 ; word <= trie->wordcount ; word++)
3492 I32 deltanext=0, minnext=0, f = 0, fake;
3493 struct regnode_charclass_class this_class;
3495 data_fake.flags = 0;
3497 data_fake.whilem_c = data->whilem_c;
3498 data_fake.last_closep = data->last_closep;
3501 data_fake.last_closep = &fake;
3503 if (flags & SCF_DO_STCLASS) {
3504 cl_init(pRExC_state, &this_class);
3505 data_fake.start_class = &this_class;
3506 f = SCF_DO_STCLASS_AND;
3508 if (flags & SCF_WHILEM_VISITED_POS)
3509 f |= SCF_WHILEM_VISITED_POS;
3511 if (trie->jump[word]) {
3513 nextbranch = tail - trie->jump[0];
3514 scan= tail - trie->jump[word];
3515 /* We go from the jump point to the branch that follows
3516 it. Note this means we need the vestigal unused branches
3517 even though they arent otherwise used.
3519 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3520 (regnode *)nextbranch, &data_fake, f,depth+1);
3522 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
3523 nextbranch= regnext((regnode*)nextbranch);
3525 if (min1 > (I32)(minnext + trie->minlen))
3526 min1 = minnext + trie->minlen;
3527 if (max1 < (I32)(minnext + deltanext + trie->maxlen))
3528 max1 = minnext + deltanext + trie->maxlen;
3529 if (deltanext == I32_MAX)
3530 is_inf = is_inf_internal = 1;
3532 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3536 if (data_fake.flags & SF_HAS_EVAL)
3537 data->flags |= SF_HAS_EVAL;
3538 data->whilem_c = data_fake.whilem_c;
3540 if (flags & SCF_DO_STCLASS)
3541 cl_or(pRExC_state, &accum, &this_class);
3544 if (flags & SCF_DO_SUBSTR) {
3545 data->pos_min += min1;
3546 data->pos_delta += max1 - min1;
3547 if (max1 != min1 || is_inf)
3548 data->longest = &(data->longest_float);
3551 delta += max1 - min1;
3552 if (flags & SCF_DO_STCLASS_OR) {
3553 cl_or(pRExC_state, data->start_class, &accum);
3555 cl_and(data->start_class, &and_with);
3556 flags &= ~SCF_DO_STCLASS;
3559 else if (flags & SCF_DO_STCLASS_AND) {
3561 cl_and(data->start_class, &accum);
3562 flags &= ~SCF_DO_STCLASS;
3565 /* Switch to OR mode: cache the old value of
3566 * data->start_class */
3567 StructCopy(data->start_class, &and_with,
3568 struct regnode_charclass_class);
3569 flags &= ~SCF_DO_STCLASS_AND;
3570 StructCopy(&accum, data->start_class,
3571 struct regnode_charclass_class);
3572 flags |= SCF_DO_STCLASS_OR;
3573 data->start_class->flags |= ANYOF_EOS;
3580 else if (PL_regkind[OP(scan)] == TRIE) {
3581 reg_trie_data *trie = (reg_trie_data*)RExC_rx->data->data[ ARG(scan) ];
3584 min += trie->minlen;
3585 delta += (trie->maxlen - trie->minlen);
3586 flags &= ~SCF_DO_STCLASS; /* xxx */
3587 if (flags & SCF_DO_SUBSTR) {
3588 scan_commit(pRExC_state,data,minlenp); /* Cannot expect anything... */
3589 data->pos_min += trie->minlen;
3590 data->pos_delta += (trie->maxlen - trie->minlen);
3591 if (trie->maxlen != trie->minlen)
3592 data->longest = &(data->longest_float);
3594 if (trie->jump) /* no more substrings -- for now /grr*/
3595 flags &= ~SCF_DO_SUBSTR;
3597 #endif /* old or new */
3598 #endif /* TRIE_STUDY_OPT */
3599 /* Else: zero-length, ignore. */
3600 scan = regnext(scan);
3605 *deltap = is_inf_internal ? I32_MAX : delta;
3606 if (flags & SCF_DO_SUBSTR && is_inf)
3607 data->pos_delta = I32_MAX - data->pos_min;
3608 if (is_par > (I32)U8_MAX)
3610 if (is_par && pars==1 && data) {
3611 data->flags |= SF_IN_PAR;
3612 data->flags &= ~SF_HAS_PAR;
3614 else if (pars && data) {
3615 data->flags |= SF_HAS_PAR;
3616 data->flags &= ~SF_IN_PAR;
3618 if (flags & SCF_DO_STCLASS_OR)
3619 cl_and(data->start_class, &and_with);
3620 if (flags & SCF_TRIE_RESTUDY)
3621 data->flags |= SCF_TRIE_RESTUDY;
3623 DEBUG_STUDYDATA(data,depth);
3629 S_add_data(RExC_state_t *pRExC_state, I32 n, const char *s)
3631 if (RExC_rx->data) {
3632 Renewc(RExC_rx->data,
3633 sizeof(*RExC_rx->data) + sizeof(void*) * (RExC_rx->data->count + n - 1),
3634 char, struct reg_data);
3635 Renew(RExC_rx->data->what, RExC_rx->data->count + n, U8);
3636 RExC_rx->data->count += n;
3639 Newxc(RExC_rx->data, sizeof(*RExC_rx->data) + sizeof(void*) * (n - 1),
3640 char, struct reg_data);
3641 Newx(RExC_rx->data->what, n, U8);
3642 RExC_rx->data->count = n;
3644 Copy(s, RExC_rx->data->what + RExC_rx->data->count - n, n, U8);
3645 return RExC_rx->data->count - n;
3648 #ifndef PERL_IN_XSUB_RE
3650 Perl_reginitcolors(pTHX)
3653 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
3655 char *t = savepv(s);
3659 t = strchr(t, '\t');
3665 PL_colors[i] = t = (char *)"";
3670 PL_colors[i++] = (char *)"";
3677 #ifdef TRIE_STUDY_OPT
3678 #define CHECK_RESTUDY_GOTO \
3680 (data.flags & SCF_TRIE_RESTUDY) \
3684 #define CHECK_RESTUDY_GOTO
3688 - pregcomp - compile a regular expression into internal code
3690 * We can't allocate space until we know how big the compiled form will be,
3691 * but we can't compile it (and thus know how big it is) until we've got a
3692 * place to put the code. So we cheat: we compile it twice, once with code
3693 * generation turned off and size counting turned on, and once "for real".
3694 * This also means that we don't allocate space until we are sure that the
3695 * thing really will compile successfully, and we never have to move the
3696 * code and thus invalidate pointers into it. (Note that it has to be in
3697 * one piece because free() must be able to free it all.) [NB: not true in perl]
3699 * Beware that the optimization-preparation code in here knows about some
3700 * of the structure of the compiled regexp. [I'll say.]
3702 #ifndef PERL_IN_XSUB_RE
3703 #define CORE_ONLY_BLOCK(c) {c}{
3704 #define RE_ENGINE_PTR &PL_core_reg_engine
3706 #define CORE_ONLY_BLOCK(c) {
3707 extern const struct regexp_engine my_reg_engine;
3708 #define RE_ENGINE_PTR &my_reg_engine
3713 Perl_pregcomp(pTHX_ char *exp, char *xend, PMOP *pm)
3716 GET_RE_DEBUG_FLAGS_DECL;
3717 DEBUG_r(if (!PL_colorset) reginitcolors());
3719 /* Dispatch a request to compile a regexp to correct
3721 HV * const table = GvHV(PL_hintgv);
3723 SV **ptr= hv_fetchs(table, "regcomp", FALSE);
3724 if (ptr && SvIOK(*ptr)) {
3725 const regexp_engine *eng=INT2PTR(regexp_engine*,SvIV(*ptr));
3727 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
3730 return CALLREGCOMP_ENG(eng, exp, xend, pm);
3741 RExC_state_t RExC_state;
3742 RExC_state_t * const pRExC_state = &RExC_state;
3743 #ifdef TRIE_STUDY_OPT
3745 RExC_state_t copyRExC_state;
3748 FAIL("NULL regexp argument");
3750 RExC_utf8 = pm->op_pmdynflags & PMdf_CMP_UTF8;
3754 SV *dsv= sv_newmortal();
3755 RE_PV_QUOTED_DECL(s, RExC_utf8,
3756 dsv, RExC_precomp, (xend - exp), 60);
3757 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
3758 PL_colors[4],PL_colors[5],s);
3760 RExC_flags = pm->op_pmflags;
3764 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
3765 RExC_seen_evals = 0;
3768 /* First pass: determine size, legality. */
3775 RExC_emit = &PL_regdummy;
3776 RExC_whilem_seen = 0;
3777 RExC_charnames = NULL;
3779 RExC_paren_names = NULL;
3781 #if 0 /* REGC() is (currently) a NOP at the first pass.
3782 * Clever compilers notice this and complain. --jhi */
3783 REGC((U8)REG_MAGIC, (char*)RExC_emit);
3785 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n"));
3786 if (reg(pRExC_state, 0, &flags,1) == NULL) {
3787 RExC_precomp = NULL;
3791 PerlIO_printf(Perl_debug_log,
3792 "Required size %"IVdf" nodes\n"
3793 "Starting second pass (creation)\n",
3796 RExC_lastparse=NULL;
3798 /* Small enough for pointer-storage convention?
3799 If extralen==0, this means that we will not need long jumps. */
3800 if (RExC_size >= 0x10000L && RExC_extralen)
3801 RExC_size += RExC_extralen;
3804 if (RExC_whilem_seen > 15)
3805 RExC_whilem_seen = 15;
3807 /* Allocate space and zero-initialize. Note, the two step process
3808 of zeroing when in debug mode, thus anything assigned has to
3809 happen after that */
3810 Newxc(r, sizeof(regexp) + (unsigned)RExC_size * sizeof(regnode),
3813 FAIL("Regexp out of space");
3815 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
3816 Zero(r, sizeof(regexp) + (unsigned)RExC_size * sizeof(regnode), char);
3818 /* initialization begins here */
3819 r->engine= RE_ENGINE_PTR;
3821 r->prelen = xend - exp;
3822 r->precomp = savepvn(RExC_precomp, r->prelen);
3824 #ifdef PERL_OLD_COPY_ON_WRITE
3825 r->saved_copy = NULL;
3827 r->reganch = pm->op_pmflags & PMf_COMPILETIME;
3828 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
3829 r->lastparen = 0; /* mg.c reads this. */
3831 r->substrs = 0; /* Useful during FAIL. */
3832 r->startp = 0; /* Useful during FAIL. */
3836 if (RExC_seen & REG_SEEN_RECURSE) {
3837 Newx(RExC_parens, RExC_npar,regnode *);
3838 SAVEFREEPV(RExC_parens);
3841 /* Useful during FAIL. */
3842 Newxz(r->offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
3844 r->offsets[0] = RExC_size;
3846 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
3847 "%s %"UVuf" bytes for offset annotations.\n",
3848 r->offsets ? "Got" : "Couldn't get",
3849 (UV)((2*RExC_size+1) * sizeof(U32))));
3853 /* Second pass: emit code. */
3854 RExC_flags = pm->op_pmflags; /* don't let top level (?i) bleed */
3859 RExC_emit_start = r->program;
3860 RExC_emit = r->program;
3861 /* Store the count of eval-groups for security checks: */
3862 RExC_emit->next_off = (RExC_seen_evals > (I32)U16_MAX) ? U16_MAX : (U16)RExC_seen_evals;
3863 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
3865 if (reg(pRExC_state, 0, &flags,1) == NULL)
3868 /* XXXX To minimize changes to RE engine we always allocate
3869 3-units-long substrs field. */
3870 Newx(r->substrs, 1, struct reg_substr_data);
3873 r->minlen = minlen = sawplus = sawopen = 0;
3874 Zero(r->substrs, 1, struct reg_substr_data);
3875 StructCopy(&zero_scan_data, &data, scan_data_t);
3877 #ifdef TRIE_STUDY_OPT
3879 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
3880 RExC_state=copyRExC_state;
3881 if (data.last_found) {
3882 SvREFCNT_dec(data.longest_fixed);
3883 SvREFCNT_dec(data.longest_float);
3884 SvREFCNT_dec(data.last_found);
3887 copyRExC_state=RExC_state;
3891 /* Dig out information for optimizations. */
3892 r->reganch = pm->op_pmflags & PMf_COMPILETIME; /* Again? */
3893 pm->op_pmflags = RExC_flags;
3895 r->reganch |= ROPT_UTF8; /* Unicode in it? */