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, 2007 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. */
105 REGEXP *rx_sv; /* The SV that is the regexp. */
106 regexp *rx; /* perl core regexp structure */
107 regexp_internal *rxi; /* internal data for regexp object pprivate field */
108 char *start; /* Start of input for compile */
109 char *end; /* End of input for compile */
110 char *parse; /* Input-scan pointer. */
111 I32 whilem_seen; /* number of WHILEM in this expr */
112 regnode *emit_start; /* Start of emitted-code area */
113 regnode *emit_bound; /* First regnode outside of the allocated space */
114 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
115 I32 naughty; /* How bad is this pattern? */
116 I32 sawback; /* Did we see \1, ...? */
118 I32 size; /* Code size. */
119 I32 npar; /* Capture buffer count, (OPEN). */
120 I32 cpar; /* Capture buffer count, (CLOSE). */
121 I32 nestroot; /* root parens we are in - used by accept */
125 regnode **open_parens; /* pointers to open parens */
126 regnode **close_parens; /* pointers to close parens */
127 regnode *opend; /* END node in program */
128 I32 utf8; /* whether the pattern is utf8 or not */
129 I32 orig_utf8; /* whether the pattern was originally in utf8 */
130 /* XXX use this for future optimisation of case
131 * where pattern must be upgraded to utf8. */
132 HV *charnames; /* cache of named sequences */
133 HV *paren_names; /* Paren names */
135 regnode **recurse; /* Recurse regops */
136 I32 recurse_count; /* Number of recurse regops */
138 char *starttry; /* -Dr: where regtry was called. */
139 #define RExC_starttry (pRExC_state->starttry)
142 const char *lastparse;
144 AV *paren_name_list; /* idx -> name */
145 #define RExC_lastparse (pRExC_state->lastparse)
146 #define RExC_lastnum (pRExC_state->lastnum)
147 #define RExC_paren_name_list (pRExC_state->paren_name_list)
151 #define RExC_flags (pRExC_state->flags)
152 #define RExC_precomp (pRExC_state->precomp)
153 #define RExC_rx_sv (pRExC_state->rx_sv)
154 #define RExC_rx (pRExC_state->rx)
155 #define RExC_rxi (pRExC_state->rxi)
156 #define RExC_start (pRExC_state->start)
157 #define RExC_end (pRExC_state->end)
158 #define RExC_parse (pRExC_state->parse)
159 #define RExC_whilem_seen (pRExC_state->whilem_seen)
160 #ifdef RE_TRACK_PATTERN_OFFSETS
161 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
163 #define RExC_emit (pRExC_state->emit)
164 #define RExC_emit_start (pRExC_state->emit_start)
165 #define RExC_emit_bound (pRExC_state->emit_bound)
166 #define RExC_naughty (pRExC_state->naughty)
167 #define RExC_sawback (pRExC_state->sawback)
168 #define RExC_seen (pRExC_state->seen)
169 #define RExC_size (pRExC_state->size)
170 #define RExC_npar (pRExC_state->npar)
171 #define RExC_nestroot (pRExC_state->nestroot)
172 #define RExC_extralen (pRExC_state->extralen)
173 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
174 #define RExC_seen_evals (pRExC_state->seen_evals)
175 #define RExC_utf8 (pRExC_state->utf8)
176 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
177 #define RExC_charnames (pRExC_state->charnames)
178 #define RExC_open_parens (pRExC_state->open_parens)
179 #define RExC_close_parens (pRExC_state->close_parens)
180 #define RExC_opend (pRExC_state->opend)
181 #define RExC_paren_names (pRExC_state->paren_names)
182 #define RExC_recurse (pRExC_state->recurse)
183 #define RExC_recurse_count (pRExC_state->recurse_count)
186 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
187 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
188 ((*s) == '{' && regcurly(s)))
191 #undef SPSTART /* dratted cpp namespace... */
194 * Flags to be passed up and down.
196 #define WORST 0 /* Worst case. */
197 #define HASWIDTH 0x01 /* Known to match non-null strings. */
198 #define SIMPLE 0x02 /* Simple enough to be STAR/PLUS operand. */
199 #define SPSTART 0x04 /* Starts with * or +. */
200 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
201 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
203 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
205 /* whether trie related optimizations are enabled */
206 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
207 #define TRIE_STUDY_OPT
208 #define FULL_TRIE_STUDY
214 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
215 #define PBITVAL(paren) (1 << ((paren) & 7))
216 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
217 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
218 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
221 /* About scan_data_t.
223 During optimisation we recurse through the regexp program performing
224 various inplace (keyhole style) optimisations. In addition study_chunk
225 and scan_commit populate this data structure with information about
226 what strings MUST appear in the pattern. We look for the longest
227 string that must appear for at a fixed location, and we look for the
228 longest string that may appear at a floating location. So for instance
233 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
234 strings (because they follow a .* construct). study_chunk will identify
235 both FOO and BAR as being the longest fixed and floating strings respectively.
237 The strings can be composites, for instance
241 will result in a composite fixed substring 'foo'.
243 For each string some basic information is maintained:
245 - offset or min_offset
246 This is the position the string must appear at, or not before.
247 It also implicitly (when combined with minlenp) tells us how many
248 character must match before the string we are searching.
249 Likewise when combined with minlenp and the length of the string
250 tells us how many characters must appear after the string we have
254 Only used for floating strings. This is the rightmost point that
255 the string can appear at. Ifset to I32 max it indicates that the
256 string can occur infinitely far to the right.
259 A pointer to the minimum length of the pattern that the string
260 was found inside. This is important as in the case of positive
261 lookahead or positive lookbehind we can have multiple patterns
266 The minimum length of the pattern overall is 3, the minimum length
267 of the lookahead part is 3, but the minimum length of the part that
268 will actually match is 1. So 'FOO's minimum length is 3, but the
269 minimum length for the F is 1. This is important as the minimum length
270 is used to determine offsets in front of and behind the string being
271 looked for. Since strings can be composites this is the length of the
272 pattern at the time it was commited with a scan_commit. Note that
273 the length is calculated by study_chunk, so that the minimum lengths
274 are not known until the full pattern has been compiled, thus the
275 pointer to the value.
279 In the case of lookbehind the string being searched for can be
280 offset past the start point of the final matching string.
281 If this value was just blithely removed from the min_offset it would
282 invalidate some of the calculations for how many chars must match
283 before or after (as they are derived from min_offset and minlen and
284 the length of the string being searched for).
285 When the final pattern is compiled and the data is moved from the
286 scan_data_t structure into the regexp structure the information
287 about lookbehind is factored in, with the information that would
288 have been lost precalculated in the end_shift field for the
291 The fields pos_min and pos_delta are used to store the minimum offset
292 and the delta to the maximum offset at the current point in the pattern.
296 typedef struct scan_data_t {
297 /*I32 len_min; unused */
298 /*I32 len_delta; unused */
302 I32 last_end; /* min value, <0 unless valid. */
305 SV **longest; /* Either &l_fixed, or &l_float. */
306 SV *longest_fixed; /* longest fixed string found in pattern */
307 I32 offset_fixed; /* offset where it starts */
308 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
309 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
310 SV *longest_float; /* longest floating string found in pattern */
311 I32 offset_float_min; /* earliest point in string it can appear */
312 I32 offset_float_max; /* latest point in string it can appear */
313 I32 *minlen_float; /* pointer to the minlen relevent to the string */
314 I32 lookbehind_float; /* is the position of the string modified by LB */
318 struct regnode_charclass_class *start_class;
322 * Forward declarations for pregcomp()'s friends.
325 static const scan_data_t zero_scan_data =
326 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
328 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
329 #define SF_BEFORE_SEOL 0x0001
330 #define SF_BEFORE_MEOL 0x0002
331 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
332 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
335 # define SF_FIX_SHIFT_EOL (0+2)
336 # define SF_FL_SHIFT_EOL (0+4)
338 # define SF_FIX_SHIFT_EOL (+2)
339 # define SF_FL_SHIFT_EOL (+4)
342 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
343 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
345 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
346 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
347 #define SF_IS_INF 0x0040
348 #define SF_HAS_PAR 0x0080
349 #define SF_IN_PAR 0x0100
350 #define SF_HAS_EVAL 0x0200
351 #define SCF_DO_SUBSTR 0x0400
352 #define SCF_DO_STCLASS_AND 0x0800
353 #define SCF_DO_STCLASS_OR 0x1000
354 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
355 #define SCF_WHILEM_VISITED_POS 0x2000
357 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
358 #define SCF_SEEN_ACCEPT 0x8000
360 #define UTF (RExC_utf8 != 0)
361 #define LOC ((RExC_flags & RXf_PMf_LOCALE) != 0)
362 #define FOLD ((RExC_flags & RXf_PMf_FOLD) != 0)
364 #define OOB_UNICODE 12345678
365 #define OOB_NAMEDCLASS -1
367 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
368 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
371 /* length of regex to show in messages that don't mark a position within */
372 #define RegexLengthToShowInErrorMessages 127
375 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
376 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
377 * op/pragma/warn/regcomp.
379 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
380 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
382 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
385 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
386 * arg. Show regex, up to a maximum length. If it's too long, chop and add
389 #define _FAIL(code) STMT_START { \
390 const char *ellipses = ""; \
391 IV len = RExC_end - RExC_precomp; \
394 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
395 if (len > RegexLengthToShowInErrorMessages) { \
396 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
397 len = RegexLengthToShowInErrorMessages - 10; \
403 #define FAIL(msg) _FAIL( \
404 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
405 msg, (int)len, RExC_precomp, ellipses))
407 #define FAIL2(msg,arg) _FAIL( \
408 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
409 arg, (int)len, RExC_precomp, ellipses))
412 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
414 #define Simple_vFAIL(m) STMT_START { \
415 const IV offset = RExC_parse - RExC_precomp; \
416 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
417 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
421 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
423 #define vFAIL(m) STMT_START { \
425 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
430 * Like Simple_vFAIL(), but accepts two arguments.
432 #define Simple_vFAIL2(m,a1) STMT_START { \
433 const IV offset = RExC_parse - RExC_precomp; \
434 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
435 (int)offset, RExC_precomp, RExC_precomp + offset); \
439 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
441 #define vFAIL2(m,a1) STMT_START { \
443 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
444 Simple_vFAIL2(m, a1); \
449 * Like Simple_vFAIL(), but accepts three arguments.
451 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
452 const IV offset = RExC_parse - RExC_precomp; \
453 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
454 (int)offset, RExC_precomp, RExC_precomp + offset); \
458 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
460 #define vFAIL3(m,a1,a2) STMT_START { \
462 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
463 Simple_vFAIL3(m, a1, a2); \
467 * Like Simple_vFAIL(), but accepts four arguments.
469 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
470 const IV offset = RExC_parse - RExC_precomp; \
471 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
472 (int)offset, RExC_precomp, RExC_precomp + offset); \
475 #define vWARN(loc,m) STMT_START { \
476 const IV offset = loc - RExC_precomp; \
477 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
478 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
481 #define vWARNdep(loc,m) STMT_START { \
482 const IV offset = loc - RExC_precomp; \
483 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
484 "%s" REPORT_LOCATION, \
485 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
489 #define vWARN2(loc, m, a1) STMT_START { \
490 const IV offset = loc - RExC_precomp; \
491 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
492 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
495 #define vWARN3(loc, m, a1, a2) STMT_START { \
496 const IV offset = loc - RExC_precomp; \
497 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
498 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
501 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
502 const IV offset = loc - RExC_precomp; \
503 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
504 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
507 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
508 const IV offset = loc - RExC_precomp; \
509 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
510 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
514 /* Allow for side effects in s */
515 #define REGC(c,s) STMT_START { \
516 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
519 /* Macros for recording node offsets. 20001227 mjd@plover.com
520 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
521 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
522 * Element 0 holds the number n.
523 * Position is 1 indexed.
525 #ifndef RE_TRACK_PATTERN_OFFSETS
526 #define Set_Node_Offset_To_R(node,byte)
527 #define Set_Node_Offset(node,byte)
528 #define Set_Cur_Node_Offset
529 #define Set_Node_Length_To_R(node,len)
530 #define Set_Node_Length(node,len)
531 #define Set_Node_Cur_Length(node)
532 #define Node_Offset(n)
533 #define Node_Length(n)
534 #define Set_Node_Offset_Length(node,offset,len)
535 #define ProgLen(ri) ri->u.proglen
536 #define SetProgLen(ri,x) ri->u.proglen = x
538 #define ProgLen(ri) ri->u.offsets[0]
539 #define SetProgLen(ri,x) ri->u.offsets[0] = x
540 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
542 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
543 __LINE__, (int)(node), (int)(byte))); \
545 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
547 RExC_offsets[2*(node)-1] = (byte); \
552 #define Set_Node_Offset(node,byte) \
553 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
554 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
556 #define Set_Node_Length_To_R(node,len) STMT_START { \
558 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
559 __LINE__, (int)(node), (int)(len))); \
561 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
563 RExC_offsets[2*(node)] = (len); \
568 #define Set_Node_Length(node,len) \
569 Set_Node_Length_To_R((node)-RExC_emit_start, len)
570 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
571 #define Set_Node_Cur_Length(node) \
572 Set_Node_Length(node, RExC_parse - parse_start)
574 /* Get offsets and lengths */
575 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
576 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
578 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
579 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
580 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
584 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
585 #define EXPERIMENTAL_INPLACESCAN
586 #endif /*RE_TRACK_PATTERN_OFFSETS*/
588 #define DEBUG_STUDYDATA(str,data,depth) \
589 DEBUG_OPTIMISE_MORE_r(if(data){ \
590 PerlIO_printf(Perl_debug_log, \
591 "%*s" str "Pos:%"IVdf"/%"IVdf \
592 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
593 (int)(depth)*2, "", \
594 (IV)((data)->pos_min), \
595 (IV)((data)->pos_delta), \
596 (UV)((data)->flags), \
597 (IV)((data)->whilem_c), \
598 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
599 is_inf ? "INF " : "" \
601 if ((data)->last_found) \
602 PerlIO_printf(Perl_debug_log, \
603 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
604 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
605 SvPVX_const((data)->last_found), \
606 (IV)((data)->last_end), \
607 (IV)((data)->last_start_min), \
608 (IV)((data)->last_start_max), \
609 ((data)->longest && \
610 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
611 SvPVX_const((data)->longest_fixed), \
612 (IV)((data)->offset_fixed), \
613 ((data)->longest && \
614 (data)->longest==&((data)->longest_float)) ? "*" : "", \
615 SvPVX_const((data)->longest_float), \
616 (IV)((data)->offset_float_min), \
617 (IV)((data)->offset_float_max) \
619 PerlIO_printf(Perl_debug_log,"\n"); \
622 static void clear_re(pTHX_ void *r);
624 /* Mark that we cannot extend a found fixed substring at this point.
625 Update the longest found anchored substring and the longest found
626 floating substrings if needed. */
629 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
631 const STRLEN l = CHR_SVLEN(data->last_found);
632 const STRLEN old_l = CHR_SVLEN(*data->longest);
633 GET_RE_DEBUG_FLAGS_DECL;
635 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
636 SvSetMagicSV(*data->longest, data->last_found);
637 if (*data->longest == data->longest_fixed) {
638 data->offset_fixed = l ? data->last_start_min : data->pos_min;
639 if (data->flags & SF_BEFORE_EOL)
641 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
643 data->flags &= ~SF_FIX_BEFORE_EOL;
644 data->minlen_fixed=minlenp;
645 data->lookbehind_fixed=0;
647 else { /* *data->longest == data->longest_float */
648 data->offset_float_min = l ? data->last_start_min : data->pos_min;
649 data->offset_float_max = (l
650 ? data->last_start_max
651 : data->pos_min + data->pos_delta);
652 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
653 data->offset_float_max = I32_MAX;
654 if (data->flags & SF_BEFORE_EOL)
656 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
658 data->flags &= ~SF_FL_BEFORE_EOL;
659 data->minlen_float=minlenp;
660 data->lookbehind_float=0;
663 SvCUR_set(data->last_found, 0);
665 SV * const sv = data->last_found;
666 if (SvUTF8(sv) && SvMAGICAL(sv)) {
667 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
673 data->flags &= ~SF_BEFORE_EOL;
674 DEBUG_STUDYDATA("commit: ",data,0);
677 /* Can match anything (initialization) */
679 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
681 ANYOF_CLASS_ZERO(cl);
682 ANYOF_BITMAP_SETALL(cl);
683 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
685 cl->flags |= ANYOF_LOCALE;
688 /* Can match anything (initialization) */
690 S_cl_is_anything(const struct regnode_charclass_class *cl)
694 for (value = 0; value <= ANYOF_MAX; value += 2)
695 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
697 if (!(cl->flags & ANYOF_UNICODE_ALL))
699 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
704 /* Can match anything (initialization) */
706 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
708 Zero(cl, 1, struct regnode_charclass_class);
710 cl_anything(pRExC_state, cl);
714 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
716 Zero(cl, 1, struct regnode_charclass_class);
718 cl_anything(pRExC_state, cl);
720 cl->flags |= ANYOF_LOCALE;
723 /* 'And' a given class with another one. Can create false positives */
724 /* We assume that cl is not inverted */
726 S_cl_and(struct regnode_charclass_class *cl,
727 const struct regnode_charclass_class *and_with)
730 assert(and_with->type == ANYOF);
731 if (!(and_with->flags & ANYOF_CLASS)
732 && !(cl->flags & ANYOF_CLASS)
733 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
734 && !(and_with->flags & ANYOF_FOLD)
735 && !(cl->flags & ANYOF_FOLD)) {
738 if (and_with->flags & ANYOF_INVERT)
739 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
740 cl->bitmap[i] &= ~and_with->bitmap[i];
742 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
743 cl->bitmap[i] &= and_with->bitmap[i];
744 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
745 if (!(and_with->flags & ANYOF_EOS))
746 cl->flags &= ~ANYOF_EOS;
748 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
749 !(and_with->flags & ANYOF_INVERT)) {
750 cl->flags &= ~ANYOF_UNICODE_ALL;
751 cl->flags |= ANYOF_UNICODE;
752 ARG_SET(cl, ARG(and_with));
754 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
755 !(and_with->flags & ANYOF_INVERT))
756 cl->flags &= ~ANYOF_UNICODE_ALL;
757 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
758 !(and_with->flags & ANYOF_INVERT))
759 cl->flags &= ~ANYOF_UNICODE;
762 /* 'OR' a given class with another one. Can create false positives */
763 /* We assume that cl is not inverted */
765 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
767 if (or_with->flags & ANYOF_INVERT) {
769 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
770 * <= (B1 | !B2) | (CL1 | !CL2)
771 * which is wasteful if CL2 is small, but we ignore CL2:
772 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
773 * XXXX Can we handle case-fold? Unclear:
774 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
775 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
777 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
778 && !(or_with->flags & ANYOF_FOLD)
779 && !(cl->flags & ANYOF_FOLD) ) {
782 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
783 cl->bitmap[i] |= ~or_with->bitmap[i];
784 } /* XXXX: logic is complicated otherwise */
786 cl_anything(pRExC_state, cl);
789 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
790 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
791 && (!(or_with->flags & ANYOF_FOLD)
792 || (cl->flags & ANYOF_FOLD)) ) {
795 /* OR char bitmap and class bitmap separately */
796 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
797 cl->bitmap[i] |= or_with->bitmap[i];
798 if (or_with->flags & ANYOF_CLASS) {
799 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
800 cl->classflags[i] |= or_with->classflags[i];
801 cl->flags |= ANYOF_CLASS;
804 else { /* XXXX: logic is complicated, leave it along for a moment. */
805 cl_anything(pRExC_state, cl);
808 if (or_with->flags & ANYOF_EOS)
809 cl->flags |= ANYOF_EOS;
811 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
812 ARG(cl) != ARG(or_with)) {
813 cl->flags |= ANYOF_UNICODE_ALL;
814 cl->flags &= ~ANYOF_UNICODE;
816 if (or_with->flags & ANYOF_UNICODE_ALL) {
817 cl->flags |= ANYOF_UNICODE_ALL;
818 cl->flags &= ~ANYOF_UNICODE;
822 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
823 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
824 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
825 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
830 dump_trie(trie,widecharmap,revcharmap)
831 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
832 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
834 These routines dump out a trie in a somewhat readable format.
835 The _interim_ variants are used for debugging the interim
836 tables that are used to generate the final compressed
837 representation which is what dump_trie expects.
839 Part of the reason for their existance is to provide a form
840 of documentation as to how the different representations function.
845 Dumps the final compressed table form of the trie to Perl_debug_log.
846 Used for debugging make_trie().
850 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
851 AV *revcharmap, U32 depth)
854 SV *sv=sv_newmortal();
855 int colwidth= widecharmap ? 6 : 4;
856 GET_RE_DEBUG_FLAGS_DECL;
859 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
860 (int)depth * 2 + 2,"",
861 "Match","Base","Ofs" );
863 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
864 SV ** const tmp = av_fetch( revcharmap, state, 0);
866 PerlIO_printf( Perl_debug_log, "%*s",
868 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
869 PL_colors[0], PL_colors[1],
870 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
871 PERL_PV_ESCAPE_FIRSTCHAR
876 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
877 (int)depth * 2 + 2,"");
879 for( state = 0 ; state < trie->uniquecharcount ; state++ )
880 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
881 PerlIO_printf( Perl_debug_log, "\n");
883 for( state = 1 ; state < trie->statecount ; state++ ) {
884 const U32 base = trie->states[ state ].trans.base;
886 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
888 if ( trie->states[ state ].wordnum ) {
889 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
891 PerlIO_printf( Perl_debug_log, "%6s", "" );
894 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
899 while( ( base + ofs < trie->uniquecharcount ) ||
900 ( base + ofs - trie->uniquecharcount < trie->lasttrans
901 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
904 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
906 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
907 if ( ( base + ofs >= trie->uniquecharcount ) &&
908 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
909 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
911 PerlIO_printf( Perl_debug_log, "%*"UVXf,
913 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
915 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
919 PerlIO_printf( Perl_debug_log, "]");
922 PerlIO_printf( Perl_debug_log, "\n" );
926 Dumps a fully constructed but uncompressed trie in list form.
927 List tries normally only are used for construction when the number of
928 possible chars (trie->uniquecharcount) is very high.
929 Used for debugging make_trie().
932 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
933 HV *widecharmap, AV *revcharmap, U32 next_alloc,
937 SV *sv=sv_newmortal();
938 int colwidth= widecharmap ? 6 : 4;
939 GET_RE_DEBUG_FLAGS_DECL;
940 /* print out the table precompression. */
941 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
942 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
943 "------:-----+-----------------\n" );
945 for( state=1 ; state < next_alloc ; state ++ ) {
948 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
949 (int)depth * 2 + 2,"", (UV)state );
950 if ( ! trie->states[ state ].wordnum ) {
951 PerlIO_printf( Perl_debug_log, "%5s| ","");
953 PerlIO_printf( Perl_debug_log, "W%4x| ",
954 trie->states[ state ].wordnum
957 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
958 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
960 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
962 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
963 PL_colors[0], PL_colors[1],
964 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
965 PERL_PV_ESCAPE_FIRSTCHAR
967 TRIE_LIST_ITEM(state,charid).forid,
968 (UV)TRIE_LIST_ITEM(state,charid).newstate
971 PerlIO_printf(Perl_debug_log, "\n%*s| ",
972 (int)((depth * 2) + 14), "");
975 PerlIO_printf( Perl_debug_log, "\n");
980 Dumps a fully constructed but uncompressed trie in table form.
981 This is the normal DFA style state transition table, with a few
982 twists to facilitate compression later.
983 Used for debugging make_trie().
986 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
987 HV *widecharmap, AV *revcharmap, U32 next_alloc,
992 SV *sv=sv_newmortal();
993 int colwidth= widecharmap ? 6 : 4;
994 GET_RE_DEBUG_FLAGS_DECL;
997 print out the table precompression so that we can do a visual check
998 that they are identical.
1001 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1003 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1004 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1006 PerlIO_printf( Perl_debug_log, "%*s",
1008 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1009 PL_colors[0], PL_colors[1],
1010 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1011 PERL_PV_ESCAPE_FIRSTCHAR
1017 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1019 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1020 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1023 PerlIO_printf( Perl_debug_log, "\n" );
1025 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1027 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1028 (int)depth * 2 + 2,"",
1029 (UV)TRIE_NODENUM( state ) );
1031 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1032 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1034 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1036 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1038 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1039 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1041 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1042 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1049 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1050 startbranch: the first branch in the whole branch sequence
1051 first : start branch of sequence of branch-exact nodes.
1052 May be the same as startbranch
1053 last : Thing following the last branch.
1054 May be the same as tail.
1055 tail : item following the branch sequence
1056 count : words in the sequence
1057 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1058 depth : indent depth
1060 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1062 A trie is an N'ary tree where the branches are determined by digital
1063 decomposition of the key. IE, at the root node you look up the 1st character and
1064 follow that branch repeat until you find the end of the branches. Nodes can be
1065 marked as "accepting" meaning they represent a complete word. Eg:
1069 would convert into the following structure. Numbers represent states, letters
1070 following numbers represent valid transitions on the letter from that state, if
1071 the number is in square brackets it represents an accepting state, otherwise it
1072 will be in parenthesis.
1074 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1078 (1) +-i->(6)-+-s->[7]
1080 +-s->(3)-+-h->(4)-+-e->[5]
1082 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1084 This shows that when matching against the string 'hers' we will begin at state 1
1085 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1086 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1087 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1088 single traverse. We store a mapping from accepting to state to which word was
1089 matched, and then when we have multiple possibilities we try to complete the
1090 rest of the regex in the order in which they occured in the alternation.
1092 The only prior NFA like behaviour that would be changed by the TRIE support is
1093 the silent ignoring of duplicate alternations which are of the form:
1095 / (DUPE|DUPE) X? (?{ ... }) Y /x
1097 Thus EVAL blocks follwing a trie may be called a different number of times with
1098 and without the optimisation. With the optimisations dupes will be silently
1099 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1100 the following demonstrates:
1102 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1104 which prints out 'word' three times, but
1106 'words'=~/(word|word|word)(?{ print $1 })S/
1108 which doesnt print it out at all. This is due to other optimisations kicking in.
1110 Example of what happens on a structural level:
1112 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1114 1: CURLYM[1] {1,32767}(18)
1125 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1126 and should turn into:
1128 1: CURLYM[1] {1,32767}(18)
1130 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1138 Cases where tail != last would be like /(?foo|bar)baz/:
1148 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1149 and would end up looking like:
1152 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1159 d = uvuni_to_utf8_flags(d, uv, 0);
1161 is the recommended Unicode-aware way of saying
1166 #define TRIE_STORE_REVCHAR \
1169 SV *zlopp = newSV(2); \
1170 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1171 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1172 SvCUR_set(zlopp, kapow - flrbbbbb); \
1175 av_push(revcharmap, zlopp); \
1177 char ooooff = (char)uvc; \
1178 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1182 #define TRIE_READ_CHAR STMT_START { \
1186 if ( foldlen > 0 ) { \
1187 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1192 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1193 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1194 foldlen -= UNISKIP( uvc ); \
1195 scan = foldbuf + UNISKIP( uvc ); \
1198 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1208 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1209 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1210 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1211 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1213 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1214 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1215 TRIE_LIST_CUR( state )++; \
1218 #define TRIE_LIST_NEW(state) STMT_START { \
1219 Newxz( trie->states[ state ].trans.list, \
1220 4, reg_trie_trans_le ); \
1221 TRIE_LIST_CUR( state ) = 1; \
1222 TRIE_LIST_LEN( state ) = 4; \
1225 #define TRIE_HANDLE_WORD(state) STMT_START { \
1226 U16 dupe= trie->states[ state ].wordnum; \
1227 regnode * const noper_next = regnext( noper ); \
1229 if (trie->wordlen) \
1230 trie->wordlen[ curword ] = wordlen; \
1232 /* store the word for dumping */ \
1234 if (OP(noper) != NOTHING) \
1235 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1237 tmp = newSVpvn_utf8( "", 0, UTF ); \
1238 av_push( trie_words, tmp ); \
1243 if ( noper_next < tail ) { \
1245 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1246 trie->jump[curword] = (U16)(noper_next - convert); \
1248 jumper = noper_next; \
1250 nextbranch= regnext(cur); \
1254 /* So it's a dupe. This means we need to maintain a */\
1255 /* linked-list from the first to the next. */\
1256 /* we only allocate the nextword buffer when there */\
1257 /* a dupe, so first time we have to do the allocation */\
1258 if (!trie->nextword) \
1259 trie->nextword = (U16 *) \
1260 PerlMemShared_calloc( word_count + 1, sizeof(U16)); \
1261 while ( trie->nextword[dupe] ) \
1262 dupe= trie->nextword[dupe]; \
1263 trie->nextword[dupe]= curword; \
1265 /* we haven't inserted this word yet. */ \
1266 trie->states[ state ].wordnum = curword; \
1271 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1272 ( ( base + charid >= ucharcount \
1273 && base + charid < ubound \
1274 && state == trie->trans[ base - ucharcount + charid ].check \
1275 && trie->trans[ base - ucharcount + charid ].next ) \
1276 ? trie->trans[ base - ucharcount + charid ].next \
1277 : ( state==1 ? special : 0 ) \
1281 #define MADE_JUMP_TRIE 2
1282 #define MADE_EXACT_TRIE 4
1285 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1288 /* first pass, loop through and scan words */
1289 reg_trie_data *trie;
1290 HV *widecharmap = NULL;
1291 AV *revcharmap = newAV();
1293 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1298 regnode *jumper = NULL;
1299 regnode *nextbranch = NULL;
1300 regnode *convert = NULL;
1301 /* we just use folder as a flag in utf8 */
1302 const U8 * const folder = ( flags == EXACTF
1304 : ( flags == EXACTFL
1311 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1312 AV *trie_words = NULL;
1313 /* along with revcharmap, this only used during construction but both are
1314 * useful during debugging so we store them in the struct when debugging.
1317 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1318 STRLEN trie_charcount=0;
1320 SV *re_trie_maxbuff;
1321 GET_RE_DEBUG_FLAGS_DECL;
1323 PERL_UNUSED_ARG(depth);
1326 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1328 trie->startstate = 1;
1329 trie->wordcount = word_count;
1330 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1331 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1332 if (!(UTF && folder))
1333 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1335 trie_words = newAV();
1338 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1339 if (!SvIOK(re_trie_maxbuff)) {
1340 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1343 PerlIO_printf( Perl_debug_log,
1344 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1345 (int)depth * 2 + 2, "",
1346 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1347 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1351 /* Find the node we are going to overwrite */
1352 if ( first == startbranch && OP( last ) != BRANCH ) {
1353 /* whole branch chain */
1356 /* branch sub-chain */
1357 convert = NEXTOPER( first );
1360 /* -- First loop and Setup --
1362 We first traverse the branches and scan each word to determine if it
1363 contains widechars, and how many unique chars there are, this is
1364 important as we have to build a table with at least as many columns as we
1367 We use an array of integers to represent the character codes 0..255
1368 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1369 native representation of the character value as the key and IV's for the
1372 *TODO* If we keep track of how many times each character is used we can
1373 remap the columns so that the table compression later on is more
1374 efficient in terms of memory by ensuring most common value is in the
1375 middle and the least common are on the outside. IMO this would be better
1376 than a most to least common mapping as theres a decent chance the most
1377 common letter will share a node with the least common, meaning the node
1378 will not be compressable. With a middle is most common approach the worst
1379 case is when we have the least common nodes twice.
1383 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1384 regnode * const noper = NEXTOPER( cur );
1385 const U8 *uc = (U8*)STRING( noper );
1386 const U8 * const e = uc + STR_LEN( noper );
1388 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1389 const U8 *scan = (U8*)NULL;
1390 U32 wordlen = 0; /* required init */
1392 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1394 if (OP(noper) == NOTHING) {
1398 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1399 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1400 regardless of encoding */
1402 for ( ; uc < e ; uc += len ) {
1403 TRIE_CHARCOUNT(trie)++;
1407 if ( !trie->charmap[ uvc ] ) {
1408 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1410 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1414 /* store the codepoint in the bitmap, and if its ascii
1415 also store its folded equivelent. */
1416 TRIE_BITMAP_SET(trie,uvc);
1418 /* store the folded codepoint */
1419 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1422 /* store first byte of utf8 representation of
1423 codepoints in the 127 < uvc < 256 range */
1424 if (127 < uvc && uvc < 192) {
1425 TRIE_BITMAP_SET(trie,194);
1426 } else if (191 < uvc ) {
1427 TRIE_BITMAP_SET(trie,195);
1428 /* && uvc < 256 -- we know uvc is < 256 already */
1431 set_bit = 0; /* We've done our bit :-) */
1436 widecharmap = newHV();
1438 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1441 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1443 if ( !SvTRUE( *svpp ) ) {
1444 sv_setiv( *svpp, ++trie->uniquecharcount );
1449 if( cur == first ) {
1452 } else if (chars < trie->minlen) {
1454 } else if (chars > trie->maxlen) {
1458 } /* end first pass */
1459 DEBUG_TRIE_COMPILE_r(
1460 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1461 (int)depth * 2 + 2,"",
1462 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1463 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1464 (int)trie->minlen, (int)trie->maxlen )
1466 trie->wordlen = (U32 *) PerlMemShared_calloc( word_count, sizeof(U32) );
1469 We now know what we are dealing with in terms of unique chars and
1470 string sizes so we can calculate how much memory a naive
1471 representation using a flat table will take. If it's over a reasonable
1472 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1473 conservative but potentially much slower representation using an array
1476 At the end we convert both representations into the same compressed
1477 form that will be used in regexec.c for matching with. The latter
1478 is a form that cannot be used to construct with but has memory
1479 properties similar to the list form and access properties similar
1480 to the table form making it both suitable for fast searches and
1481 small enough that its feasable to store for the duration of a program.
1483 See the comment in the code where the compressed table is produced
1484 inplace from the flat tabe representation for an explanation of how
1485 the compression works.
1490 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1492 Second Pass -- Array Of Lists Representation
1494 Each state will be represented by a list of charid:state records
1495 (reg_trie_trans_le) the first such element holds the CUR and LEN
1496 points of the allocated array. (See defines above).
1498 We build the initial structure using the lists, and then convert
1499 it into the compressed table form which allows faster lookups
1500 (but cant be modified once converted).
1503 STRLEN transcount = 1;
1505 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1506 "%*sCompiling trie using list compiler\n",
1507 (int)depth * 2 + 2, ""));
1509 trie->states = (reg_trie_state *)
1510 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1511 sizeof(reg_trie_state) );
1515 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1517 regnode * const noper = NEXTOPER( cur );
1518 U8 *uc = (U8*)STRING( noper );
1519 const U8 * const e = uc + STR_LEN( noper );
1520 U32 state = 1; /* required init */
1521 U16 charid = 0; /* sanity init */
1522 U8 *scan = (U8*)NULL; /* sanity init */
1523 STRLEN foldlen = 0; /* required init */
1524 U32 wordlen = 0; /* required init */
1525 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1527 if (OP(noper) != NOTHING) {
1528 for ( ; uc < e ; uc += len ) {
1533 charid = trie->charmap[ uvc ];
1535 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1539 charid=(U16)SvIV( *svpp );
1542 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1549 if ( !trie->states[ state ].trans.list ) {
1550 TRIE_LIST_NEW( state );
1552 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1553 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1554 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1559 newstate = next_alloc++;
1560 TRIE_LIST_PUSH( state, charid, newstate );
1565 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1569 TRIE_HANDLE_WORD(state);
1571 } /* end second pass */
1573 /* next alloc is the NEXT state to be allocated */
1574 trie->statecount = next_alloc;
1575 trie->states = (reg_trie_state *)
1576 PerlMemShared_realloc( trie->states,
1578 * sizeof(reg_trie_state) );
1580 /* and now dump it out before we compress it */
1581 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1582 revcharmap, next_alloc,
1586 trie->trans = (reg_trie_trans *)
1587 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1594 for( state=1 ; state < next_alloc ; state ++ ) {
1598 DEBUG_TRIE_COMPILE_MORE_r(
1599 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1603 if (trie->states[state].trans.list) {
1604 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1608 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1609 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1610 if ( forid < minid ) {
1612 } else if ( forid > maxid ) {
1616 if ( transcount < tp + maxid - minid + 1) {
1618 trie->trans = (reg_trie_trans *)
1619 PerlMemShared_realloc( trie->trans,
1621 * sizeof(reg_trie_trans) );
1622 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1624 base = trie->uniquecharcount + tp - minid;
1625 if ( maxid == minid ) {
1627 for ( ; zp < tp ; zp++ ) {
1628 if ( ! trie->trans[ zp ].next ) {
1629 base = trie->uniquecharcount + zp - minid;
1630 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1631 trie->trans[ zp ].check = state;
1637 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1638 trie->trans[ tp ].check = state;
1643 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1644 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1645 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1646 trie->trans[ tid ].check = state;
1648 tp += ( maxid - minid + 1 );
1650 Safefree(trie->states[ state ].trans.list);
1653 DEBUG_TRIE_COMPILE_MORE_r(
1654 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1657 trie->states[ state ].trans.base=base;
1659 trie->lasttrans = tp + 1;
1663 Second Pass -- Flat Table Representation.
1665 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1666 We know that we will need Charcount+1 trans at most to store the data
1667 (one row per char at worst case) So we preallocate both structures
1668 assuming worst case.
1670 We then construct the trie using only the .next slots of the entry
1673 We use the .check field of the first entry of the node temporarily to
1674 make compression both faster and easier by keeping track of how many non
1675 zero fields are in the node.
1677 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1680 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1681 number representing the first entry of the node, and state as a
1682 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1683 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1684 are 2 entrys per node. eg:
1692 The table is internally in the right hand, idx form. However as we also
1693 have to deal with the states array which is indexed by nodenum we have to
1694 use TRIE_NODENUM() to convert.
1697 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1698 "%*sCompiling trie using table compiler\n",
1699 (int)depth * 2 + 2, ""));
1701 trie->trans = (reg_trie_trans *)
1702 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1703 * trie->uniquecharcount + 1,
1704 sizeof(reg_trie_trans) );
1705 trie->states = (reg_trie_state *)
1706 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1707 sizeof(reg_trie_state) );
1708 next_alloc = trie->uniquecharcount + 1;
1711 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1713 regnode * const noper = NEXTOPER( cur );
1714 const U8 *uc = (U8*)STRING( noper );
1715 const U8 * const e = uc + STR_LEN( noper );
1717 U32 state = 1; /* required init */
1719 U16 charid = 0; /* sanity init */
1720 U32 accept_state = 0; /* sanity init */
1721 U8 *scan = (U8*)NULL; /* sanity init */
1723 STRLEN foldlen = 0; /* required init */
1724 U32 wordlen = 0; /* required init */
1725 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1727 if ( OP(noper) != NOTHING ) {
1728 for ( ; uc < e ; uc += len ) {
1733 charid = trie->charmap[ uvc ];
1735 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1736 charid = svpp ? (U16)SvIV(*svpp) : 0;
1740 if ( !trie->trans[ state + charid ].next ) {
1741 trie->trans[ state + charid ].next = next_alloc;
1742 trie->trans[ state ].check++;
1743 next_alloc += trie->uniquecharcount;
1745 state = trie->trans[ state + charid ].next;
1747 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1749 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1752 accept_state = TRIE_NODENUM( state );
1753 TRIE_HANDLE_WORD(accept_state);
1755 } /* end second pass */
1757 /* and now dump it out before we compress it */
1758 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1760 next_alloc, depth+1));
1764 * Inplace compress the table.*
1766 For sparse data sets the table constructed by the trie algorithm will
1767 be mostly 0/FAIL transitions or to put it another way mostly empty.
1768 (Note that leaf nodes will not contain any transitions.)
1770 This algorithm compresses the tables by eliminating most such
1771 transitions, at the cost of a modest bit of extra work during lookup:
1773 - Each states[] entry contains a .base field which indicates the
1774 index in the state[] array wheres its transition data is stored.
1776 - If .base is 0 there are no valid transitions from that node.
1778 - If .base is nonzero then charid is added to it to find an entry in
1781 -If trans[states[state].base+charid].check!=state then the
1782 transition is taken to be a 0/Fail transition. Thus if there are fail
1783 transitions at the front of the node then the .base offset will point
1784 somewhere inside the previous nodes data (or maybe even into a node
1785 even earlier), but the .check field determines if the transition is
1789 The following process inplace converts the table to the compressed
1790 table: We first do not compress the root node 1,and mark its all its
1791 .check pointers as 1 and set its .base pointer as 1 as well. This
1792 allows to do a DFA construction from the compressed table later, and
1793 ensures that any .base pointers we calculate later are greater than
1796 - We set 'pos' to indicate the first entry of the second node.
1798 - We then iterate over the columns of the node, finding the first and
1799 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1800 and set the .check pointers accordingly, and advance pos
1801 appropriately and repreat for the next node. Note that when we copy
1802 the next pointers we have to convert them from the original
1803 NODEIDX form to NODENUM form as the former is not valid post
1806 - If a node has no transitions used we mark its base as 0 and do not
1807 advance the pos pointer.
1809 - If a node only has one transition we use a second pointer into the
1810 structure to fill in allocated fail transitions from other states.
1811 This pointer is independent of the main pointer and scans forward
1812 looking for null transitions that are allocated to a state. When it
1813 finds one it writes the single transition into the "hole". If the
1814 pointer doesnt find one the single transition is appended as normal.
1816 - Once compressed we can Renew/realloc the structures to release the
1819 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1820 specifically Fig 3.47 and the associated pseudocode.
1824 const U32 laststate = TRIE_NODENUM( next_alloc );
1827 trie->statecount = laststate;
1829 for ( state = 1 ; state < laststate ; state++ ) {
1831 const U32 stateidx = TRIE_NODEIDX( state );
1832 const U32 o_used = trie->trans[ stateidx ].check;
1833 U32 used = trie->trans[ stateidx ].check;
1834 trie->trans[ stateidx ].check = 0;
1836 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1837 if ( flag || trie->trans[ stateidx + charid ].next ) {
1838 if ( trie->trans[ stateidx + charid ].next ) {
1840 for ( ; zp < pos ; zp++ ) {
1841 if ( ! trie->trans[ zp ].next ) {
1845 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1846 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1847 trie->trans[ zp ].check = state;
1848 if ( ++zp > pos ) pos = zp;
1855 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1857 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1858 trie->trans[ pos ].check = state;
1863 trie->lasttrans = pos + 1;
1864 trie->states = (reg_trie_state *)
1865 PerlMemShared_realloc( trie->states, laststate
1866 * sizeof(reg_trie_state) );
1867 DEBUG_TRIE_COMPILE_MORE_r(
1868 PerlIO_printf( Perl_debug_log,
1869 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1870 (int)depth * 2 + 2,"",
1871 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1874 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1877 } /* end table compress */
1879 DEBUG_TRIE_COMPILE_MORE_r(
1880 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1881 (int)depth * 2 + 2, "",
1882 (UV)trie->statecount,
1883 (UV)trie->lasttrans)
1885 /* resize the trans array to remove unused space */
1886 trie->trans = (reg_trie_trans *)
1887 PerlMemShared_realloc( trie->trans, trie->lasttrans
1888 * sizeof(reg_trie_trans) );
1890 /* and now dump out the compressed format */
1891 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
1893 { /* Modify the program and insert the new TRIE node*/
1894 U8 nodetype =(U8)(flags & 0xFF);
1898 regnode *optimize = NULL;
1899 #ifdef RE_TRACK_PATTERN_OFFSETS
1902 U32 mjd_nodelen = 0;
1903 #endif /* RE_TRACK_PATTERN_OFFSETS */
1904 #endif /* DEBUGGING */
1906 This means we convert either the first branch or the first Exact,
1907 depending on whether the thing following (in 'last') is a branch
1908 or not and whther first is the startbranch (ie is it a sub part of
1909 the alternation or is it the whole thing.)
1910 Assuming its a sub part we conver the EXACT otherwise we convert
1911 the whole branch sequence, including the first.
1913 /* Find the node we are going to overwrite */
1914 if ( first != startbranch || OP( last ) == BRANCH ) {
1915 /* branch sub-chain */
1916 NEXT_OFF( first ) = (U16)(last - first);
1917 #ifdef RE_TRACK_PATTERN_OFFSETS
1919 mjd_offset= Node_Offset((convert));
1920 mjd_nodelen= Node_Length((convert));
1923 /* whole branch chain */
1925 #ifdef RE_TRACK_PATTERN_OFFSETS
1928 const regnode *nop = NEXTOPER( convert );
1929 mjd_offset= Node_Offset((nop));
1930 mjd_nodelen= Node_Length((nop));
1934 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1935 (int)depth * 2 + 2, "",
1936 (UV)mjd_offset, (UV)mjd_nodelen)
1939 /* But first we check to see if there is a common prefix we can
1940 split out as an EXACT and put in front of the TRIE node. */
1941 trie->startstate= 1;
1942 if ( trie->bitmap && !widecharmap && !trie->jump ) {
1944 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1948 const U32 base = trie->states[ state ].trans.base;
1950 if ( trie->states[state].wordnum )
1953 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1954 if ( ( base + ofs >= trie->uniquecharcount ) &&
1955 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1956 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1958 if ( ++count > 1 ) {
1959 SV **tmp = av_fetch( revcharmap, ofs, 0);
1960 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1961 if ( state == 1 ) break;
1963 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1965 PerlIO_printf(Perl_debug_log,
1966 "%*sNew Start State=%"UVuf" Class: [",
1967 (int)depth * 2 + 2, "",
1970 SV ** const tmp = av_fetch( revcharmap, idx, 0);
1971 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1973 TRIE_BITMAP_SET(trie,*ch);
1975 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1977 PerlIO_printf(Perl_debug_log, (char*)ch)
1981 TRIE_BITMAP_SET(trie,*ch);
1983 TRIE_BITMAP_SET(trie,folder[ *ch ]);
1984 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
1990 SV **tmp = av_fetch( revcharmap, idx, 0);
1992 char *ch = SvPV( *tmp, len );
1994 SV *sv=sv_newmortal();
1995 PerlIO_printf( Perl_debug_log,
1996 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
1997 (int)depth * 2 + 2, "",
1999 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2000 PL_colors[0], PL_colors[1],
2001 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2002 PERL_PV_ESCAPE_FIRSTCHAR
2007 OP( convert ) = nodetype;
2008 str=STRING(convert);
2011 STR_LEN(convert) += len;
2017 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2023 regnode *n = convert+NODE_SZ_STR(convert);
2024 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2025 trie->startstate = state;
2026 trie->minlen -= (state - 1);
2027 trie->maxlen -= (state - 1);
2029 /* At least the UNICOS C compiler choked on this
2030 * being argument to DEBUG_r(), so let's just have
2033 #ifdef PERL_EXT_RE_BUILD
2039 regnode *fix = convert;
2040 U32 word = trie->wordcount;
2042 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2043 while( ++fix < n ) {
2044 Set_Node_Offset_Length(fix, 0, 0);
2047 SV ** const tmp = av_fetch( trie_words, word, 0 );
2049 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2050 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2052 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2060 NEXT_OFF(convert) = (U16)(tail - convert);
2061 DEBUG_r(optimize= n);
2067 if ( trie->maxlen ) {
2068 NEXT_OFF( convert ) = (U16)(tail - convert);
2069 ARG_SET( convert, data_slot );
2070 /* Store the offset to the first unabsorbed branch in
2071 jump[0], which is otherwise unused by the jump logic.
2072 We use this when dumping a trie and during optimisation. */
2074 trie->jump[0] = (U16)(nextbranch - convert);
2077 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
2078 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2080 OP( convert ) = TRIEC;
2081 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2082 PerlMemShared_free(trie->bitmap);
2085 OP( convert ) = TRIE;
2087 /* store the type in the flags */
2088 convert->flags = nodetype;
2092 + regarglen[ OP( convert ) ];
2094 /* XXX We really should free up the resource in trie now,
2095 as we won't use them - (which resources?) dmq */
2097 /* needed for dumping*/
2098 DEBUG_r(if (optimize) {
2099 regnode *opt = convert;
2101 while ( ++opt < optimize) {
2102 Set_Node_Offset_Length(opt,0,0);
2105 Try to clean up some of the debris left after the
2108 while( optimize < jumper ) {
2109 mjd_nodelen += Node_Length((optimize));
2110 OP( optimize ) = OPTIMIZED;
2111 Set_Node_Offset_Length(optimize,0,0);
2114 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2116 } /* end node insert */
2117 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2119 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2120 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2122 SvREFCNT_dec(revcharmap);
2126 : trie->startstate>1
2132 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2134 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
2136 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2137 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2140 We find the fail state for each state in the trie, this state is the longest proper
2141 suffix of the current states 'word' that is also a proper prefix of another word in our
2142 trie. State 1 represents the word '' and is the thus the default fail state. This allows
2143 the DFA not to have to restart after its tried and failed a word at a given point, it
2144 simply continues as though it had been matching the other word in the first place.
2146 'abcdgu'=~/abcdefg|cdgu/
2147 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2148 fail, which would bring use to the state representing 'd' in the second word where we would
2149 try 'g' and succeed, prodceding to match 'cdgu'.
2151 /* add a fail transition */
2152 const U32 trie_offset = ARG(source);
2153 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2155 const U32 ucharcount = trie->uniquecharcount;
2156 const U32 numstates = trie->statecount;
2157 const U32 ubound = trie->lasttrans + ucharcount;
2161 U32 base = trie->states[ 1 ].trans.base;
2164 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2165 GET_RE_DEBUG_FLAGS_DECL;
2167 PERL_UNUSED_ARG(depth);
2171 ARG_SET( stclass, data_slot );
2172 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2173 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2174 aho->trie=trie_offset;
2175 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2176 Copy( trie->states, aho->states, numstates, reg_trie_state );
2177 Newxz( q, numstates, U32);
2178 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2181 /* initialize fail[0..1] to be 1 so that we always have
2182 a valid final fail state */
2183 fail[ 0 ] = fail[ 1 ] = 1;
2185 for ( charid = 0; charid < ucharcount ; charid++ ) {
2186 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2188 q[ q_write ] = newstate;
2189 /* set to point at the root */
2190 fail[ q[ q_write++ ] ]=1;
2193 while ( q_read < q_write) {
2194 const U32 cur = q[ q_read++ % numstates ];
2195 base = trie->states[ cur ].trans.base;
2197 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2198 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2200 U32 fail_state = cur;
2203 fail_state = fail[ fail_state ];
2204 fail_base = aho->states[ fail_state ].trans.base;
2205 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2207 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2208 fail[ ch_state ] = fail_state;
2209 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2211 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2213 q[ q_write++ % numstates] = ch_state;
2217 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2218 when we fail in state 1, this allows us to use the
2219 charclass scan to find a valid start char. This is based on the principle
2220 that theres a good chance the string being searched contains lots of stuff
2221 that cant be a start char.
2223 fail[ 0 ] = fail[ 1 ] = 0;
2224 DEBUG_TRIE_COMPILE_r({
2225 PerlIO_printf(Perl_debug_log,
2226 "%*sStclass Failtable (%"UVuf" states): 0",
2227 (int)(depth * 2), "", (UV)numstates
2229 for( q_read=1; q_read<numstates; q_read++ ) {
2230 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2232 PerlIO_printf(Perl_debug_log, "\n");
2235 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2240 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2241 * These need to be revisited when a newer toolchain becomes available.
2243 #if defined(__sparc64__) && defined(__GNUC__)
2244 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2245 # undef SPARC64_GCC_WORKAROUND
2246 # define SPARC64_GCC_WORKAROUND 1
2250 #define DEBUG_PEEP(str,scan,depth) \
2251 DEBUG_OPTIMISE_r({if (scan){ \
2252 SV * const mysv=sv_newmortal(); \
2253 regnode *Next = regnext(scan); \
2254 regprop(RExC_rx, mysv, scan); \
2255 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2256 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2257 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2264 #define JOIN_EXACT(scan,min,flags) \
2265 if (PL_regkind[OP(scan)] == EXACT) \
2266 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2269 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2270 /* Merge several consecutive EXACTish nodes into one. */
2271 regnode *n = regnext(scan);
2273 regnode *next = scan + NODE_SZ_STR(scan);
2277 regnode *stop = scan;
2278 GET_RE_DEBUG_FLAGS_DECL;
2280 PERL_UNUSED_ARG(depth);
2282 #ifndef EXPERIMENTAL_INPLACESCAN
2283 PERL_UNUSED_ARG(flags);
2284 PERL_UNUSED_ARG(val);
2286 DEBUG_PEEP("join",scan,depth);
2288 /* Skip NOTHING, merge EXACT*. */
2290 ( PL_regkind[OP(n)] == NOTHING ||
2291 (stringok && (OP(n) == OP(scan))))
2293 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2295 if (OP(n) == TAIL || n > next)
2297 if (PL_regkind[OP(n)] == NOTHING) {
2298 DEBUG_PEEP("skip:",n,depth);
2299 NEXT_OFF(scan) += NEXT_OFF(n);
2300 next = n + NODE_STEP_REGNODE;
2307 else if (stringok) {
2308 const unsigned int oldl = STR_LEN(scan);
2309 regnode * const nnext = regnext(n);
2311 DEBUG_PEEP("merg",n,depth);
2314 if (oldl + STR_LEN(n) > U8_MAX)
2316 NEXT_OFF(scan) += NEXT_OFF(n);
2317 STR_LEN(scan) += STR_LEN(n);
2318 next = n + NODE_SZ_STR(n);
2319 /* Now we can overwrite *n : */
2320 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2328 #ifdef EXPERIMENTAL_INPLACESCAN
2329 if (flags && !NEXT_OFF(n)) {
2330 DEBUG_PEEP("atch", val, depth);
2331 if (reg_off_by_arg[OP(n)]) {
2332 ARG_SET(n, val - n);
2335 NEXT_OFF(n) = val - n;
2342 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2344 Two problematic code points in Unicode casefolding of EXACT nodes:
2346 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2347 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2353 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2354 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2356 This means that in case-insensitive matching (or "loose matching",
2357 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2358 length of the above casefolded versions) can match a target string
2359 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2360 This would rather mess up the minimum length computation.
2362 What we'll do is to look for the tail four bytes, and then peek
2363 at the preceding two bytes to see whether we need to decrease
2364 the minimum length by four (six minus two).
2366 Thanks to the design of UTF-8, there cannot be false matches:
2367 A sequence of valid UTF-8 bytes cannot be a subsequence of
2368 another valid sequence of UTF-8 bytes.
2371 char * const s0 = STRING(scan), *s, *t;
2372 char * const s1 = s0 + STR_LEN(scan) - 1;
2373 char * const s2 = s1 - 4;
2374 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2375 const char t0[] = "\xaf\x49\xaf\x42";
2377 const char t0[] = "\xcc\x88\xcc\x81";
2379 const char * const t1 = t0 + 3;
2382 s < s2 && (t = ninstr(s, s1, t0, t1));
2385 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2386 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2388 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2389 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2397 n = scan + NODE_SZ_STR(scan);
2399 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2406 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2410 /* REx optimizer. Converts nodes into quickier variants "in place".
2411 Finds fixed substrings. */
2413 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2414 to the position after last scanned or to NULL. */
2416 #define INIT_AND_WITHP \
2417 assert(!and_withp); \
2418 Newx(and_withp,1,struct regnode_charclass_class); \
2419 SAVEFREEPV(and_withp)
2421 /* this is a chain of data about sub patterns we are processing that
2422 need to be handled seperately/specially in study_chunk. Its so
2423 we can simulate recursion without losing state. */
2425 typedef struct scan_frame {
2426 regnode *last; /* last node to process in this frame */
2427 regnode *next; /* next node to process when last is reached */
2428 struct scan_frame *prev; /*previous frame*/
2429 I32 stop; /* what stopparen do we use */
2433 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2435 #define CASE_SYNST_FNC(nAmE) \
2437 if (flags & SCF_DO_STCLASS_AND) { \
2438 for (value = 0; value < 256; value++) \
2439 if (!is_ ## nAmE ## _cp(value)) \
2440 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2443 for (value = 0; value < 256; value++) \
2444 if (is_ ## nAmE ## _cp(value)) \
2445 ANYOF_BITMAP_SET(data->start_class, value); \
2449 if (flags & SCF_DO_STCLASS_AND) { \
2450 for (value = 0; value < 256; value++) \
2451 if (is_ ## nAmE ## _cp(value)) \
2452 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2455 for (value = 0; value < 256; value++) \
2456 if (!is_ ## nAmE ## _cp(value)) \
2457 ANYOF_BITMAP_SET(data->start_class, value); \
2464 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2465 I32 *minlenp, I32 *deltap,
2470 struct regnode_charclass_class *and_withp,
2471 U32 flags, U32 depth)
2472 /* scanp: Start here (read-write). */
2473 /* deltap: Write maxlen-minlen here. */
2474 /* last: Stop before this one. */
2475 /* data: string data about the pattern */
2476 /* stopparen: treat close N as END */
2477 /* recursed: which subroutines have we recursed into */
2478 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2481 I32 min = 0, pars = 0, code;
2482 regnode *scan = *scanp, *next;
2484 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2485 int is_inf_internal = 0; /* The studied chunk is infinite */
2486 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2487 scan_data_t data_fake;
2488 SV *re_trie_maxbuff = NULL;
2489 regnode *first_non_open = scan;
2490 I32 stopmin = I32_MAX;
2491 scan_frame *frame = NULL;
2493 GET_RE_DEBUG_FLAGS_DECL;
2496 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2500 while (first_non_open && OP(first_non_open) == OPEN)
2501 first_non_open=regnext(first_non_open);
2506 while ( scan && OP(scan) != END && scan < last ){
2507 /* Peephole optimizer: */
2508 DEBUG_STUDYDATA("Peep:", data,depth);
2509 DEBUG_PEEP("Peep",scan,depth);
2510 JOIN_EXACT(scan,&min,0);
2512 /* Follow the next-chain of the current node and optimize
2513 away all the NOTHINGs from it. */
2514 if (OP(scan) != CURLYX) {
2515 const int max = (reg_off_by_arg[OP(scan)]
2517 /* I32 may be smaller than U16 on CRAYs! */
2518 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2519 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2523 /* Skip NOTHING and LONGJMP. */
2524 while ((n = regnext(n))
2525 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2526 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2527 && off + noff < max)
2529 if (reg_off_by_arg[OP(scan)])
2532 NEXT_OFF(scan) = off;
2537 /* The principal pseudo-switch. Cannot be a switch, since we
2538 look into several different things. */
2539 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2540 || OP(scan) == IFTHEN) {
2541 next = regnext(scan);
2543 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2545 if (OP(next) == code || code == IFTHEN) {
2546 /* NOTE - There is similar code to this block below for handling
2547 TRIE nodes on a re-study. If you change stuff here check there
2549 I32 max1 = 0, min1 = I32_MAX, num = 0;
2550 struct regnode_charclass_class accum;
2551 regnode * const startbranch=scan;
2553 if (flags & SCF_DO_SUBSTR)
2554 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2555 if (flags & SCF_DO_STCLASS)
2556 cl_init_zero(pRExC_state, &accum);
2558 while (OP(scan) == code) {
2559 I32 deltanext, minnext, f = 0, fake;
2560 struct regnode_charclass_class this_class;
2563 data_fake.flags = 0;
2565 data_fake.whilem_c = data->whilem_c;
2566 data_fake.last_closep = data->last_closep;
2569 data_fake.last_closep = &fake;
2571 data_fake.pos_delta = delta;
2572 next = regnext(scan);
2573 scan = NEXTOPER(scan);
2575 scan = NEXTOPER(scan);
2576 if (flags & SCF_DO_STCLASS) {
2577 cl_init(pRExC_state, &this_class);
2578 data_fake.start_class = &this_class;
2579 f = SCF_DO_STCLASS_AND;
2581 if (flags & SCF_WHILEM_VISITED_POS)
2582 f |= SCF_WHILEM_VISITED_POS;
2584 /* we suppose the run is continuous, last=next...*/
2585 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2587 stopparen, recursed, NULL, f,depth+1);
2590 if (max1 < minnext + deltanext)
2591 max1 = minnext + deltanext;
2592 if (deltanext == I32_MAX)
2593 is_inf = is_inf_internal = 1;
2595 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2597 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2598 if ( stopmin > minnext)
2599 stopmin = min + min1;
2600 flags &= ~SCF_DO_SUBSTR;
2602 data->flags |= SCF_SEEN_ACCEPT;
2605 if (data_fake.flags & SF_HAS_EVAL)
2606 data->flags |= SF_HAS_EVAL;
2607 data->whilem_c = data_fake.whilem_c;
2609 if (flags & SCF_DO_STCLASS)
2610 cl_or(pRExC_state, &accum, &this_class);
2612 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2614 if (flags & SCF_DO_SUBSTR) {
2615 data->pos_min += min1;
2616 data->pos_delta += max1 - min1;
2617 if (max1 != min1 || is_inf)
2618 data->longest = &(data->longest_float);
2621 delta += max1 - min1;
2622 if (flags & SCF_DO_STCLASS_OR) {
2623 cl_or(pRExC_state, data->start_class, &accum);
2625 cl_and(data->start_class, and_withp);
2626 flags &= ~SCF_DO_STCLASS;
2629 else if (flags & SCF_DO_STCLASS_AND) {
2631 cl_and(data->start_class, &accum);
2632 flags &= ~SCF_DO_STCLASS;
2635 /* Switch to OR mode: cache the old value of
2636 * data->start_class */
2638 StructCopy(data->start_class, and_withp,
2639 struct regnode_charclass_class);
2640 flags &= ~SCF_DO_STCLASS_AND;
2641 StructCopy(&accum, data->start_class,
2642 struct regnode_charclass_class);
2643 flags |= SCF_DO_STCLASS_OR;
2644 data->start_class->flags |= ANYOF_EOS;
2648 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2651 Assuming this was/is a branch we are dealing with: 'scan' now
2652 points at the item that follows the branch sequence, whatever
2653 it is. We now start at the beginning of the sequence and look
2660 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2662 If we can find such a subseqence we need to turn the first
2663 element into a trie and then add the subsequent branch exact
2664 strings to the trie.
2668 1. patterns where the whole set of branch can be converted.
2670 2. patterns where only a subset can be converted.
2672 In case 1 we can replace the whole set with a single regop
2673 for the trie. In case 2 we need to keep the start and end
2676 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2677 becomes BRANCH TRIE; BRANCH X;
2679 There is an additional case, that being where there is a
2680 common prefix, which gets split out into an EXACT like node
2681 preceding the TRIE node.
2683 If x(1..n)==tail then we can do a simple trie, if not we make
2684 a "jump" trie, such that when we match the appropriate word
2685 we "jump" to the appopriate tail node. Essentailly we turn
2686 a nested if into a case structure of sorts.
2691 if (!re_trie_maxbuff) {
2692 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2693 if (!SvIOK(re_trie_maxbuff))
2694 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2696 if ( SvIV(re_trie_maxbuff)>=0 ) {
2698 regnode *first = (regnode *)NULL;
2699 regnode *last = (regnode *)NULL;
2700 regnode *tail = scan;
2705 SV * const mysv = sv_newmortal(); /* for dumping */
2707 /* var tail is used because there may be a TAIL
2708 regop in the way. Ie, the exacts will point to the
2709 thing following the TAIL, but the last branch will
2710 point at the TAIL. So we advance tail. If we
2711 have nested (?:) we may have to move through several
2715 while ( OP( tail ) == TAIL ) {
2716 /* this is the TAIL generated by (?:) */
2717 tail = regnext( tail );
2722 regprop(RExC_rx, mysv, tail );
2723 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2724 (int)depth * 2 + 2, "",
2725 "Looking for TRIE'able sequences. Tail node is: ",
2726 SvPV_nolen_const( mysv )
2732 step through the branches, cur represents each
2733 branch, noper is the first thing to be matched
2734 as part of that branch and noper_next is the
2735 regnext() of that node. if noper is an EXACT
2736 and noper_next is the same as scan (our current
2737 position in the regex) then the EXACT branch is
2738 a possible optimization target. Once we have
2739 two or more consequetive such branches we can
2740 create a trie of the EXACT's contents and stich
2741 it in place. If the sequence represents all of
2742 the branches we eliminate the whole thing and
2743 replace it with a single TRIE. If it is a
2744 subsequence then we need to stitch it in. This
2745 means the first branch has to remain, and needs
2746 to be repointed at the item on the branch chain
2747 following the last branch optimized. This could
2748 be either a BRANCH, in which case the
2749 subsequence is internal, or it could be the
2750 item following the branch sequence in which
2751 case the subsequence is at the end.
2755 /* dont use tail as the end marker for this traverse */
2756 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2757 regnode * const noper = NEXTOPER( cur );
2758 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2759 regnode * const noper_next = regnext( noper );
2763 regprop(RExC_rx, mysv, cur);
2764 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2765 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2767 regprop(RExC_rx, mysv, noper);
2768 PerlIO_printf( Perl_debug_log, " -> %s",
2769 SvPV_nolen_const(mysv));
2772 regprop(RExC_rx, mysv, noper_next );
2773 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2774 SvPV_nolen_const(mysv));
2776 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2777 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2779 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2780 : PL_regkind[ OP( noper ) ] == EXACT )
2781 || OP(noper) == NOTHING )
2783 && noper_next == tail
2788 if ( !first || optype == NOTHING ) {
2789 if (!first) first = cur;
2790 optype = OP( noper );
2796 Currently we assume that the trie can handle unicode and ascii
2797 matches fold cased matches. If this proves true then the following
2798 define will prevent tries in this situation.
2800 #define TRIE_TYPE_IS_SAFE (UTF || optype==EXACT)
2802 #define TRIE_TYPE_IS_SAFE 1
2803 if ( last && TRIE_TYPE_IS_SAFE ) {
2804 make_trie( pRExC_state,
2805 startbranch, first, cur, tail, count,
2808 if ( PL_regkind[ OP( noper ) ] == EXACT
2810 && noper_next == tail
2815 optype = OP( noper );
2825 regprop(RExC_rx, mysv, cur);
2826 PerlIO_printf( Perl_debug_log,
2827 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2828 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2832 if ( last && TRIE_TYPE_IS_SAFE ) {
2833 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2834 #ifdef TRIE_STUDY_OPT
2835 if ( ((made == MADE_EXACT_TRIE &&
2836 startbranch == first)
2837 || ( first_non_open == first )) &&
2839 flags |= SCF_TRIE_RESTUDY;
2840 if ( startbranch == first
2843 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2853 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2854 scan = NEXTOPER(NEXTOPER(scan));
2855 } else /* single branch is optimized. */
2856 scan = NEXTOPER(scan);
2858 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2859 scan_frame *newframe = NULL;
2864 if (OP(scan) != SUSPEND) {
2865 /* set the pointer */
2866 if (OP(scan) == GOSUB) {
2868 RExC_recurse[ARG2L(scan)] = scan;
2869 start = RExC_open_parens[paren-1];
2870 end = RExC_close_parens[paren-1];
2873 start = RExC_rxi->program + 1;
2877 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
2878 SAVEFREEPV(recursed);
2880 if (!PAREN_TEST(recursed,paren+1)) {
2881 PAREN_SET(recursed,paren+1);
2882 Newx(newframe,1,scan_frame);
2884 if (flags & SCF_DO_SUBSTR) {
2885 SCAN_COMMIT(pRExC_state,data,minlenp);
2886 data->longest = &(data->longest_float);
2888 is_inf = is_inf_internal = 1;
2889 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
2890 cl_anything(pRExC_state, data->start_class);
2891 flags &= ~SCF_DO_STCLASS;
2894 Newx(newframe,1,scan_frame);
2897 end = regnext(scan);
2902 SAVEFREEPV(newframe);
2903 newframe->next = regnext(scan);
2904 newframe->last = last;
2905 newframe->stop = stopparen;
2906 newframe->prev = frame;
2916 else if (OP(scan) == EXACT) {
2917 I32 l = STR_LEN(scan);
2920 const U8 * const s = (U8*)STRING(scan);
2921 l = utf8_length(s, s + l);
2922 uc = utf8_to_uvchr(s, NULL);
2924 uc = *((U8*)STRING(scan));
2927 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2928 /* The code below prefers earlier match for fixed
2929 offset, later match for variable offset. */
2930 if (data->last_end == -1) { /* Update the start info. */
2931 data->last_start_min = data->pos_min;
2932 data->last_start_max = is_inf
2933 ? I32_MAX : data->pos_min + data->pos_delta;
2935 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2937 SvUTF8_on(data->last_found);
2939 SV * const sv = data->last_found;
2940 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2941 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2942 if (mg && mg->mg_len >= 0)
2943 mg->mg_len += utf8_length((U8*)STRING(scan),
2944 (U8*)STRING(scan)+STR_LEN(scan));
2946 data->last_end = data->pos_min + l;
2947 data->pos_min += l; /* As in the first entry. */
2948 data->flags &= ~SF_BEFORE_EOL;
2950 if (flags & SCF_DO_STCLASS_AND) {
2951 /* Check whether it is compatible with what we know already! */
2955 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2956 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2957 && (!(data->start_class->flags & ANYOF_FOLD)
2958 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2961 ANYOF_CLASS_ZERO(data->start_class);
2962 ANYOF_BITMAP_ZERO(data->start_class);
2964 ANYOF_BITMAP_SET(data->start_class, uc);
2965 data->start_class->flags &= ~ANYOF_EOS;
2967 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2969 else if (flags & SCF_DO_STCLASS_OR) {
2970 /* false positive possible if the class is case-folded */
2972 ANYOF_BITMAP_SET(data->start_class, uc);
2974 data->start_class->flags |= ANYOF_UNICODE_ALL;
2975 data->start_class->flags &= ~ANYOF_EOS;
2976 cl_and(data->start_class, and_withp);
2978 flags &= ~SCF_DO_STCLASS;
2980 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
2981 I32 l = STR_LEN(scan);
2982 UV uc = *((U8*)STRING(scan));
2984 /* Search for fixed substrings supports EXACT only. */
2985 if (flags & SCF_DO_SUBSTR) {
2987 SCAN_COMMIT(pRExC_state, data, minlenp);
2990 const U8 * const s = (U8 *)STRING(scan);
2991 l = utf8_length(s, s + l);
2992 uc = utf8_to_uvchr(s, NULL);
2995 if (flags & SCF_DO_SUBSTR)
2997 if (flags & SCF_DO_STCLASS_AND) {
2998 /* Check whether it is compatible with what we know already! */
3002 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3003 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3004 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
3006 ANYOF_CLASS_ZERO(data->start_class);
3007 ANYOF_BITMAP_ZERO(data->start_class);
3009 ANYOF_BITMAP_SET(data->start_class, uc);
3010 data->start_class->flags &= ~ANYOF_EOS;
3011 data->start_class->flags |= ANYOF_FOLD;
3012 if (OP(scan) == EXACTFL)
3013 data->start_class->flags |= ANYOF_LOCALE;
3016 else if (flags & SCF_DO_STCLASS_OR) {
3017 if (data->start_class->flags & ANYOF_FOLD) {
3018 /* false positive possible if the class is case-folded.
3019 Assume that the locale settings are the same... */
3021 ANYOF_BITMAP_SET(data->start_class, uc);
3022 data->start_class->flags &= ~ANYOF_EOS;
3024 cl_and(data->start_class, and_withp);
3026 flags &= ~SCF_DO_STCLASS;
3028 else if (strchr((const char*)PL_varies,OP(scan))) {
3029 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3030 I32 f = flags, pos_before = 0;
3031 regnode * const oscan = scan;
3032 struct regnode_charclass_class this_class;
3033 struct regnode_charclass_class *oclass = NULL;
3034 I32 next_is_eval = 0;
3036 switch (PL_regkind[OP(scan)]) {
3037 case WHILEM: /* End of (?:...)* . */
3038 scan = NEXTOPER(scan);
3041 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3042 next = NEXTOPER(scan);
3043 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3045 maxcount = REG_INFTY;
3046 next = regnext(scan);
3047 scan = NEXTOPER(scan);
3051 if (flags & SCF_DO_SUBSTR)
3056 if (flags & SCF_DO_STCLASS) {
3058 maxcount = REG_INFTY;
3059 next = regnext(scan);
3060 scan = NEXTOPER(scan);
3063 is_inf = is_inf_internal = 1;
3064 scan = regnext(scan);
3065 if (flags & SCF_DO_SUBSTR) {
3066 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3067 data->longest = &(data->longest_float);
3069 goto optimize_curly_tail;
3071 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3072 && (scan->flags == stopparen))
3077 mincount = ARG1(scan);
3078 maxcount = ARG2(scan);
3080 next = regnext(scan);
3081 if (OP(scan) == CURLYX) {
3082 I32 lp = (data ? *(data->last_closep) : 0);
3083 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3085 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3086 next_is_eval = (OP(scan) == EVAL);
3088 if (flags & SCF_DO_SUBSTR) {
3089 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3090 pos_before = data->pos_min;
3094 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3096 data->flags |= SF_IS_INF;
3098 if (flags & SCF_DO_STCLASS) {
3099 cl_init(pRExC_state, &this_class);
3100 oclass = data->start_class;
3101 data->start_class = &this_class;
3102 f |= SCF_DO_STCLASS_AND;
3103 f &= ~SCF_DO_STCLASS_OR;
3105 /* These are the cases when once a subexpression
3106 fails at a particular position, it cannot succeed
3107 even after backtracking at the enclosing scope.
3109 XXXX what if minimal match and we are at the
3110 initial run of {n,m}? */
3111 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
3112 f &= ~SCF_WHILEM_VISITED_POS;
3114 /* This will finish on WHILEM, setting scan, or on NULL: */
3115 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3116 last, data, stopparen, recursed, NULL,
3118 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3120 if (flags & SCF_DO_STCLASS)
3121 data->start_class = oclass;
3122 if (mincount == 0 || minnext == 0) {
3123 if (flags & SCF_DO_STCLASS_OR) {
3124 cl_or(pRExC_state, data->start_class, &this_class);
3126 else if (flags & SCF_DO_STCLASS_AND) {
3127 /* Switch to OR mode: cache the old value of
3128 * data->start_class */
3130 StructCopy(data->start_class, and_withp,
3131 struct regnode_charclass_class);
3132 flags &= ~SCF_DO_STCLASS_AND;
3133 StructCopy(&this_class, data->start_class,
3134 struct regnode_charclass_class);
3135 flags |= SCF_DO_STCLASS_OR;
3136 data->start_class->flags |= ANYOF_EOS;
3138 } else { /* Non-zero len */
3139 if (flags & SCF_DO_STCLASS_OR) {
3140 cl_or(pRExC_state, data->start_class, &this_class);
3141 cl_and(data->start_class, and_withp);
3143 else if (flags & SCF_DO_STCLASS_AND)
3144 cl_and(data->start_class, &this_class);
3145 flags &= ~SCF_DO_STCLASS;
3147 if (!scan) /* It was not CURLYX, but CURLY. */
3149 if ( /* ? quantifier ok, except for (?{ ... }) */
3150 (next_is_eval || !(mincount == 0 && maxcount == 1))
3151 && (minnext == 0) && (deltanext == 0)
3152 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3153 && maxcount <= REG_INFTY/3 /* Complement check for big count */
3154 && ckWARN(WARN_REGEXP))
3157 "Quantifier unexpected on zero-length expression");
3160 min += minnext * mincount;
3161 is_inf_internal |= ((maxcount == REG_INFTY
3162 && (minnext + deltanext) > 0)
3163 || deltanext == I32_MAX);
3164 is_inf |= is_inf_internal;
3165 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3167 /* Try powerful optimization CURLYX => CURLYN. */
3168 if ( OP(oscan) == CURLYX && data
3169 && data->flags & SF_IN_PAR
3170 && !(data->flags & SF_HAS_EVAL)
3171 && !deltanext && minnext == 1 ) {
3172 /* Try to optimize to CURLYN. */
3173 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3174 regnode * const nxt1 = nxt;
3181 if (!strchr((const char*)PL_simple,OP(nxt))
3182 && !(PL_regkind[OP(nxt)] == EXACT
3183 && STR_LEN(nxt) == 1))
3189 if (OP(nxt) != CLOSE)
3191 if (RExC_open_parens) {
3192 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3193 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3195 /* Now we know that nxt2 is the only contents: */
3196 oscan->flags = (U8)ARG(nxt);
3198 OP(nxt1) = NOTHING; /* was OPEN. */
3201 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3202 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
3203 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
3204 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3205 OP(nxt + 1) = OPTIMIZED; /* was count. */
3206 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
3211 /* Try optimization CURLYX => CURLYM. */
3212 if ( OP(oscan) == CURLYX && data
3213 && !(data->flags & SF_HAS_PAR)
3214 && !(data->flags & SF_HAS_EVAL)
3215 && !deltanext /* atom is fixed width */
3216 && minnext != 0 /* CURLYM can't handle zero width */
3218 /* XXXX How to optimize if data == 0? */
3219 /* Optimize to a simpler form. */
3220 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3224 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3225 && (OP(nxt2) != WHILEM))
3227 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3228 /* Need to optimize away parenths. */
3229 if (data->flags & SF_IN_PAR) {
3230 /* Set the parenth number. */
3231 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3233 if (OP(nxt) != CLOSE)
3234 FAIL("Panic opt close");
3235 oscan->flags = (U8)ARG(nxt);
3236 if (RExC_open_parens) {
3237 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3238 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3240 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3241 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3244 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3245 OP(nxt + 1) = OPTIMIZED; /* was count. */
3246 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3247 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3250 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3251 regnode *nnxt = regnext(nxt1);
3254 if (reg_off_by_arg[OP(nxt1)])
3255 ARG_SET(nxt1, nxt2 - nxt1);
3256 else if (nxt2 - nxt1 < U16_MAX)
3257 NEXT_OFF(nxt1) = nxt2 - nxt1;
3259 OP(nxt) = NOTHING; /* Cannot beautify */
3264 /* Optimize again: */
3265 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3266 NULL, stopparen, recursed, NULL, 0,depth+1);
3271 else if ((OP(oscan) == CURLYX)
3272 && (flags & SCF_WHILEM_VISITED_POS)
3273 /* See the comment on a similar expression above.
3274 However, this time it not a subexpression
3275 we care about, but the expression itself. */
3276 && (maxcount == REG_INFTY)
3277 && data && ++data->whilem_c < 16) {
3278 /* This stays as CURLYX, we can put the count/of pair. */
3279 /* Find WHILEM (as in regexec.c) */
3280 regnode *nxt = oscan + NEXT_OFF(oscan);
3282 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3284 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3285 | (RExC_whilem_seen << 4)); /* On WHILEM */
3287 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3289 if (flags & SCF_DO_SUBSTR) {
3290 SV *last_str = NULL;
3291 int counted = mincount != 0;
3293 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3294 #if defined(SPARC64_GCC_WORKAROUND)
3297 const char *s = NULL;
3300 if (pos_before >= data->last_start_min)
3303 b = data->last_start_min;
3306 s = SvPV_const(data->last_found, l);
3307 old = b - data->last_start_min;
3310 I32 b = pos_before >= data->last_start_min
3311 ? pos_before : data->last_start_min;
3313 const char * const s = SvPV_const(data->last_found, l);
3314 I32 old = b - data->last_start_min;
3318 old = utf8_hop((U8*)s, old) - (U8*)s;
3321 /* Get the added string: */
3322 last_str = newSVpvn_utf8(s + old, l, UTF);
3323 if (deltanext == 0 && pos_before == b) {
3324 /* What was added is a constant string */
3326 SvGROW(last_str, (mincount * l) + 1);
3327 repeatcpy(SvPVX(last_str) + l,
3328 SvPVX_const(last_str), l, mincount - 1);
3329 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3330 /* Add additional parts. */
3331 SvCUR_set(data->last_found,
3332 SvCUR(data->last_found) - l);
3333 sv_catsv(data->last_found, last_str);
3335 SV * sv = data->last_found;
3337 SvUTF8(sv) && SvMAGICAL(sv) ?
3338 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3339 if (mg && mg->mg_len >= 0)
3340 mg->mg_len += CHR_SVLEN(last_str) - l;
3342 data->last_end += l * (mincount - 1);
3345 /* start offset must point into the last copy */
3346 data->last_start_min += minnext * (mincount - 1);
3347 data->last_start_max += is_inf ? I32_MAX
3348 : (maxcount - 1) * (minnext + data->pos_delta);
3351 /* It is counted once already... */
3352 data->pos_min += minnext * (mincount - counted);
3353 data->pos_delta += - counted * deltanext +
3354 (minnext + deltanext) * maxcount - minnext * mincount;
3355 if (mincount != maxcount) {
3356 /* Cannot extend fixed substrings found inside
3358 SCAN_COMMIT(pRExC_state,data,minlenp);
3359 if (mincount && last_str) {
3360 SV * const sv = data->last_found;
3361 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3362 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3366 sv_setsv(sv, last_str);
3367 data->last_end = data->pos_min;
3368 data->last_start_min =
3369 data->pos_min - CHR_SVLEN(last_str);
3370 data->last_start_max = is_inf
3372 : data->pos_min + data->pos_delta
3373 - CHR_SVLEN(last_str);
3375 data->longest = &(data->longest_float);
3377 SvREFCNT_dec(last_str);
3379 if (data && (fl & SF_HAS_EVAL))
3380 data->flags |= SF_HAS_EVAL;
3381 optimize_curly_tail:
3382 if (OP(oscan) != CURLYX) {
3383 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3385 NEXT_OFF(oscan) += NEXT_OFF(next);
3388 default: /* REF and CLUMP only? */
3389 if (flags & SCF_DO_SUBSTR) {
3390 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3391 data->longest = &(data->longest_float);
3393 is_inf = is_inf_internal = 1;
3394 if (flags & SCF_DO_STCLASS_OR)
3395 cl_anything(pRExC_state, data->start_class);
3396 flags &= ~SCF_DO_STCLASS;
3400 else if (OP(scan) == LNBREAK) {
3401 if (flags & SCF_DO_STCLASS) {
3403 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3404 if (flags & SCF_DO_STCLASS_AND) {
3405 for (value = 0; value < 256; value++)
3406 if (!is_VERTWS_cp(value))
3407 ANYOF_BITMAP_CLEAR(data->start_class, value);
3410 for (value = 0; value < 256; value++)
3411 if (is_VERTWS_cp(value))
3412 ANYOF_BITMAP_SET(data->start_class, value);
3414 if (flags & SCF_DO_STCLASS_OR)
3415 cl_and(data->start_class, and_withp);
3416 flags &= ~SCF_DO_STCLASS;
3420 if (flags & SCF_DO_SUBSTR) {
3421 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3423 data->pos_delta += 1;
3424 data->longest = &(data->longest_float);
3428 else if (OP(scan) == FOLDCHAR) {
3429 int d = ARG(scan)==0xDF ? 1 : 2;
3430 flags &= ~SCF_DO_STCLASS;
3433 if (flags & SCF_DO_SUBSTR) {
3434 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3436 data->pos_delta += d;
3437 data->longest = &(data->longest_float);
3440 else if (strchr((const char*)PL_simple,OP(scan))) {
3443 if (flags & SCF_DO_SUBSTR) {
3444 SCAN_COMMIT(pRExC_state,data,minlenp);
3448 if (flags & SCF_DO_STCLASS) {
3449 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3451 /* Some of the logic below assumes that switching
3452 locale on will only add false positives. */
3453 switch (PL_regkind[OP(scan)]) {
3457 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3458 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3459 cl_anything(pRExC_state, data->start_class);
3462 if (OP(scan) == SANY)
3464 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3465 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3466 || (data->start_class->flags & ANYOF_CLASS));
3467 cl_anything(pRExC_state, data->start_class);
3469 if (flags & SCF_DO_STCLASS_AND || !value)
3470 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3473 if (flags & SCF_DO_STCLASS_AND)
3474 cl_and(data->start_class,
3475 (struct regnode_charclass_class*)scan);
3477 cl_or(pRExC_state, data->start_class,
3478 (struct regnode_charclass_class*)scan);
3481 if (flags & SCF_DO_STCLASS_AND) {
3482 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3483 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3484 for (value = 0; value < 256; value++)
3485 if (!isALNUM(value))
3486 ANYOF_BITMAP_CLEAR(data->start_class, value);
3490 if (data->start_class->flags & ANYOF_LOCALE)
3491 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3493 for (value = 0; value < 256; value++)
3495 ANYOF_BITMAP_SET(data->start_class, value);
3500 if (flags & SCF_DO_STCLASS_AND) {
3501 if (data->start_class->flags & ANYOF_LOCALE)
3502 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3505 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3506 data->start_class->flags |= ANYOF_LOCALE;
3510 if (flags & SCF_DO_STCLASS_AND) {
3511 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3512 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3513 for (value = 0; value < 256; value++)
3515 ANYOF_BITMAP_CLEAR(data->start_class, value);
3519 if (data->start_class->flags & ANYOF_LOCALE)
3520 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3522 for (value = 0; value < 256; value++)
3523 if (!isALNUM(value))
3524 ANYOF_BITMAP_SET(data->start_class, value);
3529 if (flags & SCF_DO_STCLASS_AND) {
3530 if (data->start_class->flags & ANYOF_LOCALE)
3531 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3534 data->start_class->flags |= ANYOF_LOCALE;
3535 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3539 if (flags & SCF_DO_STCLASS_AND) {
3540 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3541 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3542 for (value = 0; value < 256; value++)
3543 if (!isSPACE(value))
3544 ANYOF_BITMAP_CLEAR(data->start_class, value);
3548 if (data->start_class->flags & ANYOF_LOCALE)
3549 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3551 for (value = 0; value < 256; value++)
3553 ANYOF_BITMAP_SET(data->start_class, value);
3558 if (flags & SCF_DO_STCLASS_AND) {
3559 if (data->start_class->flags & ANYOF_LOCALE)
3560 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3563 data->start_class->flags |= ANYOF_LOCALE;
3564 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3568 if (flags & SCF_DO_STCLASS_AND) {
3569 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3570 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3571 for (value = 0; value < 256; value++)
3573 ANYOF_BITMAP_CLEAR(data->start_class, value);
3577 if (data->start_class->flags & ANYOF_LOCALE)
3578 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3580 for (value = 0; value < 256; value++)
3581 if (!isSPACE(value))
3582 ANYOF_BITMAP_SET(data->start_class, value);
3587 if (flags & SCF_DO_STCLASS_AND) {
3588 if (data->start_class->flags & ANYOF_LOCALE) {
3589 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3590 for (value = 0; value < 256; value++)
3591 if (!isSPACE(value))
3592 ANYOF_BITMAP_CLEAR(data->start_class, value);
3596 data->start_class->flags |= ANYOF_LOCALE;
3597 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3601 if (flags & SCF_DO_STCLASS_AND) {
3602 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3603 for (value = 0; value < 256; value++)
3604 if (!isDIGIT(value))
3605 ANYOF_BITMAP_CLEAR(data->start_class, value);
3608 if (data->start_class->flags & ANYOF_LOCALE)
3609 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3611 for (value = 0; value < 256; value++)
3613 ANYOF_BITMAP_SET(data->start_class, value);
3618 if (flags & SCF_DO_STCLASS_AND) {
3619 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3620 for (value = 0; value < 256; value++)
3622 ANYOF_BITMAP_CLEAR(data->start_class, value);
3625 if (data->start_class->flags & ANYOF_LOCALE)
3626 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3628 for (value = 0; value < 256; value++)
3629 if (!isDIGIT(value))
3630 ANYOF_BITMAP_SET(data->start_class, value);
3634 CASE_SYNST_FNC(VERTWS);
3635 CASE_SYNST_FNC(HORIZWS);
3638 if (flags & SCF_DO_STCLASS_OR)
3639 cl_and(data->start_class, and_withp);
3640 flags &= ~SCF_DO_STCLASS;
3643 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3644 data->flags |= (OP(scan) == MEOL
3648 else if ( PL_regkind[OP(scan)] == BRANCHJ
3649 /* Lookbehind, or need to calculate parens/evals/stclass: */
3650 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3651 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3652 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3653 || OP(scan) == UNLESSM )
3655 /* Negative Lookahead/lookbehind
3656 In this case we can't do fixed string optimisation.
3659 I32 deltanext, minnext, fake = 0;
3661 struct regnode_charclass_class intrnl;
3664 data_fake.flags = 0;
3666 data_fake.whilem_c = data->whilem_c;
3667 data_fake.last_closep = data->last_closep;
3670 data_fake.last_closep = &fake;
3671 data_fake.pos_delta = delta;
3672 if ( flags & SCF_DO_STCLASS && !scan->flags
3673 && OP(scan) == IFMATCH ) { /* Lookahead */
3674 cl_init(pRExC_state, &intrnl);
3675 data_fake.start_class = &intrnl;
3676 f |= SCF_DO_STCLASS_AND;
3678 if (flags & SCF_WHILEM_VISITED_POS)
3679 f |= SCF_WHILEM_VISITED_POS;
3680 next = regnext(scan);
3681 nscan = NEXTOPER(NEXTOPER(scan));
3682 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
3683 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
3686 FAIL("Variable length lookbehind not implemented");
3688 else if (minnext > (I32)U8_MAX) {
3689 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3691 scan->flags = (U8)minnext;
3694 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3696 if (data_fake.flags & SF_HAS_EVAL)
3697 data->flags |= SF_HAS_EVAL;
3698 data->whilem_c = data_fake.whilem_c;
3700 if (f & SCF_DO_STCLASS_AND) {
3701 const int was = (data->start_class->flags & ANYOF_EOS);
3703 cl_and(data->start_class, &intrnl);
3705 data->start_class->flags |= ANYOF_EOS;
3708 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3710 /* Positive Lookahead/lookbehind
3711 In this case we can do fixed string optimisation,
3712 but we must be careful about it. Note in the case of
3713 lookbehind the positions will be offset by the minimum
3714 length of the pattern, something we won't know about
3715 until after the recurse.
3717 I32 deltanext, fake = 0;
3719 struct regnode_charclass_class intrnl;
3721 /* We use SAVEFREEPV so that when the full compile
3722 is finished perl will clean up the allocated
3723 minlens when its all done. This was we don't
3724 have to worry about freeing them when we know
3725 they wont be used, which would be a pain.
3728 Newx( minnextp, 1, I32 );
3729 SAVEFREEPV(minnextp);
3732 StructCopy(data, &data_fake, scan_data_t);
3733 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3736 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
3737 data_fake.last_found=newSVsv(data->last_found);
3741 data_fake.last_closep = &fake;
3742 data_fake.flags = 0;
3743 data_fake.pos_delta = delta;
3745 data_fake.flags |= SF_IS_INF;
3746 if ( flags & SCF_DO_STCLASS && !scan->flags
3747 && OP(scan) == IFMATCH ) { /* Lookahead */
3748 cl_init(pRExC_state, &intrnl);
3749 data_fake.start_class = &intrnl;
3750 f |= SCF_DO_STCLASS_AND;
3752 if (flags & SCF_WHILEM_VISITED_POS)
3753 f |= SCF_WHILEM_VISITED_POS;
3754 next = regnext(scan);
3755 nscan = NEXTOPER(NEXTOPER(scan));
3757 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
3758 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
3761 FAIL("Variable length lookbehind not implemented");
3763 else if (*minnextp > (I32)U8_MAX) {
3764 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3766 scan->flags = (U8)*minnextp;
3771 if (f & SCF_DO_STCLASS_AND) {
3772 const int was = (data->start_class->flags & ANYOF_EOS);
3774 cl_and(data->start_class, &intrnl);
3776 data->start_class->flags |= ANYOF_EOS;
3779 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3781 if (data_fake.flags & SF_HAS_EVAL)
3782 data->flags |= SF_HAS_EVAL;
3783 data->whilem_c = data_fake.whilem_c;
3784 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3785 if (RExC_rx->minlen<*minnextp)
3786 RExC_rx->minlen=*minnextp;
3787 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
3788 SvREFCNT_dec(data_fake.last_found);
3790 if ( data_fake.minlen_fixed != minlenp )
3792 data->offset_fixed= data_fake.offset_fixed;
3793 data->minlen_fixed= data_fake.minlen_fixed;
3794 data->lookbehind_fixed+= scan->flags;
3796 if ( data_fake.minlen_float != minlenp )
3798 data->minlen_float= data_fake.minlen_float;
3799 data->offset_float_min=data_fake.offset_float_min;
3800 data->offset_float_max=data_fake.offset_float_max;
3801 data->lookbehind_float+= scan->flags;
3810 else if (OP(scan) == OPEN) {
3811 if (stopparen != (I32)ARG(scan))
https://perl5.git.perl.org / perl5.git / blob