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; /* perl core regexp structure */
106 regexp_internal *rxi; /* internal data for regexp object pprivate field */
107 char *start; /* Start of input for compile */
108 char *end; /* End of input for compile */
109 char *parse; /* Input-scan pointer. */
110 I32 whilem_seen; /* number of WHILEM in this expr */
111 regnode *emit_start; /* Start of emitted-code area */
112 regnode *emit_bound; /* First regnode outside of the allocated space */
113 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
114 I32 naughty; /* How bad is this pattern? */
115 I32 sawback; /* Did we see \1, ...? */
117 I32 size; /* Code size. */
118 I32 npar; /* Capture buffer count, (OPEN). */
119 I32 cpar; /* Capture buffer count, (CLOSE). */
120 I32 nestroot; /* root parens we are in - used by accept */
124 regnode **open_parens; /* pointers to open parens */
125 regnode **close_parens; /* pointers to close parens */
126 regnode *opend; /* END node in program */
127 I32 utf8; /* whether the pattern is utf8 or not */
128 I32 orig_utf8; /* whether the pattern was originally in utf8 */
129 /* XXX use this for future optimisation of case
130 * where pattern must be upgraded to utf8. */
131 HV *charnames; /* cache of named sequences */
132 HV *paren_names; /* Paren names */
134 regnode **recurse; /* Recurse regops */
135 I32 recurse_count; /* Number of recurse regops */
137 char *starttry; /* -Dr: where regtry was called. */
138 #define RExC_starttry (pRExC_state->starttry)
141 const char *lastparse;
143 AV *paren_name_list; /* idx -> name */
144 #define RExC_lastparse (pRExC_state->lastparse)
145 #define RExC_lastnum (pRExC_state->lastnum)
146 #define RExC_paren_name_list (pRExC_state->paren_name_list)
150 #define RExC_flags (pRExC_state->flags)
151 #define RExC_precomp (pRExC_state->precomp)
152 #define RExC_rx (pRExC_state->rx)
153 #define RExC_rxi (pRExC_state->rxi)
154 #define RExC_start (pRExC_state->start)
155 #define RExC_end (pRExC_state->end)
156 #define RExC_parse (pRExC_state->parse)
157 #define RExC_whilem_seen (pRExC_state->whilem_seen)
158 #ifdef RE_TRACK_PATTERN_OFFSETS
159 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
161 #define RExC_emit (pRExC_state->emit)
162 #define RExC_emit_start (pRExC_state->emit_start)
163 #define RExC_emit_bound (pRExC_state->emit_bound)
164 #define RExC_naughty (pRExC_state->naughty)
165 #define RExC_sawback (pRExC_state->sawback)
166 #define RExC_seen (pRExC_state->seen)
167 #define RExC_size (pRExC_state->size)
168 #define RExC_npar (pRExC_state->npar)
169 #define RExC_nestroot (pRExC_state->nestroot)
170 #define RExC_extralen (pRExC_state->extralen)
171 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
172 #define RExC_seen_evals (pRExC_state->seen_evals)
173 #define RExC_utf8 (pRExC_state->utf8)
174 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
175 #define RExC_charnames (pRExC_state->charnames)
176 #define RExC_open_parens (pRExC_state->open_parens)
177 #define RExC_close_parens (pRExC_state->close_parens)
178 #define RExC_opend (pRExC_state->opend)
179 #define RExC_paren_names (pRExC_state->paren_names)
180 #define RExC_recurse (pRExC_state->recurse)
181 #define RExC_recurse_count (pRExC_state->recurse_count)
184 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
185 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
186 ((*s) == '{' && regcurly(s)))
189 #undef SPSTART /* dratted cpp namespace... */
192 * Flags to be passed up and down.
194 #define WORST 0 /* Worst case. */
195 #define HASWIDTH 0x01 /* Known to match non-null strings. */
196 #define SIMPLE 0x02 /* Simple enough to be STAR/PLUS operand. */
197 #define SPSTART 0x04 /* Starts with * or +. */
198 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
199 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
201 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
203 /* whether trie related optimizations are enabled */
204 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
205 #define TRIE_STUDY_OPT
206 #define FULL_TRIE_STUDY
212 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
213 #define PBITVAL(paren) (1 << ((paren) & 7))
214 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
215 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
216 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
219 /* About scan_data_t.
221 During optimisation we recurse through the regexp program performing
222 various inplace (keyhole style) optimisations. In addition study_chunk
223 and scan_commit populate this data structure with information about
224 what strings MUST appear in the pattern. We look for the longest
225 string that must appear for at a fixed location, and we look for the
226 longest string that may appear at a floating location. So for instance
231 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
232 strings (because they follow a .* construct). study_chunk will identify
233 both FOO and BAR as being the longest fixed and floating strings respectively.
235 The strings can be composites, for instance
239 will result in a composite fixed substring 'foo'.
241 For each string some basic information is maintained:
243 - offset or min_offset
244 This is the position the string must appear at, or not before.
245 It also implicitly (when combined with minlenp) tells us how many
246 character must match before the string we are searching.
247 Likewise when combined with minlenp and the length of the string
248 tells us how many characters must appear after the string we have
252 Only used for floating strings. This is the rightmost point that
253 the string can appear at. Ifset to I32 max it indicates that the
254 string can occur infinitely far to the right.
257 A pointer to the minimum length of the pattern that the string
258 was found inside. This is important as in the case of positive
259 lookahead or positive lookbehind we can have multiple patterns
264 The minimum length of the pattern overall is 3, the minimum length
265 of the lookahead part is 3, but the minimum length of the part that
266 will actually match is 1. So 'FOO's minimum length is 3, but the
267 minimum length for the F is 1. This is important as the minimum length
268 is used to determine offsets in front of and behind the string being
269 looked for. Since strings can be composites this is the length of the
270 pattern at the time it was commited with a scan_commit. Note that
271 the length is calculated by study_chunk, so that the minimum lengths
272 are not known until the full pattern has been compiled, thus the
273 pointer to the value.
277 In the case of lookbehind the string being searched for can be
278 offset past the start point of the final matching string.
279 If this value was just blithely removed from the min_offset it would
280 invalidate some of the calculations for how many chars must match
281 before or after (as they are derived from min_offset and minlen and
282 the length of the string being searched for).
283 When the final pattern is compiled and the data is moved from the
284 scan_data_t structure into the regexp structure the information
285 about lookbehind is factored in, with the information that would
286 have been lost precalculated in the end_shift field for the
289 The fields pos_min and pos_delta are used to store the minimum offset
290 and the delta to the maximum offset at the current point in the pattern.
294 typedef struct scan_data_t {
295 /*I32 len_min; unused */
296 /*I32 len_delta; unused */
300 I32 last_end; /* min value, <0 unless valid. */
303 SV **longest; /* Either &l_fixed, or &l_float. */
304 SV *longest_fixed; /* longest fixed string found in pattern */
305 I32 offset_fixed; /* offset where it starts */
306 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
307 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
308 SV *longest_float; /* longest floating string found in pattern */
309 I32 offset_float_min; /* earliest point in string it can appear */
310 I32 offset_float_max; /* latest point in string it can appear */
311 I32 *minlen_float; /* pointer to the minlen relevent to the string */
312 I32 lookbehind_float; /* is the position of the string modified by LB */
316 struct regnode_charclass_class *start_class;
320 * Forward declarations for pregcomp()'s friends.
323 static const scan_data_t zero_scan_data =
324 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
326 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
327 #define SF_BEFORE_SEOL 0x0001
328 #define SF_BEFORE_MEOL 0x0002
329 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
330 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
333 # define SF_FIX_SHIFT_EOL (0+2)
334 # define SF_FL_SHIFT_EOL (0+4)
336 # define SF_FIX_SHIFT_EOL (+2)
337 # define SF_FL_SHIFT_EOL (+4)
340 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
341 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
343 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
344 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
345 #define SF_IS_INF 0x0040
346 #define SF_HAS_PAR 0x0080
347 #define SF_IN_PAR 0x0100
348 #define SF_HAS_EVAL 0x0200
349 #define SCF_DO_SUBSTR 0x0400
350 #define SCF_DO_STCLASS_AND 0x0800
351 #define SCF_DO_STCLASS_OR 0x1000
352 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
353 #define SCF_WHILEM_VISITED_POS 0x2000
355 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
356 #define SCF_SEEN_ACCEPT 0x8000
358 #define UTF (RExC_utf8 != 0)
359 #define LOC ((RExC_flags & RXf_PMf_LOCALE) != 0)
360 #define FOLD ((RExC_flags & RXf_PMf_FOLD) != 0)
362 #define OOB_UNICODE 12345678
363 #define OOB_NAMEDCLASS -1
365 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
366 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
369 /* length of regex to show in messages that don't mark a position within */
370 #define RegexLengthToShowInErrorMessages 127
373 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
374 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
375 * op/pragma/warn/regcomp.
377 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
378 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
380 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
383 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
384 * arg. Show regex, up to a maximum length. If it's too long, chop and add
387 #define _FAIL(code) STMT_START { \
388 const char *ellipses = ""; \
389 IV len = RExC_end - RExC_precomp; \
392 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
393 if (len > RegexLengthToShowInErrorMessages) { \
394 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
395 len = RegexLengthToShowInErrorMessages - 10; \
401 #define FAIL(msg) _FAIL( \
402 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
403 msg, (int)len, RExC_precomp, ellipses))
405 #define FAIL2(msg,arg) _FAIL( \
406 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
407 arg, (int)len, RExC_precomp, ellipses))
410 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
412 #define Simple_vFAIL(m) STMT_START { \
413 const IV offset = RExC_parse - RExC_precomp; \
414 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
415 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
419 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
421 #define vFAIL(m) STMT_START { \
423 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
428 * Like Simple_vFAIL(), but accepts two arguments.
430 #define Simple_vFAIL2(m,a1) STMT_START { \
431 const IV offset = RExC_parse - RExC_precomp; \
432 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
433 (int)offset, RExC_precomp, RExC_precomp + offset); \
437 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
439 #define vFAIL2(m,a1) STMT_START { \
441 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
442 Simple_vFAIL2(m, a1); \
447 * Like Simple_vFAIL(), but accepts three arguments.
449 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
450 const IV offset = RExC_parse - RExC_precomp; \
451 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
452 (int)offset, RExC_precomp, RExC_precomp + offset); \
456 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
458 #define vFAIL3(m,a1,a2) STMT_START { \
460 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
461 Simple_vFAIL3(m, a1, a2); \
465 * Like Simple_vFAIL(), but accepts four arguments.
467 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
468 const IV offset = RExC_parse - RExC_precomp; \
469 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
470 (int)offset, RExC_precomp, RExC_precomp + offset); \
473 #define vWARN(loc,m) STMT_START { \
474 const IV offset = loc - RExC_precomp; \
475 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
476 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
479 #define vWARNdep(loc,m) STMT_START { \
480 const IV offset = loc - RExC_precomp; \
481 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
482 "%s" REPORT_LOCATION, \
483 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
487 #define vWARN2(loc, m, a1) STMT_START { \
488 const IV offset = loc - RExC_precomp; \
489 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
490 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
493 #define vWARN3(loc, m, a1, a2) STMT_START { \
494 const IV offset = loc - RExC_precomp; \
495 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
496 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
499 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
500 const IV offset = loc - RExC_precomp; \
501 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
502 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
505 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
506 const IV offset = loc - RExC_precomp; \
507 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
508 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
512 /* Allow for side effects in s */
513 #define REGC(c,s) STMT_START { \
514 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
517 /* Macros for recording node offsets. 20001227 mjd@plover.com
518 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
519 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
520 * Element 0 holds the number n.
521 * Position is 1 indexed.
523 #ifndef RE_TRACK_PATTERN_OFFSETS
524 #define Set_Node_Offset_To_R(node,byte)
525 #define Set_Node_Offset(node,byte)
526 #define Set_Cur_Node_Offset
527 #define Set_Node_Length_To_R(node,len)
528 #define Set_Node_Length(node,len)
529 #define Set_Node_Cur_Length(node)
530 #define Node_Offset(n)
531 #define Node_Length(n)
532 #define Set_Node_Offset_Length(node,offset,len)
533 #define ProgLen(ri) ri->u.proglen
534 #define SetProgLen(ri,x) ri->u.proglen = x
536 #define ProgLen(ri) ri->u.offsets[0]
537 #define SetProgLen(ri,x) ri->u.offsets[0] = x
538 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
540 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
541 __LINE__, (int)(node), (int)(byte))); \
543 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
545 RExC_offsets[2*(node)-1] = (byte); \
550 #define Set_Node_Offset(node,byte) \
551 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
552 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
554 #define Set_Node_Length_To_R(node,len) STMT_START { \
556 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
557 __LINE__, (int)(node), (int)(len))); \
559 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
561 RExC_offsets[2*(node)] = (len); \
566 #define Set_Node_Length(node,len) \
567 Set_Node_Length_To_R((node)-RExC_emit_start, len)
568 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
569 #define Set_Node_Cur_Length(node) \
570 Set_Node_Length(node, RExC_parse - parse_start)
572 /* Get offsets and lengths */
573 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
574 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
576 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
577 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
578 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
582 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
583 #define EXPERIMENTAL_INPLACESCAN
584 #endif /*RE_TRACK_PATTERN_OFFSETS*/
586 #define DEBUG_STUDYDATA(str,data,depth) \
587 DEBUG_OPTIMISE_MORE_r(if(data){ \
588 PerlIO_printf(Perl_debug_log, \
589 "%*s" str "Pos:%"IVdf"/%"IVdf \
590 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
591 (int)(depth)*2, "", \
592 (IV)((data)->pos_min), \
593 (IV)((data)->pos_delta), \
594 (UV)((data)->flags), \
595 (IV)((data)->whilem_c), \
596 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
597 is_inf ? "INF " : "" \
599 if ((data)->last_found) \
600 PerlIO_printf(Perl_debug_log, \
601 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
602 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
603 SvPVX_const((data)->last_found), \
604 (IV)((data)->last_end), \
605 (IV)((data)->last_start_min), \
606 (IV)((data)->last_start_max), \
607 ((data)->longest && \
608 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
609 SvPVX_const((data)->longest_fixed), \
610 (IV)((data)->offset_fixed), \
611 ((data)->longest && \
612 (data)->longest==&((data)->longest_float)) ? "*" : "", \
613 SvPVX_const((data)->longest_float), \
614 (IV)((data)->offset_float_min), \
615 (IV)((data)->offset_float_max) \
617 PerlIO_printf(Perl_debug_log,"\n"); \
620 static void clear_re(pTHX_ void *r);
622 /* Mark that we cannot extend a found fixed substring at this point.
623 Update the longest found anchored substring and the longest found
624 floating substrings if needed. */
627 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
629 const STRLEN l = CHR_SVLEN(data->last_found);
630 const STRLEN old_l = CHR_SVLEN(*data->longest);
631 GET_RE_DEBUG_FLAGS_DECL;
633 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
634 SvSetMagicSV(*data->longest, data->last_found);
635 if (*data->longest == data->longest_fixed) {
636 data->offset_fixed = l ? data->last_start_min : data->pos_min;
637 if (data->flags & SF_BEFORE_EOL)
639 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
641 data->flags &= ~SF_FIX_BEFORE_EOL;
642 data->minlen_fixed=minlenp;
643 data->lookbehind_fixed=0;
645 else { /* *data->longest == data->longest_float */
646 data->offset_float_min = l ? data->last_start_min : data->pos_min;
647 data->offset_float_max = (l
648 ? data->last_start_max
649 : data->pos_min + data->pos_delta);
650 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
651 data->offset_float_max = I32_MAX;
652 if (data->flags & SF_BEFORE_EOL)
654 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
656 data->flags &= ~SF_FL_BEFORE_EOL;
657 data->minlen_float=minlenp;
658 data->lookbehind_float=0;
661 SvCUR_set(data->last_found, 0);
663 SV * const sv = data->last_found;
664 if (SvUTF8(sv) && SvMAGICAL(sv)) {
665 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
671 data->flags &= ~SF_BEFORE_EOL;
672 DEBUG_STUDYDATA("commit: ",data,0);
675 /* Can match anything (initialization) */
677 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
679 ANYOF_CLASS_ZERO(cl);
680 ANYOF_BITMAP_SETALL(cl);
681 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
683 cl->flags |= ANYOF_LOCALE;
686 /* Can match anything (initialization) */
688 S_cl_is_anything(const struct regnode_charclass_class *cl)
692 for (value = 0; value <= ANYOF_MAX; value += 2)
693 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
695 if (!(cl->flags & ANYOF_UNICODE_ALL))
697 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
702 /* Can match anything (initialization) */
704 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
706 Zero(cl, 1, struct regnode_charclass_class);
708 cl_anything(pRExC_state, cl);
712 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
714 Zero(cl, 1, struct regnode_charclass_class);
716 cl_anything(pRExC_state, cl);
718 cl->flags |= ANYOF_LOCALE;
721 /* 'And' a given class with another one. Can create false positives */
722 /* We assume that cl is not inverted */
724 S_cl_and(struct regnode_charclass_class *cl,
725 const struct regnode_charclass_class *and_with)
728 assert(and_with->type == ANYOF);
729 if (!(and_with->flags & ANYOF_CLASS)
730 && !(cl->flags & ANYOF_CLASS)
731 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
732 && !(and_with->flags & ANYOF_FOLD)
733 && !(cl->flags & ANYOF_FOLD)) {
736 if (and_with->flags & ANYOF_INVERT)
737 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
738 cl->bitmap[i] &= ~and_with->bitmap[i];
740 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
741 cl->bitmap[i] &= and_with->bitmap[i];
742 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
743 if (!(and_with->flags & ANYOF_EOS))
744 cl->flags &= ~ANYOF_EOS;
746 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
747 !(and_with->flags & ANYOF_INVERT)) {
748 cl->flags &= ~ANYOF_UNICODE_ALL;
749 cl->flags |= ANYOF_UNICODE;
750 ARG_SET(cl, ARG(and_with));
752 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
753 !(and_with->flags & ANYOF_INVERT))
754 cl->flags &= ~ANYOF_UNICODE_ALL;
755 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
756 !(and_with->flags & ANYOF_INVERT))
757 cl->flags &= ~ANYOF_UNICODE;
760 /* 'OR' a given class with another one. Can create false positives */
761 /* We assume that cl is not inverted */
763 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
765 if (or_with->flags & ANYOF_INVERT) {
767 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
768 * <= (B1 | !B2) | (CL1 | !CL2)
769 * which is wasteful if CL2 is small, but we ignore CL2:
770 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
771 * XXXX Can we handle case-fold? Unclear:
772 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
773 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
775 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
776 && !(or_with->flags & ANYOF_FOLD)
777 && !(cl->flags & ANYOF_FOLD) ) {
780 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
781 cl->bitmap[i] |= ~or_with->bitmap[i];
782 } /* XXXX: logic is complicated otherwise */
784 cl_anything(pRExC_state, cl);
787 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
788 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
789 && (!(or_with->flags & ANYOF_FOLD)
790 || (cl->flags & ANYOF_FOLD)) ) {
793 /* OR char bitmap and class bitmap separately */
794 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
795 cl->bitmap[i] |= or_with->bitmap[i];
796 if (or_with->flags & ANYOF_CLASS) {
797 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
798 cl->classflags[i] |= or_with->classflags[i];
799 cl->flags |= ANYOF_CLASS;
802 else { /* XXXX: logic is complicated, leave it along for a moment. */
803 cl_anything(pRExC_state, cl);
806 if (or_with->flags & ANYOF_EOS)
807 cl->flags |= ANYOF_EOS;
809 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
810 ARG(cl) != ARG(or_with)) {
811 cl->flags |= ANYOF_UNICODE_ALL;
812 cl->flags &= ~ANYOF_UNICODE;
814 if (or_with->flags & ANYOF_UNICODE_ALL) {
815 cl->flags |= ANYOF_UNICODE_ALL;
816 cl->flags &= ~ANYOF_UNICODE;
820 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
821 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
822 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
823 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
828 dump_trie(trie,widecharmap,revcharmap)
829 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
830 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
832 These routines dump out a trie in a somewhat readable format.
833 The _interim_ variants are used for debugging the interim
834 tables that are used to generate the final compressed
835 representation which is what dump_trie expects.
837 Part of the reason for their existance is to provide a form
838 of documentation as to how the different representations function.
843 Dumps the final compressed table form of the trie to Perl_debug_log.
844 Used for debugging make_trie().
848 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
849 AV *revcharmap, U32 depth)
852 SV *sv=sv_newmortal();
853 int colwidth= widecharmap ? 6 : 4;
854 GET_RE_DEBUG_FLAGS_DECL;
857 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
858 (int)depth * 2 + 2,"",
859 "Match","Base","Ofs" );
861 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
862 SV ** const tmp = av_fetch( revcharmap, state, 0);
864 PerlIO_printf( Perl_debug_log, "%*s",
866 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
867 PL_colors[0], PL_colors[1],
868 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
869 PERL_PV_ESCAPE_FIRSTCHAR
874 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
875 (int)depth * 2 + 2,"");
877 for( state = 0 ; state < trie->uniquecharcount ; state++ )
878 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
879 PerlIO_printf( Perl_debug_log, "\n");
881 for( state = 1 ; state < trie->statecount ; state++ ) {
882 const U32 base = trie->states[ state ].trans.base;
884 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
886 if ( trie->states[ state ].wordnum ) {
887 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
889 PerlIO_printf( Perl_debug_log, "%6s", "" );
892 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
897 while( ( base + ofs < trie->uniquecharcount ) ||
898 ( base + ofs - trie->uniquecharcount < trie->lasttrans
899 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
902 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
904 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
905 if ( ( base + ofs >= trie->uniquecharcount ) &&
906 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
907 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
909 PerlIO_printf( Perl_debug_log, "%*"UVXf,
911 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
913 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
917 PerlIO_printf( Perl_debug_log, "]");
920 PerlIO_printf( Perl_debug_log, "\n" );
924 Dumps a fully constructed but uncompressed trie in list form.
925 List tries normally only are used for construction when the number of
926 possible chars (trie->uniquecharcount) is very high.
927 Used for debugging make_trie().
930 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
931 HV *widecharmap, AV *revcharmap, U32 next_alloc,
935 SV *sv=sv_newmortal();
936 int colwidth= widecharmap ? 6 : 4;
937 GET_RE_DEBUG_FLAGS_DECL;
938 /* print out the table precompression. */
939 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
940 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
941 "------:-----+-----------------\n" );
943 for( state=1 ; state < next_alloc ; state ++ ) {
946 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
947 (int)depth * 2 + 2,"", (UV)state );
948 if ( ! trie->states[ state ].wordnum ) {
949 PerlIO_printf( Perl_debug_log, "%5s| ","");
951 PerlIO_printf( Perl_debug_log, "W%4x| ",
952 trie->states[ state ].wordnum
955 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
956 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
958 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
960 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
961 PL_colors[0], PL_colors[1],
962 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
963 PERL_PV_ESCAPE_FIRSTCHAR
965 TRIE_LIST_ITEM(state,charid).forid,
966 (UV)TRIE_LIST_ITEM(state,charid).newstate
969 PerlIO_printf(Perl_debug_log, "\n%*s| ",
970 (int)((depth * 2) + 14), "");
973 PerlIO_printf( Perl_debug_log, "\n");
978 Dumps a fully constructed but uncompressed trie in table form.
979 This is the normal DFA style state transition table, with a few
980 twists to facilitate compression later.
981 Used for debugging make_trie().
984 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
985 HV *widecharmap, AV *revcharmap, U32 next_alloc,
990 SV *sv=sv_newmortal();
991 int colwidth= widecharmap ? 6 : 4;
992 GET_RE_DEBUG_FLAGS_DECL;
995 print out the table precompression so that we can do a visual check
996 that they are identical.
999 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1001 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1002 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1004 PerlIO_printf( Perl_debug_log, "%*s",
1006 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1007 PL_colors[0], PL_colors[1],
1008 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1009 PERL_PV_ESCAPE_FIRSTCHAR
1015 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1017 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1018 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1021 PerlIO_printf( Perl_debug_log, "\n" );
1023 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1025 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1026 (int)depth * 2 + 2,"",
1027 (UV)TRIE_NODENUM( state ) );
1029 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1030 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1032 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1034 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1036 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1037 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1039 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1040 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1047 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1048 startbranch: the first branch in the whole branch sequence
1049 first : start branch of sequence of branch-exact nodes.
1050 May be the same as startbranch
1051 last : Thing following the last branch.
1052 May be the same as tail.
1053 tail : item following the branch sequence
1054 count : words in the sequence
1055 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1056 depth : indent depth
1058 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1060 A trie is an N'ary tree where the branches are determined by digital
1061 decomposition of the key. IE, at the root node you look up the 1st character and
1062 follow that branch repeat until you find the end of the branches. Nodes can be
1063 marked as "accepting" meaning they represent a complete word. Eg:
1067 would convert into the following structure. Numbers represent states, letters
1068 following numbers represent valid transitions on the letter from that state, if
1069 the number is in square brackets it represents an accepting state, otherwise it
1070 will be in parenthesis.
1072 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1076 (1) +-i->(6)-+-s->[7]
1078 +-s->(3)-+-h->(4)-+-e->[5]
1080 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1082 This shows that when matching against the string 'hers' we will begin at state 1
1083 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1084 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1085 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1086 single traverse. We store a mapping from accepting to state to which word was
1087 matched, and then when we have multiple possibilities we try to complete the
1088 rest of the regex in the order in which they occured in the alternation.
1090 The only prior NFA like behaviour that would be changed by the TRIE support is
1091 the silent ignoring of duplicate alternations which are of the form:
1093 / (DUPE|DUPE) X? (?{ ... }) Y /x
1095 Thus EVAL blocks follwing a trie may be called a different number of times with
1096 and without the optimisation. With the optimisations dupes will be silently
1097 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1098 the following demonstrates:
1100 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1102 which prints out 'word' three times, but
1104 'words'=~/(word|word|word)(?{ print $1 })S/
1106 which doesnt print it out at all. This is due to other optimisations kicking in.
1108 Example of what happens on a structural level:
1110 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1112 1: CURLYM[1] {1,32767}(18)
1123 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1124 and should turn into:
1126 1: CURLYM[1] {1,32767}(18)
1128 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1136 Cases where tail != last would be like /(?foo|bar)baz/:
1146 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1147 and would end up looking like:
1150 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1157 d = uvuni_to_utf8_flags(d, uv, 0);
1159 is the recommended Unicode-aware way of saying
1164 #define TRIE_STORE_REVCHAR \
1166 SV *tmp = newSVpvs(""); \
1167 if (UTF) SvUTF8_on(tmp); \
1168 Perl_sv_catpvf( aTHX_ tmp, "%c", (int)uvc ); \
1169 av_push( revcharmap, tmp ); \
1172 #define TRIE_READ_CHAR STMT_START { \
1176 if ( foldlen > 0 ) { \
1177 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1182 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1183 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1184 foldlen -= UNISKIP( uvc ); \
1185 scan = foldbuf + UNISKIP( uvc ); \
1188 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1198 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1199 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1200 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1201 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1203 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1204 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1205 TRIE_LIST_CUR( state )++; \
1208 #define TRIE_LIST_NEW(state) STMT_START { \
1209 Newxz( trie->states[ state ].trans.list, \
1210 4, reg_trie_trans_le ); \
1211 TRIE_LIST_CUR( state ) = 1; \
1212 TRIE_LIST_LEN( state ) = 4; \
1215 #define TRIE_HANDLE_WORD(state) STMT_START { \
1216 U16 dupe= trie->states[ state ].wordnum; \
1217 regnode * const noper_next = regnext( noper ); \
1219 if (trie->wordlen) \
1220 trie->wordlen[ curword ] = wordlen; \
1222 /* store the word for dumping */ \
1224 if (OP(noper) != NOTHING) \
1225 tmp = newSVpvn(STRING(noper), STR_LEN(noper)); \
1227 tmp = newSVpvn( "", 0 ); \
1228 if ( UTF ) SvUTF8_on( tmp ); \
1229 av_push( trie_words, tmp ); \
1234 if ( noper_next < tail ) { \
1236 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1237 trie->jump[curword] = (U16)(noper_next - convert); \
1239 jumper = noper_next; \
1241 nextbranch= regnext(cur); \
1245 /* So it's a dupe. This means we need to maintain a */\
1246 /* linked-list from the first to the next. */\
1247 /* we only allocate the nextword buffer when there */\
1248 /* a dupe, so first time we have to do the allocation */\
1249 if (!trie->nextword) \
1250 trie->nextword = (U16 *) \
1251 PerlMemShared_calloc( word_count + 1, sizeof(U16)); \
1252 while ( trie->nextword[dupe] ) \
1253 dupe= trie->nextword[dupe]; \
1254 trie->nextword[dupe]= curword; \
1256 /* we haven't inserted this word yet. */ \
1257 trie->states[ state ].wordnum = curword; \
1262 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1263 ( ( base + charid >= ucharcount \
1264 && base + charid < ubound \
1265 && state == trie->trans[ base - ucharcount + charid ].check \
1266 && trie->trans[ base - ucharcount + charid ].next ) \
1267 ? trie->trans[ base - ucharcount + charid ].next \
1268 : ( state==1 ? special : 0 ) \
1272 #define MADE_JUMP_TRIE 2
1273 #define MADE_EXACT_TRIE 4
1276 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1279 /* first pass, loop through and scan words */
1280 reg_trie_data *trie;
1281 HV *widecharmap = NULL;
1282 AV *revcharmap = newAV();
1284 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1289 regnode *jumper = NULL;
1290 regnode *nextbranch = NULL;
1291 regnode *convert = NULL;
1292 /* we just use folder as a flag in utf8 */
1293 const U8 * const folder = ( flags == EXACTF
1295 : ( flags == EXACTFL
1302 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1303 AV *trie_words = NULL;
1304 /* along with revcharmap, this only used during construction but both are
1305 * useful during debugging so we store them in the struct when debugging.
1308 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1309 STRLEN trie_charcount=0;
1311 SV *re_trie_maxbuff;
1312 GET_RE_DEBUG_FLAGS_DECL;
1314 PERL_UNUSED_ARG(depth);
1317 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1319 trie->startstate = 1;
1320 trie->wordcount = word_count;
1321 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1322 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1323 if (!(UTF && folder))
1324 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1326 trie_words = newAV();
1329 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1330 if (!SvIOK(re_trie_maxbuff)) {
1331 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1334 PerlIO_printf( Perl_debug_log,
1335 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1336 (int)depth * 2 + 2, "",
1337 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1338 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1342 /* Find the node we are going to overwrite */
1343 if ( first == startbranch && OP( last ) != BRANCH ) {
1344 /* whole branch chain */
1347 /* branch sub-chain */
1348 convert = NEXTOPER( first );
1351 /* -- First loop and Setup --
1353 We first traverse the branches and scan each word to determine if it
1354 contains widechars, and how many unique chars there are, this is
1355 important as we have to build a table with at least as many columns as we
1358 We use an array of integers to represent the character codes 0..255
1359 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1360 native representation of the character value as the key and IV's for the
1363 *TODO* If we keep track of how many times each character is used we can
1364 remap the columns so that the table compression later on is more
1365 efficient in terms of memory by ensuring most common value is in the
1366 middle and the least common are on the outside. IMO this would be better
1367 than a most to least common mapping as theres a decent chance the most
1368 common letter will share a node with the least common, meaning the node
1369 will not be compressable. With a middle is most common approach the worst
1370 case is when we have the least common nodes twice.
1374 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1375 regnode * const noper = NEXTOPER( cur );
1376 const U8 *uc = (U8*)STRING( noper );
1377 const U8 * const e = uc + STR_LEN( noper );
1379 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1380 const U8 *scan = (U8*)NULL;
1381 U32 wordlen = 0; /* required init */
1383 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1385 if (OP(noper) == NOTHING) {
1389 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1390 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1391 regardless of encoding */
1393 for ( ; uc < e ; uc += len ) {
1394 TRIE_CHARCOUNT(trie)++;
1398 if ( !trie->charmap[ uvc ] ) {
1399 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1401 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1405 /* store the codepoint in the bitmap, and if its ascii
1406 also store its folded equivelent. */
1407 TRIE_BITMAP_SET(trie,uvc);
1409 /* store the folded codepoint */
1410 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1413 /* store first byte of utf8 representation of
1414 codepoints in the 127 < uvc < 256 range */
1415 if (127 < uvc && uvc < 192) {
1416 TRIE_BITMAP_SET(trie,194);
1417 } else if (191 < uvc ) {
1418 TRIE_BITMAP_SET(trie,195);
1419 /* && uvc < 256 -- we know uvc is < 256 already */
1422 set_bit = 0; /* We've done our bit :-) */
1427 widecharmap = newHV();
1429 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1432 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1434 if ( !SvTRUE( *svpp ) ) {
1435 sv_setiv( *svpp, ++trie->uniquecharcount );
1440 if( cur == first ) {
1443 } else if (chars < trie->minlen) {
1445 } else if (chars > trie->maxlen) {
1449 } /* end first pass */
1450 DEBUG_TRIE_COMPILE_r(
1451 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1452 (int)depth * 2 + 2,"",
1453 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1454 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1455 (int)trie->minlen, (int)trie->maxlen )
1457 trie->wordlen = (U32 *) PerlMemShared_calloc( word_count, sizeof(U32) );
1460 We now know what we are dealing with in terms of unique chars and
1461 string sizes so we can calculate how much memory a naive
1462 representation using a flat table will take. If it's over a reasonable
1463 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1464 conservative but potentially much slower representation using an array
1467 At the end we convert both representations into the same compressed
1468 form that will be used in regexec.c for matching with. The latter
1469 is a form that cannot be used to construct with but has memory
1470 properties similar to the list form and access properties similar
1471 to the table form making it both suitable for fast searches and
1472 small enough that its feasable to store for the duration of a program.
1474 See the comment in the code where the compressed table is produced
1475 inplace from the flat tabe representation for an explanation of how
1476 the compression works.
1481 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1483 Second Pass -- Array Of Lists Representation
1485 Each state will be represented by a list of charid:state records
1486 (reg_trie_trans_le) the first such element holds the CUR and LEN
1487 points of the allocated array. (See defines above).
1489 We build the initial structure using the lists, and then convert
1490 it into the compressed table form which allows faster lookups
1491 (but cant be modified once converted).
1494 STRLEN transcount = 1;
1496 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1497 "%*sCompiling trie using list compiler\n",
1498 (int)depth * 2 + 2, ""));
1500 trie->states = (reg_trie_state *)
1501 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1502 sizeof(reg_trie_state) );
1506 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1508 regnode * const noper = NEXTOPER( cur );
1509 U8 *uc = (U8*)STRING( noper );
1510 const U8 * const e = uc + STR_LEN( noper );
1511 U32 state = 1; /* required init */
1512 U16 charid = 0; /* sanity init */
1513 U8 *scan = (U8*)NULL; /* sanity init */
1514 STRLEN foldlen = 0; /* required init */
1515 U32 wordlen = 0; /* required init */
1516 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1518 if (OP(noper) != NOTHING) {
1519 for ( ; uc < e ; uc += len ) {
1524 charid = trie->charmap[ uvc ];
1526 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1530 charid=(U16)SvIV( *svpp );
1533 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1540 if ( !trie->states[ state ].trans.list ) {
1541 TRIE_LIST_NEW( state );
1543 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1544 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1545 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1550 newstate = next_alloc++;
1551 TRIE_LIST_PUSH( state, charid, newstate );
1556 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1560 TRIE_HANDLE_WORD(state);
1562 } /* end second pass */
1564 /* next alloc is the NEXT state to be allocated */
1565 trie->statecount = next_alloc;
1566 trie->states = (reg_trie_state *)
1567 PerlMemShared_realloc( trie->states,
1569 * sizeof(reg_trie_state) );
1571 /* and now dump it out before we compress it */
1572 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1573 revcharmap, next_alloc,
1577 trie->trans = (reg_trie_trans *)
1578 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1585 for( state=1 ; state < next_alloc ; state ++ ) {
1589 DEBUG_TRIE_COMPILE_MORE_r(
1590 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1594 if (trie->states[state].trans.list) {
1595 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1599 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1600 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1601 if ( forid < minid ) {
1603 } else if ( forid > maxid ) {
1607 if ( transcount < tp + maxid - minid + 1) {
1609 trie->trans = (reg_trie_trans *)
1610 PerlMemShared_realloc( trie->trans,
1612 * sizeof(reg_trie_trans) );
1613 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1615 base = trie->uniquecharcount + tp - minid;
1616 if ( maxid == minid ) {
1618 for ( ; zp < tp ; zp++ ) {
1619 if ( ! trie->trans[ zp ].next ) {
1620 base = trie->uniquecharcount + zp - minid;
1621 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1622 trie->trans[ zp ].check = state;
1628 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1629 trie->trans[ tp ].check = state;
1634 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1635 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1636 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1637 trie->trans[ tid ].check = state;
1639 tp += ( maxid - minid + 1 );
1641 Safefree(trie->states[ state ].trans.list);
1644 DEBUG_TRIE_COMPILE_MORE_r(
1645 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1648 trie->states[ state ].trans.base=base;
1650 trie->lasttrans = tp + 1;
1654 Second Pass -- Flat Table Representation.
1656 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1657 We know that we will need Charcount+1 trans at most to store the data
1658 (one row per char at worst case) So we preallocate both structures
1659 assuming worst case.
1661 We then construct the trie using only the .next slots of the entry
1664 We use the .check field of the first entry of the node temporarily to
1665 make compression both faster and easier by keeping track of how many non
1666 zero fields are in the node.
1668 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1671 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1672 number representing the first entry of the node, and state as a
1673 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1674 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1675 are 2 entrys per node. eg:
1683 The table is internally in the right hand, idx form. However as we also
1684 have to deal with the states array which is indexed by nodenum we have to
1685 use TRIE_NODENUM() to convert.
1688 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1689 "%*sCompiling trie using table compiler\n",
1690 (int)depth * 2 + 2, ""));
1692 trie->trans = (reg_trie_trans *)
1693 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1694 * trie->uniquecharcount + 1,
1695 sizeof(reg_trie_trans) );
1696 trie->states = (reg_trie_state *)
1697 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1698 sizeof(reg_trie_state) );
1699 next_alloc = trie->uniquecharcount + 1;
1702 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1704 regnode * const noper = NEXTOPER( cur );
1705 const U8 *uc = (U8*)STRING( noper );
1706 const U8 * const e = uc + STR_LEN( noper );
1708 U32 state = 1; /* required init */
1710 U16 charid = 0; /* sanity init */
1711 U32 accept_state = 0; /* sanity init */
1712 U8 *scan = (U8*)NULL; /* sanity init */
1714 STRLEN foldlen = 0; /* required init */
1715 U32 wordlen = 0; /* required init */
1716 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1718 if ( OP(noper) != NOTHING ) {
1719 for ( ; uc < e ; uc += len ) {
1724 charid = trie->charmap[ uvc ];
1726 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1727 charid = svpp ? (U16)SvIV(*svpp) : 0;
1731 if ( !trie->trans[ state + charid ].next ) {
1732 trie->trans[ state + charid ].next = next_alloc;
1733 trie->trans[ state ].check++;
1734 next_alloc += trie->uniquecharcount;
1736 state = trie->trans[ state + charid ].next;
1738 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1740 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1743 accept_state = TRIE_NODENUM( state );
1744 TRIE_HANDLE_WORD(accept_state);
1746 } /* end second pass */
1748 /* and now dump it out before we compress it */
1749 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1751 next_alloc, depth+1));
1755 * Inplace compress the table.*
1757 For sparse data sets the table constructed by the trie algorithm will
1758 be mostly 0/FAIL transitions or to put it another way mostly empty.
1759 (Note that leaf nodes will not contain any transitions.)
1761 This algorithm compresses the tables by eliminating most such
1762 transitions, at the cost of a modest bit of extra work during lookup:
1764 - Each states[] entry contains a .base field which indicates the
1765 index in the state[] array wheres its transition data is stored.
1767 - If .base is 0 there are no valid transitions from that node.
1769 - If .base is nonzero then charid is added to it to find an entry in
1772 -If trans[states[state].base+charid].check!=state then the
1773 transition is taken to be a 0/Fail transition. Thus if there are fail
1774 transitions at the front of the node then the .base offset will point
1775 somewhere inside the previous nodes data (or maybe even into a node
1776 even earlier), but the .check field determines if the transition is
1780 The following process inplace converts the table to the compressed
1781 table: We first do not compress the root node 1,and mark its all its
1782 .check pointers as 1 and set its .base pointer as 1 as well. This
1783 allows to do a DFA construction from the compressed table later, and
1784 ensures that any .base pointers we calculate later are greater than
1787 - We set 'pos' to indicate the first entry of the second node.
1789 - We then iterate over the columns of the node, finding the first and
1790 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1791 and set the .check pointers accordingly, and advance pos
1792 appropriately and repreat for the next node. Note that when we copy
1793 the next pointers we have to convert them from the original
1794 NODEIDX form to NODENUM form as the former is not valid post
1797 - If a node has no transitions used we mark its base as 0 and do not
1798 advance the pos pointer.
1800 - If a node only has one transition we use a second pointer into the
1801 structure to fill in allocated fail transitions from other states.
1802 This pointer is independent of the main pointer and scans forward
1803 looking for null transitions that are allocated to a state. When it
1804 finds one it writes the single transition into the "hole". If the
1805 pointer doesnt find one the single transition is appended as normal.
1807 - Once compressed we can Renew/realloc the structures to release the
1810 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1811 specifically Fig 3.47 and the associated pseudocode.
1815 const U32 laststate = TRIE_NODENUM( next_alloc );
1818 trie->statecount = laststate;
1820 for ( state = 1 ; state < laststate ; state++ ) {
1822 const U32 stateidx = TRIE_NODEIDX( state );
1823 const U32 o_used = trie->trans[ stateidx ].check;
1824 U32 used = trie->trans[ stateidx ].check;
1825 trie->trans[ stateidx ].check = 0;
1827 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1828 if ( flag || trie->trans[ stateidx + charid ].next ) {
1829 if ( trie->trans[ stateidx + charid ].next ) {
1831 for ( ; zp < pos ; zp++ ) {
1832 if ( ! trie->trans[ zp ].next ) {
1836 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1837 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1838 trie->trans[ zp ].check = state;
1839 if ( ++zp > pos ) pos = zp;
1846 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1848 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1849 trie->trans[ pos ].check = state;
1854 trie->lasttrans = pos + 1;
1855 trie->states = (reg_trie_state *)
1856 PerlMemShared_realloc( trie->states, laststate
1857 * sizeof(reg_trie_state) );
1858 DEBUG_TRIE_COMPILE_MORE_r(
1859 PerlIO_printf( Perl_debug_log,
1860 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1861 (int)depth * 2 + 2,"",
1862 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1865 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1868 } /* end table compress */
1870 DEBUG_TRIE_COMPILE_MORE_r(
1871 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1872 (int)depth * 2 + 2, "",
1873 (UV)trie->statecount,
1874 (UV)trie->lasttrans)
1876 /* resize the trans array to remove unused space */
1877 trie->trans = (reg_trie_trans *)
1878 PerlMemShared_realloc( trie->trans, trie->lasttrans
1879 * sizeof(reg_trie_trans) );
1881 /* and now dump out the compressed format */
1882 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
1884 { /* Modify the program and insert the new TRIE node*/
1885 U8 nodetype =(U8)(flags & 0xFF);
1889 regnode *optimize = NULL;
1890 #ifdef RE_TRACK_PATTERN_OFFSETS
1893 U32 mjd_nodelen = 0;
1894 #endif /* RE_TRACK_PATTERN_OFFSETS */
1895 #endif /* DEBUGGING */
1897 This means we convert either the first branch or the first Exact,
1898 depending on whether the thing following (in 'last') is a branch
1899 or not and whther first is the startbranch (ie is it a sub part of
1900 the alternation or is it the whole thing.)
1901 Assuming its a sub part we conver the EXACT otherwise we convert
1902 the whole branch sequence, including the first.
1904 /* Find the node we are going to overwrite */
1905 if ( first != startbranch || OP( last ) == BRANCH ) {
1906 /* branch sub-chain */
1907 NEXT_OFF( first ) = (U16)(last - first);
1908 #ifdef RE_TRACK_PATTERN_OFFSETS
1910 mjd_offset= Node_Offset((convert));
1911 mjd_nodelen= Node_Length((convert));
1914 /* whole branch chain */
1916 #ifdef RE_TRACK_PATTERN_OFFSETS
1919 const regnode *nop = NEXTOPER( convert );
1920 mjd_offset= Node_Offset((nop));
1921 mjd_nodelen= Node_Length((nop));
1925 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1926 (int)depth * 2 + 2, "",
1927 (UV)mjd_offset, (UV)mjd_nodelen)
1930 /* But first we check to see if there is a common prefix we can
1931 split out as an EXACT and put in front of the TRIE node. */
1932 trie->startstate= 1;
1933 if ( trie->bitmap && !widecharmap && !trie->jump ) {
1935 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1939 const U32 base = trie->states[ state ].trans.base;
1941 if ( trie->states[state].wordnum )
1944 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1945 if ( ( base + ofs >= trie->uniquecharcount ) &&
1946 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1947 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1949 if ( ++count > 1 ) {
1950 SV **tmp = av_fetch( revcharmap, ofs, 0);
1951 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1952 if ( state == 1 ) break;
1954 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1956 PerlIO_printf(Perl_debug_log,
1957 "%*sNew Start State=%"UVuf" Class: [",
1958 (int)depth * 2 + 2, "",
1961 SV ** const tmp = av_fetch( revcharmap, idx, 0);
1962 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1964 TRIE_BITMAP_SET(trie,*ch);
1966 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1968 PerlIO_printf(Perl_debug_log, (char*)ch)
1972 TRIE_BITMAP_SET(trie,*ch);
1974 TRIE_BITMAP_SET(trie,folder[ *ch ]);
1975 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
1981 SV **tmp = av_fetch( revcharmap, idx, 0);
1983 char *ch = SvPV( *tmp, len );
1985 SV *sv=sv_newmortal();
1986 PerlIO_printf( Perl_debug_log,
1987 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
1988 (int)depth * 2 + 2, "",
1990 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
1991 PL_colors[0], PL_colors[1],
1992 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1993 PERL_PV_ESCAPE_FIRSTCHAR
1998 OP( convert ) = nodetype;
1999 str=STRING(convert);
2002 STR_LEN(convert) += len;
2008 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2014 regnode *n = convert+NODE_SZ_STR(convert);
2015 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2016 trie->startstate = state;
2017 trie->minlen -= (state - 1);
2018 trie->maxlen -= (state - 1);
2020 regnode *fix = convert;
2021 U32 word = trie->wordcount;
2023 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2024 while( ++fix < n ) {
2025 Set_Node_Offset_Length(fix, 0, 0);
2028 SV ** const tmp = av_fetch( trie_words, word, 0 );
2030 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2031 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2033 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2040 NEXT_OFF(convert) = (U16)(tail - convert);
2041 DEBUG_r(optimize= n);
2047 if ( trie->maxlen ) {
2048 NEXT_OFF( convert ) = (U16)(tail - convert);
2049 ARG_SET( convert, data_slot );
2050 /* Store the offset to the first unabsorbed branch in
2051 jump[0], which is otherwise unused by the jump logic.
2052 We use this when dumping a trie and during optimisation. */
2054 trie->jump[0] = (U16)(nextbranch - convert);
2057 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
2058 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2060 OP( convert ) = TRIEC;
2061 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2062 PerlMemShared_free(trie->bitmap);
2065 OP( convert ) = TRIE;
2067 /* store the type in the flags */
2068 convert->flags = nodetype;
2072 + regarglen[ OP( convert ) ];
2074 /* XXX We really should free up the resource in trie now,
2075 as we won't use them - (which resources?) dmq */
2077 /* needed for dumping*/
2078 DEBUG_r(if (optimize) {
2079 regnode *opt = convert;
2081 while ( ++opt < optimize) {
2082 Set_Node_Offset_Length(opt,0,0);
2085 Try to clean up some of the debris left after the
2088 while( optimize < jumper ) {
2089 mjd_nodelen += Node_Length((optimize));
2090 OP( optimize ) = OPTIMIZED;
2091 Set_Node_Offset_Length(optimize,0,0);
2094 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2096 } /* end node insert */
2097 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2099 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2100 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2102 SvREFCNT_dec(revcharmap);
2106 : trie->startstate>1
2112 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2114 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
2116 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2117 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2120 We find the fail state for each state in the trie, this state is the longest proper
2121 suffix of the current states 'word' that is also a proper prefix of another word in our
2122 trie. State 1 represents the word '' and is the thus the default fail state. This allows
2123 the DFA not to have to restart after its tried and failed a word at a given point, it
2124 simply continues as though it had been matching the other word in the first place.
2126 'abcdgu'=~/abcdefg|cdgu/
2127 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2128 fail, which would bring use to the state representing 'd' in the second word where we would
2129 try 'g' and succeed, prodceding to match 'cdgu'.
2131 /* add a fail transition */
2132 const U32 trie_offset = ARG(source);
2133 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2135 const U32 ucharcount = trie->uniquecharcount;
2136 const U32 numstates = trie->statecount;
2137 const U32 ubound = trie->lasttrans + ucharcount;
2141 U32 base = trie->states[ 1 ].trans.base;
2144 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2145 GET_RE_DEBUG_FLAGS_DECL;
2147 PERL_UNUSED_ARG(depth);
2151 ARG_SET( stclass, data_slot );
2152 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2153 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2154 aho->trie=trie_offset;
2155 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2156 Copy( trie->states, aho->states, numstates, reg_trie_state );
2157 Newxz( q, numstates, U32);
2158 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2161 /* initialize fail[0..1] to be 1 so that we always have
2162 a valid final fail state */
2163 fail[ 0 ] = fail[ 1 ] = 1;
2165 for ( charid = 0; charid < ucharcount ; charid++ ) {
2166 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2168 q[ q_write ] = newstate;
2169 /* set to point at the root */
2170 fail[ q[ q_write++ ] ]=1;
2173 while ( q_read < q_write) {
2174 const U32 cur = q[ q_read++ % numstates ];
2175 base = trie->states[ cur ].trans.base;
2177 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2178 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2180 U32 fail_state = cur;
2183 fail_state = fail[ fail_state ];
2184 fail_base = aho->states[ fail_state ].trans.base;
2185 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2187 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2188 fail[ ch_state ] = fail_state;
2189 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2191 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2193 q[ q_write++ % numstates] = ch_state;
2197 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2198 when we fail in state 1, this allows us to use the
2199 charclass scan to find a valid start char. This is based on the principle
2200 that theres a good chance the string being searched contains lots of stuff
2201 that cant be a start char.
2203 fail[ 0 ] = fail[ 1 ] = 0;
2204 DEBUG_TRIE_COMPILE_r({
2205 PerlIO_printf(Perl_debug_log,
2206 "%*sStclass Failtable (%"UVuf" states): 0",
2207 (int)(depth * 2), "", (UV)numstates
2209 for( q_read=1; q_read<numstates; q_read++ ) {
2210 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2212 PerlIO_printf(Perl_debug_log, "\n");
2215 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2220 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2221 * These need to be revisited when a newer toolchain becomes available.
2223 #if defined(__sparc64__) && defined(__GNUC__)
2224 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2225 # undef SPARC64_GCC_WORKAROUND
2226 # define SPARC64_GCC_WORKAROUND 1
2230 #define DEBUG_PEEP(str,scan,depth) \
2231 DEBUG_OPTIMISE_r({if (scan){ \
2232 SV * const mysv=sv_newmortal(); \
2233 regnode *Next = regnext(scan); \
2234 regprop(RExC_rx, mysv, scan); \
2235 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2236 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2237 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2244 #define JOIN_EXACT(scan,min,flags) \
2245 if (PL_regkind[OP(scan)] == EXACT) \
2246 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2249 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2250 /* Merge several consecutive EXACTish nodes into one. */
2251 regnode *n = regnext(scan);
2253 regnode *next = scan + NODE_SZ_STR(scan);
2257 regnode *stop = scan;
2258 GET_RE_DEBUG_FLAGS_DECL;
2260 PERL_UNUSED_ARG(depth);
2262 #ifndef EXPERIMENTAL_INPLACESCAN
2263 PERL_UNUSED_ARG(flags);
2264 PERL_UNUSED_ARG(val);
2266 DEBUG_PEEP("join",scan,depth);
2268 /* Skip NOTHING, merge EXACT*. */
2270 ( PL_regkind[OP(n)] == NOTHING ||
2271 (stringok && (OP(n) == OP(scan))))
2273 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2275 if (OP(n) == TAIL || n > next)
2277 if (PL_regkind[OP(n)] == NOTHING) {
2278 DEBUG_PEEP("skip:",n,depth);
2279 NEXT_OFF(scan) += NEXT_OFF(n);
2280 next = n + NODE_STEP_REGNODE;
2287 else if (stringok) {
2288 const unsigned int oldl = STR_LEN(scan);
2289 regnode * const nnext = regnext(n);
2291 DEBUG_PEEP("merg",n,depth);
2294 if (oldl + STR_LEN(n) > U8_MAX)
2296 NEXT_OFF(scan) += NEXT_OFF(n);
2297 STR_LEN(scan) += STR_LEN(n);
2298 next = n + NODE_SZ_STR(n);
2299 /* Now we can overwrite *n : */
2300 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2308 #ifdef EXPERIMENTAL_INPLACESCAN
2309 if (flags && !NEXT_OFF(n)) {
2310 DEBUG_PEEP("atch", val, depth);
2311 if (reg_off_by_arg[OP(n)]) {
2312 ARG_SET(n, val - n);
2315 NEXT_OFF(n) = val - n;
2322 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2324 Two problematic code points in Unicode casefolding of EXACT nodes:
2326 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2327 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2333 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2334 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2336 This means that in case-insensitive matching (or "loose matching",
2337 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2338 length of the above casefolded versions) can match a target string
2339 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2340 This would rather mess up the minimum length computation.
2342 What we'll do is to look for the tail four bytes, and then peek
2343 at the preceding two bytes to see whether we need to decrease
2344 the minimum length by four (six minus two).
2346 Thanks to the design of UTF-8, there cannot be false matches:
2347 A sequence of valid UTF-8 bytes cannot be a subsequence of
2348 another valid sequence of UTF-8 bytes.
2351 char * const s0 = STRING(scan), *s, *t;
2352 char * const s1 = s0 + STR_LEN(scan) - 1;
2353 char * const s2 = s1 - 4;
2354 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2355 const char t0[] = "\xaf\x49\xaf\x42";
2357 const char t0[] = "\xcc\x88\xcc\x81";
2359 const char * const t1 = t0 + 3;
2362 s < s2 && (t = ninstr(s, s1, t0, t1));
2365 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2366 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2368 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2369 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2377 n = scan + NODE_SZ_STR(scan);
2379 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2386 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2390 /* REx optimizer. Converts nodes into quickier variants "in place".
2391 Finds fixed substrings. */
2393 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2394 to the position after last scanned or to NULL. */
2396 #define INIT_AND_WITHP \
2397 assert(!and_withp); \
2398 Newx(and_withp,1,struct regnode_charclass_class); \
2399 SAVEFREEPV(and_withp)
2401 /* this is a chain of data about sub patterns we are processing that
2402 need to be handled seperately/specially in study_chunk. Its so
2403 we can simulate recursion without losing state. */
2405 typedef struct scan_frame {
2406 regnode *last; /* last node to process in this frame */
2407 regnode *next; /* next node to process when last is reached */
2408 struct scan_frame *prev; /*previous frame*/
2409 I32 stop; /* what stopparen do we use */
2413 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2415 #define CASE_SYNST_FNC(nAmE) \
2417 if (flags & SCF_DO_STCLASS_AND) { \
2418 for (value = 0; value < 256; value++) \
2419 if (!is_ ## nAmE ## _cp(value)) \
2420 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2423 for (value = 0; value < 256; value++) \
2424 if (is_ ## nAmE ## _cp(value)) \
2425 ANYOF_BITMAP_SET(data->start_class, value); \
2429 if (flags & SCF_DO_STCLASS_AND) { \
2430 for (value = 0; value < 256; value++) \
2431 if (is_ ## nAmE ## _cp(value)) \
2432 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2435 for (value = 0; value < 256; value++) \
2436 if (!is_ ## nAmE ## _cp(value)) \
2437 ANYOF_BITMAP_SET(data->start_class, value); \
2444 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2445 I32 *minlenp, I32 *deltap,
2450 struct regnode_charclass_class *and_withp,
2451 U32 flags, U32 depth)
2452 /* scanp: Start here (read-write). */
2453 /* deltap: Write maxlen-minlen here. */
2454 /* last: Stop before this one. */
2455 /* data: string data about the pattern */
2456 /* stopparen: treat close N as END */
2457 /* recursed: which subroutines have we recursed into */
2458 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2461 I32 min = 0, pars = 0, code;
2462 regnode *scan = *scanp, *next;
2464 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2465 int is_inf_internal = 0; /* The studied chunk is infinite */
2466 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2467 scan_data_t data_fake;
2468 SV *re_trie_maxbuff = NULL;
2469 regnode *first_non_open = scan;
2470 I32 stopmin = I32_MAX;
2471 scan_frame *frame = NULL;
2473 GET_RE_DEBUG_FLAGS_DECL;
2476 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2480 while (first_non_open && OP(first_non_open) == OPEN)
2481 first_non_open=regnext(first_non_open);
2486 while ( scan && OP(scan) != END && scan < last ){
2487 /* Peephole optimizer: */
2488 DEBUG_STUDYDATA("Peep:", data,depth);
2489 DEBUG_PEEP("Peep",scan,depth);
2490 JOIN_EXACT(scan,&min,0);
2492 /* Follow the next-chain of the current node and optimize
2493 away all the NOTHINGs from it. */
2494 if (OP(scan) != CURLYX) {
2495 const int max = (reg_off_by_arg[OP(scan)]
2497 /* I32 may be smaller than U16 on CRAYs! */
2498 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2499 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2503 /* Skip NOTHING and LONGJMP. */
2504 while ((n = regnext(n))
2505 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2506 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2507 && off + noff < max)
2509 if (reg_off_by_arg[OP(scan)])
2512 NEXT_OFF(scan) = off;
2517 /* The principal pseudo-switch. Cannot be a switch, since we
2518 look into several different things. */
2519 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2520 || OP(scan) == IFTHEN) {
2521 next = regnext(scan);
2523 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2525 if (OP(next) == code || code == IFTHEN) {
2526 /* NOTE - There is similar code to this block below for handling
2527 TRIE nodes on a re-study. If you change stuff here check there
2529 I32 max1 = 0, min1 = I32_MAX, num = 0;
2530 struct regnode_charclass_class accum;
2531 regnode * const startbranch=scan;
2533 if (flags & SCF_DO_SUBSTR)
2534 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2535 if (flags & SCF_DO_STCLASS)
2536 cl_init_zero(pRExC_state, &accum);
2538 while (OP(scan) == code) {
2539 I32 deltanext, minnext, f = 0, fake;
2540 struct regnode_charclass_class this_class;
2543 data_fake.flags = 0;
2545 data_fake.whilem_c = data->whilem_c;
2546 data_fake.last_closep = data->last_closep;
2549 data_fake.last_closep = &fake;
2551 data_fake.pos_delta = delta;
2552 next = regnext(scan);
2553 scan = NEXTOPER(scan);
2555 scan = NEXTOPER(scan);
2556 if (flags & SCF_DO_STCLASS) {
2557 cl_init(pRExC_state, &this_class);
2558 data_fake.start_class = &this_class;
2559 f = SCF_DO_STCLASS_AND;
2561 if (flags & SCF_WHILEM_VISITED_POS)
2562 f |= SCF_WHILEM_VISITED_POS;
2564 /* we suppose the run is continuous, last=next...*/
2565 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2567 stopparen, recursed, NULL, f,depth+1);
2570 if (max1 < minnext + deltanext)
2571 max1 = minnext + deltanext;
2572 if (deltanext == I32_MAX)
2573 is_inf = is_inf_internal = 1;
2575 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2577 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2578 if ( stopmin > minnext)
2579 stopmin = min + min1;
2580 flags &= ~SCF_DO_SUBSTR;
2582 data->flags |= SCF_SEEN_ACCEPT;
2585 if (data_fake.flags & SF_HAS_EVAL)
2586 data->flags |= SF_HAS_EVAL;
2587 data->whilem_c = data_fake.whilem_c;
2589 if (flags & SCF_DO_STCLASS)
2590 cl_or(pRExC_state, &accum, &this_class);
2592 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2594 if (flags & SCF_DO_SUBSTR) {
2595 data->pos_min += min1;
2596 data->pos_delta += max1 - min1;
2597 if (max1 != min1 || is_inf)
2598 data->longest = &(data->longest_float);
2601 delta += max1 - min1;
2602 if (flags & SCF_DO_STCLASS_OR) {
2603 cl_or(pRExC_state, data->start_class, &accum);
2605 cl_and(data->start_class, and_withp);
2606 flags &= ~SCF_DO_STCLASS;
2609 else if (flags & SCF_DO_STCLASS_AND) {
2611 cl_and(data->start_class, &accum);
2612 flags &= ~SCF_DO_STCLASS;
2615 /* Switch to OR mode: cache the old value of
2616 * data->start_class */
2618 StructCopy(data->start_class, and_withp,
2619 struct regnode_charclass_class);
2620 flags &= ~SCF_DO_STCLASS_AND;
2621 StructCopy(&accum, data->start_class,
2622 struct regnode_charclass_class);
2623 flags |= SCF_DO_STCLASS_OR;
2624 data->start_class->flags |= ANYOF_EOS;
2628 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2631 Assuming this was/is a branch we are dealing with: 'scan' now
2632 points at the item that follows the branch sequence, whatever
2633 it is. We now start at the beginning of the sequence and look
2640 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2642 If we can find such a subseqence we need to turn the first
2643 element into a trie and then add the subsequent branch exact
2644 strings to the trie.
2648 1. patterns where the whole set of branch can be converted.
2650 2. patterns where only a subset can be converted.
2652 In case 1 we can replace the whole set with a single regop
2653 for the trie. In case 2 we need to keep the start and end
2656 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2657 becomes BRANCH TRIE; BRANCH X;
2659 There is an additional case, that being where there is a
2660 common prefix, which gets split out into an EXACT like node
2661 preceding the TRIE node.
2663 If x(1..n)==tail then we can do a simple trie, if not we make
2664 a "jump" trie, such that when we match the appropriate word
2665 we "jump" to the appopriate tail node. Essentailly we turn
2666 a nested if into a case structure of sorts.
2671 if (!re_trie_maxbuff) {
2672 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2673 if (!SvIOK(re_trie_maxbuff))
2674 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2676 if ( SvIV(re_trie_maxbuff)>=0 ) {
2678 regnode *first = (regnode *)NULL;
2679 regnode *last = (regnode *)NULL;
2680 regnode *tail = scan;
2685 SV * const mysv = sv_newmortal(); /* for dumping */
2687 /* var tail is used because there may be a TAIL
2688 regop in the way. Ie, the exacts will point to the
2689 thing following the TAIL, but the last branch will
2690 point at the TAIL. So we advance tail. If we
2691 have nested (?:) we may have to move through several
2695 while ( OP( tail ) == TAIL ) {
2696 /* this is the TAIL generated by (?:) */
2697 tail = regnext( tail );
2702 regprop(RExC_rx, mysv, tail );
2703 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2704 (int)depth * 2 + 2, "",
2705 "Looking for TRIE'able sequences. Tail node is: ",
2706 SvPV_nolen_const( mysv )
2712 step through the branches, cur represents each
2713 branch, noper is the first thing to be matched
2714 as part of that branch and noper_next is the
2715 regnext() of that node. if noper is an EXACT
2716 and noper_next is the same as scan (our current
2717 position in the regex) then the EXACT branch is
2718 a possible optimization target. Once we have
2719 two or more consequetive such branches we can
2720 create a trie of the EXACT's contents and stich
2721 it in place. If the sequence represents all of
2722 the branches we eliminate the whole thing and
2723 replace it with a single TRIE. If it is a
2724 subsequence then we need to stitch it in. This
2725 means the first branch has to remain, and needs
2726 to be repointed at the item on the branch chain
2727 following the last branch optimized. This could
2728 be either a BRANCH, in which case the
2729 subsequence is internal, or it could be the
2730 item following the branch sequence in which
2731 case the subsequence is at the end.
2735 /* dont use tail as the end marker for this traverse */
2736 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2737 regnode * const noper = NEXTOPER( cur );
2738 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2739 regnode * const noper_next = regnext( noper );
2743 regprop(RExC_rx, mysv, cur);
2744 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2745 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2747 regprop(RExC_rx, mysv, noper);
2748 PerlIO_printf( Perl_debug_log, " -> %s",
2749 SvPV_nolen_const(mysv));
2752 regprop(RExC_rx, mysv, noper_next );
2753 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2754 SvPV_nolen_const(mysv));
2756 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2757 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2759 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2760 : PL_regkind[ OP( noper ) ] == EXACT )
2761 || OP(noper) == NOTHING )
2763 && noper_next == tail
2768 if ( !first || optype == NOTHING ) {
2769 if (!first) first = cur;
2770 optype = OP( noper );
2776 make_trie( pRExC_state,
2777 startbranch, first, cur, tail, count,
2780 if ( PL_regkind[ OP( noper ) ] == EXACT
2782 && noper_next == tail
2787 optype = OP( noper );
2797 regprop(RExC_rx, mysv, cur);
2798 PerlIO_printf( Perl_debug_log,
2799 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2800 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2804 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2805 #ifdef TRIE_STUDY_OPT
2806 if ( ((made == MADE_EXACT_TRIE &&
2807 startbranch == first)
2808 || ( first_non_open == first )) &&
2810 flags |= SCF_TRIE_RESTUDY;
2811 if ( startbranch == first
2814 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2824 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2825 scan = NEXTOPER(NEXTOPER(scan));
2826 } else /* single branch is optimized. */
2827 scan = NEXTOPER(scan);
2829 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2830 scan_frame *newframe = NULL;
2835 if (OP(scan) != SUSPEND) {
2836 /* set the pointer */
2837 if (OP(scan) == GOSUB) {
2839 RExC_recurse[ARG2L(scan)] = scan;
2840 start = RExC_open_parens[paren-1];
2841 end = RExC_close_parens[paren-1];
2844 start = RExC_rxi->program + 1;
2848 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
2849 SAVEFREEPV(recursed);
2851 if (!PAREN_TEST(recursed,paren+1)) {
2852 PAREN_SET(recursed,paren+1);
2853 Newx(newframe,1,scan_frame);
2855 if (flags & SCF_DO_SUBSTR) {
2856 SCAN_COMMIT(pRExC_state,data,minlenp);
2857 data->longest = &(data->longest_float);
2859 is_inf = is_inf_internal = 1;
2860 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
2861 cl_anything(pRExC_state, data->start_class);
2862 flags &= ~SCF_DO_STCLASS;
2865 Newx(newframe,1,scan_frame);
2868 end = regnext(scan);
2873 SAVEFREEPV(newframe);
2874 newframe->next = regnext(scan);
2875 newframe->last = last;
2876 newframe->stop = stopparen;
2877 newframe->prev = frame;
2887 else if (OP(scan) == EXACT) {
2888 I32 l = STR_LEN(scan);
2891 const U8 * const s = (U8*)STRING(scan);
2892 l = utf8_length(s, s + l);
2893 uc = utf8_to_uvchr(s, NULL);
2895 uc = *((U8*)STRING(scan));
2898 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2899 /* The code below prefers earlier match for fixed
2900 offset, later match for variable offset. */
2901 if (data->last_end == -1) { /* Update the start info. */
2902 data->last_start_min = data->pos_min;
2903 data->last_start_max = is_inf
2904 ? I32_MAX : data->pos_min + data->pos_delta;
2906 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2908 SvUTF8_on(data->last_found);
2910 SV * const sv = data->last_found;
2911 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2912 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2913 if (mg && mg->mg_len >= 0)
2914 mg->mg_len += utf8_length((U8*)STRING(scan),
2915 (U8*)STRING(scan)+STR_LEN(scan));
2917 data->last_end = data->pos_min + l;
2918 data->pos_min += l; /* As in the first entry. */
2919 data->flags &= ~SF_BEFORE_EOL;
2921 if (flags & SCF_DO_STCLASS_AND) {
2922 /* Check whether it is compatible with what we know already! */
2926 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2927 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2928 && (!(data->start_class->flags & ANYOF_FOLD)
2929 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2932 ANYOF_CLASS_ZERO(data->start_class);
2933 ANYOF_BITMAP_ZERO(data->start_class);
2935 ANYOF_BITMAP_SET(data->start_class, uc);
2936 data->start_class->flags &= ~ANYOF_EOS;
2938 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2940 else if (flags & SCF_DO_STCLASS_OR) {
2941 /* false positive possible if the class is case-folded */
2943 ANYOF_BITMAP_SET(data->start_class, uc);
2945 data->start_class->flags |= ANYOF_UNICODE_ALL;
2946 data->start_class->flags &= ~ANYOF_EOS;
2947 cl_and(data->start_class, and_withp);
2949 flags &= ~SCF_DO_STCLASS;
2951 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
2952 I32 l = STR_LEN(scan);
2953 UV uc = *((U8*)STRING(scan));
2955 /* Search for fixed substrings supports EXACT only. */
2956 if (flags & SCF_DO_SUBSTR) {
2958 SCAN_COMMIT(pRExC_state, data, minlenp);
2961 const U8 * const s = (U8 *)STRING(scan);
2962 l = utf8_length(s, s + l);
2963 uc = utf8_to_uvchr(s, NULL);
2966 if (flags & SCF_DO_SUBSTR)
2968 if (flags & SCF_DO_STCLASS_AND) {
2969 /* Check whether it is compatible with what we know already! */
2973 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2974 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2975 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2977 ANYOF_CLASS_ZERO(data->start_class);
2978 ANYOF_BITMAP_ZERO(data->start_class);
2980 ANYOF_BITMAP_SET(data->start_class, uc);
2981 data->start_class->flags &= ~ANYOF_EOS;
2982 data->start_class->flags |= ANYOF_FOLD;
2983 if (OP(scan) == EXACTFL)
2984 data->start_class->flags |= ANYOF_LOCALE;
2987 else if (flags & SCF_DO_STCLASS_OR) {
2988 if (data->start_class->flags & ANYOF_FOLD) {
2989 /* false positive possible if the class is case-folded.
2990 Assume that the locale settings are the same... */
2992 ANYOF_BITMAP_SET(data->start_class, uc);
2993 data->start_class->flags &= ~ANYOF_EOS;
2995 cl_and(data->start_class, and_withp);
2997 flags &= ~SCF_DO_STCLASS;
2999 else if (strchr((const char*)PL_varies,OP(scan))) {
3000 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3001 I32 f = flags, pos_before = 0;
3002 regnode * const oscan = scan;
3003 struct regnode_charclass_class this_class;
3004 struct regnode_charclass_class *oclass = NULL;
3005 I32 next_is_eval = 0;
3007 switch (PL_regkind[OP(scan)]) {
3008 case WHILEM: /* End of (?:...)* . */
3009 scan = NEXTOPER(scan);
3012 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3013 next = NEXTOPER(scan);
3014 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3016 maxcount = REG_INFTY;
3017 next = regnext(scan);
3018 scan = NEXTOPER(scan);
3022 if (flags & SCF_DO_SUBSTR)
3027 if (flags & SCF_DO_STCLASS) {
3029 maxcount = REG_INFTY;
3030 next = regnext(scan);
3031 scan = NEXTOPER(scan);
3034 is_inf = is_inf_internal = 1;
3035 scan = regnext(scan);
3036 if (flags & SCF_DO_SUBSTR) {
3037 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3038 data->longest = &(data->longest_float);
3040 goto optimize_curly_tail;
3042 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3043 && (scan->flags == stopparen))
3048 mincount = ARG1(scan);
3049 maxcount = ARG2(scan);
3051 next = regnext(scan);
3052 if (OP(scan) == CURLYX) {
3053 I32 lp = (data ? *(data->last_closep) : 0);
3054 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3056 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3057 next_is_eval = (OP(scan) == EVAL);
3059 if (flags & SCF_DO_SUBSTR) {
3060 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3061 pos_before = data->pos_min;
3065 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3067 data->flags |= SF_IS_INF;
3069 if (flags & SCF_DO_STCLASS) {
3070 cl_init(pRExC_state, &this_class);
3071 oclass = data->start_class;
3072 data->start_class = &this_class;
3073 f |= SCF_DO_STCLASS_AND;
3074 f &= ~SCF_DO_STCLASS_OR;
3076 /* These are the cases when once a subexpression
3077 fails at a particular position, it cannot succeed
3078 even after backtracking at the enclosing scope.
3080 XXXX what if minimal match and we are at the
3081 initial run of {n,m}? */
3082 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
3083 f &= ~SCF_WHILEM_VISITED_POS;
3085 /* This will finish on WHILEM, setting scan, or on NULL: */
3086 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3087 last, data, stopparen, recursed, NULL,
3089 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3091 if (flags & SCF_DO_STCLASS)
3092 data->start_class = oclass;
3093 if (mincount == 0 || minnext == 0) {
3094 if (flags & SCF_DO_STCLASS_OR) {
3095 cl_or(pRExC_state, data->start_class, &this_class);
3097 else if (flags & SCF_DO_STCLASS_AND) {
3098 /* Switch to OR mode: cache the old value of
3099 * data->start_class */
3101 StructCopy(data->start_class, and_withp,
3102 struct regnode_charclass_class);
3103 flags &= ~SCF_DO_STCLASS_AND;
3104 StructCopy(&this_class, data->start_class,
3105 struct regnode_charclass_class);
3106 flags |= SCF_DO_STCLASS_OR;
3107 data->start_class->flags |= ANYOF_EOS;
3109 } else { /* Non-zero len */
3110 if (flags & SCF_DO_STCLASS_OR) {
3111 cl_or(pRExC_state, data->start_class, &this_class);
3112 cl_and(data->start_class, and_withp);
3114 else if (flags & SCF_DO_STCLASS_AND)
3115 cl_and(data->start_class, &this_class);
3116 flags &= ~SCF_DO_STCLASS;
3118 if (!scan) /* It was not CURLYX, but CURLY. */
3120 if ( /* ? quantifier ok, except for (?{ ... }) */
3121 (next_is_eval || !(mincount == 0 && maxcount == 1))
3122 && (minnext == 0) && (deltanext == 0)
3123 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3124 && maxcount <= REG_INFTY/3 /* Complement check for big count */
3125 && ckWARN(WARN_REGEXP))
3128 "Quantifier unexpected on zero-length expression");
3131 min += minnext * mincount;
3132 is_inf_internal |= ((maxcount == REG_INFTY
3133 && (minnext + deltanext) > 0)
3134 || deltanext == I32_MAX);
3135 is_inf |= is_inf_internal;
3136 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3138 /* Try powerful optimization CURLYX => CURLYN. */
3139 if ( OP(oscan) == CURLYX && data
3140 && data->flags & SF_IN_PAR
3141 && !(data->flags & SF_HAS_EVAL)
3142 && !deltanext && minnext == 1 ) {
3143 /* Try to optimize to CURLYN. */
3144 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3145 regnode * const nxt1 = nxt;
3152 if (!strchr((const char*)PL_simple,OP(nxt))
3153 && !(PL_regkind[OP(nxt)] == EXACT
3154 && STR_LEN(nxt) == 1))
3160 if (OP(nxt) != CLOSE)
3162 if (RExC_open_parens) {
3163 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3164 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3166 /* Now we know that nxt2 is the only contents: */
3167 oscan->flags = (U8)ARG(nxt);
3169 OP(nxt1) = NOTHING; /* was OPEN. */
3172 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3173 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
3174 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
3175 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3176 OP(nxt + 1) = OPTIMIZED; /* was count. */
3177 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
3182 /* Try optimization CURLYX => CURLYM. */
3183 if ( OP(oscan) == CURLYX && data
3184 && !(data->flags & SF_HAS_PAR)
3185 && !(data->flags & SF_HAS_EVAL)
3186 && !deltanext /* atom is fixed width */
3187 && minnext != 0 /* CURLYM can't handle zero width */
3189 /* XXXX How to optimize if data == 0? */
3190 /* Optimize to a simpler form. */
3191 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3195 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3196 && (OP(nxt2) != WHILEM))
3198 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3199 /* Need to optimize away parenths. */
3200 if (data->flags & SF_IN_PAR) {
3201 /* Set the parenth number. */
3202 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3204 if (OP(nxt) != CLOSE)
3205 FAIL("Panic opt close");
3206 oscan->flags = (U8)ARG(nxt);
3207 if (RExC_open_parens) {
3208 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3209 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3211 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3212 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3215 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3216 OP(nxt + 1) = OPTIMIZED; /* was count. */
3217 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3218 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3221 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3222 regnode *nnxt = regnext(nxt1);
3225 if (reg_off_by_arg[OP(nxt1)])
3226 ARG_SET(nxt1, nxt2 - nxt1);
3227 else if (nxt2 - nxt1 < U16_MAX)
3228 NEXT_OFF(nxt1) = nxt2 - nxt1;
3230 OP(nxt) = NOTHING; /* Cannot beautify */
3235 /* Optimize again: */
3236 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3237 NULL, stopparen, recursed, NULL, 0,depth+1);
3242 else if ((OP(oscan) == CURLYX)
3243 && (flags & SCF_WHILEM_VISITED_POS)
3244 /* See the comment on a similar expression above.
3245 However, this time it not a subexpression
3246 we care about, but the expression itself. */
3247 && (maxcount == REG_INFTY)
3248 && data && ++data->whilem_c < 16) {
3249 /* This stays as CURLYX, we can put the count/of pair. */
3250 /* Find WHILEM (as in regexec.c) */
3251 regnode *nxt = oscan + NEXT_OFF(oscan);
3253 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3255 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3256 | (RExC_whilem_seen << 4)); /* On WHILEM */
3258 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3260 if (flags & SCF_DO_SUBSTR) {
3261 SV *last_str = NULL;
3262 int counted = mincount != 0;
3264 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3265 #if defined(SPARC64_GCC_WORKAROUND)
3268 const char *s = NULL;
3271 if (pos_before >= data->last_start_min)
3274 b = data->last_start_min;
3277 s = SvPV_const(data->last_found, l);
3278 old = b - data->last_start_min;
3281 I32 b = pos_before >= data->last_start_min
3282 ? pos_before : data->last_start_min;
3284 const char * const s = SvPV_const(data->last_found, l);
3285 I32 old = b - data->last_start_min;
3289 old = utf8_hop((U8*)s, old) - (U8*)s;
3292 /* Get the added string: */
3293 last_str = newSVpvn(s + old, l);
3295 SvUTF8_on(last_str);
3296 if (deltanext == 0 && pos_before == b) {
3297 /* What was added is a constant string */
3299 SvGROW(last_str, (mincount * l) + 1);
3300 repeatcpy(SvPVX(last_str) + l,
3301 SvPVX_const(last_str), l, mincount - 1);
3302 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3303 /* Add additional parts. */
3304 SvCUR_set(data->last_found,
3305 SvCUR(data->last_found) - l);
3306 sv_catsv(data->last_found, last_str);
3308 SV * sv = data->last_found;
3310 SvUTF8(sv) && SvMAGICAL(sv) ?
3311 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3312 if (mg && mg->mg_len >= 0)
3313 mg->mg_len += CHR_SVLEN(last_str) - l;
3315 data->last_end += l * (mincount - 1);
3318 /* start offset must point into the last copy */
3319 data->last_start_min += minnext * (mincount - 1);
3320 data->last_start_max += is_inf ? I32_MAX
3321 : (maxcount - 1) * (minnext + data->pos_delta);
3324 /* It is counted once already... */
3325 data->pos_min += minnext * (mincount - counted);
3326 data->pos_delta += - counted * deltanext +
3327 (minnext + deltanext) * maxcount - minnext * mincount;
3328 if (mincount != maxcount) {
3329 /* Cannot extend fixed substrings found inside
3331 SCAN_COMMIT(pRExC_state,data,minlenp);
3332 if (mincount && last_str) {
3333 SV * const sv = data->last_found;
3334 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3335 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3339 sv_setsv(sv, last_str);
3340 data->last_end = data->pos_min;
3341 data->last_start_min =
3342 data->pos_min - CHR_SVLEN(last_str);
3343 data->last_start_max = is_inf
3345 : data->pos_min + data->pos_delta
3346 - CHR_SVLEN(last_str);
3348 data->longest = &(data->longest_float);
3350 SvREFCNT_dec(last_str);
3352 if (data && (fl & SF_HAS_EVAL))
3353 data->flags |= SF_HAS_EVAL;
3354 optimize_curly_tail:
3355 if (OP(oscan) != CURLYX) {
3356 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3358 NEXT_OFF(oscan) += NEXT_OFF(next);
3361 default: /* REF and CLUMP only? */
3362 if (flags & SCF_DO_SUBSTR) {
3363 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3364 data->longest = &(data->longest_float);
3366 is_inf = is_inf_internal = 1;
3367 if (flags & SCF_DO_STCLASS_OR)
3368 cl_anything(pRExC_state, data->start_class);
3369 flags &= ~SCF_DO_STCLASS;
3373 else if (OP(scan) == LNBREAK) {
3374 if (flags & SCF_DO_STCLASS) {
3376 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3377 if (flags & SCF_DO_STCLASS_AND) {
3378 for (value = 0; value < 256; value++)
3379 if (!is_VERTWS_cp(value))
3380 ANYOF_BITMAP_CLEAR(data->start_class, value);
3383 for (value = 0; value < 256; value++)
3384 if (is_VERTWS_cp(value))
3385 ANYOF_BITMAP_SET(data->start_class, value);
3387 if (flags & SCF_DO_STCLASS_OR)
3388 cl_and(data->start_class, and_withp);
3389 flags &= ~SCF_DO_STCLASS;
3393 if (flags & SCF_DO_SUBSTR) {
3394 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3396 data->pos_delta += 1;
3397 data->longest = &(data->longest_float);
3401 else if (OP(scan) == FOLDCHAR) {
3402 int d = ARG(scan)==0xDF ? 1 : 2;
3403 flags &= ~SCF_DO_STCLASS;
3406 if (flags & SCF_DO_SUBSTR) {
3407 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3409 data->pos_delta += d;
3410 data->longest = &(data->longest_float);
3413 else if (strchr((const char*)PL_simple,OP(scan))) {
3416 if (flags & SCF_DO_SUBSTR) {
3417 SCAN_COMMIT(pRExC_state,data,minlenp);
3421 if (flags & SCF_DO_STCLASS) {
3422 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3424 /* Some of the logic below assumes that switching
3425 locale on will only add false positives. */
3426 switch (PL_regkind[OP(scan)]) {
3430 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3431 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3432 cl_anything(pRExC_state, data->start_class);
3435 if (OP(scan) == SANY)
3437 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3438 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3439 || (data->start_class->flags & ANYOF_CLASS));
3440 cl_anything(pRExC_state, data->start_class);
3442 if (flags & SCF_DO_STCLASS_AND || !value)
3443 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3446 if (flags & SCF_DO_STCLASS_AND)
3447 cl_and(data->start_class,
3448 (struct regnode_charclass_class*)scan);
3450 cl_or(pRExC_state, data->start_class,
3451 (struct regnode_charclass_class*)scan);
3454 if (flags & SCF_DO_STCLASS_AND) {
3455 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3456 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3457 for (value = 0; value < 256; value++)
3458 if (!isALNUM(value))
3459 ANYOF_BITMAP_CLEAR(data->start_class, value);
3463 if (data->start_class->flags & ANYOF_LOCALE)
3464 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3466 for (value = 0; value < 256; value++)
3468 ANYOF_BITMAP_SET(data->start_class, value);
3473 if (flags & SCF_DO_STCLASS_AND) {
3474 if (data->start_class->flags & ANYOF_LOCALE)
3475 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3478 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3479 data->start_class->flags |= ANYOF_LOCALE;
3483 if (flags & SCF_DO_STCLASS_AND) {
3484 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3485 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3486 for (value = 0; value < 256; value++)
3488 ANYOF_BITMAP_CLEAR(data->start_class, value);
3492 if (data->start_class->flags & ANYOF_LOCALE)
3493 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3495 for (value = 0; value < 256; value++)
3496 if (!isALNUM(value))
3497 ANYOF_BITMAP_SET(data->start_class, value);
3502 if (flags & SCF_DO_STCLASS_AND) {
3503 if (data->start_class->flags & ANYOF_LOCALE)
3504 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3507 data->start_class->flags |= ANYOF_LOCALE;
3508 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3512 if (flags & SCF_DO_STCLASS_AND) {
3513 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3514 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3515 for (value = 0; value < 256; value++)
3516 if (!isSPACE(value))
3517 ANYOF_BITMAP_CLEAR(data->start_class, value);
3521 if (data->start_class->flags & ANYOF_LOCALE)
3522 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3524 for (value = 0; value < 256; value++)
3526 ANYOF_BITMAP_SET(data->start_class, value);
3531 if (flags & SCF_DO_STCLASS_AND) {
3532 if (data->start_class->flags & ANYOF_LOCALE)
3533 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3536 data->start_class->flags |= ANYOF_LOCALE;
3537 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3541 if (flags & SCF_DO_STCLASS_AND) {
3542 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3543 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3544 for (value = 0; value < 256; value++)
3546 ANYOF_BITMAP_CLEAR(data->start_class, value);
3550 if (data->start_class->flags & ANYOF_LOCALE)
3551 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3553 for (value = 0; value < 256; value++)
3554 if (!isSPACE(value))
3555 ANYOF_BITMAP_SET(data->start_class, value);
3560 if (flags & SCF_DO_STCLASS_AND) {
3561 if (data->start_class->flags & ANYOF_LOCALE) {
3562 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3563 for (value = 0; value < 256; value++)
3564 if (!isSPACE(value))
3565 ANYOF_BITMAP_CLEAR(data->start_class, value);
3569 data->start_class->flags |= ANYOF_LOCALE;
3570 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3574 if (flags & SCF_DO_STCLASS_AND) {
3575 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3576 for (value = 0; value < 256; value++)
3577 if (!isDIGIT(value))
3578 ANYOF_BITMAP_CLEAR(data->start_class, value);
3581 if (data->start_class->flags & ANYOF_LOCALE)
3582 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3584 for (value = 0; value < 256; value++)
3586 ANYOF_BITMAP_SET(data->start_class, value);
3591 if (flags & SCF_DO_STCLASS_AND) {
3592 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3593 for (value = 0; value < 256; value++)
3595 ANYOF_BITMAP_CLEAR(data->start_class, value);
3598 if (data->start_class->flags & ANYOF_LOCALE)
3599 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3601 for (value = 0; value < 256; value++)
3602 if (!isDIGIT(value))
3603 ANYOF_BITMAP_SET(data->start_class, value);
3607 CASE_SYNST_FNC(VERTWS);
3608 CASE_SYNST_FNC(HORIZWS);
3611 if (flags & SCF_DO_STCLASS_OR)
3612 cl_and(data->start_class, and_withp);
3613 flags &= ~SCF_DO_STCLASS;
3616 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3617 data->flags |= (OP(scan) == MEOL
3621 else if ( PL_regkind[OP(scan)] == BRANCHJ
3622 /* Lookbehind, or need to calculate parens/evals/stclass: */
3623 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3624 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3625 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3626 || OP(scan) == UNLESSM )
3628 /* Negative Lookahead/lookbehind
3629 In this case we can't do fixed string optimisation.
3632 I32 deltanext, minnext, fake = 0;
3634 struct regnode_charclass_class intrnl;
3637 data_fake.flags = 0;
3639 data_fake.whilem_c = data->whilem_c;
3640 data_fake.last_closep = data->last_closep;
3643 data_fake.last_closep = &fake;
3644 data_fake.pos_delta = delta;
3645 if ( flags & SCF_DO_STCLASS && !scan->flags
3646 && OP(scan) == IFMATCH ) { /* Lookahead */
3647 cl_init(pRExC_state, &intrnl);
3648 data_fake.start_class = &intrnl;
3649 f |= SCF_DO_STCLASS_AND;
3651 if (flags & SCF_WHILEM_VISITED_POS)
3652 f |= SCF_WHILEM_VISITED_POS;
3653 next = regnext(scan);
3654 nscan = NEXTOPER(NEXTOPER(scan));
3655 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
3656 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
3659 FAIL("Variable length lookbehind not implemented");
3661 else if (minnext > (I32)U8_MAX) {
3662 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3664 scan->flags = (U8)minnext;
3667 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3669 if (data_fake.flags & SF_HAS_EVAL)
3670 data->flags |= SF_HAS_EVAL;
3671 data->whilem_c = data_fake.whilem_c;
3673 if (f & SCF_DO_STCLASS_AND) {
3674 const int was = (data->start_class->flags & ANYOF_EOS);
3676 cl_and(data->start_class, &intrnl);
3678 data->start_class->flags |= ANYOF_EOS;
3681 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3683 /* Positive Lookahead/lookbehind
3684 In this case we can do fixed string optimisation,
3685 but we must be careful about it. Note in the case of
3686 lookbehind the positions will be offset by the minimum
3687 length of the pattern, something we won't know about
3688 until after the recurse.
3690 I32 deltanext, fake = 0;
3692 struct regnode_charclass_class intrnl;
3694 /* We use SAVEFREEPV so that when the full compile
3695 is finished perl will clean up the allocated
3696 minlens when its all done. This was we don't
3697 have to worry about freeing them when we know
3698 they wont be used, which would be a pain.
3701 Newx( minnextp, 1, I32 );
3702 SAVEFREEPV(minnextp);
3705 StructCopy(data, &data_fake, scan_data_t);
3706 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3709 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
3710 data_fake.last_found=newSVsv(data->last_found);
3714 data_fake.last_closep = &fake;
3715 data_fake.flags = 0;
3716 data_fake.pos_delta = delta;
3718 data_fake.flags |= SF_IS_INF;
3719 if ( flags & SCF_DO_STCLASS && !scan->flags
3720 && OP(scan) == IFMATCH ) { /* Lookahead */
3721 cl_init(pRExC_state, &intrnl);
3722 data_fake.start_class = &intrnl;
3723 f |= SCF_DO_STCLASS_AND;
3725 if (flags & SCF_WHILEM_VISITED_POS)
3726 f |= SCF_WHILEM_VISITED_POS;
3727 next = regnext(scan);
3728 nscan = NEXTOPER(NEXTOPER(scan));
3730 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
3731 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
3734 FAIL("Variable length lookbehind not implemented");
3736 else if (*minnextp > (I32)U8_MAX) {
3737 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3739 scan->flags = (U8)*minnextp;
3744 if (f & SCF_DO_STCLASS_AND) {
3745 const int was = (data->start_class->flags & ANYOF_EOS);
3747 cl_and(data->start_class, &intrnl);
3749 data->start_class->flags |= ANYOF_EOS;
3752 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3754 if (data_fake.flags & SF_HAS_EVAL)
3755 data->flags |= SF_HAS_EVAL;
3756 data->whilem_c = data_fake.whilem_c;
3757 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3758 if (RExC_rx->minlen<*minnextp)
3759 RExC_rx->minlen=*minnextp;
3760 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
3761 SvREFCNT_dec(data_fake.last_found);
3763 if ( data_fake.minlen_fixed != minlenp )
3765 data->offset_fixed= data_fake.offset_fixed;
3766 data->minlen_fixed= data_fake.minlen_fixed;
3767 data->lookbehind_fixed+= scan->flags;
3769 if ( data_fake.minlen_float != minlenp )
3771 data->minlen_float= data_fake.minlen_float;
3772 data->offset_float_min=data_fake.offset_float_min;
3773 data->offset_float_max=data_fake.offset_float_max;
3774 data->lookbehind_float+= scan->flags;
3783 else if (OP(scan) == OPEN) {
3784 if (stopparen != (I32)ARG(scan))
3787 else if (OP(scan) == CLOSE) {
3788 if (stopparen == (I32)ARG(scan)) {
3791 if ((I32)ARG(scan) == is_par) {
3792 next = regnext(scan);
3794 if ( next && (OP(next) != WHILEM) && next < last)
3795 is_par = 0; /* Disable optimization */
3798 *(data->last_closep) = ARG(scan);
3800 else if (OP(scan) == EVAL) {
3802 data->flags |= SF_HAS_EVAL;
3804 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
3805 if (flags & SCF_DO_SUBSTR) {
3806 SCAN_COMMIT(pRExC_state,data,minlenp);
3807 flags &= ~SCF_DO_SUBSTR;
3809 if (data && OP(scan)==ACCEPT) {
3810 data->flags |= SCF_SEEN_ACCEPT;
3815 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */